Rob Fab in Arc Art and Des 201

Dagmar Reinhardt
Rob Saunders
Jane Burry
Editors
Robotic Fabrication
in Architecture,
Art and Design
2016
Robotic Fabrication in Architecture,

Art and Design 2016
Dagmar Reinhardt Rob Saunders

Jane Burry
Editors
Robotic Fabrication
in Architecture, Art
and Design 2016
Foreword by Sigrid Brell-okcan and Johannes Braumann,
Association for Robots in Architecture
with contributions by Marjo Niemel
123
Editors
Dagmar Reinhardt
Faculty of Architecture, Design
and Planning
University of Sydney
Sydney, NSW
Australia
Jane Burry
School of Architecture and Design
RMIT University
Melbourne, NSW
Australia
Rob Saunders
Faculty of Architecture, Design
and Planning
University of Sydney
Sydney, NSW
Australia
Funded by KUKA Robotics and the Association for Robots in Architecture

ISBN 978-3-319-26376-2
DOI 10.1007/978-3-319-26378-6
ISBN 978-3-319-26378-6
(eBook)
Library of Congress Control Number: 2015955855

Springer Cham Heidelberg New York Dordrecht London
Springer International Publishing Switzerland 2016
Chapter 2 is published with kind permission of the John Wiley & Sons Ltd. 2014.
This work is subject to copyright. All rights are reserved by the Publisher, whether the whole or part
of the material is concerned, specically the rights of translation, reprinting, reuse of illustrations,
recitation, broadcasting, reproduction on microlms or in any other physical way, and transmission
or information storage and retrieval, electronic adaptation, computer software, or by similar or dissimilar
methodology now known or hereafter developed.
The use of general descriptive names, registered names, trademarks, service marks, etc. in this
publication does not imply, even in the absence of a specic statement, that such names are exempt from
the relevant protective laws and regulations and therefore free for general use.
The publisher, the authors and the editors are safe to assume that the advice and information in this
book are believed to be true and accurate at the date of publication. Neither the publisher nor the
authors or the editors give a warranty, express or implied, with respect to the material contained herein or
for any errors or omissions that may have been made.
Printed on acid-free paper
Springer International Publishing AG Switzerland is part of Springer Science+Business Media
(www.springer.com)
Foreword by the Association for Robots

in Architecture
It is incredible to conceive that only four years and two conferences have passed
since the very rst Rob|Arch in Vienna in 2012. At that time, the largest part of the
audience were people with a general interest in new technologies, innovative
workflows, and large-scale machinesbut without any actual hands-on experience
with robotic arms. At that time, it was also not clear if the use of industrial robots in
the creative industry would be a short-lived hype, or something sustainable and
permanent. With the number of robot installations growing nearly exponentially,
robotic arms have now reached a critical mass within the creative industry.
The Australian community especially showcases this extraordinary shift in
architectural education and fabrication. While in 2013 only RMIT and Sydney
University had just started to incorporate robots in their academic curriculum and
research, at the time of Rob|Arch 2016 in Sydney, 50 % of the 20 Australian
architecture schools will have industrial robots as part of their curriculum.
We believe that an important step in this development have been hands-on robot
workshops that demonstrate that architects, artists, designers, researchers, and
educators how to realize robotic projects, using accessible software interfaces that
are already well established within the creative industry. As such, these workshops
have been the centerpiece of the biennial Rob|Arch conference series, involving
cutting-edge research institutions, as well as partners from industry.
While previously the success of many robot-labs relied on the specialized
know-how of a single key-engineer, the community has now become mostly
self-sustaining and educates their own roboticists with creative backgrounds.
Already now, many students of architecture and design schools worldwide are
graduating with a deep knowledge of robotics and fabrication, and are putting it to
good use both within academia as well as in exciting new start-ups and established
companies such as Bot & Dolly, now owned by Google.
Ultimately, this deep robotic know-how allows the creative industry to work on
an equal level with robotic engineers and industry partners, towards developing
innovative solutions in a collaborative and creative environment.
vi
Foreword by the Association for Robots in Architecture
With the increasing proliferation of previously arcane robotic knowledge, new

and exciting companies are being established that build upon the multifunctionality
and flexibility of robotic arms to realize new products and technologies; Odico1
develops their own software to efciently create formwork using hotwire cutting,
Branch Technology2 uses a robot on a linear rail to fabricate large-scale structures
via 3D printing and Artis Engineering3 even put their heavy-payload KUKA on a
ceiling-mounted rail system. At the same time, signicant research funding is
flowing into institutions such as the newly established National Centre for
Competence in Research (NCCR) for Digital Fabrication dfab4 in Switzerland and
new elds of research into individualized building production are established at
Rheinisch-Westflische Technische Hochschule (RWTH) Aachen University5 in
Germany.
This book showcases many of these new developments and covers a large scope
from purely academic pursuits to highly applied technologies. While flying drones
are building bridges, robotic arms structure stone surfaces with both traditional as
well as digital methods, move around on custom-built platforms, or form metal
sheets. The industry section provides a preview into new technologies that are
expected to shape both the traditional robotic industries, as well as the creative
eld. Entirely new robot systems such as the ABB YuMi and KUKA iiwa allow for
the rst time actual humanmachine collaboration, without having to rely on
complicated workflows or external sensor systems. While initially targeted at elds
such as electronics assembly, these strategies will also be incredibly useful for the
creative industry where robots are often set up within existing workplaces, rather
than along robot-exclusive assembly lines in factory buildings. KUKA also presents
a new interface that will allow the streaming of data from basically any device to a
KUKA robot, building upon regular UDP communication, so that an expensive,
industry-PLC can be replaced by basically any PC. Finally, Erne showcases an
entirely new 7-axis machine developed by Gdel that was custom-t to their special
wood fabrication requirements.
Following the 2012 conference in Vienna and the 2014 conference in Michigan,
2016 marks another large geographic shift to Sydney, Australiaa country and
continent with a highly active creative and technology sector that is also very close
to the robotic hotspots of China and Japan. Australian universities were extremely
quick to acquire robots for teaching and research, with Rob|Arch 2016 host
University of Sydney, as well as Rob|Arch partners RMIT, Bond University,
University of Technology Sydney, University of New South Wales Sydney,
Monash University, and the University of Tasmania now using one or more robotic
arms in the elds of architecture, art, and design.
www.odico.dk.
www.branch.technology.
3
www.artisengineering.de.
4
www.dfab.ch.
5
www.ip.rwth-aachen.de.
2
Foreword by the Association for Robots in Architecture
vii
We want to express our thanks to the conference chairs Dagmar Reinhardt and
Rob Saunders, co-editor Jane Burry, workshop chair Marjo Niemel, the scientic
committee, and everyone else who contributed to the success of the Rob|Arch
community and the 2016 conference. The Rob|Arch conferences and workshops
would not be possible without our long-term industry partners, spearheaded by
KUKA represented by Alois Buchstab and Greg Sale and ABB with Frank-Peter
Kirgis and Peter Katsos as the main sponsors of the conference. Beyond their
robotic support, their promotion of the conference series within automation and
industry has led to many exciting and fruitful collaborations with the Rob|Arch
community. A new main sponsor from the software side is Autodesk, represented
by a team around Matt Jezyk. This shows the new impact of the software-driven
Rob|Arch community and the demand for industrial robots as design and fabrication devices for future CAD software developments. And as a main workshop
sponsor, local partner BVN represents a major architecture practice engagement.
Finally, we want to thank the entire Rob|Arch community for their enthusiasm to
share and discuss their projects and research, thus advancing our common cause of
making robots accessible to the creative industry and other new users!
Sigrid Brell-okcan
Johannes Braumann
Preface
From Robotic Fabrication to Creative Robotics

The adoption of digital fabrication in the creative industries continues to accelerate
as the potential for innovation and creative expression using robotics is harnessed.
Following the conference theme of Trajectories the research presented in this
book demonstrates the continuing evolution of robotic fabrication and creative
robotics in architecture, art, and designtoward the integration of humanrobot
interactions informed by sensor input and real-time feedback under diverse environmental conditions.
Developed for factory automation, industrial robots offer accuracy, flexibility,
and reliability with reduced operational costs. For these reasons, artists and
designers seeking to explore and expand the possibilities of computational design,
parametric modeling, and real-time sensor feedback have enthusiastically adopted
industrial robots. The efforts of early pioneers in the eld and the adoption of open
standards for programming and connectivity by manufacturers have lowered the
barriers to exploring the creative application of industrial robotics, allowing even
more creative practitioners to get involved. Digital fabrication combined with
open-source hardware and software has opened up the development of novel
technologies, interfaces, and methods to interdisciplinary teams of designers, artists,
and engineers.
Creative robotics offers new insights into the potential of robotics as researchers
and practitioners explore novel approaches to fabrication and interaction with
robotics. The flexible nature of industrial robotics presents an opportunity to
reconsider the entire design-to-production process, while the integration of
real-time sensor feedback has created opportunities for working with new materials
and processes that bring design and production closer.
ix
Preface
Rob|Arch 2016
Initiated by the Association for Robots in Architecture as a conference series
focusing on the use of robotic fabrication within a design-driven context, the now
established biennial Rob|Arch conference for Robotic Fabrication in Architecture,
Art, and Design provides an opportunity to foster a dialog between leading
members of the industrial robotic industry and cutting-edge research institutions in
architecture, design, and the arts. Launched by its founders Sigrid Brell-okcan and
Johannes Braumann in Vienna, Austria as Rob|Arch 2012, the Rob|Arch 2014
conference traveled to University of Michigan, Taubman College of Architecture
and Urban Planning, USA. In its third iteration, the Rob|Arch 2016 conference
arrives in Australia, hosted by the Faculty of Architecture, Design and Planning,
The University of Sydney: Australias rst university and academically known for
its multidisciplinary design education and long-standing tradition of knowledge
creation.
Research and Discussions

Robotic Fabrication in Architecture, Art, and Design is rst and foremost a platform
for sharing research developed across the eld of robotics, undertaken by designers,
artists, and architects, by researchers and educators, in academia and in practice and
industry, and by innovative forms and start-ups. Consequently, the call for papers,
together with the call for workshops and invited keynotes and industry papers,
reflects this open agenda.
Keynotes, Research and Project Papers

This publication presents a series of key sections that contribute to the Rob|Arch
2016 conference and workshop discussions; the invited keynotes, the research and
project papers, the workshop papers, and the industry papers.
The two keynotes address the conference theme of Trajectories in a dialectical
discourse. Mark Burry traces the development of material practices in Gauds
Sagrada Famlia from the century-old art of stereotomy toward 7-axis robotic
stone-cutting. Francois Roches mediation of new prosthetic entities breaches the
territory of conventions and expectations for robots as we have previously understood them toward uncertain processes and protocols of interactive behavior.
In the research and project papers, a number of themes emerge with the scientic
paper submissions for this years conference, which continue robotic fabrication
research and expand toward sensor-enabled processes and robotic interaction.
Amongst those, the aspect of engineering structures as a direct transfer from
Preface
xi
computational modeling to robotic manufacturing of joints, deposition of material,

or assembly plays an increasing role.
Assembly strategies are discussed here that include strategies of aerial robotic
construction; autonomous robotic assembly with virtual storage of material data; the
location-aware robotic laying of a brick wall; a cable robot system with visual
dynamic feedback for on-site construction; robotic assembly prototyping with
sensor-enabled material selection of components; and include a survey of scanning
techniques for surface descriptions in robotic assembly.
New material and work techniques developments with strong interest in structure
include a further advancement of robotic incremental sheet forming toward a
frameless stressed skin structure; the robotic smocking of sheet metal as developable surfaces; approaches for robotic enabled stress-line additive manufacturing
in curved surfaces; or the robotic 3D printing as compression-based material
deposition; and the robotically assisted welding of a grid shell structure.
Flexible and adaptive additive manufacturing strategies are presented as the
robotic multidimensional weaving print based on structural performance; and the
robotic positioning of a flexible fabric formwork. Fabrication and specically the
detailing of joints for stabs, plates, and modules are discussed; in the folded plate
shell with double through tenon joints; as multidimensional nger-joint in a rib
structure module; in the topology optimization and robotic fabrication of timber
space-frame structures; as wave jointed stereotomic construction; and extended
toward industrial application in a robotic fabrication of a free-form ceiling structure.
Toolpath planning plays an increased role for customizable stone structuring patterns; and for the micro-design of acoustically efcient disks.
The robotic hot blade cutting of double-curved geometries is presented here in
an 18-axis, tri-robot system for the cutting of doubly curved surfaces; as robotic
blade cutting of customized components for acoustic panels. Processes for
mass-customizable formwork for free-form geometries for mold fabrication include
the testing of reusable adaptive production strategies for concrete constructive
elements; the continued fabrication of formworks for deposition of ceramic 3D
printing, and the 3D printing of interlocking modules.
And nally, real-time and recursive feedback between algorithms, robot, and
material are explored as multiple-agent and robotic fabrication prototypes; the
integration of smart technologies and sensor loops in a multidisciplinary, open
design, and collaboration platform.
Workshop Papers
The practical hands-on workshops are a core part of the Rob|Arch conferences,
allowing individuals and teams from around the world to collaborate with leading
researchers and practitioners from academia and industry. The workshops foster
research networks across international teams of researchers and practitioners to
exchange knowledge about this exciting eld and to speculate on future trajectories.
xii
Preface
Many of the workshops represent cutting-edge research and practice under development and the workshop papers in this book present the ideas behind these
workshops in detail.
The selected workshops cover a diverse set of experimental approaches to
robotic fabrication from real-time humanmachine interaction to novel form nding
strategies. The workshop from the Southern California Institute of Architecture
explores user interfaces for live robotic control such that creative practitioners can
continuously engage and adapt to an evolving context. A collaborative team from
HAL Robotics, Bond University and Soundisplay presents an approach to natural
humanmachine interactions for context-specic object recognition in collaborative
robotics via ad hoc communication using voice and gestures. The paper from RMIT
explores an agent-based model of generative fabrication to imbue physical material
with digital agency with the aim of collapsing design and fabrication processes. The
paper from IAAC, Make It Locally and The University of Sydney presents a
framework for understanding the evolution of feedback loops within humanmachinematerial interactions in robotic fabrication workflows as sensors are introduced into processes. The workshop from the team from the University of
Technology Sydney and the University of Michigan explores the potential for
expanding traditional architectural form nding processes by embedding algorithmic design methods and robotic fabrication strategies.
Industry Papers
As in previous years, the industry papers provide a strong voice from the Rob|Arch
2016 industry, partners and sponsors, and more excitingly so, this time bridging
between continents, from Europe over to the United States to Australia. These
papers offer insights into the most recent industry developments in the context of
robotic fabrication. KUKA presents the LBR iiwa, and the new mxAutomation
interface that allows direct robot control in interaction with modern industrial
real-time communication, and thus enables entirely new, flexible workflows from
design to production towards fabricating highly customizable products in the creative industry. ABB showcases collaborative robotics with its new series of sensitive robotsthe ABB YuMithat can assemble and collaborate safely with
humans due to the ABB ForceControl which allows robots to react to the forces that
are applied to their end-effectors. Autodesk presents the application of its novel
visual scripting interface Dynamo for a collaboration and robotic fabrication
workflow of the construction of spatial structures. SCHUNK shows an approach for
efcient creation of form-tting and flexible gripper design. ERNE Holzbau reports
on one of the largest robots for building component manufacturing in Europe; a
multifunctional 7-axis machine that can manufacture large building components on
an industrial scale. Delcam presents the PowerMILL Robot, a software system that
provides an easy to use computer interface allowing the programmer to design,
analyze, and simulate in a single virtual environment.
Preface
xiii
Outlook
Reflecting the theme of the conference, a number of trajectories can be identied
from the work presented in this volume that offer insights into the future of robotic
fabrication in architecture, art and design.
The integration of sensors into robotic fabrication continues to be a theme that
has run through the Rob|Arch conferences to date, from enabling tolerance for
material and environmental variability in the fabrication process, essential for
on-site construction robotics, to assisting with the planning of workflows and the
live control of robotic systems. Beyond the integration of sensors, the increasing
application of intelligent control systems take advantage of the feedback at different
scales, increasing the level of autonomy of the robotic systems and opening up new
materials and processes that require constant monitoring and adaptation. Beyond
the current state-of-the-art of robotic fabrication there lies creative robotics, the
evolution of an embodied computational creativity capable of sustained creative
practices without human interventionopening up new horizons for human
machine collaboration.
Advances in sensors and intelligent control systems have highlighted the agency
of materials in the fabrication process, both as a result of the increased sensitivity of
robotic systems to the effects of machining and the ability of robotic fabrication
systems to react to changing environmental conditions. Perhaps seeing the start of a
generational change in thinking about construction automation, we are increasingly
seeing researchers radically rethinking how we shape our environment through
different forms of robotic fabrication. The scale and ambition of these approaches is
increasing with every conference cycle and we can look ahead to a future where
robotic fabrication will change every aspect of our built environment from the
smallest component to whole cityscapes.
Industrial robots provide the ideal platform for experimenting with fabrication
processes being both flexible to changing requirements and standardized across
working environments to support the transfer of new knowledge between research
groups. In this third iteration of the Rob|Arch conference, we continue to see the
benets of the knowledge transfer between researchers, practitioners and industry
partners. In particular, the opening up of industrial robotics to experimental
approaches and creative explorations is dramatically accelerating progress in the
eld. As new robotics technologies are developed, researchers and innovators in the
creative industries will continue to adapt these tools and transform practices to take
advantage of the opportunities they bring. The potential of adopting new automation methods, especially robotics, has become part of the public discourse on the
future of work as part of a second industrial revolution. Within the creative
industries robotics fabrication is affecting the whole design process, to the point of
challenging what it means to design.
xiv
Preface
Acknowledgments
The conference chairs would like to thank KUKA Robotics who devoted themselves to make this conference and scientic book possible, ABB, Autodesk and
BVN and our other sponsors for their main support, as well as our advisory board,
and the Association for Robots in Architecture for the opportunity to organize the
conference. In addition we would like to thank the Scientic Committee, composed
of architects, engineers, designers, and robotic experts; without their help it would
not have been possible to develop the quality of work presented within.
The Architecture Robotics Lab since its rst initiation in 2013 with the Robot
Symposium in collaboration with the Association would not have been possible
without the visionary engagement of our Faculty Dean, John Redmond, and
Associate Dean of Research, Peter Phibbs, to whom we express our gratitude for
their continued and generous support. Special thanks go to our Workshop Co-chair,
Marjo Niemel, to Sarah Breen Lovett, and the entire team at DMaF for their
tireless support. We would like to extend this to the entire team at the Faculty of
Architecture, Design and Planning, including both staff and faculty, who have
supported the development of the conference. We would also like to thank our team
partners in peer institutions who collaborate with their international partners to host
the workshops, our local friends and colleagues at Bond, UTS, UNSW, Monash,
RMIT, and our international partners at University of Michigan, ICD Stuttgart,
IAAC Barcelona, and Harvard, and industry workshop partners Odico and Hal.
Finally, special thanks to Springer Engineering for their assistance in editing and
publishing these proceedings.
December 2015
Dagmar Reinhardt
Rob Saunders
Jane Burry
Contents
Part I
Keynotes
Robots at the Sagrada Famlia Basilica: A Brief History

of Robotised Stone-Cutting . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Mark Burry
Psychaestenia . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Francois Roche, Camille Lacadee and Stephan Henrich
Part II
3
17
Scientic Papers
Building a Bridge with Flying Robots . . . . . . . . . . . . . . . . . . . . . . . . .

Ammar Mirjan, Federico Augugliaro, Raffaello DAndrea,
Fabio Gramazio and Matthias Kohler
35
Autonomous Robotic Assembly with Variable Material Properties. . . . .

Michael Jeffers
49
An Integrated Modelling and Toolpathing Approach

for a Frameless Stressed Skin Structure, Fabricated Using
Robotic Incremental Sheet Forming . . . . . . . . . . . . . . . . . . . . . . . . . . .
Paul Nicholas, David Stasiuk, Esben Nrgaard, Christopher Hutchinson
and Mette Ramsgaard Thomsen
Robotic Lattice Smock . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Andrew Saunders and Gregory Epps
Robotic Multi-dimensional Printing Based on Structural
Performance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Philip F. Yuan, Hao Meng, Lei Yu and Liming Zhang
63
79
93
Fabric Forms: The Robotic Positioning of Fabric Formwork . . . . . . . . 107

Ron Culver, Julia Koerner and Joseph Saraan
xv
xvi
Contents
Path Planning for Robotic Artistic Stone Surface Production . . . . . . . . 123

Gregor Steinhagen, Johannes Braumann, Jan Brninghaus,
Matthias Neuhaus, Sigrid Brell-Cokcan and Bernd Kuhlenktter
Towards a Micro Design of Acoustic Surfaces . . . . . . . . . . . . . . . . . . . 137
Dagmar Reinhardt, Densil Cabrera, Alexander Jung and Rod Watt
Robotic Hot-Blade Cutting . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 151
Asbjrn Sndergaard, Jelle Feringa, Toke Nrbjerg, Kasper Steenstrup,
David Brander, Jens Graversen, Steen Markvorsen, Andreas Brentzen,
Kiril Petkov, Jesper Hattel, Kenn Clausen, Kasper Jensen, Lars Knudsen
and Jacob Kortbek
Part III
Projects
Fabrication-Aware Design of Timber Folded Plate Shells

with Double Through Tenon Joints . . . . . . . . . . . . . . . . . . . . . . . . . . . 167
Christopher Robeller and Yves Weinand
RBDM_Robodome: Complex Curved Geometries
with Robotically Fabricated Joints . . . . . . . . . . . . . . . . . . . . . . . . . . . . 179
Alexander Jung, Dagmar Reinhardt and Rod Watt
Topology Optimization and Robotic Fabrication of Advanced
Timber Space-Frame Structures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 191
Asbjrn Sndergaard, Oded Amir, Phillip Eversmann, Luka Piskorec,
Florin Stan, Fabio Gramazio and Matthias Kohler
Mobile Robotic Brickwork . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 205
Kathrin Drfler, Timothy Sandy, Markus Giftthaler, Fabio Gramazio,
Matthias Kohler and Jonas Buchli
Closeness: On the Relationship of Multi-agent Algorithms
and Robotic Fabrication . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 219
Roland Snooks and Gwyllim Jahn
The SPIDERobot: A Cable-Robot System for On-site
Construction in Architecture . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 231
Jos Pedro Sousa, Cristina Gass Palop, Eduardo Moreira,
Andry Maykol Pinto, Jos Lima, Paulo Costa, Pedro Costa,
Germano Veiga and A. Paulo Moreira
Developing Architectural Geometry Through Robotic
Assembly and Material Sensing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 241
Kaicong Wu and Axel Kilian
Contents
xvii
BotBar: A Platform for Multi-disciplinary Design Education. . . . . . . . . 251

Marjo Niemel, Samantha Horlyck, Susana Alarcon-Licona,
Dylan Wozniak-OConnor, Gabriele Ulacco, Rodney Watt
and Rob Saunders
RECONstruction. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 263
Joshua Bard, Richard Tursky and Michael Jeffers
Robotics-Based Prefabrication in Architecture . . . . . . . . . . . . . . . . . . . 275
Xun Li, DongHan Shin, JinHo Park and HyungUk Ahn
Stereotomy of Wave Jointed Blocks . . . . . . . . . . . . . . . . . . . . . . . . . . . 285
Simon Weir, Dion Moult and Shayani Fernando
Crafting Robustness: Rapidly Fabricating Ruled Surface
Acoustic Panels . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 295
Nicholas Williams and John Cherrey
From Analysis to Production and Back Attempts and Results
of Reusable Adaptive Freeform Production Strategies for Double
Curved Concrete Construction Elements . . . . . . . . . . . . . . . . . . . . . . . 305
Felix Amtsberg, Gernot Parmann, Andreas Trummer and Stefan Peters
Free Form Clay Deposition in Custom Generated Molds . . . . . . . . . . . 317
Kate Dunn, Dylan Wozniak OConnor, Marjo Niemel
and Gabriele Ulacco
Solar Bytes Pavilion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 327
Brian Peters
Materially Informed Design to Robotic Production: A Robotic 3D
Printing System for Informed Material Deposition . . . . . . . . . . . . . . . . 339
Sina Mostafavi and Henriette Bier
Robotics-Enabled Stress Line Additive Manufacturing . . . . . . . . . . . . . 351
Kam-Ming Mark Tam, James R. Coleman, Nicholas W. Fine
and Caitlin T. Mueller
BUILD-ing the MASS Lo-Fab Pavilion . . . . . . . . . . . . . . . . . . . . . . . . 363
Nathan King, Nathan Melenbrink, Nick Cote and Gustav Fagerstrm
Part IV
Workshops
Robot UI . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 377
Curime Batliner, Michael Jake Newsum and M. Casey Rehm
Towards On-site Collaborative Robotics. . . . . . . . . . . . . . . . . . . . . . . . 389
Thibault Schwartz, Sebastian Andraos, Jonathan Nelson, Chris Knapp
and Bertrand Arnold
xviii
Contents
Stigmergic Accretion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 399

Sensors and Workow Evolutions: Developing a Framework
for Instant Robotic Toolpath Revision . . . . . . . . . . . . . . . . . . . . . . . . . 411
Alexandre Dubor, Guillem Camprodom, Gabriel Bello Diaz,
Dagmar Reinhardt, Rob Saunders, Kate Dunn, Marjo Niemel,
Samantha Horlyck, Susana Alarcon-Licona, Dylan Wozniak-OConnor
and Rodney Watt
Towards Real-Time Adaptive Fabrication-Aware Form Finding
in Architecture . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 427
Dave Pigram, Iain Maxwell and Wes McGee
Part V
Industry Papers
Direct Robot Control with mxAutomation: A New Approach

to Simple Software Integration of Robots in Production Machinery,
Automation Systems, and New Parametric Environments . . . . . . . . . . . 441
Heinrich Munz, Johannes Braumann and Sigrid Brell-Cokcan
Collaborative Robotics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 449
Frank-Peter Kirgis, Peter Katsos and Martin Kohlmaier
Flexible Gripper Design Through Additive Manufacturing . . . . . . . . . . 455
Marcel Nagel, Felix Giese and Ralf Becker
Individual Serialism Through the Use of Robotics in the Production
of Large-Scale Building Components . . . . . . . . . . . . . . . . . . . . . . . . . . 461
Martin Krammer
PowerMILL Robot: Simplifying the Complex Issue
of Robotic Programming . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 469
Robert Simunic
Erratum to: The SPIDERobot: A Cable-Robot System
for On-site Construction in Architecture . . . . . . . . . . . . . . . . . . . . . . .
E1
Scientic Committee . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 473
Part I
Keynotes
Until 1989 nearly all stone was dressed on site using saws where possible, and manual techniques
to complete the task: highly skilled work, tough on the muscles as well as the ears and lungs
Robots at the Sagrada Famlia Basilica:

A Brief History of Robotised
Stone-Cutting
Mark Burry
Abstract Cutting stone by hand to the architects precise measurements is an

ancient craft using one of the oldest materials known to humankindtraditionally it
is a highly laborious undertaking. Curiously the efforts taken to continue constructing Gauds magnum opus long after his death in 1926 included the introduction of 2D robots to the project in 1989, preceding the introduction of
computer-aided design a little later. Meeting the challenges of speeding-up the
preparation of dressed stone took precedence over modernising to include digitally
assisted stereotomy (the geometrical art of efcient stone-cutting) in the design
studio. This paper highlights the extraordinary leaps that have been made in the
intervening quarter century. From humble but early adoption of relatively primitive
equipment this has led to 7-axis robot stone preparation in use now for over a decade
at the time of writing. The particulars of this adoption and adaptation emphasise the
advantages that designers have when they embrace emerging technology as closely
as possible at the rst opportunity by setting aside over anxiety about risk.
1 Enter the Robots: The Nave Columns (1989 Onwards)

First time visitors to the interior of the Sagrada Famlia Basilicainaugurated by
the then Pope in 2010, frequently remark on the quality and use of the stonework,
not least the colours of the various stones used for the extraordinary branching
columns that dominate the space (Fig. 1). Gaud had scientically tested the
load-bearing capabilities of many local and more exotically sourced stones seeking
optimal strength for the various loadings that these, the main supports to the
building, would be subjected. There is a spatial hierarchy to the columns used
throughout the building, which coincides with the different loadings. Four principal
M. Burry (&)
Melbourne School of Design, University of Melbourne, Melbourne, Australia
e-mail: mark.burry@unimelb.edu.au
URL: http://www.mcburry.net
D. Reinhardt et al. (eds.), Robotic Fabrication in Architecture,
Art and Design 2016, DOI 10.1007/978-3-319-26378-6_1
M. Burry
Fig. 1 Nave columns as built and a geometrical description of their derivation
columns in the centre of the building outline the crossing and do the bulk of the
work in supporting the main tower that will reach 172.5 m in height when completed. These columns are made from porphyry, the most resistant stone found in
nature, visually exploiting its unique burgundy colour. This group of four columns
branches up to support the ceiling vaults 60 m above, and are flanked by a ring of 8
columns made from dark grey basalt, which also share some of the work supporting
the central dome as well as the four towers that encircle the main tower.
The crossing has a lateral transept on each side, and a longer nave extending
back towards the Glory Fa adethe main entrance, the design for which will be
nalised in 2016. The central nave and transept vaults are 45 m high with clerestory
lighting running along their lengths, with 30 m high lateral nave vaults on either
side. The taller columns dening the central aisle space are made from pure grey
granite and the shortest columns of this series, supporting the lateral nave vaults, are
made from local Montjuc sandstone, ivory-pink in colour.
There is a slenderness ratio for all of the columns of 1:10 meaning that the tallest
columns, supporting the most weight, are the broader of this series of four, and the
shortest are therefore the slimmest. The generation of the column geometry is highly
explicit, and beguilingly simple given the apparent complexity of the resulting
column geometry. Gaud used the same generative approach for all columns, all
parametrically variable to distinguish each column stone type and height from the
others. All columns are based on intersected polygons with its consequential apex
and re-entrant angularity softened by using respectively cotangential convex and
concave parabolas to form cotangentially the distinctive undulating base prole
found in all the columns. The shorter column is based on two intersecting triangles
counter-rotated by 60the 6-sided sandstone column. The eight-sided granite
Robots at the Sagrada Famlia Basilica
column prole is generated from two intersecting squares counter-rotated by 45, the
basalt column from two pentagons, and the porphyry columns from two intersected
hexagons (Fig. 1).
At the column bases the prole comprises two proles superimposed, but each
prole subtly rotates as the column gets higher, one clockwise and the other to the
equivalent degree anticlockwise. The horizontal section of the column at any point
is the intersection of the counter-rotated proles, which accounts for the gradual
appearance of vertical arises on the column surfaces as they rise from the floor,
eventually forming the characteristic Doric fluting higher-up.
This is, of course, the briefest of description albeit rather laboured in the context of
this paper, but it is necessary in order to explain, in part, what was behind the introduction of a 2D robot saw to assist in the production of the column drum elements.
Every section of the columnwhether taken through the base or at the top, or
anywhere in between, is comprised of a series of similar convex and concave
parabolas: always the same curve. The g-code generation of the saw paths was
therefore relatively easy to generate: the same piece of code that factored in a slight
rotation for each incremental 5.5 mm shift up that stones length as the procedure
progressed, which was the width of the saw blade.
The cutting speed varies obviously with the relative hardness of the stone being
dressed. For the granite columns there are eight pieces of dressed stone for each
metre high drum. The column is 14 m high requiring 112 pieces in total per column,
taking 36 hours of robotised cutting per piece. Each column therefore took up to
168 days of round-the-clock cutting, tying-up the machine24/7. Frustratingly, for
every 3 minutes the saw blade was in operation only 1 minute was employed
actually in contact with the stone; the rest of the time it was waving around in the
air in some complicated robotic choreographic prance seeking accurate alignment
before contacting the stone anew.
There were three methods deployed to speed-up the process, because although
far quicker than the alternative, which was hand dressing, the sheer number of the
columns meant that these rates of progress were too slow. The rst method was to
remove as much of the stone as possible from relatively inexact hand operated
machine-cutting (Fig. 2). The second was to deploy robotic duplicators; once each
Fig. 2 Preparing the stone elements by saw to be as close as possible to the nished sizes prior to
being offered to the robot saw drawing on hand-drawn digitised data
M. Burry
piece was cut by the 2D robot, it was copied into several pieces at once by a robot
with multiple saw blades using a 1:1 mechanical pantograph to guide them. The
third method was the more interesting one with respect to digital fabrication and
programming: direct intervention with the machine code.
Bizarrely the information required to guide the saw blade was extracted from 1:1
hand drawn proles using a 2D point digitiser and a drawing board next to the robot
saw onsite (Fig. 2). These coordinates fed into the g-code generator, a software
program provided by the saw manufacturer. Unfortunately the manufacturer
obliged the operation to be undertaken purely through the use of their own protected software without giving the technical team on site any opportunity to
intervene directly with the machine path coding, making it impossible therefore to
have any direct intervention with the way the robot went about its business. The
problem was actually quite a simple one: no matter where it was positioned in
relationship to the height of the column, the saw behaved as if it was still cutting the
base prole, which it displaced through slight incremental shift in rotation angle as
it made its way up the column.
As the column pieces become more and more fluted, there is far less parabola
length being pursued yet the machine saw kept on cutting virtual stone as if indeed
that were its task (Fig. 3). The manufacturers appeared not to have provided any
means for the g-code generator to optimise its path as it progressed up the columns.
Ironically this greatly encouraged the technical team to not only look at alternatives
to what felt like enslavement to a software that allowed for no direct intervention
but also to seek a different kind of contractual engagement with quarries and
stonemasons, one that would combine the digital representations skills of the
architects involved, with the digital fabrication skills of the stonemasons. But rst
there was an intermediary step on the path to innovation based on human ingenuity
alone and no robots at all: the production of the Passion Faade rose window.
Fig. 3 Robot saw cutting and the nal result (8 sided central nave column)
2 Lessons from the Stonemasons Ingenuity: Passion

Faade Rose Window (20002001)
In order to celebrate the Christian millennium, a decision was taken to produce a
single spectacular element in record time: the rose window in the Passion Faade,
the west facing transept side entrance (Fig. 4). Internally measuring 18 m high and
8 m wide, externally it was 25 m highcomprising 18 courses of stereotomic
jigsaw puzzling to exploit the constituent hyperbolic geometries deployed for this
window, as had been used for all of Gauds nal designs deployed elsewhere in the
building (Fig. 5). The quarry providing the granite was based near Lugo in Galicia,
a region in Spain 1,200 km from site. The late proprietor, Sr. Manuel Mallo, was a
master at reducing technical challenges from being apparently impossible to relatively achievable through a deep-seated knowledge and pragmatism.
Fig. 4
The rose window is designed for the space between the Passion Faade towers
Fig. 5 Virtual sculpting of the rose window using Boolean Operations (left to right)
M. Burry
Fig. 6 Sr. M Mallo and his hand-made polystyrene rapid prototypes (left), and the rst digitally
produced rapid prototypes made from wax (right)
This period coincided neatly with the introduction of rapid prototyping to

architecture studios for the rst time. Rapid prototyping proved to be indispensible
for the Sagrada Famlia Basilica project at the time, and still is today, but not for
Sr. Mallo, who found it quicker, cheaper, more convenient, and far more revelatory
to prototype manually using polystyrene and a hot wire. In fact he adapted this
process subsequently to 1:1 scale in order to provide physical facsimiles to guide
the masons engaged in producing each piece (Fig. 6).
The role of the technical ofce was to provide over 780 A0 sized 1:1 templates
that the quarry used to build the polystyrene facsimiles. Innovation was required at
both ends of the operation: parametric modelling drawn from the aeronautical
industries, introducing scripting to relieve the burden on the template producers
based in Australia, and abstraction of the constituent spatial geometries to full-scale
templates by the architects, and interpretation of the templatessterotomic traits
extending from the great French descriptive geometers who preceded the Industrial
Revolution, by the masons based on site (Fig. 7). The project was based on a
just-in-time workflow which, at one point, had the lowest quarter of the window
being installed on site, the next quarter being prepared at the distant quarry while
the next quarter up was still being templated in Australia. Meanwhile the top quarter
Fig. 7 CADDS5 parametric model for rose window driven directly from an Excel
spreadsheet leading to the 780 traits (full-size templates) produced as A0 drawings using
Autocad
of the window was still being negotiated between the university-based architectural
teams based on opposite ends of the globe.
In terms of production, this was the rst time Gauds profusion of intersected
hyperbolic paraboloids and hyperboloids of revolution had been produced directly
in stone at this scale, so the masons were obliged to innovate along the way;
previously these geometries were mostly deployed towards mould-making for the
production of articial stone elements. The stonemasons rst operation was to cut
the stone as it emerged from the quarry to a block as close to the nal surface as
possible, principally through blasting. They then reduced the stone to within a
centimetre or two of its nished surface using a diamond-encrusted wire saw. The
wire was pulled between two wheels with the stone block to be cut placed in
between. The cutting wire was calibrated to drop through the stone at a speed that
equalled that of the wires ability benches. Sr. Mallo was able to exploit fully the
doubly ruled surfaces that Gaud deployed: two operatives twisted the stone block
in space while the wire descended vertically through its cutting plane. The architects templates were used to scribe the prole at each horizontal joint, which were
precisely cut planes at each end of the block. The operative at each side of the stone
simply twisted their end of the block as the cutting wire dropped to ensure that it
passed as close to the nal ruled surface as possible leaving the last centimetre to be
cut by stonemasons by hand (Fig. 8).
The successful completion of this project on time nevertheless induced cause for
reflection to all involved (Fig. 9). This particular quarry, as well as others close to
hand, drew lessons from this hybrid digital-analogue experience and were prompted
to consider more acutely the relative advantages of pursuing a path to fully integrated digital fabrication (Fig. 10).
Fig. 8 Digitally assisted stone masonry whereby the templates scribed onto the ends of the stone
block for the operatives to use to guide the diamond encrusted cutting wire as it descends
Fig. 9 This method efciently produced 18 courses of masonry with absolute precision
10
M. Burry
Fig. 10 The precision of this production method can be seen in the 10 mm gaps between each
element, a tolerance hardly required (left hand), with images of the completed interior
3 7-Axis Robotic Stone Cutting: The Passion Faade

Narthex Columns (20012015)
As had been the case with the Passion Faade rose window, the design team for the
upper regions of the narthex located immediately below the window described
above collaborated with the stonemasons from the point of the rst decision, in this
case Jordi Barbany, based 40 km to the northeast of Barcelona: somewhat more
convenient to site than Lugo in Galicia.
Jordi Barbany is a sculptor as well as third generation stonemason. He is also
both a consummate innovator as well as a master of 7-axis robotics. Collaborating
with the design team, stone-cutting options could be considered in concert with the
design development of a complex spatial assembly more sculptural than architectural in its nature.
The team experimented virtually with controlling a diamond encrusted cutting
wire with robots rather than human operatives guiding the stone block in space
while the wire did its cutting as successfully deployed for the Passion Faade rose
window described above. Despite being able to cut more than 95 % of the
prospective pieces using this method it in fact turned-out to be more efcient to
deploy a 7-axis robot to do the cutting and dressing. The potential of this approach
was further augmented during the life of the project: the columns, approximately
9 m high, were originally conceived and developed as base and capital, with the
column shaft being prepared in three 2 m high sequential elements5 elements in
total. Having chosen to use the 7-axis robot the process rened to the extent that the
previously envisaged three column elements eventuated as a single piece cut with
remarkable precision from a piece of granite, measuring more than 6 m in length.
The resulting columns are therefore comprised of three elements in totalbase,
11
shaft, and capital. Despite the sculptural complexity of the faade it was constructed
with absolute precision to the millimetre, a testament to the technical skill,
inventiveness and dedication of Jordi Barbany and his stonemasons (Figs. 11, 12,
13, 14 and 15).
Fig. 11 Except from Gauds original sketch, early efforts to nd a suitable interpretation using
hand-modelling techniques, and the initial parametric digital model
Fig. 12 Early wax rapid prototype (left hand side), the nal digitally designed column (centre
image), completed parametric design model
12
M. Burry
Fig. 13 The process whereby stone is cut to order at the quarry at Tarn (France) leading to its
whittling-down using a diamond wire, then nishing with the 7-Axis robot saw arrangement
Fig. 14 Human interaction between diverse members of the team with a variety of media: Jordi
Barbany (master mason), Dr Jane Burry (project architect), Toni Caminal (technical architect),
Jordi Barbany, Jordi Bonet (Chief Architect), and Jordi Faul (Assistant Chief Architect) (left to
right)
13
Fig. 15 Columns under construction (2014) with three 1:1 column prototypes in position for
testing the visual effect shown in right hand side image (2006)
4 Concluding Remarks
This brief account offers insight into the advantages of a coherent design schema
pervading an entire project (the use of doubly ruled surfaces throughout the project)
and the internalization of contractual arrangements such that all involved innovate
together as a single family. Most of all, this is an account of how technological
innovation is spawned along the route of highly risky design of this nature: in this
case representational and constructional adversity being the mother of innovation in
sponsoring the extended use of robots to assist in the production of highly complex
pieces of masonry, the worlds oldest building material in every sense.
Clearly the Sagrada Famlia Basilica has the advantage of being a uniquely
slow-burning project, offering the luxury of contingent innovation based on
necessity tempered by the flow and continuity of a project presenting extraordinary
design challenges with a long duration. Making even better use of available
robotised machinery beyond their manufacturers expectations through direct digital CAAD input from the architects avoided externally contracted coding agency.
This was stimulated by the need to make the most of the digital fabrication
opportunities that the robot presented on site, not least through the motivation to
limit dependence on the robot manufacturers mechatronic coding. Seen as a mutual
challenge to both designer and maker the project continues to be at the cusp of
technological advances in the production of dressed stone for building (Fig. 16).
14
M. Burry
Fig. 16 Jordi Barbany and

team standing next to one of
the completed Passion Faade
columns, made in three pieces
Acknowledgments The design team for the Passion Faade narthex consisted of Jordi Bonet and
Xisco Llabrs in Barcelona, and Mark Burry and Jane Burry in Melbourne. Much of the design
research reported here was signicantly funded by the Australian Research Council. This additional support has been within the framework of a longstanding commitment by the Sagrada
Famlia Basilica Foundation to commission university design research teams to assist investigations on site into innovative design, design representation, and digital fabrication.
Credits for stone production: Sagrada Famlia Basilica should be noted as follows:
Stonemasons Marbres Juyol: Sr. Alfons Juyol i Arenas, (LHospitalet de LlobregatBarcelona)
produced the granite columns for the central nave, transepts and apse, as well as basalt columns
and other major interior stone elements (principal altar).
Stonemasons Talleres de Manuel Mallo: Sr. Manuel Mallo (RbadeLugo) produced the rose
window for the Passion Faade, and other signicant interior parts (sandstone columns in the
lateral nave).
15
Stonemasons PCM Granitos Moldurados: Sr. Octavio Vazquez (RbadeLugo) produced the
crest to the Passion Faade narthex and other major elements for the building exterior.
Stonemasons Granits Barbany: Sr. Jordi Barbany i Triad (Llinars del VallsBarcelona)
produced the columns for the Passion Faade and major external elements for the emerging towers
in the centre of the building.
Psychaestenia
Psycho-Case Studies by New Territories/M4
Francois Roche, Camille Lacadee and Stephan Henrich
Abstract At the end of the eighteenth century, as the French Revolution challenged Europes political order and the Industrial Revolution transformed the world
economy, an English merchant and political activist named James Tilly Matthews
became convinced that his mind was being controlled by a machine. According to
Matthews, a gang of radical French Jacobins had inltrated England, bringing with
them the knowledge and means to construct a mechanism called the Air Loom. By
producing invisible gasses and magnetic elds, the machine could manipulate a
victims mind and body from afar. Matthews described the Air Looms effects and
inner workings to anyone who would listen, detailing how it could make him speak
like a puppet, or force his brain to accept an idea, all with the simple pull of a lever.
Psychologists have since reported that schizophrenics and autistic children often
employ mechanistic imagery to articulate basic psychic experiences seemingly
outside of their control (James Tilly Matthews, The Airloom).
Keywords Psycho-robotic protocols and mind machine fabrication
apparatuses
Pataphysics
1 Introduction
There are many machines, so many desirable machines that in fact pretend to do
more than they are doing. In the pursuit of pataphysicsthe branch of philosophy
that deals with the imaginary realmthey never reveal their deep natures: whether it
is their lineage or their illusionary appearance, their genuine qualities or their sham
features. Simultaneously speculative, ctional and accurately and efciently productive, these machines navigate the world of Yestertomorrowday, with happiness
and innocence, walking briskly through the mountains of rubbish of the 21st century
F. Roche (&) C. Lacadee S. Henrich
Lab M4 (mindmachinemakingmyth), Bangkok, Thailand
e-mail: fr@new-territories.com; francoisroche61@gmail.com
URL: http://www.new-territories.com/props.htm
John Wiley & Sons Ltd. 2014. Published by Springer International Publishing Switzerland 2016.
All Rights Reserved.
17
18
F. Roche et al.
and beyond. These pataphysical machines articulate symmetricallythrough weird

apparatusesdifferent arrows of time, different layers of knowledge, but more
efciently they negotiate the endless limits of what we could consider the territory of
absurdity, where illogical behaviour is protocolised with an extreme logic of
emerging design and geometry, where input and output are described by mathematical rules
Neither a satire of this and other worlds, nor a techno-pessimism or a
techno-derision, these machines reside at the very limits of the dystopian or they
constitute the limit between the territory of conventions, of certainties and stabilities, where it is comfortable to consider everything legitimated by an order, or an
intuition of an order, and by other territories; all the other paranoid, phantasm-like
imaginings reported back by travellers.
In a casual and basic sense, machines have always been used to elaborate
technicism as the extension of the hand, through its replacement, its improving, its
acceleration of the speed and powers of transformation, of production. However, it
seems very naive to reduce the machine to this obvious objective dimension, in a
purely functional and mechanical approach; limiting it exclusively to a Cartesian
notion of productive power, located in the visible spectrum of appearance and fact.
In parallel, machines are producing artefacts, assemblages, multiple associations
and desires, and are inltrating the very raison dtre of our own bodies and minds
that are codependent on our own biotopes or habitats. Fundamentally, everywhere
in nature, at the origin of all exchange processes, in the transaction of any substances, they are the guarantee of its vitalism. Machines coexistence with nature
renders them in effect a paradigm of the body. This is true of all processes, protocols and apparatuses, where transitory and transactional substances constitute and
affect simultaneously all the species, where machines identities and outputs are
both object and subject.
In pursuit of this polyphonic approach, we cannot pass over the notion of the
bachelor machine as a tentative attempt to integrate mechanical apparatuses in a
narrative transaction and transmutation (in the mode of the alchemist). This is the
opposite approach to a headlong critique or denouncement of capitalism that
highlights the substitution of craftsmen with unskilled workers manning machines
(the natural consequence of this now being a mechanical system without workers).
Walter Benjamin described this shift as a move from the singularity of production
to mass production. This contrasts with the nostalgic romanticism that bachelor
machines evoke through our fascination with their sophisticated human-made-like
construction: their eroticism or even barbaric eroticism. The impulsive urge and
gut-wrenching repulsion they generate means they exist in a permanent state of
schizophrenia, vacillating between the simultaneous potential of production and
destruction. Both positive and negative processes are the product of the same
industrial system; their genesis is consubstantial, and their collateral effect diametrically in opposition. They are both dependent on this schizoid potential.
Following are a few examples of those pathological strategies of narrationproduction.
Psychaestenia
19
2 Machines
2.1
Darwinian StarGate
Instructions: Stand up and face the ghosts in the depths of your private garden! Rent
this vehicle to transport yourself from a seated, peaceful, sleepy archaic body
posture to a standing, lucid awakened position that induces bravery in those faced
with the present. Powered by photovoltaic cells, the Darwinian Star-gates arms
unfold on their way from a panoptical to a worrisome heterotopic space you would
normally refuse to see (Fig. 1).
The star-gate machine introduces the passage of time between two constructions
of different origins and periods. As a strategy for questioning the orientation of the
arrow of time, it is able to quieten the anxiety of misunderstanding provoked by the
shifts between the Modern to Postmodern, Postmodern to Digital, and Digital to
Robotic Computational extension, in a beam me up Scotty rhyzomatic shortcut.
The travel could take an evolutionary and/or regressive trajectory. But mainly this
machine is most efciently used as a vector of discovery that reaches a point of
uncertainties, of un-determinism, to escape from a zone where everything has
already been flattened, classied and validated. Its rst use and development was
for thebroomwitch experiment.
Precautions for use: Using the vehicle too often might cause a sensation of time
deprivation and sometimes immortality, but also ultimately a good excuse for denial
Fig. 1 Stargate. New-Territories/Francois Roche-Stephan Henrich/Broomwitch/2009/ http://

www.new-territories.com/broomwitch.htm
20
F. Roche et al.
of your duties in a given time. You might also lose the sense of time passing, which
can signicantly impair synchronising motor actions. Abuse of the device can be
extremely dangerous for mental health and seriously affect the users temporal
perception, especially in regards to the notion of a specious present. Ultimately, it
can cause memory loss. On the contrary, overexposure to the present time (staying
in one time or another) might cause the user depression, cynical behaviour or other
pathological distresses, which the vehicle shall not be held responsible for The
device does not work for French architecture, which already confuses its origins.
2.2
Antipersonal Nymphomanic Wanderer
Instructions: Rent this machine to brave the danger and retrieve on her back rotten
species, decomposed biomass, from any no-mans land. The Wanderer (Fig. 2)
can be transformed for collecting other materiel. All robot tuning of terminations,
articulated arms, legs and tips is authorised, under the condition that you return the
machine in its initial state. The machine collects any ingredients to be recycled in a
new productive use. This grants a second life to the waste, and the trash in polluted
Fig. 2 Wanderer. New-Territories/Francois Roche-Stephan Henrich/He shoot me down/2009/

http://www.new-territories.com/he%20shoot%20me%20down.htm
Psychaestenia
21
areas such as post-military zones with unreachable infrastructure interstices.

Legends and fairytales are simultaneously transported out of the deepness of those
abandoned situations, as in a Stalker experiment to touch the unknown. Please
take care of the backlash of those creatures. Its rst use and development was for
the itshootmedown experiment.
Precautions for use: The machine is originally built with a very high
self-estimation sensorial device, as well as a danger-blinding component, both
necessary for its brave actions and responsibilities. However, depending on the
environment it is exposed to, the machine could be subject to sudden and violent
changes in self-esteem. In case of failure or small breakdown (if the danger-blind
component gets hurt), the machine will exhaust itself until suicide. If you notice that
the machine repeats a very high exposure to dangerous situations, put it off to avoid
risks of suicidal tendencies disguised into bravery.
2.3
Introverted Eczemental Recycler
Instructions: Transforming informal heaps into deformed ones, this machine

recycles waste from metallurgic and construction sites into potential troglodyte
morphologies. This machine (Fig. 3) is still in development. It is thus available for
rent under a special discount as a beta test. The provision of sufcient energy levels
Fig. 3 Recycler. New-Territories Academia/Francois Roche-Stephan Henrich/Adam Orlinski/

LAB Anguewandte/ 2009
22
F. Roche et al.
for the effective gathering of steel has not yet been accurately gauged, and dysfunctions may easily occur when the Recycler is in operation. We recommend for
this machine only to be rented in parallel with the crane that is able to stabilise its
agenda and positioning. We require feedback from customers to improve the reasons behind its design, which appear for now weak. This machine will be removed
from the catalogue if there are no further reasons for its existence.
Precautions for use: Due to its lack of resolution, this machine is especially
vulnerable. Protect it from the feeling of identity loss by engaging with it on a
private levelotherwise it might show a tendency to confuse its own being with the
built environment, and develop skin camouflage diseases in order to disappear
inside its own construction. An early stage of depersonalisation can be spotted by
its tendency towards metallic somatisation.
2.4
Difference and Repetition/Intricate Randomizer
Instructions: Rent this device to populate a surface that will be revealed by the
trajectory you convey by impulse to the machine (Fig. 4). Its multiple arms will
follow a dance of intricacy in compulsive articulated movements, giving ideal
programmable empirical shape and outcomes. This machine has to be rented with a
specic number of components (only available in packages of 500 units) to be
populated in any condition, any situation. The individual component is developed
as a Velcro termination, self-attached by a comb-feather design, with variable
positions able to assume, at your convenience, polyphonic structures; be they
massive, fluid, opaque or transparent. The machine is able to be packed in a pick-up
of 3 2 1 m (10 7 3 ft) including the tracks that are 10 m (33 ft) long. Please
Fig. 4 Randomizer. New-Territories Academia/Francois Roche-Stephan Henrich/Guo and

Wang/LAB USC/2010/ http://www.new-territories.com/blog/usc2/
Psychaestenia
23
refer to the installation instructions for ascertaining the dual positions of the
machine/component on the ground. You will be trained in the inverse cinematic
process that will enable you to draw rst the structure manually, as curves in space
by manipulating the machine tips, and secondly discover how the footprint of your
handy movement is becoming the trajectory of the components stacking, automatically repeated and assembled by the machine (4 m/13 ft high maximum). The
intricate packing fabrication will follow the isocurves you dened in the space in a
repetitive adaptation.
Precautions for use: Due to the requirement of unpredictability of its work, the
machine is subject to bipolar disorder, alternating manic, hypomanic and depressive
episodes of varying lengths. Although these episodes are necessary to the nature of
the random intricacy process, they might in the long run cause side effects such as
racing thoughts and rest (mode OFF) deprivation.
Take care of possible exhaustion of the machine, as well as of the feeling of
impuissance in front of its never-ending chore. On the contrary, if the machine
shows repetitive, ordered or systematic combination processes, bring it back to the
shop immediately for emergency reprogramming.
2.5
Body. Builder. Shitter
Instructions: Rent an agile hyper-proteined device, shitting liquid concrete in a

vertical phallic extrusion, which is turned into coagulations that it stands on to
continue the construction process in deance of gravity. The Shitter is only made
available to rent to a minimum of 30 families, dedicated and driven by a bio-politic
decision. The device is a usable, operative machine for a self-organised
micro-urbanism conditioned by a bottom-up system. The 30-plus families, called
the multitude, are able to drive the entropy of their own system of construction,
their own system of vivre ensemble. It is based on the potential offered by contemporary bioscience, the rereading of human corporality in terms of physiology
and chemical balance to make palpable and perceptible the emotional transactions
of the animal body, the headless body, the bodys chemistry, and information
about individuals adaptation, sympathy, empathy and conflict when confronted
with a particular situation and environment. The construction process developed
through machinismindeterminate and unpredictable behaviourwith the creation of a secretive and weaving machine that can generate a vertical structure by
means of extrusion and sintering (full-size 3D printing) using a hybrid raw material
(a bio-plastic cement) that chemically agglomerates to physically constitute the
computational trajectories. This structural calligraphy works like a machinist stereotomy composed of successive geometrics according to a strategy of permanent
production of anomalies: with no standardisation, no repetition, except for the
procedures and protocols at the base of this technoid slums emergence. Its rst use
and development was for the anarchitecture_deshumeurs experiment.
24
F. Roche et al.
Fig. 5 Shitter. New-Territories/Francois Roche-Stephan Henrich/An architecture des humeurs/

2011/ http://www.new-territories.com/une%20architectures%20des%20humeurs%20page.htm
Precautions for use: The machine is set in between anal and foecal stages
(Fig. 5), leaving both unresolved in order to achieve full development of its construction capacities. Anal expulsive behaviours, as well as exhibitionism, are frequent phenomena of the machine and are to be considered as signs of good health.
Placed in an extremely social zone, these behaviours could later develop into
paraphilia: manifesting in hyperbolic intensications, distortions, monstrous fruits
of erotic expression outside of normal eroticism. It is strongly recommended,
therefore, not to place it in public zones (i.e. outside of your own multitude). The
device is also slightly narcissistic, which could provoke strong reactions in similar
devices of different multitudes.
2.6
OCD Packer
Instructions: Rent this extremely efcient packing, ordering, classifying, numerating and xyz-positioning machine, for an endless stacking and staggering. The
Packer is only available for long-term rent. The machine works to extend existing
construction, by testing the possibility of wrapping, smearing and invading a previous situation to develop a surrounding maze with multiple uncertain trajectories
and parcours. The morphological trap it creates is both a jail and a protection
apparatus (Fig. 6). This dual strategy avoids the occupant perceiving their own
madness and protects others from their own pathologies. Participants require a
personal agreement and discharge to play this game as a voluntary prisoner, lost in
the permanent entropy of packing. In any case you could use, if necessary, RFIDs
on PDAs to rediscover positioningbut at your own risk. Its rst use and development was for the Olzweg experiment.
Psychaestenia
25
Fig. 6 OCD Packer. New-Territories/Francois Roche-Stephan Henrich/Welostit/2006/ http://

www.new-territories.com/welostit.htm
Precautions for use: In order to achieve high efciencies in ordering, numbering,

arranging, checking, cleaning etc., the machine was implanted with intrusive
thoughts that can produce uneasiness, apprehension, fear and worry. The repetitive
behaviours aimed at reducing these anxieties can also manifest in an aversion to
particular numbers or in the absurd repetition of nervous rituals. In case you notice
such signs of obsessive compulsive disorder, please bring the machine back to the
shop immediately for a diminution of input anxieties.
2.7
Algae-Sacher-Cyclothymia
Instructions: Rent this under-seawater device that behaves as an extractor removing

algae and extracting chemicals (calcite) and particles from the water in order to
agglomerate a masochism structure. The progressive accumulation is condemned to
be pulled and pushed by the current and tide, which drives the orientation and the
progression of the crystallisation without a forecasted positioning agenda. The
machine is usable only in seawater, which contains approximately 400 mg/L of
calcium and represents 1.6 tons/km3. The calcium is obtained from dissolving rocks
such as limestone, marble, calcite, dolomite, gypsum, fluorite and apatite. Before
renting you need to request a survey to conrm the quantity of calcium in your
location. We can provide this expertise. In order to function, the device requires a
water depth of between 6 and 20 m (20 and 65 feet). The extraction, transformation
processes are patented. The chemistry ltering and reaction cannot be divulgated in
26
F. Roche et al.
Fig. 7 Cyclothymia. New-Territories Academia/Francois Roche-Stephan Henrich LAB

Anguewandte/Mirko Daneluzzo and Martina Johan/2009/
these instructions of uses. Please do not open the sealed core of the machine; it is
toxic.
Precautions for use: The machine is built with a total submission to external
factors such as currents, tides and lunar eccentricities (Fig. 7). The more it is
Psychaestenia
27
ill-treated by the water, the better it will work. The machine is also cyclothymic,
subject to mood swings, and is voluble in its responses to the water humiliations.
Due to the mixture of these characteristics, the device is susceptible to construct
totally useless structures, and cannot be held responsible for the unusable nature of
the structures built. You are renting it at your own risk. In extreme cases of
maltreatment, where the machine is overexposed to water or other environmental
factors, it could become self-defeating or suicidal. Ultimately, it could completely
stop functioning.
2.8
Bulimic Enclosure Weaver
Instructions: Rent this silk cocoon-weaving devicepreciously precise and accurateto create temporary buildings, camping sites, outdoor workshops or garden
parties. Do not complain that this machine is both the producer and the structure of
the production, trapped in its own net (Fig. 8). It is its own process of know-how.
The silk membrane could be waterproof or not. Please refer to the density of
knitting in the machines instructions. The wire is the product of bio-production,
starch and flax. Its lifespan is around 10 days before it degrades and loses its
structural resistance. This melting condition is 100 % natural, and the process of
necrosis will provide nitrogen and nutritional elements to the ground. Do not be
afraid of the ostensible pollution it seems to generate. Different time spans for
synthetic silks are available on application. The synthetic silk wire is provided by a
bobbin of 10 km (6 miles).
Fig. 8 Enclosure Weaver. New-Territories Academia/Francois Roche-Stephan Henrich/Qi Su and

Shenyuan Guo/LAB USC/2009
28
F. Roche et al.
Precautions for use: The machine is conditioned to have a lack of bodily feeling
in its surfaces in order to keep it endlessly weaving surfaces. However, it can
unexpectedly reject the surface and return to its body, inducing the formation of a
protective cocoon around itself. This bulimic tendency to recreate a virtual
dimension of potential traits, connections, affects, and movements around its own
body are symptoms of the future loss of the machine inside its own production. At
this stage nothing should be attempted to stop the claustrophobic process.
2.9
Astrolabe Stutterer
Instructions: Rent this devicethe Astrolabe Stuttererto ascertain the level of

threat posed by two discrete planets, the sun and the moon, and the human
pathologies they produce. This machine detects any potential harm that these
planets threaten, and secures your negotiation with the celestial vault: protecting
you from magnetic storms and radiation from the sun, and the psycho-lupus
affliction of the moon. The Astrolabe Stutterer (Fig. 9) cannot simultaneously
maintain an equal position between the sun and the moon, except during eclipses,
every 6,585.32 days, exactly 18 years, 10 or 11 days and 8 h, depending on the
occurrence of leap years.
The part of the machine dedicated to the sun indicates the planets celestial
cycle. It particularly highlights any gap in the suns position and the degree of
protection afforded by the interaction of solar rays with the ozone layer, which has
had its impact depleted by UV emissions. The device can be used in correspondence with uranium powder, which has a natural afterglow that indicates the
intensity of UV emissions. The uranium powder is provided with special conditions, because of the emission of alpha rays (below the administrative threshold),
which have been agreed by legal settlements. This machine has the potential for a
double paranoia: one harmful substance acts as sensor to another, providing a chain
of pastpresentfuture industrial collateral effects. The moon part of the device
points to the symptoms of the moon: the forces of attraction, and fear of transformation (real or illusionary). It works as a vector of science of the imaginary,
through a pataphysical approach. Nothing seems real, but everything in fact affects
your metabolism. Its rst use and development was for the Building which never
dies experiment.
Precautions for use: Due to the dual nature of the object to be read, schizophrenic episodes of low intensity are normal and even necessary to the effective
functioning of the device. If placed under ambiguous coordinates, the clock is
subject to delusive and paranoiac interpretations of the astral movements and this
will induce disorganised reports and drawings, impossible to be read or understood
under normal circumstances.
Hallucinatory episodes may include the creation of a third aster or the predictions of absurd mortal events such as over-exposure to the moons dangerous UV
light or again a daytime invasion of werewolves. If pushed to its extreme, the device
Psychaestenia
29
Fig. 9 Astrolabe Stutterer. New-Territories/Francois Roche-Stephan Henrich/TBWND/2010

2012/ http://www.new-territories.com/laboratoryoight.htm. All others pictures/New-Territories Lab M4/More credits and informations on http://www.new-territories.com/props.htm
30
F. Roche et al.
will run from purposeless agitation and motions to complete catatonia in which case
it is recommended to unplug it.
Could we integrate the machines fabrication, means and meaning in the pursuit
of Desirable Machine according to Gilles Deleuze or Bachelor Machine to
quote Michel Carrouges.
At the opposite, we remember how the misunderstanding, or misreading of the
Fold has been exclusively limited by topological fetishism, within the fold of the
substances, of the matter, of the geometry, reduced and eviscerated of its dangerous
and unknown part the Folds of the Soul, the tilt of the soul, able to generate and
negotiate our animality, as a cryptographic science able to decipher the Catatonia
Crisis of Antonin Artaud
Are we able to embrace, 20 years after, a new phase of technology, facing their
own convulsions and ambivalence.in other way than the revival of idealistic
neo-positivism, with the same pathology of misunderstanding and misreading from
philosophical concept/new Mac Gufn for lazy boys and girls/in the comfortable
alibi of speculative materialismin the nihilist neo-conservatism After Finitude
malentendu.
Part II
Scientific Papers
Building a Bridge with Flying Robots

Ammar Mirjan, Federico Augugliaro, Raffaello DAndrea,
Fabio Gramazio and Matthias Kohler
Abstract The research presented here investigates techniques and tools for design
and fabrication of tensile structures with flying robots. Tensile aggregations are
described as a concatenation of nodes and links. Computational tools provide the
designer of such a structure with the necessary aid to simulate, sequence and
evaluate a design before fabrication. Using a prototypical suspension footbridge as
an example, this paper describes the techniques and challenges for implementing
the construction method on a full-scale, loadbearing, architectural artefact. Firstly, a
series of tensile links is fabricated at dened lengths between two distant support
structures to build the primary elements of the bridge. Secondly, cooperating flying
robots brace the assembly by braiding the primary elements to one another. And
nally, the structure is stabilized through the fabrication of additional connections
by robots flying around existing elements within the porous structure.
Keywords Aerial robotic construction
fabrication
Tensile structures
Cooperative
A. Mirjan (&) F. Gramazio M. Kohler

Gramazio Kohler Research, Chair of Architecture and Digital Fabrication, ETH, Zurich,
Switzerland
e-mail: mirjan@arch.ethz.ch
F. Gramazio
e-mail: gramazio@arch.ethz.ch
M. Kohler
e-mail: kohler@arch.ethz.ch
F. Augugliaro R. DAndrea
Institute for Dynamic Systems and Control, ETH, Zurich, Switzerland
e-mail: faugugliaro@ethz.ch
R. DAndrea
e-mail: rdandrea@ethz.ch
35
36
A. Mirjan et al.
1 Introduction
Today, digital fabrication is predominately realized with devices that are xed to
the ground. The solid fastening of a robotic arm or the movable parts of a
CNC-machine to a base ensure precision in material manipulation. A static environment is assumed in order to calculate the spatial situation of an end-effector for
trajectory generation and position control. Recent developments in sensing, computation and control, however, allow the creation of autonomous construction
machines that are mobile and have the ability to localize themselves in unstructured
environments. Flying robots are examples of this type of machine and are becoming
increasingly relevant in robotic construction (Willmann et al. 2012). Aerial robots
can be used to move construction elements to locations not accessible by ground
robots, they can manoeuvre around existing objects to aggregate construction elements and they can fly in or around already built structures and manipulate them
(Mirjan et al. 2014a). Flying robots have profoundly different capabilities to
established mechanical fabrication devices and, as such, may disrupt the conditions
for how architecture is designed and materialized (Kohler 2012).
From the assembly of an architectural scale tower structure built from discrete
foam elements (Augugliaro et al. 2014), the construction of experimental truss
structures (Lindsey and Kumar 2013), the assembly of space frame structures,1 to
the 3D printing of structures (Hunt et al. 2014), the past few years have seen various
robotic construction projects incorporating flying robots. The authors of this paper
are specically interested in the fabrication of tensile structures with hover capable
Unmanned Aerial Vehicles (UAVs), such as quadrocopters (Augugliaro et al.
2013). This fabrication method fully exploits the capabilities of the flying machines
and allows fabricating loadbearing architectural structures that no other machine
could build (Mirjan et al. 2014b) (Fig. 1).
Flying vehicles such as helicopters are applied on construction sites since the
1950s (Mirjan et al. 2013). They are used in bridge construction to transport prefabricated building elements to the site and to string pilot cables between the two
sides to later pull suspension cables across. Here, robotic control of flying machines
offers potential in the realisation of structures. The machine not only substitutes a
crane with almost unlimited reach, but, through the linking of computational design
with fabrication, allows new forms of material interaction (Mirjan et al. 2014a, b).
2 Techniques
In natural formations, as well as in manmade constructions, tensile structures, such
as cable net structures, usually act continuous in material but are structurally
non-linear. This is also the case in this project, where a continuous, flexible building
1
www.arcas-project.eu.
37
Fig. 1 A suspension footbridge structure obtained through a simulation tool
material is used to aggregate an interconnected assembly, however, the research

presented here regards the actual fabrication of such a structure as a sequence of
discrete building modules (Mirjan et al. 2013), connected to one another. The aerial
aggregation of tensile structures can therefore be summarised as choreography of
two basic modules, a concatenation of nodes and links. Here, we describe their
parameters, as well as computational tools that enable the design of aerially
buildable structures.
2.1
Nodes
A node is a point of intersection where a tensile construction element, such as a

rope or a cable, interacts with another object or with itself. A node can be a solid
fastening to an element, a knot, or it can be a sliding connection. In previous work
(Augugliaro et al. 2015), we have presented a general framework that permits a
descriptions of nodes that can be realized by flying machines. This framework
consists of three parts. Firstly, in the knot theory,2 a branch of topology that studies
mathematical knots, a node is represented as a knot diagram, where the node is
projected onto a plane and crossings are identied and numbered. These numbers
listed in a matrix dene the node as code.
Secondly, since knot theory only mathematically describes a closed knot with
joined ends, the knot representation has to be modied in order to take into account
the actual fabrication of the node as a sequence of moves on a support element
(Fig. 2). This topological representation of the node does not incorporate spatial
2
http://katlas.org/wiki/DT_(Dowker-Thistlethwaite)_Codes.
38
A. Mirjan et al.
Fig. 2 Sequence of knot diagrams for the Munter Hitch with the respective numerical notations,
outlining the fabrication instruction
information (such as scale, position and orientation) of the node, which is required
to generate the trajectory for the flying machines. Hence, nally, a
three-dimensional trajectory is generated from the knot code by incorporating the
following parameters: the position of the knot (3D-point), the orientation of the
support element (3D vector) and the approach direction of the flying machine (3D
vector).
These three parameters dene a 3D plane and add the needed spatial information
to the node. Following the knot code, a node is realised as a series of circular
movements around the centre point. The described generalised method allows the
fabrication of any node with constant tension (winding knots), assuming there is
enough space to fly the node. An interesting challenge will be to incorporate the
realisation of nodes with loops (looping knots), where the vehicle creates a hanging
segment to fly through.
2.2
Links
A rope spanned between two nodes generates a link that resembles a catenary
curve. The parameters of a link are its two support points and its length or tension,
dening the sagging of the link. While a node, once it is built, does not change its
characteristics, a link might change its shape over time, since every intersection
with another link results in a new equilibrium state.
39
The flying machine is equipped with an active rope dispenser; a motorized roller,
which allows the force on the rope to be controlled during its deployment. This
allows links to be created with different lengths or tensions. The payload capacities
of flying machines are constrained and their manoeuvrability is greatly influenced
by the load. These mechanical constraints limit the solution space of buildable links
in relation to the weight of the rope and the tension applied (the flying machines
used here can apply approximately 2 N3 N of force while still being controllable).
The shorter the link, the higher tension can be applied and vice versa.
2.3
Irregular Modules
Most tensile structures buildable by flying machines can be realized as a concatenation of nodes and links. As described above, they can be generally dened as
basic building modules. The amalgamation of these modules permits the creation of
manifold physical realizations. The unique capabilities of the flying machine,
however, also allow the fabrication of tensile elements that are neither nodes nor
links, they are situated in between, requiring specic methods and tools to be
utilized.
One example of these elements is the aerial preforming of tensile weaving. In
Mirjan et al. (2013), we demonstrated a similar approach to construct a zigzag
surface (Fig. 3). The density of this surface was limited by the size of the flying
machine and the distance between the support points. In contrast to this earlier
work, the weaving module makes use of the ability of the material to slide.
Therefore, in this method, the flying machine creates a surface-like structure by
flying a gure eight-like trajectory around already built structural members (similar
to the zigzag) with an additional pullback movement after every crossing. This
pullback movement closes the space that was needed for the machine to fly through,
while creating a dense lling. The density of the lling can be adjusted by flying
Fig. 3 Zigzag structure on the left and zigzag structure with pullback movement (right). The
dashed curve represents the trajectory of the vehicle (left)
40
A. Mirjan et al.
Fig. 4 Braids in relation to the minimum amount of vehicles needed for their realization. A braid
with three strands, for example, can either be built with two vehicles around an already fabricated
strand, resulting in a deformed arrangement, or simultaneously with three vehicles, creating a
symmetrical braid
additional circles around the support elements, similar to woven vinyl cord of the
Acapulco Chair.
Another building module for the aerial fabrication of tensile structures, the
braiding module, is both a link and a node. A braid is formed by interlacing
multiple strands of tensile material, overlapping and crossing each other, in an
intertwined, often linear, manner (Larsen 1987). While most of the elements
described above can be built by one vehicle, the fabrication of braids usually
requires multivehicle cooperation. The construction of a braid with more than two
strands cannot be sequenced for a single vehicle and therefore requires the interaction of minimum two flying machines (Fig. 4).
2.4
Simulation, Sequencing and Evaluation
In order to be able to design tensile structures that are buildable with flying robots, a
series of computational tools have been developed, specically addressing the
characteristics of the building method. These tools allow the simulation, sequencing
and evaluation of structures incorporating material, machine and fabrication
constraints.
2.4.1
41
Simulation Tool
Designing
the form of an active structure, like a cable net structure, is

challenging since its shape is not known in advance. Designing with
linear tensile elements, such as ropes or cables, demands the aid of
form finding techniques to statically determinate the structure acting
in pure tension under self-weight (Kilian and Ochsendorf 2005).
Furthermore, a tensile structure might contain nodes that are not fixed
(a simple turn node), sliding on structural members. Hence, this work
proposes a tool for the simultaneous physical simulation of tensile
elements, as they act under gravity and collisions. The tool combines
a design environment with a physics engine: the design information,
such as support points, link length or node type, are defined in
McNeel Rhinoceros 3D, while the physical simulation in Maya
Nucleus (Stam 2009) runs in the background.
2.4.2
Sequencing Tool
The fabrication of tensile structures with flying robots does not require building
from the ground up. The order of when and where a link is constructed does not
have to be linear. This design freedom implies additional complexity. The form of
an active structure changes its shape with every newly built interacting link. The
spatial situation alters over time, constraining the path a vehicle can take. The
design therefore has to incorporate the spatiotemporal performance of the structure
and simulate it step by step. To take this into account, we propose designing aerially
buildable tensile structures sequentially, according to the actual fabrication order
(not as a global representation). A design usually starts with a single node on a
support point, followed by a link to another support point, and so on. The whole
structure is simulated with each new link, creating a time-fabrication based digital
model of the artefact (Fig. 5). All steps are recorded, allowing jumping back to
specic steps in order to implement design changes.
2.4.3
Evaluation Tool
Prior to a material realisation of the design, the digital model has to be evaluated on
whether it is buildable according to environmental, physical and mechanical constraints. These constraints influence the design of the structure and have to be
integrated into the design process.
Each node type has a specic solution space in relation to the orientation of the
support element and the approach direction. Alongside the sequencing and the
simulation of the structure, each node is evaluated rstly regarding its orientation
and secondly regarding collisions with the environment and already fabricated
42
A. Mirjan et al.
Fig. 5 Step-by-step simulation of a bridge conguration
elements, respectively. The vehicle size and manoeuvrability influence the solution
space.
3 Implementation
The building modules and design tools described above have been individually
tested and adopted in separate experiments. Here, the research takes the important
step of interlinking these single elements and testing their synthesis in a prototypical architectural context. Multiple flying robots fabricate a full scale, loadbearing footbridge, spanning 7.5 m.
3.1
Experimental Setup
The bridge is constructed in the Flying Machine Arena,3 a 10 10 10 m indoor

space for aerial robotic research. The space is equipped with a motion capture
system that provides vehicle position and attitude measurements. This information
is sent to a PC, which runs algorithms and control strategies and sends commands
to the quadrocopters (acceleration and body rates) (Lupashin et al. 2014). As
previously described, the vehicles are equipped with a motorized rope dispenser,
enabling the dynamical adaption of rope tension during its deployment, as well as
estimating the length of already placed rope. The bridge is fabricated from
Ultrahigh-molecular weight polyethylene rope (Dyneema) with 3 mm and 4 mm
diameters. The material distinguishes itself for aerial manipulation due to its high
http://yingmachinearena.org/.
43
strength and low weight. Its weight-to-strength ratio is around 815 times lower
(better) than that of steel. A 100 m long rope with a diameter of 4 mm weighs 700 g
and can support 1300 kg. Its low stretch and positive durability properties (water,
chemical and UV-resistance) make it useable for architectural applications.
Additionally, Dyneema has a low coefcient of friction, allowing the material to
slide easily against itself. This is benecial when the rope has to slide under load to
nd a structural equilibrium, but it is also challenging during the fabrication of a
node when minimum sliding is required. The white cube nature of the building
space offers few options for support points for building. In response to that, two
scaffolding towers were erected and solidly fastened on either side of the space. The
round horizontal, vertical and diagonal bars offer multiple options for support,
allowing the realization of a variety of structures in the space.
3.2
Realisation
The bridge is fabricated in three consecutive steps. Firstly, three tensile links are
erected at dened lengths between the two distant support structures to build the
primary elements of the bridge, responsible for sustaining the majority of the loads.
Three vehicles simultaneously construct one tensile link each, using a variation of
the boom hitch as a start node (Fig. 6), flying to the other side (Fig. 7) and fastening
the link by constructing a Munter hitch. These primary tensile links could have been
fabricated by a single machine, building one link after another, however, parallelization of the task substantially speeds up the fabrication process. When viewed
in section, the links are arranged in a V-shape. The bottom link supports the feet of
a user when crossing, while the two parallel upper links provide the handrails. As
mentioned, the building material is low stretch, however, when a link is loaded the
nodes tighten, resulting in the sagging of the link. Therefore, the three links are
Fig. 6 Boom hitch fabrication at the scaffolding crossing for the central link
44
A. Mirjan et al.
Fig. 7 Erection of the three primary links between the scaffolding towers
constructed with the maximum possible tension to make the crossing more
comfortable.
After the fabrication of the main links, two vehicles brace the assembly by
braiding the primary elements to one another. First, the machines simultaneously
erect a node on the handrail support points on one of the scaffolding towers. Then, a
series of braids at the centre link and single turn nodes at the handrail links are
realized. The vehicles navigate to the centre link and construct an 800 mm long
braid (Fig. 4, braid 3), crossing each other (Fig. 8), before making a turn node
around the respective handrail link. This secondary structure braces the bridge and
joins all the elements to a structural whole. The connections between the central
link and the handrails distribute the forces when the bridge is loaded, while the
sliding of the single turns at the handrails allow the bridge to dynamically adapt its
shape and nd an equilibrium according to different load cases. Finally, in the last
step, the bridge is stabilised by adding additional links to the structure. First, two
links are erected between the scaffolding towers below the central link. Then these
two connections are joined to the central link one after another by flying a zigzag
trajectory around them and through the openings between the central link and the
Fig. 8 Braiding maneuver on the main support with two interacting vehicles
45
Fig. 9 Crossing the footbridge
handrail links. The size of these openings, dened by the length of the braids and
the gaps between them (Fig. 8), is dimensioned for the vehicles to fly through.
The bridge can be crossed securely without the addition of the stabilizers.
However, integrating them helps to absorb non-uniform loads, and possible lateral
and uplift forces. It alters the artefact from a passable structure to a usable footbridge (Fig. 9).
4 Conclusion
The work presented here demonstrates for the rst time the use of flying robots for
the construction of a full scale, loadbearing architectural structure. A framework for
representing and building tensile joints, a method to fabricate links at dened
lengths, as well as computational tools that allow the simulating, sequencing and
evaluation of structures enable the design and fabrication of aerially buildable
suspension structures. The prototypical artefact described in this paper showcases
the ability of the vehicles to architecturally aggregate material independently of the
ground conditions and the machine size.
The ability of the vehicles to fly in and around existing objects is utilized to
interconnect existing members of the assembly and create a structural ensemble.
Cooperation between machines through parallelization accelerates the production
process, while the cooperation through interaction allows the fabrication of structures that a single vehicle could not realize.
The V-shape of the footbridge allows the safe crossing of the structure.
However, the traversing could be enhanced with the integration of the winding
module by introducing a surface structure to walk onto, instead of a linear tensile
46
A. Mirjan et al.
element. A further interesting challenge is the aerial robotic fabrication of architectural structures in an outdoor environment.
The work demonstrates that flying robots are not constrained to aggregate
material layer by layer from the ground up and proposes thinking about robotic
construction as a nonlinear, sequential set of operations in material interaction. As
such, the realised bridge structure does not mimic the usual manual process of
building such a structure but reinterprets it using the unique abilities of the flying
machine.
47
Acknowledgments The research presented here is based on a collaboration between the Institute
for Dynamic Systems and Control and Gramazio Kohler Research, ETH Zurich. The experiments
shown here are performed in the Flying Machine Arena [3] at the Institute for Dynamic Systems
and Control at ETH Zurich. The work is supported by the Hartmann Mueller-Fonds on ETH
Research Grant ETH-30 12-1. A special thanks goes to Augusto Gandia and Maximilian Schulz
who have contributed to the work.
References
Augugliaro, F, Zarfati, E, Mirjan, A and DAndrea, R 2015, Knot-tying with Flying Machines for
Aerial Construction, IEEE/RSJ International Conference on Intelligent Robots and Systems.
Augugliaro, F, Lupashin, S, Hamer, M, Male, C, Hehn, M, Mueller, MW, Will-mann, J,
Gramazio, F, Kohler, M and DAndrea, R 2014, The Flight Assembled Architecture
installation: Cooperative construction with flying machines, IEEE Control Systems, vol. 34,
no. 4, pp. 4664.
Augugliaro, F, Mirjan, A, Gramazio, F, Kohler, M and DAndrea, R 2013, Building tensile
structures with flying machines, IEEE/RSJ International Conference on Intelligent Robots and
Systems, pp. 34873492.
Hunt, G, Mitzalis, F, Alhinai, T, Hooper, PA and Kovac, M 2014, 3D printing with flying robots,
IEEE International Conference on Robotics and Automation, pp. 44934499.
Kilian, A and Ochsendorf, J 2005, Particle Spring Systems for Structural Form Finding, Journal
of the International Association for Shell and Spatial Structures, vol. 46, no. 2, pp. 7784.
Kohler, M 2012, Aerial Architecture, LOG, no. 25, pp. 2330.
Larsen, JL 1987, Interlacing: The Elemental Fabric, Kodansha International, New York.
Lindsey, Q and Kumar, V 2013, Distributed Construction of Truss Structures, in Frazzoli, E,
Lozano-Perez, T, Roy, N and Rus, D (eds), Algorithmic Foundations of Robotics X, ser.
Springer Tracts, Advanced Robotics, vol. 86, pp. 209225.
Lupashin, S, Hehn, M, Mueller, MW, Schoellig, AP, Sherback, M and DAndrea, R 2014,
A platform for aerial robotics research and demonstration: The Flying Machine Arena,
Mechatronics, vol. 24, no. 1, pp. 154.
Mirjan, A, Gramazio, F and Kohler, M 2014a, Building with Flying Robots in Gramazio, F,
Kohler, M and Langenberg, S (eds), Fabricate: Negotiating Design and Making, gta-Verlag,
Zurich, pp. 266271.
Mirjan, A, Willmann, J, Gramazio, F and Kohler, M 2014b, Designing Behaviour: Materializing
Architecture with Flying Machines, GAM, no. 10, pp. 23647.
Mirjan, A, Gramazio, F, Kohler, M, Augugliaro, F, DAndrea, R 2013, Architectural fabrication
of tensile structures with flying machines, in Brtolo, H et al. (eds), Green Design, Materials
and Manufacturing Process, CRC Press, Boca Raton FL, pp. 513518.
Stam, J 2009, Nucleus: Towards a unied dynamics solver for computer graphics,
Computer-Aided Design and Computer Graphics, 11th IEEE International Conference,
pp. 111, 1921.
Willmann, J, Augugliaro, F, Cadalbert, T, DAndrea, R, Gramazio, F and Kohler, M 2012, Aerial
Robotic Construction Towards a New Field of Architectural Re-search, International Journal
of Architectural Computing, vol. 10, no. 3, pp. 439460.
Autonomous Robotic Assembly

with Variable Material Properties
Michael Jeffers
Abstract This paper discusses the problems within autonomous robotic assembly
workflows as they encounter a variable property of assembly parts or
materials. This is shown through a case study with an industrial robot in an enclosed
work cell and a simple assembly task with wooden sticks of variable lengths,
designed as an adaptive feedback control system. To perform the study, the
development of a virtual model for the persistent storage of material data and
computation of next build-actions is required. Different sensing strategies are used
to address issues of substantial, and minute, material variabilities of dimensional
properties as they deviate from a predictive virtual model. Establishing communication strategies for a live-control pipeline as the infrastructure for this system
allows the system to respond to pre-build scans of part dimensions, as well as
update the virtual model when post-build scanning detected deviation. In the
worst-case scenarioif preconditions were not metthe post-build scan would be
unsuccessful and the system would self-terminate. Otherwise, deviations would
update and influence future actions. This influence is what leads to the indeterminate nature of the resultant forms.
Keywords Autonomous robotics
Material variability Assembly
Adaptive control system Sensor feedback
1 Introduction
This paper illustrates the problems within and methods for addressing autonomous
robotic assembly using a case study and, by its current implementation, a set of
constraints as proxies for real world conditions. The purpose of the case study is to
demonstrate viability in projected real-world conditions. Incorporating wood
materials, eye-in-hand sensing, and adaptive feedback control within autonomous
M. Jeffers (&)
Carnegie Mellon University, Pittsburgh, USA
e-mail: mike.s.jeffers@gmail.com
49
50
M. Jeffers
robotic assembly processes are each targets for providing evidence and testing of
some advantages and disadvantages of this approach.
The case for autonomous robotics has largely been made in the context of
dangerous or difcult tasks where human actors are supplanted by robots (Jung
et al. 2013). This has resulted in certain materials and processes being quickly
adopted and heavily engineered into a precise process to minimize error.
The increasing need for varying tolerance or adaptability of assembly on-site
(Elashry and Glynn 2014) limits the implementation of robotics in construction
applications. Fewer factors can be ensured as reliable preconditions (Bruyninckx
et al. 2001) and are usually compensated by on-site decision-making based on
human observation and measurements of actual conditions. If the task is known but
the nature of subtasks includes factors that vary this can still be a closed system
with feedback (Bruyninckx et al. 2001; Vasey et al. 2014). This is the problem
considered in this paper: How can an autonomous robotic system accomplish a
known task/goal if some subtasks are known to be variable? Other questions that
have emerged and worth considering are: Can we assemble a known form with
unknown parts? Or rather, can we assemble unknown parts in a known manner?
The latter is used to isolate behavior, although ensuring that local rules are satised
leads to difculty in determining outcome.
1.1
Material Variance
This case study is designed to use length as a highly variable factor, and to include
other properties and uncertainties of the wood material which might introduce noise
to the system that could neither be preemptively determined nor engineered out.
The domain of variability is the variance in property. Member length has high
variance, whereas slight irregularities along the prole have low variance. It is
found that highly variant properties, those that exceed design tolerances of the
system, must be known prior to build-action. Low variances, on the other hand,
might not be detectable within the resolution of the sensing techniques or devices.
In the case of assembly, error may accumulate beyond tolerances and therefore
requires the introduction of post-build verication and error recovery. Similar work
(Drfler et al. 2012) explores the same single dimensional variance but with focus
only on the initial analysis of the part and its placement in order to minimize
subsequent error.
This case study demonstrates a number of methods to overcome both high and
low material variance. An experiment was designed to replicate practical scenarios
of standard cuts or drops of material that are delivered within some tolerance;
nominal versus actual dimension. Handling, moisture, and storage life can contribute to more deviation between expected and actual dimensional properties
especially in the case of wood (Lyth et al. 1995). The robotic process, even when
standardized, has to adapt to such alterations if tooling and manipulation requires
Autonomous Robotic Assembly with Variable Material Properties
51
Fig. 1 Placement of stick at generated location based on its length and positions of other sticks
non-compliant registration. The difference here is that length is highly variable and
the system is designed to take into account any length (Fig. 1). This value, once
known, strongly characterizes subsequent actions that depend on this data. The low
variance of material irregularities plays itself out only over the course of the
assembly task.
2 Methodology
2.1
Communication
An autonomous system with feedback needs the ability to compute next actions
based on the feedback. Communication between the external computer (client) and
the robot controller (server) is via TCP/IP. Other projects, for example Hal,1
Scorpion2 (Elashry and Glynn 2014), Robo.Op,3 the server and client described in
Interlacing (Drfler et al. 2013), and ROS4 tackle the same functionality with
each adapting to different client-side or robot controller environments. The client in
this case processes and communicates with the sensor devices, parses inputs,
computes a virtual model, and produces commands for the robot. This was
developed in Java using the standard library with graphical elements leveraging the
Processing5 API and associated libraries. An important criterion in the design and
implementation is that this communication must be synchronous. The server is
often occupied with executing motion commands, while the client could easily
http://www.grasshopper3d.com/group/scorpion.
http://hal.thibaultschwartz.com.
3
https://github.com/ros-industrial/abb/tree/groovy-devel/abb_common/rapid.
4
https://github.com/peopleplusrobots/robo-op.
5
https://processing.org.
2
52
M. Jeffers
overflow a queue of sent messages. A send-receive pattern is enforced by a

handshake protocol.
2.2
Virtual Model
The virtual model is the persistent data structure used to store and update data
relevant to the assembly building process (Fig. 2). The build-algorithm parses this
data and determines next actions. Its design is strongly linked to the nature of the
assembly process. The more abstract the model, the fewer input variables that can
be detected and taken into account. It is therefore assumed to be constant. The
stacking model is 2.5D. At each level, including the base-plane, construction is
planar and parallel. There are a number of factors about wooden sticks, which
obscure and defeat this ideal. This discrepancy serves as a proxy for other on-site
issues of dynamic environments, materials, and noise in the system that post-build
scanning processes are designed to compensate for and maintain validity of the
virtual model. Consequently, the purpose of the case study is to determine what
data can be predicted, to what extent, and when is revision required.
Fig. 2 Client-side interface with tool-side camera, virtual model representations, and communication logs with robot controller and sensor device
53
For the virtual representation, any stick is reduced to: a pair of endpoints, a
length, and gripper location as a factor of length. This is sufcient for generating
higher-level information for subsequent calculations. Endpoints are rst determined
and then used to iteratively test potential placements of the stick for a valid
assembly. Stick length is the seed for this process. Boundary geometry, for representation or collision detection, is generated from this information.
Within these rules different geometries are used to perform different relational
calculations. The centerline is used when trying to determine bearing conditions,
whereas a scaled prole is used for collision predictions. The gripper boundary is
computed for collision only for the current stick that is being added to the assembly.
This multiplicity of representations allows for more economical computation of
otherwise complex 3D relationships by leveraging the simplest representation
required (Fig. 3).
The build-algorithm (Fig. 4) produces not only tall but also relatively stable
stacks. Height was a simple metric that would present more challenges with respect
to stability when faced with accumulating error from sticks below. Looking for a
maximum span condition helps reduce cantilevered ends. Otherwise, should sticks
accumulate heavily to one side, it will cause the stack to topple. This would produce
an error-state that would later be detected in scanning. The current rules and
build-algorithm only take into account the current stick as the object that is
manipulated. Thus, the remaining virtual model is effectively frozen. This is
designed to allow for subsequent incremental mechanical fastening of parts in the
assembly to reflect more closely a real assembly process and to reduce the
search-space for scanning processes. Economizing what data needed to be
Fig. 3 Algorithm for enforcing assembly rules
54
M. Jeffers
Fig. 4 Algorithm for determining placement of new stick
recovered, updated, and stored is vital for reducing the complexity of the scanning
problem and the resources it requires.
2.3
Scanning Methods
Pre-Build: Measuring. Sensor feedback is isolated to two specic steps in the

build-cycle: pre- and post-build. Pre-build scanning is used to acquire stick length, a
process that involves signal processing to ensure that certain preconditions are met
before the start of each robot action. A signal is tripped when registering a
material or part against a known location. Typically, changes in input signal can be
used to stop a robot motion, or extract its current position etc. To calculate stick
length, if location of the limit switch and robot approach vector are known, one can
compute the difference in distance from where the switch is and the distance away
55
where the Tool Center Point (TCP) is located at the time the switch is tripped
(Fig. 5). Measuring both sides of the stick gives us the length and position of the
TCP as the dispenser does not center the stick at the location of grasping.
Post-Build: Verify and Search. Different devices afford different advantages,
however some require developing a process to acquire higher-level data beyond
what the given sensor is tailored to provide. Positioning these within the build
process (pre- or post-) is weighted according to their sensor attributes. The Parallax
Laser Range Finder (LRF) or Sharp GP2Y0A02YK0F Infra-Red (IR) sensor provides depth from the point of emission, within a degree of delity and error,
delivering new readings at an approximate rate. Hosting sensor devices at the End
of Arm (EOA) has the additional advantage of taking a single 1-Dimensional sensor
and allowing it to acquire data about any non-occluded surface in the robots work
envelope.
To verify the depth of a stick on its centerline, the LRF is used to hit multiple
points on the length of the most recently placed stick (Fig. 6). Comparatively, the
LRF gives more reliable measure of real depth than the IR sensor. The latter
Fig. 5 Moment of contact on limit switch (Omron Z-15GQ-B) during pre-build measurement
Fig. 6 LRF pinging stick midpoint to verify presence in expected location
56
M. Jeffers
Fig. 7 Unit tests of grid-to-endpoint algorithm development. Purple line is regression x- over y-;
Cyan is y- over x-; and Yellow line with circle indicates the solution with endpoint approximation
fluctuates with the analog signal processing and is slightly affected by the color of
the target. However, the LRF scans at about once per second, which is slow for a
single reading. As a result, IR sensing is used to populate depth values in a grid
over a known location, despite its increased output noise. There are two major
components to the depth grid. First, previous as for example Stock Finder (Bard
et al. 2014) utilizes a quad-tree like method of self-subdivision that has proven
advantageous for selectively rened grid-based processing. Notably, statistical
culling of outliers and determining which cells are signicant for further examination are of value when paired with an expected depth value. The second is to
derive a new endpoint of a stick that is either within the search radius or not. If it is
not present, a threshold needs to be determined at which the search should abort. If
it is present, the question then is how to use a matrix of values to indicate an
endpoint of a stick.
Of all the implementations, linear regression proved the most reliable at low
resolutions. Linear regression is used to model trends in the data assuming a
dependence of one value along an axis against an independent variable on the
perpendicular axis. Understanding this assumption does not hold, linear regression
could be performed over both the x- and the y-axis, then compared for the best
match or averaged as they converge (Fig. 7). The line segment generated better
approximated an endpoint at lower resolutions than other algorithms like the
convex-hull, which required higher levels of renement and further analysis to
determine the endpoint.
3 Results
The system is able to construct stacks of sticks without collision, with proper
bearing conditions, and able to determine at each step if and where it could build
higher. Furthermore, it can detect and compensate for disturbances, either as a result
of accumulated error from subtle material variance or from a dynamic environment.
Tolerances of the system are relatively low, as rst pass post-build scanning would
57
allow for certain amount of shifting. This tolerance however is incorporated on the
front-end of the build-algorithm when detecting collision conditions. Establishing
this tolerance range as a baseline lets the system decide when further scanning is
needed.
Scanning processes are targeted to the scope of material variance they are
designed to compensate for. Pre-build scanning is used to acquire the length of the
stick, which largely determines where such a shape could be placed. Post-build
scanning had two phases, the rst of which is to verify the placed-stick. Pinging
from a known depth, the virtual model produces an expected distance that should be
acquirable from the top of any given part in the assembly. Should measured distances deviate from expected distances, a second-phase recovery process is initiated
because the stick is not present in the expected location. In its current implementation, this can only account for the stick being slightly rotated or translated from its
intended location. Should the stick exist within the search-space of the recovery, the
data acquired is processed to create revised endpoints of the new location of the
stick object.
This subsequently updates the virtual model and a rst phase verication scan
then conrms the update. When the stick is not found within this search-space, we
cannot guarantee the success nor safety of further actions and therefore the system
enters a failure-state (Figs. 8 and 9).
Fig. 8 System in failure-state after stick not found in post-build search due to its removal
58
M. Jeffers
Fig. 9 Placement of stick after measurement
4 Discussion
Design processes cannot predict on-site realities, but they can provide a framework
for a decision-making process with respect to these variables. The role of the
designer here is not one of user interaction, but of establishing the rules that govern
the autonomous system. Furthermore, the design here has no formal value as the
parts are assumed to be entirely unknown. This indeterminacy of resultant
assemblies relates directly to the rules that govern the stacking and the recovery
from deviations that occur within the process as a result of material irregularities.
Therefore trying to force this to achieve a desired form is futile, but providing
metrics instead has the opposite effect.
4.1
Future Development
The case study demonstrates the systems scope of behavior in the context of
variable materials. Variable context, or dynamic environment, demands attention
next to fully address the problem of on-site robotics. Sensing in this case expands
from analyzing a task at hand to constant monitoring of surroundings. Many
variables can be ruled out with enclosed work cells for repeatable success and
safety.
Short of developing a mobile platform that can interface with just one task
on-site (Chu et al. 2013), there is additional work with creating automated workflows involving multiple tasks. Integrating complementary processes such as
automated mechanical fastening will complete the ambitions of a truly autonomous
assembly process. Working with standard and reclaimed wood products will also
59
push the precision sensing of this system forward to be able to adapt to highly
variable materials.
Even within the problems found in the case study, the boundary where unrecoverable failure occurs can be pushed further and additional cases for error
recovery can be adopted. Understanding this failure-state as a moving boundary
between the known, the unknown, and up to the unknowable is a critical observation for analyzing the viability, repeatability, and safety of an autonomous
system.
5 Conclusion
This paper presents a case study, which demonstrates an implementation of a closed
adaptive feedback control system for autonomous robotic assembly with known
material variance. The four fundamental components of this system are: memory,
feedback, decision, and actuation. Construction processes can be described through
this framework. Construction documents (memory) detail a task at hand. One may
observe discrepancies (feedback) that are not present in the document. Action based
on information given and observed must be taken to best satisfy any directives in
the document (decision and actuation). The system developed for the case study
follows the same form. Memory is contained in the virtual model. Feedback comes
from the scanning processes that supply the real-data to revise model information.
Decision-making is shown as the build-algorithms.
This framework can therefore apply to a construction task that can be described
in the same manner. On a high level, most construction tasks may sound like
procedures. When examined, they are described more precisely as an algorithm
with many low level rules that may be referred to as intuition or skill. If this
methodology is adopted and additional sources of variance can be identied and
incorporated, autonomous robotic systems in construction becomes a more immediate future. This opens the possibility of not just precise execution, but rather
autonomous construction systems that can operate with little oversightobserving
and adapting to a dynamic environment. The result may satisfy the task, but the
exact nature of the form will be a consequence of on-site decisions, as it is with
construction today.
60
M. Jeffers
61
Acknowledgments This research was generously supported by Carnegie Mellon Universitys

School of Architecture and associated staff and faculty. The authors would like to express their
gratitude to Josh Bard, Ramesh Krishnamurti, and Richard Tursky for their generous advice and
support.
References
Bard, J, Gannon, M, Jacobson-Weaver, Z, Jeffers M, Smith, B and Contre-ras, 2014, Seeing Is
Doing: Synthetic Tools For Robotically Augmented Fabrication In High-Skill Domains,
ACADIA 14: Design Agency, pp. 409416.
Bruyninckx, H, Lefebvre, T, Mihaylova, L, Staffetti, E, De Schutter, J and Xiao, J 2001, A
Roadmap For Autonomous Robotic Assembly, Proceedings of the 2001 IEEE International
Symposium on Assembly and Task Planning (ISATP2001). Assembly and Disassembly in the
Twenty-rst Century. (Cat. No.01TH8560), pp. 4954.
Chu, B, Jung, K, Lim, M and Hong, D 2013, Robot-Based Construction Automation: An
Application To Steel Beam Assembly (Part I), Automation in Construction, vol. 32, pp. 4661.
Drfler, K, Rist, F and Rust, R 2012, Interlacing in Brell-Cokcan, S and Braumann, J (eds),
Robotic Fabrication in Architecture, Art and Design, Springer Wien, New York, pp. 8291.
Elashry, K and Glynn, R 2014, An Approach To Automated Construction Using Adaptive
Programing, in McGee, W and Ponce de Leon, M (eds), Robotic Fabrication in Architecture,
Art and Design 2014, Springer International Publishing, Switzerland, pp. 5166.
Jung, K, Chu, B and Hong, D 2013, Robot-Based Construction Automation: An Application To
Steel Beam Assembly (Part II), Automation in Construction, vol. 32, pp. 6279.
Lyth, D and Rabiej, R 1995, Critical Variables In Wood Manufacturings Process Capability:
Species, Structure, and Moisture Content, Quality Engineering, vol. 8, no. 2, pp. 275281.
Vasey, L, Maxwell, I and Pigram, D 2014, Adaptive Part Variation, Robotic Fabrication in
Architecture, Art and Design, pp. 291304.
An Integrated Modelling and Toolpathing

Approach for a Frameless Stressed Skin
Structure, Fabricated Using Robotic
Incremental Sheet Forming
Paul Nicholas, David Stasiuk, Esben Nrgaard,
Christopher Hutchinson and Mette Ramsgaard Thomsen
Abstract For structural assemblies that depend upon robotic incremental sheet
forming (ISF) the rigidity, connectivity, customization and aesthetics play an
important role for an integrated and accurate modeling process. Furthermore, it is
critical to consider fabrication and forming parameters jointly with performance
implications at material, element and structural scales. This paper briefly presents
ISF as a method of fabrication, and introduces the context of structures where the
skin plays an integral role. It describes the development of an integrated approach
for the modelling and fabrication of Stressed Skins, an incrementally formed sheet
metal structure. The paper then focus upon the use of prototypes and empirical
testing as means to inform digital models about fabrication and material parameters
including: material forming limits and thinning; the parameterisation of macro and
meso simulations with calculated and observed micro behaviour; the organisation
and extraction of toolpaths; and rig setup logics for fabrication. Finally, the validity
of these models is evaluated for structural performance, and for geometric accuracy
at multiple scales.
Keywords Incremental sheet forming Mass customisation Robotic fabrication

Toolpath optimization
P. Nicholas (&) D. Stasiuk E. Nrgaard M.R. Thomsen

School of Architecture, Royal Danish Academy of Fine Arts, Copenhagen, Denmark
e-mail: paul.nicholas@kadk.dk
D. Stasiuk
e-mail: dsas@kadk.dk
E. Nrgaard
e-mail: enor@kadk.dk
M.R. Thomsen
e-mail: mette.thomsen@kadk.dk
C. Hutchinson
Monash University, Melbourne, Australia
e-mail: christopher.hutchinson@monash.edu
63
64
P. Nicholas et al.
1 ISF
Incremental sheet forming (ISF) is a fabrication method that imparts 3D form onto
2D metal sheets. It is driven by 3D CAD models and has been developed for the
purpose of industrial prototyping within the automotive industry. In the most
typical ISF method, a ball-head tool is moved over the surface of a thin metal sheet,
causing a progression of localised plastic deformation (Jeswiet et al. 2005). ISF is
useful for three reasons. First, it negates the need for time-intensive creation of
costly dies (negative forming), instead directly machining semi-nished pieces of
metal. Secondly, because forming is highly localized, the force required does not
increase with scale, meaning that there is no theoretical limit to formed sheet size
(Tisza 2012). Lastly, ISF extends the formability of metals beyond conventional
methods, such as stamping or deep drawing (Bagudanch et al. 2013). Drivers of
new research in this eld include the exploration of larger scale applications,
typically in the automotive and aerospace industries (Amino et al. 2014; U.S.
Department of Energy 2013), and improving forming accuracy. The geometry
change impacted on the steel sheet is achieved through a local tensile or biaxial
stretching of the metal, and is dependent upon a connection between geometric
considerations, processing parameters and material properties. As it is stretched the
metal undergoes strain hardening or cold working, which increases its strength
locally through the accumulation of plastic deformation. This metallurgical transformation attends geometric change, as the sectional thickness of the sheet
diminishes relative to stretching. In the context of a lightweight skin, these changes
are not insignicant. For example, in Stressed Skinswhich uses low carbon mild
steel formed at room temperaturesectional thickness reduces in places from 0.5 to
0.15 mm, and strengths increase from 220 to 410 MPa.
2 The Architectural Relevance of ISF

Transferred into architecture, ISF graduates from a prototyping to a production
technology that supports mass customization. As has been noted, potential architectural applications include for example folded plate thin metal sheet structures
(Trautz and Herkrath 2009). We have further identied an application for ISF in
customised, load-adapted architectural designs. Architects use thin metal sheets as
cladding panels to provide integrated enclosure, structure and form. As loads vary
in building system, so do performance requirements, so that the customization of
elements becomes a key concern. Using ISF on pre-cut metal cladding panels to add
features that locally stiffen the panel (in the locations and to the extents needed) can
signicantly increase efciencies of material use and reductions for supporting
structural systems.
65
3 ISF for Stressed-Skin Structures

In this research, the ISF process is used to fabricate an architectural stressed skin
structure. Such systems are typically a hybrid assembly in which a thin skin is
structurally active, bearing both planar and shear forces and providing signicant
rigidity by continuously wrapping an underlying, compressive frame. They are an
intermediate between monocoque and rigid frame assemblies, and have been particularly associated with the early application of metals in lightweight structures. In
their design, rigidity is a central concern at multiple scales: rigidity against instability in the whole structure, against local buckling of the components that carry
compressive load, and against micro buckling or wrinkling.
The research Stressed Skins (Fig. 1) develops a structural approach in which the
skin carries planar and shear forces, but without the use of an additional framing
system, at the scale of a pavilion. Research at RWTH Aachen has established ISF as
structurally feasible at this scale (Bailly et al. 2014), in the case of formed panels
spanning between a hexagonal continuous framing structure. Recent research
explores doubly curved sheet metal panels for free-form metal skins (Kalo and
Newsum 2014) and self-supporting structures (Bailly et al. 2015), which utilize
cone geometries as means to reach from one skin to another. These have been
developed to prototype scale. Stressed Skins is designed as an asymmetric tunnel,
which cantilevers at one end. The structure consists of 186 unique planar, pentagonal panels. These are arranged into an inner and outer skin. The framing system
for typical stressed skin assemblies is replaced by the introduction of geometric
features for resisting local buckling and structural connections, both continuous
within each skin and for managing shear across inner and outer skins. These are
produced through the custom robotic ISF of individual panels.
Fig. 1 The research structure Stressed Skins: structure and details
66
P. Nicholas et al.
4 Prototype, Modelling and Fabrication: Integrated

Approach
Stressed Skins is developed through multiple iterations of physical prototypes and
computational models, and integrates observations from physical prototypes towards
the digital environment to addresses the multiscale nature of the forming process and the
structural assembly. In this approach, different computational models, specic to particular scales of parameterisation, behaviour and decision making, are made critically
interdependent upon one another. A number of considerations are addressed here:
4.1
Computation
The digital modelling of Stressed Skins is informed by parameters and limits

derived directly from these physical prototypes. Three modeling scalesmacro,
meso and microare considered to be markers along a structural continuum
(Fig. 2). In general, the macro scale refers to overall geometric congurations and
predictions of its structural performance. The meso scale considers the level of the
panel and its detailing, and implements geometric transformations related to connectivity, stiffening, and component-level tectonic expression. Finally, the micro
scale relates to the calculation of material implications at the most discrete level,
which includes the thinning and hardening of the steel sheet that results from
forming. The modeling does not include the actual simulation of the ISF fabrication
process, but only the expected material transformations introduced through it.
The multi-scale modelling approach is thus comprised of multiple techniques
that enable the information generated at each scale to flow both up and down the
Fig. 2 Interscalar relations: macro (left), meso (center) and micro (right)
67
continuum. Here, an adaptive mesh renement method is used to support localized

variations in resolution and information flow. From a perspective of design
development, these include the overall form-nding and panelization operations;
global structural analysis and adaptive specication of connectivity arrays; and
recursive local tectonic pattern formation which depends upon nite element
analyses and is further informed through the calculation of forming strains and
material thinning. The features of a half edge meshits vertices, half-edges and
facesare coupled with a series of lists, dictionaries and Grasshopper data trees
that effectively bundle within mesh elements critical design data related to: topology, form-nding and geometry; structural behavior; material characteristics;
connection detailing; and patterning and tectonic expression (Nicholas et al. 2015).
4.2
Physical Considerations
Physical prototypes have played a key role in determining the parameters and limits
of mass customization. Parameters that informed the digital modelling include
individual panel constraints related to size, orientation, and formable territory; the
development of connection and assembly strategies; change in material properties;
and forming limits in regards to both feature geometry and tooling time. Samples
were prepared to systematically vary multiple processing parameters: tool type
(either hammer or point), tool movement speed, feature angle, and to measure strain
hardening and thinning. To integrate factors inherent to the fabrication setup, all
samples are produced using the same rig used for nal production.
4.3
ISF Setup
An ABB IRB 140 six-axis multipurpose industrial robot is used to fabricate

Stressed Skins. The wall-mounted robot arm is situated above a flat table, which
bears a clamped, re-orientable MDF jig (Fig. 3). MDF dies are xed to this jig and
where necessary are supported from below using a collection of standard elements.
The dies are laser-cut templates that dene the outlines of desired formed
geometries, where those geometries cut the plane, and provide resistance for the
steel sheets in areas intended to remain planar. Steel sheet blanks are xed to the
dies along their edges with bolted MDF blocks.
4.4
Position of Features
Through empirical testing, the capacity of the ABB IRB 140 to exert a downward
force was established for the working area (Fig. 4). Because of this varying strength
68
P. Nicholas et al.
Fig. 3 ISF setup
Fig. 4 Force capacity of

robot (colourised) and area
prone to singularity (shown in
grid)
capacity, the position of features proved to have an impact on the forming accuracy,
and in some cases resulted in the robots failure to apply sufcient force to form the
steel. To counter this problem, which is to a large extent linked to the size and
specication of the robot system, 6 different forming positions were dened. An
analysis of target locations per feature enabled toolpath distribution across locations
(Fig. 5). As a result, forming was concentrated in areas of maximum strength.
4.5
Production of Samples
Samples were prepared using two alternative forming methods. The rst of these was
a pneumatic hammering tool, where force was generated via the stroke of the
tooltip. The second method, a single-point pressing approach that utilizes the robots
69
Fig. 5 Reorientation logic and jig positions
capacity to impart force, was understood and developed based on access to an

ISF-designated CNC setup at DTU Mekanik. Several processing parameters were
varied in a systematic fashion across ve samplesthe forming method, the tool
speed and the wall angle. Both forming methods were tested in order to understand
their relation to strain rate. The relation between tool speed and wall angle was tested
to understand how these parameters affected formability. The speeds varied between
2065 mm/s, and the wall angle was tested at 15, 35 and 50.
4.6
Testing of Samples
Within the range of speeds and angles tested, all samples were successfully formed.
Visual monitoring of the grains and measurement of thickness at the same points
was achieved using optical microscopy at 5 points on the cross sectional thickness
of each sample. Their local increase in strength was monitored using Vickers
70
P. Nicholas et al.
Fig. 6 a Initial grain size, b grain deformation induced by forming, c relationship between local
strains and measured yield strength calculated from Vickers hardness measurements
Hardness tests with a 5 kg load measured at 50 points on the cross sectional

thickness of the sample (Fig. 6). The resulting hardness were converted to estimated
flow stresses and correlated with the local strains. Flow stress is the yield strength of
the metal as a function of strain, and describes the point at which the material enters
plastic deformation. The conversion between hardness and flow stress is stress
(MPa) 3VHN, where VHN is the Vickers hardness number (Tabor 1951). Yield
strength equates to flow stress.
4.7
Results
The hammering technique imparts plastic deformation through the rapid sequential
impacting of a tool against the surface. This involves a much higher speed of tool
motion and therefore produces larger strain rates and leads to a greater straining of
the sheet but not, in the samples tested, to higher flow stresses. In this respect,
hammering does not offer obvious advantages compared to the pressing technique.
Furthermore, according to our observations hammering was associated with greater
springback, and we were able to better achieve the target form through the pressing
technique. Based on these observations as well as results from the results of the
Vickers hardness tests, the pressing mode of forming was identied as the fabrication method of interest. It was found that in the range of 3060 mm/s the tooling
speed had a large impact in both the surface quality and formability of the steel
sheets and that the amount of impact had a direct connection to the wall angle.
However there was no observed impact on the material properties and that at this
range, the impact of tool speed is negligible to resultant strain hardening.
Equation 1 is then derived from the measured relationship between local strains
induced by forming and the measured flow stress by tting the logarithmic curve to
the observed data points (Fig. 6c):
ry 43; 407ln 431; 23
71
The measured thickness correlates with results based on the calculation of thickness
strain (3) and can be calculated due to the rule of volume constancy:
1 2 3 0
5 Synthesis of Computational and Physical Considerations

in the Example of a Panel
The digital model and processes of physical prototyping are synthesized in the
development of geometric features within a panel, toolpath extraction and organization methods.
5.1
Panel and Connection Arrays and Global FE

Simulation
The panel arrays for both layers of Stressed Skins were developed through a
stepwise accumulation of panels onto two respective target surfaces according to a
pentagonal planar tiling strategy. The targets were derived as being variably offset
from a baseline surface that was developed both to accommodate occupancy
requirements on the site and to generate suitably challenging structural performance
demands through its spanning and cantilevering. The variable offset was calculated
based on an initial shell FEA (Finite Element Analysis), with greater offsetsand
structural depthsassigned to areas of high utilization. A constraint-based dynamic
form nding systema beta scripting library of the Grasshopper plug-in
Kangaroo2was used to adapt and planarize each panel as it adhered to its
respective offset target design surface. Following this initial panel organization, a
series of connection cones were solved between the two skins. These cones form
the primary basis for managing the structural shear requirements, taking on much of
the role played by the compressive frame in traditional stressed skin structures.
Here they were located to maximize diverse connectivity across multiple panels
between the upper and lower skins, and oriented in response to a second FEA that
identied shear force vector lengths and directions. A third FEA was performed
following this precise locating of the connection cones, and translation and rotation
nodal displacements were extracted from the model at all connection points
between panels, both within and across skins.
72
5.2
P. Nicholas et al.
Informed Patterning and Local FEA Simulation
The prior understanding of the relationship between geometric forming and the
consequential material hardening was then integrated with these connection point
translation and rotation vectors into an iterative feedback design and analysis cycle
for the purpose of locally introducing performance improvements into individual
panels specically to resist in-panel bending forces. This was achieved through the
variable-depth forming of a pattern onto each panel, integrated with the base
inter-skin connection cones that provide primary structural depth and accommodate
the transfer of shear forces within the assembly.
The pattern was rst generated as a flat, graphic element over all panels on the
mesh. An implementation of the Gray-Scott reaction-diffusion algorithm on the
design mesh was used to achieve this. Beyond its aesthetic, this algorithm was
selected for a two key reasons: its generally isotropic nature enabled resistance to
bending in multiple directions, and its form could reliably be cut into the MDF dies
used during forming. This baseline pattern worked as a scaffold to receive additional depth. This additional depth was realised through an iterative process (diagrammed in the second full-page image). First, each panel begins as flat in all areas
except for the features used for both inter and intra skin panel connections. Over
each panel sub-mesh, an adaptive quad mesh is arrayed and inscribed ellipses used
to determine the local strain introduced in the forming of these baseline features.
Equation 1 is deployed to locally differentiate yield-strength material settings for
each face in the primary mesh, and an FEA is executed for the individual panel,
using the translation (blue) and rotation (green) force vectors derived from the
global FEA. Resulting utilisations are extracted. Here, note that areas within each
panel that have been hardened due to forming (as in the deep connection cones)
tend to have signicantly higher strengths, and therefore lower utilisations. High
utilisation areas then drive the local introduction of incremental depth, which is here
visualised as incremental changes from baseline features, with black being zero
change. This process of transforming material settings, applying connection nodal
translations, and adding local pattern depth is then iterated up to fteen times per
panel, resulting in a steady decrease in utilisations for each panel due to strain
hardening, and greater bending energy due to geometric stiffening resulting from
added depth where it is useful. Finally, each panel is subdivided to a ner level of
resolution, and initial contours are extracted for toolpath generation.
5.3
Fabrication and Toolpath Generation
An algorithm was developed to derive tool paths from the model geometry, based
on contours cut from a design mesh. The progression of the algorithm was:
73
Fig. 7 Grouping of features
1. Grouping of features, 2. Position of features, 3. Tooling speed in relation to wall

angle. This order was developed to rstly calculate the entire toolpath, then divide it
into modules, and lastly to add information about speed to the target positions. To
improve control over tooling time and surface quality, the toolpathing algorithm is
developed based on the creation of spirals, an established approach for ISF (Filice
et al. 2002; Jeswiet et al. 2005). The grouping and position of features, and the
tooling speed in relation to wall angle are discussed in the following:
1. Grouping of Features. With the basis being horizontal contours, a strategy for
grouping the section curves was developed (Fig. 7). By checking the section
curves for inclusion in the domains of the previous layer, groups of curves were
created in order to make multiple continuous spiraling toolpaths that allowed
forming of complex geometry, but ensured no metal was being formed
unconsciously or tried forming twice. The gure shows an example of a colour
coded grouping, revealing the complexity in both geometry and toolpath.
2. Position of Features. To ensure features being formed in optimal position in
relation to the strength of the robot, each spiraling toolpath was divided into
1000 points. Each point was checked for inclusion in predened areas. The
toolpath was then transformed into the area that recorded the highest percentage
of inclusion.
3. Tooling Speed in Relation to Wall Angle. The production schedule for Stressed
Skins required working at the maximum permissible speed, with consideration to
formability and surface quality. Building on the knowledge obtained from the
tested samples, further exploration showed that if the wall angle did not exceed 45
degrees a speed of 65 mm/s could be used, but as wall angle got higher the speed
needed to be lower to ensure both surface quality and formability. With our setup
74
P. Nicholas et al.
we reached a limit of a 60 wall angle which could be achieved with a tool speed of
30 mm/s. Based on this testing, a linear relation between angle and tool speed was
used to set a unique tool speed for each target point along a toolpath:
Smm=s 2:333 A
6 Assembly and Evaluation

The research extended towards a research project, with the fabrication of 187 panels
for structural assembly (Danish Design Museum 2015). The structure is characterized by a high degree of connectivity for successful assembly, and thus relied
upon accurate forming and low tolerances. Geometric accuracy has been a key
concern regarding ISF since its inception, with typical geometric tolerances of more
than 3 mm within a part (Allwood et al. 2005). Though traditional architectural
practice accepts such tolerances, in the case of Stressed Skins these needed to be
tightly managed and not extrapolated over the entire structure.
The research thus evaluated the built structure for its performance. A Faro Focus
3D 120 scanner was used as a means to measure geometric accuracy and structural
performance. At the scale of panel, forming accuracies were measured to have a
2 mm standard deviation (Fig. 8). This was accounted for via the inclusion of a
4 mm spacer at connection points between inner and outer skin. After assembly and
a setting period of one month, the structure was scanned and evaluated. The
maximum deviation to the geometry predicted using the nite element model was
10 cm, recorded at the extremity of the cantilever (Fig. 9), which illustrates the
contours colour-coded to distinguish unique groupings of curves for creating
individual spirals, enabling the translation of complex geometries to discrete, sortable toolpaths.
Fig. 8 Range of deviation between predictive model and formed panel
75
Fig. 9 Contours colour-coded to distinguish unique groupings of curves for creating individual
spirals
7 Results and Conclusion

This research has discussed the modeling and fabrication of an incrementally
formed, stressed skin architectural structure. A robotic ISF process has been used to
increase rigidity through geometries within a surface of thin steel panels, and
through connections between those panels. ISF possesses signicant architectural
potential in the area of mass customisation, but as has been discussed, this requires
a tight coupling between fabrication process, material properties and the digital
design model.
This research paper aimed to contribute to ongoing research in the robotic
fabrication of single elements, towards highly integrated structural assemblies. In
doing so, it aimed to extend the scope of architectural applications by developing a
highly integrated structural assembly and an associated digital modelling method.
The paper discussed a modelling method that is informed by fabrication parameters
and material properties, which are established through prototyping and empirical
testing. This method incorporates the specics of a fabrication environment, and
integrates empirically derived material hardening and thinning data. Empirical data
have been used to dene parameters of macro and meso nite element simulations
with calculated micro behavior; in order to set and extract toolpath information; and
to inform rig setup logics for fabrication.
The high level of integration between modelling and prototyping enabled the
simulation to incorporate a level of information not typical within architectural
modeling, and a fabrication process where the relationship between tool speed and
wall angle was optimised. The successful assembly of the panels, some of which
support up to ten unique connections, demonstrates that incrementally formed
frameless structural assemblies can be made available at an architectural scale.
76
P. Nicholas et al.
77
Acknowledgements This research was undertaken as part of the Sapere Aude Advanced Grant
research project Complex Modelling, supported by The Danish Council for Independent
Research (DFF). The authors would like to acknowledge the collaboration of Bollinger Grohmann
consulting engineers, Daniel Piker and Will Pearson, the research departments DTU Mekanik and
Monash Materials Science and Engineering, and the robot command and control software HAL.
References
Allwood, J, King, G and Duflou, J 2005, A structured search for applications of the incremental
sheet-forming process by product segmentation, Proceedings of the Institution of Mechanical
Engineers, Part B: Journal of Engineering Manufacture, vol. 219, pp. 239244.
Amino, M, Mizoguchi, M, Terauchi, Y and Maki, T 2014, Current Status of Dieless Aminos
Incremental Forming, in Procedia Engineering: 11th International Conference on Technology
of Plasticity, vol. 81, pp. 5462.
Bagudanch, I, Centeno, G, Vallellano, C and Garcia-Romeu, M 2013, Forming force in Single
Point Incremental Forming under different bending conditions, in MESIC, vol. 63,
pp. 354360.
Bailly, D, Bambach, M, Hirt, G, Pofahl, T, Herkrath, R, Heyden, H and Trautz, M 2014,
Manufacturing of Innovative Self-supporting Sheet-metal Structures Representing Freeform
Surfaces, in Procedia CIRP, vol. 18. pp. 5156.
Bailly, D, Bambach, M, Hirt, G, Pofahl, T, Della Puppa, G and Trautz, M 2015, Flexible
Manufacturing of Double-Curved Sheet Metal Panels for the Realization of Self-Supporting
Freeform Structures, in Key Engineering Materials, vol. 639, pp. 4148.
Filice, L, Fratini, L and Micari, F 2002, Analysis of Material Formability, Incremental Forming,
CIRP Annals Manufacturing Technology, vol. 54, no. 1, pp. 199202.
Jeswiet, J, Micari, Hirt, G, Bramley, A, Duflou, J and Allwood, J 2005, Asymmetric Single Point
Incremental Forming of Sheet Metal, in CIRP Annals-Manufacturing Technology, vol. 54, no.
2. pp. 88114.
Kalo, A and Newsum, M 2014, An Investigation of Robotic Incremental Sheet Metal Forming as
a Method for Prototyping Parametric Architectural Skins in Robotic Fabrication in
Architecture, Art and Design 2014, pp. 3349.
Nicholas, P, Stasiuk, D, Nrgaard, E, Hutchinson, C and Thomsen, MR 2015, A Multiscale
Adaptive Mesh Renement Approach to Architectured Steel Specication in The Design Of A
Frameless Stressed Skin Structure in Design Modelling Symposium: Modelling Behaviour,
2015.
Tabor, D 1951, Hardness of Metals, Clarendon Press, Oxford.
Tisza, M 2012, General Overview of Sheet Incremental Forming, Journal of Achievements in
Materials & Manufacturing Engineering: vol. 55, no. 1, pp. 113120.
Trautz, M and Herkrath, R 2009, The application of folded plate principles on spatial structures
with regular, irregular and free-form geometries in International Association for Shell and
Spatial Structures (IASS), pp. 10191031.
Robotic Lattice Smock

A Method for Transposing Pliable Textile
Smocking Techniques Through Robotic Curved
Folding and Bending of Sheet Metal
Andrew Saunders and Gregory Epps
Abstract Architect Gottfried Semper built a discourse on architectural aesthetics
based on his belief that textiles were the mother of all arts, and the initial motivation
for all architectural form. Inherent in this evolutionary premise is the concept that
cultural development begins with pliable and easily manipulated materials, and can
be extended and transformed through technological advances for crafting more robust
and permanent materials. As a contemporary projection of this framework, Robotic
Lattice Smock (RLS) presents a method for transposing pliable fabric folding techniques of smocking to an architectural scale through robotic bending and folding of
rigid planar sheet metal. Building on the limitations of three-axis CAD/CAM fabrication techniques for unfolding and cutting planar pieces, RLS explores the process
of six-axis robotic curved folding and bending to gather or smock planar
developable surfaces to overcome brute force assembly, build volume through more
efcient material use of planar sheet material and generate novel material aesthetics
through the hard constraint of disciplined material transposition.
Keywords Robotic fabrication Smocking Curved folding Textiles Gottfried

semper Transposition Physical computing Developable surface Torsal ruled
surface Rulings Aesthetics

1 Introduction: Semper, Textiles and Transposition

The nineteenth-century German architect and historian asserted that textiles are the
mother of all arts, influencing every branch of the technical arts, thus the origins of
all basic architectural forms. His lexicon Der Stil (transl. Style in the Technical and
A. Saunders (&)
University of Pennsylvania, Penndesign, Philadelphia, USA
e-mail: AndrewSaunders@design.upenn.edu
G. Epps
RoboFold Ltd, London, UK
e-mail: gregory.epps@robofold.com
79
80
A. Saunders and G. Epps
Tectonic Arts; or, Practical Aesthetics) methodically traces the influence of textile
motifs on various forms and manufacturing procedures for more permeant material
including ceramics, tectonics (carpentry) and stereotomy (Semper et al. 2004). The
motivation was twofold: (1) to prove polychromatic ornamental covering of Greek
architecture through the persistence of certain symbolic motifs carried over from
nomadic textiles; and (2) to critique the cheap 19th century industrial simulation of
one material by another, specically casting, stamping, and molding. For Semper,
these fabrication processes were paradoxically indifferent to the symbolic continuity
essential to the recreation of tectonic form (Frampton and Cava 1995).
Sempers criticism nds new relevance as the current integration of industrial
robotic arms in bespoke design and fabrication gives rise to new methods for
challenging industrial standardization and construction processes at the architectural
scale (Gramazio et al. 2014). The expanded scope of six-axis movement offers
increased agility to work material to the full extent of its expressive scope (McGee
et al. 2014). As a contemporary projection of Sempers tectonic framework, Robotic
Lattice Smock (RLS) explores the expressive qualities of folded sheet metal through
the transposition of pliable textile smocking techniques to robotic folding routines.
2 Lattice Smocking
It is interesting to note the etymological game that Semper plays when choosing
architectural terminology. In Die vier Elemente der Baukunst (transl. The Four
Elements of Architecture) he uses the term die Wand for enclosure or wall, a word
that is strikingly similar to Gewand and winden (German for dress and embroidery, Frampton and Cava 1995). Building on this, we selected the technique of
smocking as one of the most basic forms of embroidery to transpose. Lattice
smocking was further chosen for the project due to its (1) flexibility, (2) deep relief
and (3) variation of motifs, as is discussed in the following:
1. Flexibility. The base grid for the lattice smock pattern can remain regular or
accommodate increase and decrease in the size of the grid modules, supporting
topological variation in the stich pattern. In anticipation of a faade system,
panels following this logic would be easily adapted to a grid or diagrid
substructure.
2. Deep relief. The relatively low density of stitches allows for more fabric to be
gathered between stiches producing deep volumetric folds with recognizable
signatures. In addition, the longer distance between stitches accepts much
thicker fabric. In anticipation of folding and bending thin gage metal, thick felt
was chosen for smocking to produce larger and more consistent folding radii.
3. Variation. Very basic grid stitch patterns of crenulations, zigzags and branching
produce amazingly complex folded formations in the fabric when gathered. The
resulting manifolds include patterns identied as lattice, lozenge, flower, and
arrow Wolff 1996. An arrow pattern was chosen due to its triangular relief. The
three points integrated with the robotic lab environment that included a station
with two 6-axis ABB robotic arms and one stationary arm.
81
3 Surface Rationalization: Geometry of Developable

Surfaces
RLS offered an opportunity to deploy six-axis robotic fabrication and physical
computation to build on the research of Sartorial Tectonics (Saunders 2012) that
similarly addressed tectonic relationships of textile analogs and three-axis computer
aided manufacturing. Although the previous research was successful at transposing
affects of fabric manipulation to more rigid materials from computational aided
manufacturing, the complexity of the developable surfaces were limited to tangential surfaces and the re-rolling of unrolled surface components was entirely by
manual brute force. The metal folding and bending process requires very specic
surface rationalization (Tachi and Epps 2011). Because both processes begin with
flat material and are manipulated to build volume and pattern, they are subjected to
the rules of developable surfaces. Developable surfaces are a subset of ruled surfaces with zero Gaussian curvature, able to be unrolled onto a flat plane. A ruled
surface is a surface generated by a straight line moving along a curve. The straight
lines that make up the surface are rulings. If the ruling has different tangent planes
at each end point, it is scrolar and the points on the scrolar ruling are hyperbolic
(non-developable). If the ruling is touched by one unique tangent plane, the ruling
is torsal and the points on a torsal ruling are parabolic (developable).1 For RLS to
be robotically folded, all rulings on the surface needed to be torsal guaranteeing the
surfaces were developable. As a developable surface, all governing prole lines
could be unrolled for cutting and scoring in planar material.
4 Methodology
The RLS faade system prototype is fabricated from flat thin gauge 1.5 mm aluminum panels with milled proles and curved fold line scores. These panels are
then placed flat on a workstation with two six-axis ABB IRB6400 industrial robotic
arms and one stationary arm. All three are equipped with vacuum end effectors to
grip the panel. The end-to-end process outlined in the following involves smocking
felt and developing unfolded patterns for the gathered smock manifold, constructing
physical mockups in cardboard, paper and at 1:1 in hand-folded aluminum, conducting physical and digital simulation to obtain three-dimensional folding and
bending paths and pinning bent panels at the fully gathered position before release
from robotic arms. The initial research had the following objectives:
to explore the novel expressive aesthetic qualities generated by transposing
textile smocking to robotic curved folding and bending,
to explore the manipulation of flat fabric as an analog model for robotic folding
and bending of planar sheet metal,
1
http://www.grad.hr/itproject_math/Links/sonja/gausseng/ruled/ruled.html.
82
to examine the explicit feedback loop between physical and digital simulation of
folding developable surfaces through robotic manufacturing,
to explore lattice smocking as a generative process for the creation of complex
torsal ruled surfaces (rationalized through partial planes, cylinders, cones or
tangent surfaces) to be robotically folded,
to examine robotic folding as a method to automate curved folding and bending
of metal panels to counter brute force assembly tactics.
4.1
Lattice Smocking Felt
As a departure point, the transfer of fabric to harder substrates was investigated in

order to improve the process of folding paper, which has a limited amount of
changes that can be made before it deteriorates. Building on this process, RLS
began by lattice smocking an arrow pattern (Figs. 1 and 2) with thread, needle and a
(40.64 cm x 22.86 cm) piece of felt. Felt was chosen due to its heavy density, and
also the dened and semi-regular fold lines and curved creases that emerge when
the textile is gathered.
Fig. 1 a Arrow lattice smock stitch pattern, b arrow pattern gathered front, and c arrow pattern
gathered back
Fig. 2 From left a lattice smock variations, b arrow pattern front, and c pattern back
83
The resulting felt arrow smocks were carefully analyzed and distinct curved
crease lines were identied and traced on the felt in its gathered state. When the
pulled stitches were released, the felt returned to a flat position revealing new
intricate curved crease patterns on the original arrow lattice smock grid pattern of
pulled stitches. Mountain and valley coloring of red and blue curves are used during
the curve acquisition as per the origami conventions, to indicate a positive
(mountain) or negative (valley) fold.
4.2
Folded Paper Mockups as Torsal Ruled Surfaces
The new flattened patterns of curved crease lines were scanned in 2D to create
digital templates. Flat heavyweight paper mockups panels were scored and cut
using a die cast paper cutting machine. The paper mock-up panels were folded to
form more rigid versions of the arrow smock. The paper mock-ups were used as
analog models for the metal folding process. They were helpful to examine bending
tension and rationalization of the surface geometry. Due to the stiffness of the
heavyweight paper, the resultant surface embodied the developable properties of a
torsal ruled surface. As the folding was enacted, regions between the curved creases
revealed composition of partial planes, cylinders, cones or tangent surfaces (Fig. 3).
The points where the gathering stitches were previously became points of tight
radii cones and were the points in the paper panel with the most intense bending
tension when folded. The greater the degree of curvature, the greater degree of surface
tension when folded. There are two options for relieving the bending pressure:
decrease in curvature of folding crease or removal of material closest to where the
ruling lines converge. By using the points of the cones as center points to construct arc
cut lines on the template, additional material was removed enabling the new paper
panel to fully fold mimicking the original gathered position in the felt smock (Fig. 4).
Fig. 3 a Torsal ruled surface with material removal overplayed with original arrow lattice smock
stich pattern b detail of individual torsal ruled surface panel
84
Fig. 4 Heavyweight paper

mock-up of six torsal ruled
surface lattice smock panels
Once a developable surface version of the lattice smock panel had been established,
the template was calibrated to accommodate specic dimensional constraints of the
robotic fabrication environment. Key criteria in adjusting the template were the
identication of planar regions for vacuum end effectors attachment; maximum
bending angles; and space for manual pinning of a fully bent metal panel.
With nal adjustments being made to the local module of one arrow fold of the
smock pattern, focus returned to the global conguration. Additional paper
mock-ups were produced to study how modules would nest as an assembly and
how edge conditions of the eld would terminate. It was important that the design
embodied a flexible rule set to accommodate edge conditions as terminus. A sole
plate was developed to provide stability for the module. The sole plate is a flat
prole traced from the gathered position of the panel edges. The sole plate panel
was attached after the robotic arms bent the panel to the fully gathered position,
keeping the panel from returning to the maximum fold angle (which is less than the
maximum bending angle). In addition, the bottom edges and sole plate extend
further at any edge condition to create a fringe condition terminating the continuity
of the lattice smock pattern (Fig. 5).
4.3
Physical Simulation of Folding and Bending Metal

Prototype
The nal physical mock-up phase included cutting and scoring of aluminum panels
and conducting manual folding and bending tests (Fig. 6).
Through the process of manually bending each metal mock-up, sharpness of
curved fold creases were nessed and termination points of fold scores were
adjusted to calibrate the hardness and fading of creases within the metal panel.
85
Fig. 5 Heavyweight paper RLS mock-ups of nested module and terminating fringe conditions
Fig. 6 Cutting and scoring aluminum panel mock-ups (left and center). Manual folding and
bending of aluminum panel mock-up and identication of nal torsal rulings (right)
To simulate the fully bent form of the panels, zip-ties were used to incrementally
gather the panel under tension into the nal position of the arrow smock. The
nal adjusted aluminum mockup was analyzed by sweeping a straight edge over the
surface to identify and trace torsal rulings between curved creases. The locations of
all major rulings were scanned and added back to re-inform the digital template.
Together with the major curved fold creases, they would form the critical network
of components necessary to construct the digital simulation model.
86
4.4
Digital Simulation of Folding and Bending
To accurately model and simulate the transition from flat planar material to nal
folded form without distortion in the digital environment, the King Kong plug-in
for Rhino Grasshoper that integrates a live physics engine was used. The curved
creases and torsal rulings traced from the physical models were dened as a network of springs and hinge forces. Similar to the parameters for simulating rigid
origami, curved creases are dened as hinge forces swinging in one of two directions, mountain or valley. Rulings are dening as springs with rest length equal to
actual length to eliminate distortion through the folding process. After establishing
a digital simulation of the folding and bending behavior, the positions of each face
of the model were tracked. The process was critical in identifying the exact
three-dimensional path of the faces dedicated for vacuum gripper contact. Once
dened, the digital simulation was used to choreograph the motion of two six-axis
robotic arms (Fig. 7).
Fig. 7 Digital simulation of two robots in sequence (top), folding panel (left), and robotic arm
choreography (right)
4.5
87
Robotic Folding and Bending of Metal Panels
The nal phase of transposing smocking techniques to curved folding and bending
sheet metal deployed robotic fabrication. The robotic arm routines were checked for
self-intersection, reach and rotation limits in the IO plug-in. Once the code was
veried as safe, the software allowed it to be uploaded and enacted. The nal
fabrication of six RLS panels involved manually placing a precut and pre-scored flat
panel in a known position on the workstation that corresponded with the simulation.
After the robotic arms had gripped, folded and bent the piece into position, the sole
plate was pinned manually to prohibit the panel from returning to the folded position
and the arms released the piece. The robot was programmed to pick up, move into
position, fold, and return to home position, with each distinct stage indicated in a
named section of the timeline component in the IO plug-in. This level of clarity
allowed for fast iterations through modication of basic robot parameters such as
speed. The timeline features two robots, which are synchronized using an algorithm
to determine equal time gaps between each individual plane. Two entire sequences
are the same length of time, but follow a complex curve as determined by the
non-linear folding animation (simply matching the overall time will not sufce).
Various aspects of the end-effectors such as the angle and position of the vacuum
grippers are adjusted through a series of sliders in Grasshopper, in order to ne tune
the location that they are imparting force on the metal surface. This adjustment is
used to determine the best location to achieve a good resultthis process occurs
through some know-how and trial and error. Figure 8 illustrates the axis angles,
where the display aids de-bugging over-rotations and other errors that occur during
complex folding procedures. Very little room for error was allowed, with only a valid
solution occurring in space vertical region of 20 cms and within a rotation of 30.
Fig. 8 Axis Angles displayed within a full simulation of the working environment
88
Fig. 9 Close up of the timeline component. Yellow areas show warnings for over-rotation
As an integral plug-in of Grasshopper (McNeel), RoboFold enables control over

different sections of the code: Mesh Import, Robot and Tool adjustment, simulation
and control and Outputs including: Code, Synchronization Codes, Fold Preview,
Axis Angles, Camera Animation, Render Mesh Baking, or Simulation Timing.
Figure 9 displays software inset warnings for over-rotation (yellow), with includes a
code warning component.
5 Analysis
In the case of bending metal, where the resulting surface must be developable, the
smocking analog presents a flexible method for developing torsal ruled surfaces,
and can thus become useful for guiding robotic fabrication of more rigid material at
the architectural scale. The act of transposition champions qualities of each material
(or medium) and fabrication processes of each specic material. Achieving the deep
undercut relief and supple intricacy of folding in fabric has presented the largest
challenge. The rst stage of this research successfully examined the local scale of
individual panels, whereby six identical arrow smocked panels were robotically
fabricated. Curved creases with high curvature and bending against the embodied
energy of the bent metal produce recognizable thresholds and limitations in the
process. The extreme pressure tested the strength of the vacuum gripper end
effectors and created dangerous conditions for the manual pinning of the gathered
position. This could be overcome by introducing more secure gripping mechanisms
or bending thinner gauge metal (or other planar material). Ideally, the entire process, even the pinning of bent panel would be automated robotically.
89
6 Conclusion
RLS presents a method for transposing textile manipulation techniques into more
rigid material at the architectural scale through robotic curved folding and bending
of planar sheet metal. By using the process of lattice smocking, flexible textiles
serve as an analog model for producing deep relief and complex curvature from a
flat sheet material. Torsal ruled surface properties are intrinsic in the surface
manipulation from the initial starting point and lead to a fluid transposition into a
rationalized surface for robotic simulation and fabrication. In terms of future
research, one area of further investigation would be topological variations of the
arrow smock panel, and the integration into a seamless fabrication workflow. In
addition, the flexibility of the lattice or grid will be examined, rst in two dimensions with warped grids, and then in three dimensions as a diagrid with the capability of accommodating predened complex forms.
As Semper discussed, this process then enables a transfer of cultural knowledge
from one medium to another. The process of transposition champions matter and
the investment of techniques and rules unique to each medium. The resulting design
is not evaluated on whether or not it looks like the but instead on whether the
metal behaves like the lattice smock. In the direct process of transposition it gains
unique aesthetic expression as the embodiment of particular geometric and physical
signatures of the material and process of fabrication.
90
Acknowledgments Robotic Lattice Smock is a collaboration of Andrew Saunders and RoboFold

Ltd. sponsored by the Rensselaer Robert S. Browns 52 Fellows Program. Design Team: Andrew
Saunders, Sahar Mihandoust, Guo Huanyu, Jessica Collier, Elizabeth Sammartino, Matthew
Vogel. RoboFold Team: Gregory Epps, Ema Epps, Florent Michel, Jeg Dudley.
References
Frampton, K and Cava, J 1995, Studies in tectonic culture. Cambridge, Mass.: MIT Press.
Gramazio, F, Kohler, M, Picon, A, Roche, F and Verebes (eds), 2014, Made by Robots:
Challenging Architecture at a Larger Scale: Architectural Design, Academy Press, London
McGee, W, once de Le n, M and Willette, A (eds), 2014, Robotic Fabrication in Architecture, Art
and Design. Springer Cham, Heidelberg, New York, Dordrecht and London, pp viix.
91
Saunders, A 2012, Sartorial Tectonics, ACADIA 12: Synthetic Digital Ecologies Projects of the
32th Annual Conference of the Association for Computer Aided Design in Architecture
(ACADIA), pp. 7277.
Semper, Gottfried. Style in the Technical and Tectonic Arts; or, Practical Aesthetics. Trans.
Harry F. Mallgrave (Santa Monica, 2004).pg 113 ISBN 0-89236-597-8
Tachi, T and Epps, G 2011, Designing One-DOF Mechanisms by Rationalizing Curved Folding,
roceedings of AL ODE 2011, Tokyo, Japan.
Wolff, C. (1996). The art of Manipulating Fabric, Chilton Book Co, Radnor
Robotic Multi-dimensional Printing Based

on Structural Performance
Philip F. Yuan, Hao Meng, Lei Yu and Liming Zhang
Abstract This paper discusses a robotic multi-dimensional printing design

methodology based on a materials structural performance. Through research on the
process of a spiders behavior, e.g., spinning and weaving, the designers simulate
natural construction principles and apply them to the optimization of traditional 3D
printing techniques. A 6-axis robot is programmed to carry a customized printing
end effector to create free-standing geometries in space. The structural behavior of
the design is optimized through the consistent negotiation between material analysis
and structural simulation in both virtual and physical environment, together with the
implementation of sensor input and real-time feedback between construction tools
and simulation interfaces. The printing tools are designed with additional extruders
and nozzles of various dimensions to adapt to different materials and design
requirements. In this way, a flexible and adaptive additive manufacturing
methodology is established, which integrates the material and structural information
with design initiatives. Displaying a high degree of spatial and structural complexity, the alliance between 3D printing and robotic technology opens new possibilities to sophisticated architectural structures.
Keywords Multi-dimensional printing

Robotic fabrication
performance Material performance Tool development
Structural
P.F. Yuan (&) L. Zhang

College of Architecture and Urban Planning, Tongji University, Shanghai, China
e-mail: philipyuan007@tongji.edu.cn
L. Zhang
e-mail: lim_zhang@foxmail.com
H. Meng
Archi-Union Architects, Shanghai, China
L. Yu
School of Architecture, Tsinghua university, Beijing, China
e-mail: asworkshop@vip.163.com
93
94
P.F. Yuan et al.
1 Introduction
1.1
From 3D Printing to Multi-dimensional Printing
3D printing has been developing for decades. So far, there are two types of commonly used and representative printing technologies, namely FDM (Fused
Deposition Modeling) and SLS (Selective Laser Sintering). FDM melts printing
material and extrudes melted material through ne printing nozzles for deposition
modeling while SLS provides selective sintering of powdered 3D printing material
for modeling and remove the un-sintered part. These two representative 3D printing
technologies have shown certain limitations in actual application. On one hand,
both of these two 3D printing technologies are rapid prototyping with given
materials. The printing process is a simple formation process and excludes inherent
structural logic of the product. Such contour printing method also has certain
limitations in printing capability, especially in FDM. On the other hand, due to the
limitation of the printing method (material deposition modeling), the products
printed using such technology will inevitably sacrice the structural strength at the
interface between different layers. Hence the strength of the material cannot be
sufciently exerted. Although industrial standard SLS 3D printers with high precision can perform printing with extreme strength, the cost is very high.
In this context, the multi-dimensional printing concept is introduced to overcome
the limitations of traditional 3D printing. This has two advantages. The rst is higher
printing freedom in space. Although the current 3D printing technology is called three
dimensional, it only refers to the three dimensional volume of end results. In terms of
formation technology, the current 3D printing mostly adopts a planer contour stacking
technique. Multi-dimensional printing is expected to realize 3D modeling in space
through multiple axes just like a person using their hands. The second advantage is the
ability to integrate various performative optimizations within design and the fabrication process. As mentioned before, the material deposition modeling will cause
structural defects in printed products. The flexible printing strategies and the constant
data feedback between the design and fabrication platform improve the structural
rigidity of the material (Budig et al. 2014).
1.2
Multi-dimensional Printing with Integration

of Structural and Material Performance
Existing studies on multi-dimensional printing are mainly focused on the following

two aspects, namely, the integration of a structural logic, and the improvement of
strength. The integration of a structural logic with the printing process has been
explored illustrated in the Mesh Mould project from the team of Gramazio & Kohler,
ETH. In this research project, the geometry to be printed was rst transformed to a
Robotic Multi-dimensional Printing
95
spatial truss, which is an optimized structural skeleton of the design.1 The designer
built a rapid cooling device inside the traditional 3D printing extruder to ensure
instant cooling and solidication of extruded printing material. Though the material
used in this project is a traditional 3D printing material, the printed geometry has
specic self-support abilities to keep its form stable after the material being cooled
down. Thus, a spatial structure can be steadily built up following the extruders
movement along the preset grid pattern (Gramazio and Kohler 2008). The Weaving
Printing project conducted by Digital Design Research Center at Tongji University
carried similar research in testing more sophisticated geometry and spatial frame
with transitional grid density (Fig. 1).
This type of research indicates the possibility of rapidly producing large structural
prototypes and architectural moulds for building components with complicated
geometry. Secondly, research has been undertaken for improving the strength of
printing materials. In the project MATAERIAL by Advanced Institute for Advanced
Architecture of Catalonia (IAAC) and Joris Laarman Studio, the research team
developed a printing material that can solidify rapidly. This material has excellent
strength after solidication.2 Thus, it can ensure accurate shaping and self-support
ability of the geometry. The signicance of this research project is the realization of
free curvilinear printing through material engineering.
1.3
From Multi-dimensional Printing to Architectural

3D Printing
In recent years, architectural 3D printing has become an avenue for novel research.
Several experimental 3D printed architectures have been produced, such as the 3D
printed residential house of Yingchuang Technology Company, Shanghai (Fig. 2).
Integrated structural reinforcement has been applied innovatively in the printing
process of the residential products, e.g. in the wall section of the building, slant
braces similar to a truss structure are printed to ensure the strength of the wall.3 Yet
there exists a signicant difference between the building industry and current 3D
printing industry: traditional 3D printing is a simple model making process,
whereas in building construction materials, structure, and construction processes
need to be considered. As a result, traditional 3D printing technology is far from
meeting the requirements of building construction when the building is expected to
be 3D printed. Therefore, only through integration of fabrication techniques in
traditional 3D printing industry may the technology be used in building construction eld. The multi-dimensional printing concept is consistent with the demand for
developing large scale 3d printing architecture. The study of multi-dimensional
1
http://gramaziokohler.arch.ethz.ch/web/e/forschung/221.html.
http://www.mataerial.com/.
3
http://www.a0club.com/news/show/575/.
2
96
P.F. Yuan et al.
Fig. 1 Weaving printing (Digital Design Research Center, Tongji University)
Fig. 2 Experimental 3D printed residence (Yingchuang Technology, Shanghai)
printing will have decisive signicance in architectural 3D printing research

(Lipson and Kurman 2013).
Albeit a good example of multi-dimensional printing, in the project
MATAERIAL the multi-dimensional printing is only realized by enhancing the
physical strength of print material, i.e., the structural logic of the built geometry is
not a primary concern in the printing process. This paper introduces a new research
that tries to incorporate structural behavior of the material into the printing process
and employs a customized printing end effector with greater freedom and precision
97
through one 6-axis robot arm to realize multi-dimensional printing of complex

geometries. This method makes it possible to create flexible and versatile nonstandard space structures with no formwork or additional support, demonstrating a
sustainable construction strategy that is materially efcient and cost effective.
2 Robotic Multi-dimensional Structure Printing

2.1
Bionic Design Inspired by Spider Silk
One of the most widely used 3D printing technology, FDM (a linear deposition of
melted material), is the basis for this project. Therefore, if the linear structure can be
strengthened during the printing process to self-support its own weight,
multi-dimensional printing can be achieved with various materials.
Researchers have looked to spider webs for answers (Fig. 3), however, the
structural rigidity of a cobweb is its resistance to tension while the printing structure
in our case is mainly subjected to self-weight and bending moment. We were also
inspired by the sectional morphology of the spider silk, data shows that the sectional diameter of the spider silk is about 1/1000 mm, but its strength is four times
stronger than that of steel in the same diameter (Zheng et al. 2010).
Fig. 3 Structure of cobweb ( Nature Publishing Group)
98
P.F. Yuan et al.
Fig. 4 Main curve and axillary curves
In terms of geometry, a spider web consists of two parts, namely a spindle-knot and a
joint.4 A spindle-knot surface has a stretched porous structure while the joint surface is
made of a random porous structure. These two different structures join together to
guarantee the structure performance of the cobweb (Gosline et al. 1986). The sectional
variation from spindle-knot to alternate joint brings insight to self-standing linear
structures. The bending moment of 3D printed geometry can be overcome by reproducing different sections like spider silk based on a structural reinforcement (Fig. 4).
Integrating the geometric morphology of the spider silk into 3D printing process
is the key to the realization of spatial printing. We have developed a new printing
prole by adding multiple wavy auxiliary curves to the main curve. The area where
auxiliary curves are in contact with the main curve forms joint, while the place
where the supplementary structures deviate from the main structure forms
spindle-knot. Therefore, the combination of primary structure and secondary auxiliary structure inherits the morphology from the spider silk, and achieves greater
structural performance in resisting bending force.
2.2
Section Optimization
In order to achieve convincing spatial prototypes, the number of auxiliary curves that
integrate with the main curve needed to be further determined. Common sense would
suggest that the more auxiliary structures are added, the better self-support ability that
the overall geometry can achieve. Excessive auxiliary support, however, may
increase dead weight of the structure and thus lead to negative effects. Considering the
combination of multiple extruders will inevitably increase the possibility of collisions
among the mechanical devices, combinations from zero to four auxiliary curves with
different compositional positions have been tested and a total number of seven possibilities are simulated and compared.
Five different load scenarios were simulated in the experiments; self-weight
load, point load of 200 N vertical force, 40 Nm external torque, point load of
4
http://www.nature.com/nature/journal/v463/n7281/g_tab/nature08729_F5.html.
99
200 N vertical external force of different sectional scale, and 40 Nm external torque
load of different sectional scale. According to the structural result simulated in
Millipede (Fig. 5, to right), TYPE F has the minimum average deformation under
the ve loading experiments. Though TYPE F is slightly inferior to TYPE G in
withstanding pressure under 40 Nm external torque. It is much stronger than
TYPE G in resisting vertical load, also the number of auxiliary curves used in
TYPE F is smaller than TYPE G, potentially alleviating self-weight and simplifying
the design of the printing device. Therefore, TYPE F was selected as the nal
prototype for section optimization.
2.3
Material Performance Study
The materials tested here are ABS and PLA. These two 3D printing materials have
their own advantagesthe main characteristic this research investigates is rapid
solidication, which is of critical importance because the time that a material takes to
solidify directly relates to whether the prole of wavy auxiliary curves can be
accurately printed. The greatest difference between PLA and ABS is that PLA is
crystal material, while ABS is non-crystal material.5 Crystal materials have a xed
melting temperature. When the crystal is heated to a specic temperature, it begins to
melt, and the temperature remains stable during the process until the crystal is
completely melted. After that, the temperature starts to rise again. The same goes for
the solidication process (Lam et al. 2002). The heat has to be signicantly
decreased in a short time for the solidication of PLA material, which is a signicant
challenge to the cooling device. ABS is a non-crystal material, which does not have a
xed melting point; it is able to gradually solidify with the decrease of temperature.
Hence, it can take more cooling time for the material solidication process.
Meanwhile, the printing temperature of ABS (over 230 C) is much higher than
PLA (200 C) and the solidied temperature of ABS is much easier to reach.
Therefore, the material performance of ABS is superior to that of PLA for this
project. Our cooling test veries this result.
In terms of material strength, the flexural behavior of ABS material is stronger
than PLA. Under the situation of bending force, PLA material is more fragile and
easier to be broken. After comprehensive consideration of various aspects, ABS
was chosen as the printing material in this study.
2.4
Mechanical System
The mechanical system refers to the device controlling the collective motion of all
the printing nozzles. The result of section optimization is the combination of one
5
http://www.maplin.co.uk/p/3doodler-abs-put-on-yo-shades-mixed-pack-n01dr.
100
P.F. Yuan et al.
Fig. 5 Structural deformation simulation of 7 different section types
primary structure plus three auxiliary structures. Therefore, one xed printing
nozzle and three flexible nozzles that can move consistently and synchronously are
required. There are two options for the design of the mechanical system: one is to
use three drivers (e.g. stepper motors) to control the motion of each movable nozzle
respectively. The other is to use one driver and a set of the linkage system to
monitor the movement of three secondary nozzles together. Since the action of the
101
Fig. 6 Printing device
three movable nozzles is not complicated, but requires relatively high consistency,
the second option is more sophisticated but avoids unnecessary deviation caused by
desynchronization of three separate drivers (Fig. 6).
The nal mechanical system consists of a central turn plate, three nozzle operators, stepping motors and a gear set (Fig. 7). The central turn plate is in hexagonal
shape with llet corners. During the printing process, one stepping motor drives the
rotation of the central turn plate through a gear set. Because the distance from the
central point of the turn plate to the six vertexes and six edges are different, the three
angle switches can be controlled to open and close periodically (Fig. 8). In this way,
one stepping motor could control the three nozzles work synchronously, and collaboratively produce wavy sub structures.
2.5
Electronic Control System
The electronic control system of the spatial printing end effector is developed based
on the structure of a conventional FDM 3D printer. The system consists of an
operating interface and a central controller (Fig. 9). The core controller includes one
main control panel (Arduino Mega 2560), one extension board (RAMPS1.4), ve
stepper motor drivers (A4988), four of which control the stepping of the motors of
the material extruders while the other one controls the stepping motor of the central
rotation plate. The operating interface includes one control panel and one display
panel. The control panel controls the rotation speed of the central turn plate, the
102
P.F. Yuan et al.
Fig. 7 Exploded axonometric view of the mechanical system
Fig. 8 Open state and close state of the end effector
switches of material extruders, central turn plate and cooling unit, and the display
panel shows the present temperature and material distribution status.
The working process of the electronic control system is as follows: after turning
on the printing switch on the control panel, the heating device rstly heats up the
four 3D printing nozzles to the melting temperature of ABS material. Then the
103
Fig. 9 Diagram of electronic control system
panel controlling material extrusion sends out signal to command printing nozzles
to extrude the melted ABS material; while the cooling device cools down and
solidify the extruded ABS simultaneously. Meanwhile, the rotation control panel
operates the stepping motor to drive the rotation of central turn plate; the rotating
hexagon takes the three extra nozzles to move towards and deviate from the central
nozzle in a regular manner. A Kuka 6-axis robot arm carries the printing end
effector steadily along the pre-designed route. As a result, a spatial geometry with
variable cross-sections is produced through the collaboration between robot and
customized printing tool.
104
P.F. Yuan et al.
Fig. 10 Deformation of structure printed at different inclinations
2.6
Printing Experiments
The designers conducted a series of multi-dimensional printing test with various

geometrical inputs, including straight curves with different inclinations and
multi-dimensional curvilinear structures with different section proles (Fig. 10).
Experimental results have shown that the printing process of this integrated robotic
system is very stable. It can realize various self-supported spatial geometries within
the accessible area of the 6-axis robot. The printed structure is also proven to be
rigid and effective. The research successfully introduces new methods in spatial
printing through designing variable cross sectional structures.
3 Conclusion and Further Research Goal

While rst testing was successful, there are challenges that will be further investigated by the research. These include material and geometry:
105
Firstly, the material in this study is traditional 3D printing material ABS with
very limited strength in resisting bending and shearing forces. Therefore, it cannot
be applied into larger scale architectural fabrication since the material itself does not
have sufcient structural rigidity compared to building material. Secondly, the
printed result of this research is a complex combination of curvilinear geometries.
The scale and complexity of the structural component limits the design development in the study of surface or volume printing. Although this research is at
preliminary stage, the result has proved its potential. Further research will be
conducted in order to apply this structure performance-oriented multi-dimensional
printing strategy to architectural fabrication. This will include further research of
printing materials, respectively high strength and rapidly solidied materials that
would expand the range and scale of design products. As long as the material
strength meets the demand in architectural industry, this printing technology could
be rapidly used for customized building components, in particular non-linear
structural systems. This structure performance-based printing technology could
easily realize customized fabrication of complex spatial structure with greater
material efciency and smaller energy consumption.
Furthermore, diversied printing strategies will be developed with new combination of rst and secondary structure.
As this research discussed, spatial printing with variable sectional prole is
feasible. However, the research should not be limited to curvilinear printing; different spatial structures could be developed based on the curvilinear geometry. With
further improvements of the mechanics of the end-effector it will be possible to
handle volumetric structure printing. The research indicates new possibilities for
robotically printed, innovative structures for architecture and the building industry.
Acknowledgments The authors would like to acknowledge project group information: Project
Name: Robotic Extrusion(Robotic 6-Axis 3D Printing); Brief Info: 3-week group work of Digital
Future Shanghai Summer Workshop 2014, Shanghai; Design Team: SHI Ji in collaboration with
LIU Xun/LUO Ruihua/CUI Yuqi; Instructor: YU Lei (Project Instructor, from Tsinghua)/
Philip. F. YUAN(Workshop Leader, form Tongji)/Panagiotis Michalatos(Software Tutorial, from
GSD); Photography (Filming) and Editting: SHI Ji.
References
Budig, M, Lim, J and Petrovic, R 2014, Integrating Robotic Fabrication in the Design Process,
Architectural Design, vol. 84, no.3, pp. 23.
Gosline, JM, DeMont, ME and Denny, MW 1986, The Structure and Properties of Spider Silk,
Endeavour, vol. 10, no. 1, pp. 3743.
Gramazio, F and Kohler, M 2008, Digital Materiality in Architecture, Lars Mller Publishers,
Baden.
Lam CXF, Moa XM, Teoha SH, Hutmacher DW (2002) Scaffold development using 3D printing
with a starch-based polymer. Mater Sci Eng 20(12):4956
Lipson H, Kurman M (2013) Fabricated: the new world of 3D printing. Wiley, London
Zheng Y, Bai H, Huang Z, Tian X, Nie F, Zhao Y, Zhai J, Jiang L (2010) Directional water
collection on wetted spider silk. Nature 463:640643
Fabric Forms: The Robotic Positioning

of Fabric Formwork
Ron Culver, Julia Koerner and Joseph Saraan
Abstract The novel, robotically-controlled system delineated by this research

facilitates a rapid and economical workflow realizing a complex network of parametric geometry. The method of concrete fabrication proposed here removes the
traditional limitations of rigid formwork and satises the need for variation in the
realization of parametric design. Lycra is stretched and positioned by robot arms as
a formwork into which concrete is poured. Thus, the flexibility of fabric is translated into flexibility in design permutations. The prototyping considers material
constraints, structural weaknesses, and load-path optimization to achieve a digitally
informed nal geometry.
Keywords Robotics Flexible fabric Casting Parametric design

Gravity simulation Nodal connectivity Mesh relaxation
Scripting
1 Introduction
Traditional rigid formwork has distinct disadvantages for casting complex forms
from concrete. Time-intensive computer numerically controlled milling and subsequent form assembly fail to adequately replicate the compound shapes and
undercuts required of complex geometries. Moreover, the casting of multiple parts
with even slight variation is often cost-prohibitive. The proposed robotic system
facilitates a faster, more precise and more economical workflow to realize complex
or truncated parametric geometry from unique cast masonry components.
R. Culver (&) J. Koerner J. Saraan

UCLA Architecture and Urban Design, Los Angeles, USA
e-mail: ronarc@ucla.edu
J. Koerner
e-mail: juliakoerner@ucla.edu
J. Saraan
e-mail: josephsaraan@ucla.edu
107
108
R. Culver et al.
Fig. 1 Tensioned fabric formwork on a pair of robotic arms
Given the shortcomings described above, this project focuses on developing a

fabrication technique utilizing motion to create 3D space and components. In this
respect, robots are used as precision, time-based tools to generate motion for
variability between the individual prototypes. This independent research project is
part of a UCLA technology seminar that focuses on achieving a robotic casting
system for the fabrication of 3D concrete component typologies (Koerner 2015).
Robotically-controlled, flexible fabric formwork is explored as a means of accurate,
replicable and cost-effective production. Geometrically complex concrete objects
can be fabricated with practically innite organic variation and texture. A pair of
six-axis robotic arms attached to identical flexible fabric sleeves acts as an adjustable formwork for concrete (Fig. 1).
The robotic arms can position the endpoints of the limbs accurately and quickly,
enabling the composition of an intricate series of unique objects as dictated by the
design. For this research, branched objects are designated as a constant starting
shape to be manipulated by the robot arms. This allows for one branched limb to be
xed to a stationary formwork at a lling point and two others to be stretched to the
desired geometry by two robotic arms working synchronously. Custom robotic
assemblies can be manufactured to meet specic project needs beyond the scope of
this research.
Fabric Forms: The Robotic Positioning of Fabric Formwork
109
Fig. 2 American cement

building, DMJM (Vessel
USA)
Fig. 3 Malcolm Leland &

faade elements (Vessel
USA)
The robotically cast components have the potential for deployment as constituents of a compound structure that can also be realized on a building scale.
Load-bearing faades, walls, glazing modules and freestanding sculptures are all
possible applications. The composition of the specic project will dictate many
factors in the design of the components including for example overall scale of the
objects, thickness of casting, or density of the fabric relative to elasticity. For the
purpose of evaluating the system, a small-scale prototype is deployed as a case
study. This self-supporting structural composition consists of 13 individual pieces
of 1:12 scale, standing 32 (81.3 cm) tall.
One relevant precedent is the American Cement Building in Los Angeles by
DMJM (Figs. 2 and 3). This building showcases the potential for precast concrete as
a structural as well as aesthetic element. The original design concept for the faade
of the Los Angeles Broad Museum by Diller Scodio + Renfro would be a closer
example of the true potential of this system (Fig. 4). The far less ambitious faade
that was actually realized is a prime example of the limitations of design/fabrication
systems currently employed and how this robotically controlled flexible fabric
system can dramatically expand the vocabulary of designers worldwide.
110
R. Culver et al.
Fig. 4 Rendering of the

Broad Museum, Los Angeles
( The Broad and Diller
Scodio + Renfro)
2 Digital Process
The proposed system allows for a complex design to be rationalized into discrete
elements that are fabricated and assembled into the nal composition. A design of
either digital or analog origins is simplied into discrete elements that are analyzed
using a parametric feedback loop for structural performance, and then re-engineered
for optimization. Using Karamba for Grasshopper3D, load paths that act on each
member in the system are calculated to understand their behavior prior to fabrication. Subsequently, each element is rened to meet the structural and performance
criteria, then re-generated. Once the components are analyzed, their endpoints are
determined and optimized. These coordinates are sent to the robotic arms, which
translate the Euclidean coordinates into physical space.
2.1
Shape Optimization
The robot lab at UCLA features the arrangement of two industrial robots in tandem.
Therefore, this prototype project is constrained to three controlled end points, one
of which is xed in space and functions as the concrete ll point, while the other
two are attached to robot arms. The shape of the object between the three end points
is optimized for volume, gravity, load path and structural stress points. Increased
volume leads to high gravitational deformation which can be alleviated by
removing either the center or perimeter mass. Removing mass at the perimeter
111
Fig. 5 Design progression of cast elements showing material and structural optimization
simplies the object, the load path and results in the least amount of material per
piece (Fig. 5). Curves are introduced between the end points to effectively reduce
structurally vulnerable sharp turns and provide additional mass where it is needed at
the connectors.
2.2
From Agency to Analytics
The design process begins by coding in the input parameters necessary for a specic
morphology. These include constraints such as the bounding box dimensions, build
volume, the size of each component and the characteristics of the casting material.
The size of the original fabric formwork is also considered, as its stretch limits will
determine the ultimate size of each component. These constraints are written into
the Grasshopper denition to form a quantitative basis for the design.
The next phase requires intuition as well as digital modeling (Lynn 1999). By
manipulating an array of nodes connected by line segments, a form emerges that
best demonstrates the unique characteristics of the fabric formwork system. Each
node represents a connective piece in the composition while the lines represent the
cast concrete elements. Each component is analyzed as a discrete agent in a larger
structural system, on local then macro levels (Fig. 6).
When the position of the nodes are established rst, the interstitial connected
network of line segments are generated by an algorithm which groups the array of
points into clusters of three-legged objects. Each set of objects can be considered in
isolation (Fig. 7).
Fig. 6 Digital modeling based on mesh relaxation and simulation using Kangaroo3D
112
R. Culver et al.
Fig. 7 Digital sorting from point cloud to discrete three-limbed objects
Once the wireframe lattice of nodal connections is built, each line segment is
given thickness independently, simulating the fabric stretching of the Lycra
formwork. Consideration is given to the fabric elasticity limits and thermochemical
curing of the cement when evaluating the composition.
Vertical load paths and bending moments within the array are evaluated and the
model is recursively adjusted as necessary to attain the design goal. A similar
technique of clustering nodes into discreet elements has been implemented in the
installation Cast Thicket by Kenneth Tracy and Christine Yogiaman. This prototype uses nodes and a connectivity network as the design generators for a
cast-in-place concrete technique using plastic formwork (Tracy 2014).
2.3
Digital to Analog
The translation from Euclidean space to physical space is achieved through BD

Move software and the use of robotic arms. Each limb endpoint is assigned its
coordinates in space and inclination angle to position its end-arm tool so that the
surface normal of the end nodes always faces the concrete center. Each position is
systematically sent to the robots as the pieces are cast (Fig. 8). Every physical
constraint is modeled digitally from the armature and its pedestal to the custom
end-arm tool. It is essential to have an accurate 3D environment to minimize error.
Any slight omission in the digital le would result in a misaligned nal product.
113
Fig. 8 Digital composition (left) and oriented objects in rig for casting (right)
2.4
Matrix Composition
Early composition studies of the design show that elements with three limbs have
limited design capabilities. Without employing weaving strategies, the resultant
matrix is predominantly two-dimensional with restricted opportunity in composition depth. To counter this, a truncated tetrahedral coupler acts as the interface
between concrete elements. This not only affords design flexibility but also facilitates a logical means of attachment by inserting bolts through the truncated faces
into sleeve inserts cast into the end of each limb. In addition, the connector adds
resilience to the structural assembly while allowing flexibility in the connective
nodes. This coupler can be built with a flexible material to allow bending moments
or it can be rigid to increase the structural integrity of the composition (Fig. 9).
The success of robotic positioned fabric formwork does not reside with this
connector, however. It is employed as part of the composition of the installation,
and designed to be a proof of concept for the system. The system itself is intended
Fig. 9 2D characteristics of three-limbed objects (left) and coupler design (right)
114
R. Culver et al.
as a vehicle to solve a myriad of design and construction challenges. It is eminently

conceivable that a composition be designed without a similar connector, where each
cast member connects directly to the adjacent ones as shown in DMJMs American
Cement Building faade. It is also possible that the cast piece itself acts as a
bespoke coupler for highly variable curtain wall outrigger systems that engage as
interstitial elements between standardized tube sections.
2.5
Robotic Armature and End-Arm Tools
An armature capable of anchoring one cast object is installed between two 6-axis
robotic arms such that the forces applied to the fabric formwork will not disturb the
armature. One fabric limb is afxed to each robotic arm and the remaining limb is
afxed to the top of the armature (Fig. 10). This top limb is attached so that it can
also serve as the cement lling point with a retaining cutout that constrains the end
to the shape of the nodal connector. After lling, the open end is capped with an
acrylic form with inserted nut attached for bolting (Fig. 11).
Each of the side limbs similarly has an end-arm tool in the shape of the matrix
connector. It includes an insert nut attached that allows the fabric to be tensioned
prior to pouring the cement mixture. The tools allow secure and rapid
attachment/removal from the robotic arms for more consistent casting (Fig. 12).
2.6
Flexible Fabric Form
While the original patent held by Joseph Shivers on segmented copolyetherester

elastomers (otherwise known as Spandex) indicates a potential 325 % elasticity
(Shivers 1962) and spandex yarn has 650 % 30 elongation (Senthilkumar et al.
Fig. 10 Typical robotic
casting arrangement
115
Fig. 11 Fill opening w/reinforcing & insert
Fig. 12 End effector on robotic arm w/reinforcing (left), with fabric (center), and clamp (right)
2012), it is invariably blended up to 80 % with other natural and man-made fabrics

in order to make it more practical in the manufacture of garments (Reisch 1999).
The ultimate elasticity of spandex materials depends on the specic blend type and
ratio. When stretched on the robots to maximum elasticity in eld tests, a seam
length of 14.25 reached a maximum length of 31.0 before failure, representing
elasticity of 217.5 %. The fabric itself could be stressed further but the practical
assembly requires stitching, so it was tested with a flexible zigzag stitching using
polyester thread. Even at maximum tension, the form lled with casting material
naturally and with sufcient volume to produce usable results.
While other fabric casting methods seek to constrain the fabric to specic shapes
without undue deviation, or tailor the fabric specically to the cast object (West and
Araya 2010), the Lycra used in this project specically enables the radical
116
R. Culver et al.
manipulation of the fabric formwork. This allows an almost innite number of

potential cast concrete geometries with a singular fabric starting prole. The use of
fabric as a formwork also plays a role in the compressive strength of the member.
Because Lycra allows moisture to wick through its fabric while retaining the
cement, a high-strength concrete is produced with minimal air pockets entrained in
the object (Orr et al. 2011).
3 Prototyping Procedure
The combination of a nodal coupler and concrete element has distinct advantages
over other flexible formwork approaches such as that employed in Crease, Fold,
Pour (Kaczynski 2013). The installation utilized a monolithic pour with flexible
formwork but had no means of avoiding form failure other than pouring in stages
with resultant cold joints that weaken the structure. By breaking the composition
down into individual pieces attached with nodal matrix connectors, the design can
be rapidly assembled and disassembled, and the node strength increases.
To achieve these advantages, the nodal matrix connector must be rapidly
mountable, easy to assemble and capable of handling temporary cantilevered
loading during assembly as each new cast piece is initially attached (Fig. 13, left).
Making the attachment screws bypass the truncated tetrahedral center and attach to
the wall adjacent to the node of the cast object enables the positioning of all bolts
through the connector center but adds torsional stress to the wall (Fig. 13, right).
As a flexible fabric that has a natural form-found rather than explicit geometry
(West and Araya 2010), Lycra formwork produces a unique shape and texture. This
system not only creates accurately positioned geometries, it also allows for organic
variation, texture and a natural unpredictability (Fig. 14).
Fig. 13 3D printed nodal connector attached to cast limb with cast nut (left) and view of struts
inside nodal connector (right)
117
Fig. 14 Discrete cast objects demonstrating organic variation and textural quality
4 Learning from Failure

4.1
Structural Weaknesses
Initial casting experiments ranged from plaster of Paris to high-strength cementitious grouts. Experiments with ready-mixed 9,000 psi construction grout yielded
good results with fast cure time of 3,000 psi within one hour, enabling removal
from the robotic arms with an initial set time of 30 min. Despite the high compressive strength, the nal cast objects were still highly susceptible to fracturing at
nite edges and breakage from sudden impact. Introduction of nylon monolament ber at the rate of 0.032 oz per pound of cement dramatically improved the
tensile strength and edge denition without compromising surface appearance
(Fig. 15).
Further reinforcements of the cast objects can be achieved in larger scale castings with a slip-jointed steel reinforcing bar installed prior to casting.
Post-tensioning can also be achieved by introducing flexible polyethylene tubing
through the objects prior to casting that will accept a steel cable strung through the
entire assembly after casting. An alternate method of support has reinforcing bars
attached to the end-arm tools drawn into the fabric forms during tensioning. The
assembly for this research employs this method of reinforcing at a 1:12 scale.
A 16.5 gauge form tie-wire at the prototype scale represents the equivalent of a 5/8
(12.7/20.32 cm) diameter reinforcing bar.
118
R. Culver et al.
Fig. 15 Unique connecting conditions with a consistent coupler element
4.2
Flexible Fabric, Gravity and Minimal Dimensions
Early casting tests reveal that restricting the fabric to less than 1 (1.54 cm) thick
leads to breakage, even with 9,000 psi cement. Objects that come to a minimal edge
are similarly problematic. Maintaining adequate dimensions throughout the entire
cast object becomes a critical requirement to successful production of the pieces.
Increasing fabric dimensions beyond our given size results in uncontrollable,
excessive masses. If the fabric is not pre-tensioned to at least 60 % of maximum
elasticity and dimensions exceed allowable limits at any point, runaway loading
occurs at the point of greatest volume. Therefore, close control of the potential
volume must be maintained through appropriate pre-tensioning. Failures of this
type occur most frequently when using 4-way (biaxial) stretch material (designated
as Spandex brand) but also occur with 2-way (monaxial) stretch material (designated as Lycra brand) when limits are exceeded.
5 Construction Implications
Product applicability includes self-supporting structural faade systems, sculptural
breezeblock or wall applications, glazing modules and freestanding sculptural
elements. Another potential use includes uniquely-shaped connectors between
119
standardized tube sections for facade or curtain wall systems. The prefabrication of
individual elements speeds time-intensive site-work and diversies the supply of
materials to a building site.
As a result, formwork costs can be dramatically reduced and construction waste
is virtually eliminated. This is a highly sustainable manufacturing method when
compared to alternative processes such as computer numerically controlled milling
of single-use forms that will subsequently be destroyed after use.
This research has further implications in architectural screen walls. Erwin Hauer
pioneered the interlocking concrete wall as a modular facade system. These were
used as interior porous partitions whose interlocking character and elegance are
considered one of the quintessential works of modernism in Domus 19281999.
Hauers exploration of the curvature, plasticity and weaving of concrete are echoed
in the potential of this system (Fig. 16).
Fig. 16 Prototype composition of flexible fabric-formed robotically-cast objects
120
R. Culver et al.
6 Conclusion
The dissemination of robotic fabric formwork into the construction industry would
not only allow for the realization of the next generation of parametric, node-based
lattice structures, but would signicantly reduce concrete construction costs for
conventional components. Multi-directional pieces can be cast without rigid
formwork and their destructive removal.
The robotically-controlled system resulting from this research proved its ability
to facilitate a rapid and economical workflow to realize complex or truncated
parametric geometry. Most importantly, this method can help the construction
industry adapt to emerging digital fabrication tools and allow for rapid design to
production cycles that go beyond rapid prototyping. The effect is a
digital-to-physical workflow that is abundantly more flexible in its dimensional
freedom and more economical than the industry standard.
Future development of this system would likely include the use of custom-built
robots tailored specically to the congurations of the pieces to be cast, allowing for
many more points and limbs if desired. Customized robotic work processes could one
day replace human labor in many professions including those on the construction site.
Acknowledgments The authors gratefully acknowledge the technical and material teaching of
Julia Koerner and Peter Vikar from conception to completion. This work would not have been
possible without their expertise at the UCLA School of Architecture and Urban Design. Special
thanks to Greg Lynn and Guvenc Ozel for their guidance towards the completion of the project.
The early experiments in casting with fabric for this project were conducted with fellow UCLA
students Shobitha Jacob, Oscar Li and Qi Zhang who were a part of the initial research.
References
Kaczynski, MP 2013, Crease, Fold, Pour: Rethinking Flexible Formwork with Digital Fabrication
and Origami Folding, ACADIA 13: Adaptive Architecture, Proceedings of the 33rd Annual
Conference of the Association for Computer Aided Design in Architecture, 2426 October,
2013, Cambridge, pp. 419420.
Koerner, J 2015, Syllabus Technology Seminar UCLA Suprastudio/Independent Study.
Lynn, G 1999, Animate Form, Princeton Architectural Press, New York.
Orr, J, Derby, A, Ibell, T, Evernden, M and Otlet M 2011, Concrete Structures Using Fabric
Formwork, University of Bath, United Kingdom.
Reisch, M 1999, Spandex, Chemical and Engineering News, American Chemical Society,
Washington, D.C.
Senthilkumar, M, Sounderraj, S and Anbumani, N 2012, Effect of Spandex Input Tension,
Spandex Linear Density and Cotton Yarn Loop Length on Dynamic Elastic Behavior of
Cotton/Spandex Knitted Fabrics, Journal of Textile and Apparel, Technology and
Management, NC State University, North Carolina.
121
Shivers J 1962, Segmented Copolyetherester Elastomers, United States Patent Ofce no.
3,023,192.
Tracy, K 2014, Textile Effects Semi-rigid Concrete Formwork, Riverside Architectural Press,
Cambridge, Ontario, Canada.
West, M and Araya, R 2010, Fabric-Formwork For Reinforced Concrete Structures And
Architecture, Centre for Architectural Structures and Technology (C.A.S.T.), University of
Manitoba, Canada.
Path Planning for Robotic Artistic Stone

Surface Production
Gregor Steinhagen, Johannes Braumann, Jan Brninghaus,
Matthias Neuhaus, Sigrid Brell-Cokcan and Bernd Kuhlenktter
Abstract Traditional artistic stone processing techniques offer vast possibilities for
nishing stone products. However, stone processing is physically highly demanding work requiring stamina as well as skill. This makes products expensive to
produce and the detailed design only accessible for skilled masons as an efcient
communication between designers and masons is difcult. We introduce a
robot-based approach to produce artistic surfaces for individualized stone products. First, distinctive traditional, manual processing techniques will be introduced
and analyzed towards enabling us to specify the necessary requirements of an
adaption to an industrial robot. These requirements are then implemented in an
automated tool and an automated path planning algorithm. Building upon a visual
programming environment we will present an accessible interface that allows the
user to apply customizable stone structuring patterns to an individual stone product.
G. Steinhagen (&) M. Neuhaus

TU Dortmund, Institut fr Produktionssysteme (IPS), Dortmund, Germany
e-mail: gregor.steinhagen@ips.tu-dortmund.de
M. Neuhaus
e-mail: matthias.neuhaus@ips.tu-dortmund.de
J. Braumann
Association for Robots in Architecture, University for Arts and Design Linz, Linz, Austria
e-mail: johannes@robotsinarchitecture.org
J. Brninghaus S. Brell-Cokcan
Association for Robots in Architecture, RWTH Aachen University, Chair of Individualized
Building Production, Aachen, Germany
e-mail: jan@robotsinarchitecture.org
S. Brell-Cokcan
e-mail: sigrid@robotsinarchitecture.org
B. Kuhlenktter
Ruhr-Universitt Bochum, Lehrstuhl fr Produktionssysteme (LPS), Bochum, Germany
e-mail: Kuhlenkoetter@lps.rub.de
123
124
G. Steinhagen et al.
1 Introduction
Stonemasonry is an old craft, which developed over centuries. It is characterised by
a high number of different techniques, which make use of both tool geometry, as
well as the tools handling. There are also a vast number of different stones with
different heterogeneous behaviour: Not every technique is applicable for every
stone. While some techniques are easy to produce, other techniques are demanding
and require high manual skills and long training.
Thus the production of artistic stone surfaces with such techniques requires
knowledge and talent. This restricts the accessibility of techniques for designers and
architectsthe sole possibility is the communication with a skilled mason.
However, the number of techniques and their variability combined with the many
different kinds of stones make it impossible to communicate the vision of ones
design idea in detail. This results in the re-application of traditional old patterns
whereas new design strategies do not often nd their way into architectural
applications. As such, the main eld of application for stone-structuring techniques
is still the restoration of old buildings.
Another drawback is that most techniques are physically demanding, so that
even a skilled mason needs a long time to produce large surfacesmaking the
application of stone structuring to building fronts very expensive.
While there exist some large specialised machines for stone structuring, they
either rely on very simple, xed chisel-strategies or simply aim to imitate traditional
techniques with grinding tools. As such, they do not allow the user to go beyond
traditional structuring strategies or to work with three-dimensional, curved, surfaces. Generic methods that can also be applied to stone structuring such as milling
(Hayes et al. 2014), sawing (Garrido et al. 2009) and waterjet-cutting (Bortolussi
et al. 2009; Ciccu and Bortolussi 2010) are being used in industry but are either
very time consuming or do not closely emulate the nish of traditional stone
structuring techniques. Current research in the eld of architecture and design
focuses mostly on cutting, rather than surface processing (Rippmann et al. 2013;
Kaczynski et al. 2011). All these problems are addressed in a new robot based
approach, where a robot-mounted tool performs the traditional techniques and then
goes beyond the scope of manual processes (Fig. 1). The freedom of the robot
allows variation in the technique parameters and results in huge design possibilities.
Furthermore, a design interface is implemented which allows designers to model
their ideas and carve them into stone as the special masonry knowledge is
embedded within the code and thus not required of the user.
First we will describe the techniques, which have been analyzed. We will then
describe the adaption of these techniques to the robot. The path planning which
connects the design interface with real robot cell will be shown. In conclusion, we
will give an overview to the design interface and show rst examples of the design
possibilities.
Path Planning for Robotic Artistic Stone Surface Production
125
Fig. 1 Regular and irregular robotically fabricated stone surface patterns
2 Dening a Stone Structuring Process

Robotic arms are highly multifunctional machines with a power and precision
thatat least in combinationby far exceed the capabilities of human arms.
However, in order to be able to program a particular robotic task, we have to be able
to clearly dene it.
This task denition is one of the main challenges towards applying robotic
labour to stone surface structuring: The most common robotic applications are well
dened and structured (e.g. pick-and-placing or spot-welding in the automotive
industry), while some more complex processes such as milling can at least be
quantied, measured, and evaluated based on static criteria such as the minimization of difference between the digital data and resulting physical output.
Stone surfacing is much less dened: Two stone masons utilizing the same
technique and an identical tool can produce very different surfaces that are the result
of small variations in force, angle, speed, and other parameters. These variations are
then not defects, but rather make up the aesthetic appeal of traditionally structured
stone surfaces. Previous research and literature reviews show that a direct translation of human movements to robotic movements is highly challenging even for
tasks that are considered easy for humans. The precise timing and force control
needed to exactly emulate a human mason was therefore considered to be beyond
126
the scope of the research project. As such, the idea was that rather than trying to
emulate human processes directly, we would attempt to create optimized robotic
processes that incorporate traditional masonry strategies but augment them in the
areas where a robot exceeds a stonemason, namely, in speed and accuracy. The rst
step towards implementing new movement strategies is to analyse existing, manual
techniques and to evaluate them in regards to compatibility with robotic
processes.
3 Analysis of Traditional Techniques

Due to the high number of different techniques, a rst selection of popular strategies
was made with the help of the masons of Bamberger Natursteinwerke, who later
also supported the analysis of the different techniques. The selection was based on
the dissemination of techniques and an estimation of their potential for automation.
Tooling, aligned tooling, punching and bush hammering were selected as rst
experiments. When not considering the complete imitation of the masons movement, the most important aspects of the techniques are the chisel movement and the
kinetic energy, which is applied to the chisel with the hammer. Both are analysed
with the high speed camera system GOM Pontos HS (Fig. 2, left) capturing
the manually-performed techniques with 20,000 fps. The videos can afterwards be
analysed with the help of markers placed on the chisel and the hammer. Thus
we can analyse the relevant movements in X and Y direction as well as the angle
around the Z axis (Fig. 2, right). The energy was estimated by the speed of the
hammer before impact and its weight. In the process of analysis we also analysed
further aspects of automation such as predictability of results and necessary positioning precision (for detailed information on analysis, see Steinhagen et al. 2015).
Similar methods have so far only been applied in the eld of anthropology and
archaeology to understand prehistoric tooling processes (Williams et al. 2010).
Fig. 2 Setup of the camera system with camera and stone specimen for chiseling (left), Picture of
a lmed chisel with the measured coordinates (right)
127
1. Tooling This technique is very common for nishing traditional stone surfaces,
creating highly characteristic rounded grooves. An example of a processed
surface can be seen in Fig. 4. It is performed with a drove chisel and a mallet.
An example of the movement can be seen in Fig. 3. While moving through the
stone the chisel is rotating slightly. At the end of the movement it leaves the
stone. This movement results in the characteristic appearance. The Energy was
estimated to be 19.46 J. This energy can be applied with a robot mounted actor.
Based on these rst measurements we can say with a high condence that the
movement of the chisel is challenging but reproducible. Positioning accuracy
and predictability of the results are manageable as well.
2. Aligned Tooling Aligned tooling is a variation of tooling which main application can be found at the edges and corners of complex stone products where
there is no space for performing the original tooling. The chisel does not leave
the stone and moves linear into the stone, resulting in triangular, rather than
rounded grooves (Fig. 5). The movement analysis exhibits only a small rotation
of the chisel. Since smaller drove chisels are commonly used, the energy level is
smaller than what we observed with normal tooling, while other values are
mostly within the same range. Resulting from all this aspects we conclude that
the automation of this technique should be even easier than normal tooling.
3. Punching This technique is performed with a punch and a hammer. There are
different variations of this technique: Either single punches are performed in
point punching or concluding punches are performed to produce a linear pattern
through in line punching (Fig. 4, center). However, especially the second one is
difcult to automate since the predictability of the punches is low. For
point-punching, the movement is relatively simple since it linearly punches into
the stone with an estimated energy of 23.99 J.
Fig. 3 Movement of the chisel for normal tooling
128
Fig. 4 Results of tooling (left), punching (center) and bush hammering (right)
Fig. 5 Comparison of tooling
(left) and aligned tooling
(right)
4. Bush Hammering This technique uses a hammer head with a number of tips
and can be either performed with a special bush hammer or pneumatic tools. The
hammer hits the surface but does not move deep into the stone, producing a
shallow pattern on the surface (Fig. 4, right). This results in a movement that is
relatively easy to perform and requires just 1.14 J, but only offers a limited
degree of variability.
4 Design Parameters for Stone Structuring

Building upon the analysis of human stone masons using high-speed cameras a
custom, a modular tool was developed that emulates the stone masons
micro-movements and decouples its chisel from the robot itself, preventing backlash to the robots gear system. Therefore, the robot acts in the macro-scale as a
129
Fig. 6 Custom-developed stone structuring tool and toolframe
high-speed, accurate spatial positioner while the modular tool interacts with the
stone. This modular approach allowed us to pursue the time-constrained process
development in two parallel tracks, while also facilitating future maintenance and
trouble-shooting.
Based on the custom-developed stone-structuring tool (Fig. 6), we have identied a number of design parameters that can be adjusted to greatly influence the
surface nish and process time. These have been grouped into three different layers,
with the rst one dening the general layout of the design and the second and third
one ne-tuning the structure and depth of the processing.
Layer 1. Tool position and rotation in the materials surface. The general tool
position has got the most signicant impact on the surface as it denes its overlaying structure, e.g. by following the isocurves of a surface, tracing lines, or being
aligned according to the raster information of an image (see below). This positioning happens exclusively in the XY plane of flat material, or alternatively on the
UV/mesh parametrization of curved surfaces.
Layer 2. Relative tool inclination. The tools inclination in relation to the
stone-surface is performed around the chisel edge by the internal tool mechanism.
Due to the optimized movement of the tool, following the predened path rather
than a straight line, the tool inclination has got a distinctive effect on the surface
nish, allowing us to accurately ne-tune the width of each stroke. Figure 7 shows a
range of possibilities the designer can use.
Layer 3. Hardware: Width and Force. The third strategy to influence the surface
nish is through adjusting the hardware parameters, most notably the tool dimensions and the force with which the tooltip is applied to the stone. In particular, the
tool-width has a signicant impact on process timehowever, while larger tools
130
Fig. 7 Variations in surface effect relating purely to tool inclination
allow for quicker processing of structures, they are also more geometrically limited
when not following the rulings of a surface.
For the initial experiments we have created a number of different strategies that
utilize the three parameters previously described, depending on the design intent.
These rst structures are based on patterns, curves, raster images, or a combination
of the above.
1. Patterns. Manual stone structuring is based on regular patterns that are applied
to a surface. As such, we have created a set of adjustable, regular patterns that
can then also be perturbed through attractor points or other parameters. In Fig. 8,
a sine curve adjusts the rotation both within the stone surface plane, as well as
around the tools X axis, resulting in a 3D form which would not be possible
with a manual process.
2. Curves. Here the toolpath is provided as a list of curves that have to be chiseled
into the stone. A signicant challenge is the division of a curve into a number of
linear segments with a xed length (according to the used tooltip). To divide the
curve according to different chisel sizes we assessed the chord, which connects the
starting point on the surface with points on the curve. The length has to correspond
to the chosen chisel length. This approach works in cases where the curvature is
not too high, and there are no complex forms, small grooves or waveforms. Based
on the chord middle the chisel position is then derived. If a long curve is divided in
different chisel segments, the overlap of two chisel positions is a parameter, which
can vary. Thus a curve can be build up from a number of chisel positions with a
xed chisel size. The inclination of the chisel is derived from the normal surface
vector of this chisel position and the design inclination dened by the user.
131
Fig. 8 Procedurally generated and robotically fabricated spatial stone structure, using the newly
developed tool and software
3. Raster Images. In previous research we have already explored the possibilities

of using raster images for surface structuring. In the case of stone structuring,
the brightness values of an image are sampled and then turned into toolpaths.
Depending on the (processed) surface of each stone, the stone becomes then
either brighter or darker where it is hit. In addition to locally adjusting the
brightness, the orientation of the tool geometry can also be informed by the
process, e.g. based on vectorized geometry.
5 Geometry and Representation

The representation of the developed strategies is a very important topic as it is
crucial for allowing a new user to estimate the result of the programmed process.
An ideal way to accurately represent the effect of the tool would be to perform
Boolean operations for each hit. However, such a process would be extremely
computationally expensive, and would thus interrupt the fluid design process. As an
alternative, the most basic representation of the tool would be a line, representing its
chisel edge. While this visualizes the effect of the rst fabrication layer (see above)
it would not show the effect of the tools inclination. We therefore implemented a
132
Fig. 9 Pre-calculated surface effect of chiseling tool depending on tool angle and depth (left),
curvature angle analysis (right)
geometric model that calculates the approximate shape of each chiseling movement
based on the tool size, inclination, force and stone without having to physically
intersect two geometries. Therefore, the geometric footprint and memory requirements are kept low, even for complex structures (Fig. 9).
However, signicant challenges arise once doubly-curved surfaces are processed
(or when the tools y-axis moves away from the straight rulings of the surface), both
in regards to the physical process and the initial preview: The contact surface
between the tool and the stone then covers only a part of the length of the tool,
which can possibly chip either tool or stone. Depending on the stone that is being
used, the user has to evaluate the effect based on the visual feedback provided by
our software, and either takes the risk with soft stone, or use an automated process
to optimize the tools orientation with the goal of lessening the irregular effect.
6 Programming Stone Structuring Design Patterns

We expect that by solving the kinematic complexity and timing of a stone structuring process with an intelligent, modular tool, the robot can be used much more
freely as a design tool with a comparably low stone-specic overhead. This
macro-path planning is however still highly challenging, as no commercial solutions for the dynamic structuring of stone are available on the market. Initial virtual
simulations of early stone-structuring approaches showed that the large-size of the
structuring tool, along with the geometric constraints of traditional chiselling tools,
can very quickly lead to unreachable positions, and singularitiesthus requiring a
capable, dynamic robot simulation environment where such processes can be
reliably dened, simulated, and optimized. As such we decided to build upon the
flexible KUKA|prc framework (Brell-Cokcan and Braumann 2010) and expanded it
at the source code level with new simulation capabilities, improved interpolation
algorithms, and multi-core optimizations to more efciently optimize stone structuring processes. KUKA|prc itself builds upon the visual programming environment
Grasshopper and expands the standard modules with a range of robot simulation
133
and control components. As such it is ideal for the usage as a path planning tool for
complex technical systems with an easy to use interface.
The path planning is divided in different steps. First, the user designs an idea as
described above and receives feedback regarding the applicability of different chisel
sizes to the curvature by color-coding each preview position. With this information
the designer can decide when to apply a distinct chisel size to a surface point.
For non-planar surfaces, the whole path planning is performed on a scanned mesh
representation of the stone. This is necessary because of high production deviations in
the machining of stone products, where the tool is continuously ground down by the
abrasive stone. These initial offline steps can be performed without a robot.
The rst step where the robot is needed is the measuring of the initial stone
position. We implemented a measurement system, which was developed in a previous research project (Mller et al. 2014). It is based on a laser triangulation
sensor, which captures a number of points on the stone. The points are than matched with a point cloud of the stone and thus the stone is virtually placed in the
robots simulated workspace.
The robots poses are then updated and checked for singularities and other
conflicts such as reachability. Due to the geometries of stone processing, the only
probable singularities in the process are poses where the fourth axis and sixth axis
are aligned. These positions can be resolved if the inclination of the chisel is
slightly changed for the corresponding movement instruction. This approach builds
upon the new monitoring function of KUKA|prc which exposes much of the
simulated robot data for analysis and optimization. For an iterative update of the
chisel positions we then coupled the robot simulation with the plugin Hoopsnake.
Thus the angle is automatically altered only as much as is absolutely necessary.
As a further step, we plan to implement a module that optimizes the sequence of
the single chisel poses in regard to the process time, so that even complex stone
products with a high number of chisel positions can be produced efciently.
7 Conclusion
In this chapter, we have presented our work on the analysis of traditional stone
processing techniques and their adaption to an industrial robot. By implementing
this knowhow and strategies into the custom soft- and hardware, our research
allows designers to be directly involved in the structuring of stone. Changeable
parameters such as tool inclination and different design patterns can be adjusted or
new stone products can be dened from scratch.
The automated path planning also enables the designer to focus on the design
rather than on the intricacies of robotic path planning such as singularities.
Thus we have developed a competitive approach that will enable a broader user
group to apply traditional stone processing techniques. Our further work will focus
on the optimisation of the shown technique, the adaption of other techniques, and
134
new robot control interfaces for dynamic processes (Braumann and Brell-Cokcan
2015). Furthermore, we will optimize the process speed by analysing and optimising the dynamic behaviour of the robot and tool.
135
Acknowledgments The research leading to these results has received funding from the European
Union Seventh Framework Programme (FP7/2007-2013) under grant agreement n606453. SME
consortium partner: Klero GmbH, Bamberger Natursteinwerk, G. Gibson & Co Ltd, II Architects.
Research consortium partners: TU Dortmund University, Association for Robots in Architecture,
Labor. Associate consortium partner: KUKA. Web: www.arosu.eu.
References
Bortolussi, A, Foldyna, J, Ciccu, R, Scucka, J, Martinec, P and Sitek, L 2009, Ornamental stone
surface treatment by pulsating water jets, Proceedings of the 9th Pacic Rim International
Conference on Water Jetting technology, Koriyama-city.
Braumann, J and Brell-Cokcan, S 2015, Adaptive Robot Control, Proceedings of the 33rd
eCAADe Conference, Vienna.
Brell-Cokcan, S and Braumann, J 2010, A New Parametric Design Tool for Robot Milling, in
Life In:FormationProceedings of the 30th Annual Conference of the Association for
Computer Aided Design in Architecture, New York, pp. 357363.
Ciccu, R and Bortolussi, A 2010, Stone Surface Finishing by Pulsed Waterjets, Proceedings of
the Global Stone Congress 2010, Alicanta.
Garrido, J, Martin, RM, Armesto, JI and Lopez, JS 2009, Stone cutting automation technology
based on features, Proceedings of the 2009 IEEE International Conference on Mechatronics,
Malaga.
Hayes, J, Fai, S, and While, P 2014, Digitally-Assisted Stone Carving on Canada's Parliament
Hill, Proceedings of the 32nd eCAADe Conference, Newcastle upon Tyne, pp. 643651.
Kaczynski, M, McGee, W and Pigram, D 2011, Robotically Fabricated Thin-shell Vaulting: A
method for the integration of multi-axis fabrication processes with algorithmic form-nding
techniques, ACADIA 11: Integration through Computation: Proceedings of the 31st Annual
Conference of the Association for Computer Aided Design in Architecture, Banff, pp. 114121.
Mller, M, Brninghaus, J and Kuhlenktter, B 2014, Konzept zur vollautomatischen
Bauteillagebestimmung von Freiformbauteilen zur Korrektur fr die Roboterbearbeitung in
Konferenzband Automation 2014, Baden-Baden.
Rippmann, M, Curry, J, Escobedo, D and Block, P 2013, Optimising Stone- Cutting Strategies for
Freeform Masonry Vaults, Proceedings of the International Shell and Spatial Structures
Symposium IASS, Wroclaw.
Steinhagen, G, Brninghaus, J and Kuhlenktter, B 2015, Robotergesttzte knstlerische
Steinbearbeitung, in Tagungsband Mechatronik 2015, Dortmund.
Williams, EM, Gordon, AD and Richmond, BG 2010, Upper Limb Kinematics and the Role of
the Wrist During Stone Tool Production, American Journal of Physical Anthropology, 2010
Sep, vol. 143, no. 1, pp. 13445.
Towards a Micro Design of Acoustic

Surfaces
Robotic Fabrication of Complex Pattern
Geometries
Dagmar Reinhardt, Densil Cabrera, Alexander Jung and Rod Watt
Abstract In the context of acoustic performance in architecture, this paper presents

research into the computational design and robotic fabrication of surfaces with
micro-geometries that can change the acoustic response of space. It explores the
design affordances for acoustically efcient patterns for sound scattering - between
complex geometries, acoustical effects, and robotic fabrication. Spline curves pose a
problem for the translation between geometry and material fabrication, specically
when a series of tests is required with a high degree of detail. Whereas 3D printed
samples are impractically small, and CNC fabrication is limited by tool path axis,
robotic fabrication enables precision for 1:10 scale model prototypes such as the
quick sampling of sound discs that can be used to analyze acoustic scattering.
Through a process of reverse engineering from parametric modeling to scale model
production to physical simulation, the acoustic reflective properties of surface
patterns are investigated for scattering coefcients, in order to derive statistical data
on acoustic properties of these surfaces, and to deduce design rules.
Keywords Subtractive manufacturing Complex geometries Parametric design

Robotic fabrication Sound scattering
D. Reinhardt (&) D. Cabrera A. Jung R. Watt

Faculty of Architecture, Design and Planning, The University of Sydney,
Sydney, Australia
e-mail: dagmar.reinhardt@sydney.edu.au
D. Cabrera
e-mail: densil.cabrera@sydney.edu.au
A. Jung
e-mail: alexander.jung@sydney.edu.au
R. Watt
e-mail: rodney.watt@sydney.edu.au
137
138
D. Reinhardt et al.
1 Introduction
Architectural surfaces shape the way in which we hear space. Sounds such as
speech can be thought of as an acoustic signal that is heard through architecture,
which forms an extended acoustic system. Architecture transforms a sound that
travels, but also influences the way in which signals are produced by people: speech
projection depends on the visual and auditory environment experienced by the
talker. On a very basic level, all architectural surfaces express acoustic space
through the way in which sound is reflected at each surface, depending on material
properties, surface geometry, and the sound elds spatial, spectral and temporal
features. Acoustic performance results from the complex combination of spatial
volume, building envelope and surface properties, which combine to affect speech
transmission, in many contexts yielding reduced intelligibility. Through diffusion,
scattering or appropriate reflection/absorption of sound (Cox and DAntonio 2009),
this degradation can be improved. And while the discourse of performance as a key
design factor of the built environment has been associated with design computation,
geometry, material, or structural performance (Kolarevic 2005, Oxman 2008),
recent studies have just begun to integrate acoustic performance (Bonwetsch et al.
2008; Burry et al. 2011; Peters and Burry 2011).
This paper reports on ongoing research into the acoustic effects of complex
architectural geometries (Reinhardt et al.2012, 2013), with a focus here on robotic
fabrication of micro-geometric surfaces that could be used to improve acoustic
performance by scattering (Fig. 1). In doing so, it expands previous research into
the robotic fabrication of spatial geometries that change the coloration of sound
(Reinhardt et al. 2014). We are presenting here ongoing research and workflow
shared between architecture and acoustics, for the design of robotically fabricated
Fig. 1 Robotic milling of micro-geometries
Towards a Micro Design of Acoustic Surfaces
139
scattering surfaces as scale models for acoustic testing. In the following, the paper
introduces parameters of sound reflections; reports about a series of preliminary
design and physical tests of acoustic patterns. It further discusses a transfer from
generative tools to 6-axis robotic fabrication, linked to the angle and cavity depth in
a surface medium that impact on acoustic performance.
2 Specular Reflections and Acoustic Scattering

Like light, the propagation of sound in space can be understood through ray and
wave theory, which is more important on a human scale because sound wavelengths
can be signicantly large compared to the size of objects, surfaces, and surface
elements in the human environment. To address this complexity, architectural
acoustics employs a variety of theoretical paradigms for modeling the behavior of
sound: including analytic, statistical and numerical methods; based on ray, wave
and particle propagation; in any or all of the domains of time, frequency and space.
In architectural spaces, sound focusing, discrete and flutter echoes, and sound
coloration can strongly detract from the spaces usability for speech communication
or music performance. These problems can be avoided without deadening the
acoustics by introducing scattering, whereby sound is reflected irregularly over a
wide range of directions. In general terms, scattering can be created by variation in
the physical surface such as curvature, relief forms or textures, and changes in
contrasts in material acoustic properties (Sabine 1964). The angle of a reflection
may be influenced by the incidence angle onto the particular part of a surface and
the incidence angle on the overall surface, which can be conceptualized as specular
reflection. Even if specular reflection occurs on a small scale, there may be phase
interference between reflections across the surface, yielding a much more complex
reflected sound eld, so that sound is scattered (Fig. 2).
Fig. 2 Specular reflections. Relationships between surface angle and depth using a ray model of
reflection: specular reflection showing two incident sound angles (left); ray paths reflected by an
irregular surface, which may yield scattering due to phase interference between differently delayed
reflections (center); sound rays reflected by an angled surface (right)
140
D. Reinhardt et al.
Non-flat surfaces introduce small time delays, which vary with the depth of the
cavity or well in the surface, and which can result in signicant phase shifts
introduced by a reflected wave front so scattering occurs. As this is non-trivial to
predict, physical measurement of prototypes plays an important role in the development of scattering surfaces. Specular acoustic reflections occur on large and
small scales, where the angle of reflection is equal to the incidence angle mirrored
by the surfaces normal (Sn) at the point of reflection. Figure 2a shows the purely
specular case, while Fig. 2b shows a surface with varying depth (which is typical
for surfaces that have been designed for acoustic scattering). While the reflected
rays radiate in the same direction, in practice this will only occur in this simple way
when the phase shift due to varying depth is smallotherwise phase interference
between reflections will yield scattering.
Considering that the wavelengths of audible sound range between 17 m (for
extremely low frequency) to 17 mm (extremely high frequency), it is evident that
for a given variable depth surface the extent of scattering is likely to be strongly
dependent on frequency. The wavelength of sound () is equal to the speed of sound
(c, typically about 344 m/s) divided by frequency ( = c/f). For example, for a
frequency of 1 kHz, the wavelength is 0.344 m: if a well depth is of this
(0.086 m), this would result in a 180 phase shift relative to a part of the surface
with no well, potentially creating local sound cancellation (and complex reflection
patterns for frequencies in the vicinity of this, and similarly at odd multiples of the
frequency). Yet for 100 Hz the wavelength is 10 times longer, so the phase change
due to this well is only 18, hence the surface can be considered to be almost flat at
that frequency, yielding an essentially specular reflection. When there are multiple
depths across a surface, the reflection pattern develops from interference between a
multitude of phase shifts, so that scattering occurs from a macroscopic perspective
(Fig. 2c). Increasing the contrast in surface depth extends scattering to lower frequencies, although in practical implementation there are limits to the available
depth (due to cost and/or available space). In the high frequency range, surface
elements may be sufciently large (relative to wavelength) for their angle to affect
the reflection angle. Periodicity in the surface reduces the complexity of the
reflected sound eld, and so aperiodic patterns are preferred for scattering surface
treatments to achieve high scattering (Bonwetsch et al. 2008).
3 Shaping Angles: Scattering Disks

The complexity of acoustic reflections from micro-geometric surfaces provides a
workflow rationale that extends from scripting surfaces towards the physical measurement of scale prototypes as an important part of the design and validation process.
In order to identify the potential of acoustic behavior, and to derive threshold criteria,
the research employed an iterative collaborative process with the following parts:
141
Fig. 3 Acoustic scattering disks (HexN/Def-S1): GH geometry, 310 mm scattering discs with
deformed hexagonal module patterns (left, center), and Flowl streams (right)
(i) specication of the architectural design parameters, along with the acoustic design
aims (e.g. scattering coefcient spectrum); (ii) computational design of specic
surface micro-geometries; (iii) fabrication of physical scale model test samples in the
form of discs; (iv) acoustic measurement and analysis of sample performance; and
(v) renement of the design with potential further iteration. This approach concatenates computational design, acoustic analysis and robotic fabrication, which expands
the potential scope of micro-geometric surfaces by integrating scripting logic, surface
angles and depth, and toolpath, thus enabling successive acoustic design variations
that can be tested for prociency.
3.1
Pattern Scripts: HexDf and Flowl
As an initial departure point, two pattern variations were generated in GH

Grasshopper (a plug-into McNeel Rhino/visual scripting environment). The rst is a
hexagonal periodic (HexNDef-S1) and deformed pattern (HexDef-S2), the second a
vector based pattern (Flowl-S1) which adopted the customized script Flowl.1 Both
develop zones of highly differentiated depth across the surface (Fig. 3).
The rst sample uses a parametric pattern of tessellated hexagons varied between
two primary points, so that the initially periodic tiling is deformed in
attraction/repulsion. As a result, individual facets vary in depth, height,
flowl is a plug-in for McNeel Rhino/GH, used to visualize a vector eld, generated trough
positive and negative point charges, with path lines calculated with the Runge-Kutta 4th Order
Method short RK4 (by mathematicians Runge and Kutta 1900). Developed by uto, http://www.
uto-lab.com and http://www.food4rhino.com/project/owl.
142
D. Reinhardt et al.
Fig. 4 Parametric design (Flowl-S1): GH protocols for change of reflective area in scattering disc
building zones around attractor points (left), and changes in vortex eld (right)
directionality, so that diversity is created for testing sound scattering and sound
diffusive properties of surfaces.
The second geometry orients curvilinear splines in relation to an increasing
number of attractor points (Fig. 4). It uses a parametric vector eld of streams
between a variable number of up to 9 of attractor points, and relative adjacency
between neighboring elds. Through control over number of attractor points in the
scripting environment, the overall depth of surface resulting from isocurves can be
manipulated and used as adjustable toolpath in KUKA|prc (a Rhino McNeel plug-in
for robotic fabrication).
3.2
Prototyping 1:1 Scale Models for Acoustic Testing
Computational prediction of performance describes sound through mathematical

models, but scale models or prototypes monitor the physical phenomenon itself
(Peters and Olesen 2010). In wave acoustics, accurate computational prediction can
be very expensive, whereas physical modeling is comparatively efcient. The test
surface samples are disks (designed circular so they can be rotated without any
change in their outer edge, which is important for scattering measurement), fabricated at a scale of 1:10, and tested in a scale model reverberation chamber (Fig. 5).
As a shared geometry base, the patterns were thus adapted to circular disks and
follow two standard equations that determine relative height and depth:
(1) height 1/16th of diameter (Peters and Olesen 2010); and (2) d < (depth of
the structure relative to surface area must be greater than half the wavelength), in
order to be effective for a certain frequency (Bonwetsch et al. 2008). The sound
discs were thus designed with 310 mm diameter and 19 mm depth, and prototyped
in subtractive cutting processes in XPS Styrofoam and wood (Fig. 4, right). While
the dimensions of the scattering discs were selected to t the acoustic scale
reverberation room, height eld variations in the prototypical scale models were
also dependent on the CNC milling path of the tooling head (4 mm/1 mm spindle
head) and cutting time, and the directionality of material, resulting in different
roughness of disc surfaces.
143
Fig. 5 Acoustic test in reverberation chamber (left), sample discs (right, from top: HexDef-S1,
Flowl, and all samples including periodic tiling)
3.3
Random Incidence Measurement
The acoustic behavior of the 1:10 scale model prototype surfaces was then tested by
random incidence measurement in a scale-model reverberant room. Each disk is
placed on a turntable, and synchronously averaged impulse responses are obtained
for different source and receiver positions from the material sample (using a long
duration maximum length sequence test signal, with AARAE software, Cabrera
et al. 2014). The acoustic performance is measured as apparent reverberation time:
with and without the sample; and in stasis and rotation, yielding a spectrum of
random incidence scattering coefcients (per ISO 17497-1:2004). From these, the
scattering coefcient is calculated, which describes the ratio of acoustic energy
reflected in a non-specular manner to the total reflected acoustic energy. This
provides a summary parameter by which fabricated prototypes can be evaluated for
their effectiveness over the frequency range of interest, and presents criteria for
further design developments. Prototypes evaluated (Fig. 6) included a
non-patterned reference disk, a non- (HexNF) and deformed hexagonal tessellation
(HexDef-S1 red), and the flowl (blue), which showed better results for the latter.
The normal incidence absorption coefcient of the material was measured (in an
impedance tube), showing values of less than 0.3 across the frequency range.
144
D. Reinhardt et al.
Fig. 6 Scaled scattering

coefcient for disc. Physical
acoustic test results (scattering
coefcient) for Hexagon and
Flowl: measurements made at
1:10, results shown at 1:1.
The sharp increase that can be
seen at 1000 Hz in the
Hexagonal undeformed grid
can be attributed to the
relationship between the
wavelength of the higher
frequencies and the well depth
of the patterns
Results measured for HexNF (Undeformed) were insignicant, which is likely to be

due to a combination of the periodicity of its pattern and the shallow depth across
each hexagon. In contrast, the second and deformed pattern (HexDef) resulted in
signicant scattering at frequencies above 1 kHz (after frequency scaling), thereby
proving to be effective in sound scattering for important sounds such as human
speech. But better scattering effects resulted from the Flowl surface, due to depth
and number of ridges/valley volumes.
The physical test series resulted here in two important determining factors for
fabrication of further surfaces in the Flowl series: rstly, results are impacted by the
structure of the discs; and secondly in relation to width, depth, and variation of
depth across each flow line of the design. This is important, because these two key
criteria can be parametrically controlled through computational design.
Furthermore, the particular benet of robotic fabrication (compared to other options
such as CNC fabrication) is the ability to develop forms using a wide range of
tool-path angles, which can expand the effective surface area of a prototype scattering sample, with the potential to increase its effective depth by cutting at angles
signicantly different to the surfaces normal.
145
4 Reverse Design of Sound: Pattern Geometries Relative

to 6-Axis Fabrication Angle
The research then continued in a reverse design process, by which the angle of
sound reflection was linked to the robotic toolpath, thereby creating depth in the
material that impacts the sound reflection. This was explored focusing on two
aspects to increased depth: rstly, by increasing depth of available angle per singular isocurve, and secondly by increasing the sum of lines resulting in the total
well area that sweep along each isocurve (Fig. 7).
In this further development of the pattern, we prioritized the Flowls geometrical
logic where multiple surfaces result from directional robotic milling of isocurves. In
combination with the controlled tooling path, this results in overall depth for the
acoustic disk. Effectively, the mother geometry is then reduced to parametric
scripting of attractor points (GH and plug-in flowl), in combination with robot
simulation in KUKA|prc. The 7-axis robotic fabrication offers degrees of freedom
over CNC (MultiCAM CNC 3-axis router) milling that allows to adopt swarf or
flank machining techniques in which the side of the tool is used to produce the
desired patterns in a single pass, the pattern adapted to the scale of the surface
texture, material and tool selection.
4.1
Robotic Fabrication: Initial Parameters
The particular Flowl prototype uses the advantage of a relatively simple geometrical
rule for deforming a collection of individual lines relative to one attractor and its
adjacent neighbors (depending on the GH denition). Each single line is a spline, but
can be directly linked to the robotic toolpath, with the angle of the milling tool predened and variable along the curve, resulting in the depth of valley that must be
Fig. 7 Robotic simulation of tooling path for isocurves (left), and prototype milling (right)
146
D. Reinhardt et al.
Fig. 8 Robotic toolpath simulation: a sweeping isocurves in 4 passes, b subdivision into relative
distance points between upper and lower curve, c location of valley and curves in attractor region,
d robotic 6axis control relative to surface area, toolpath angle and depth
achieved to provide scattering. Instead of a workflow with multiple passes along

splines, a pair of 2 passes can produce the valley and thus effect. Our initial
studies focused on robotic milling of a single well resulting from two parallel
isocurves (Fig. 8).
The robotic parameters include here toolpath and dened angle of the milling
tool; multiple passes along isocurve; distance of endeffector to material surface,
depth and surface angles of well variable along the curve. The robotic cutting path
is set as a series of paths between: (a) two edge curves (top and bottom surface),
with (b) subdivision between points on each curve with distance 20 mm, and the
dened limit of 19 mm cutting depth relative to the 3.2 mm ballpoint toolbit
(120 mm tooldepth).
4.2
Workflow and Fabrication Process
This system was then parametrized for the disc system, and tested on the more
complex geometry of the Flowl series. The robotically fabricated prototype includes
414 faces that are again resulting from isocurves from XPS extruded Styrofoam,
which allows precise and ne milling due to its close packed and non-directional
material characteristics. The transfer from the scripting in GH (Grasshopper)
towards KUKA|prc took into account several adaptations to the robotic fabrication
147
Fig. 9 Adaptations of scripting geometry to robotic fabrication: a surface orientation relative to

robotic zero point, b faces and 34 manually adjusted, c singular valley by two milling passes,
d attractor points and relation of surface valleys relative to producible robotic angles
process: The positioning of desired object surface in relation to the robotic zero
point required a vertical orientation of the material (EPSX Styrofoam), due to the
local positioning of the robot endeffector. This proved to be benecial as material
cut-offs were spontaneously removed from the toolpath by gravity. Out of the total
number of faces, 34 faces needed to be adjusted manually (>8 %), due to the
intersections of isocurves in concentration areas around attractor points. Each face
is controlled in kuka|prc as each double curved surface path with 4 stepovers, in
closely packed linear moves. All faces are cut at a maximum 40 angle of the toolbit
relative to the surfaces normal (Fig. 9).
To date, the research has continued the robotic manufacturing of a series of
sound effective acoustic discs as scale model prototypes for acoustic analysis in
different degrees of overall surface depth. In addition, the parametric GH code and
KUKA|prc have been aligned, and thus enabled an upscaling towards three times
the original size (930 mm), cut to a maximum size of milling depth of 65 mm with a
6 mm ballpoint toolbit. In a continued work series, patterns can be further manufactured in plaster, using a KUKA KR10 onsite, with again toolpaths scaled to the
required robotic reach and work envelope.
148
D. Reinhardt et al.
5 Conclusion and Future Research

Acoustic surface design is a specialized area of architecture, which has traditionally
involved relatively simple structures exploiting the bulk properties of materials and
resonance phenomena. The conceptual framework and pattern language available
for acoustic scattering discs can be signicantly expanded through highly flexible
robotic fabrication that allows the efcient production of 1:10 scale model prototypes. This research project has presented an example of this, applied to acoustic
scattering. Scattering is not the only acoustic surface characteristic that can benet
from high degrees of freedom digital fabrication. The micro-design of special
patterns of reflection, highly tuned absorption, and potentially other unconventional
acoustic surface behaviors can be investigated in similar ways. However, scattering
is of particular interest as a case study because there is a standard and efcient
method to measure it.
The scattering coefcient is a single number (as a function of frequency) relating
to statistically-dened (random) sound elds. Future research will study reflections
more comprehensively with a 196-hemispherical loudspeaker array in a sound
absorptive room, which, in conjunction with a microphone array, could comprehensively describe reflection phenomena (including direction-specic absorption
and diffusion) from micro-geometric surface designs (Cabrera et al. 2015). Future
research can then extend criteria, conceptual framework and robotic fabrication
processes to the acoustic surface treatment of existing surface geometries, or to the
conditioning of complex curved surfaces that can be sound effective on a larger
scale in architectural space(s).
149
Acknowledgments This project is part of an ongoing interdisciplinary research collaboration

(Architecture and Audio & Acoustics into complex curved geometries and their acoustic behavior,
undertaken at the Faculty of Architecture, Design and Planning, The University of Sydney. Research
assistance for coding of geometric patterns by Iain Blampied (2013/4) and Mitchell R Page (2015),
for digital fabrication by Celeste Raanoja (2014), with acoustic behavior and sound measuring
undertaken by James R Colla, Jesse H Loweke and David S OBrien (2015). The research has been
supported by a 2014 ECR SEED fund, and was produced at DMaF, The University of Sydney.
References
Bonwetsch, T, Baertschi, R and Oesterle, S 2008, Adding Performance Criteria to Digital
Fabrication Room-Acoustical Information of Diffuse Respondent Panels. Acadia Conference
Proceedings, Non Standard Production Techniques Tools, Techniques and TechnologiesAdding Performance Criteria to Digital Fabrication, Minneapolis, Minnesota, pp. 364369.
Burry, J, Davis, D, Peters, B, Ayres, P, Klein, J, Pena de Leon, A and Burry, M 2011, Modelling
Hyperboloid Sound Scattering: The Challenge of Simulating, Fabricating and Measuring, in
Gengnagel, C, Kilian, A, Palz, N and Scheurer, F (eds), Modelling Symposium,
Springer-Verlag, Berlin, pp. 8996.
Cabrera, D, Jimenez, D and Martens, W 2014, Audio and Acoustical Response Analysis
Environment (AARAE): a tool to support education and research in acoustics. 43rd
International Congress on Noise Control Engineering , Melbourne, Australia.
Cabrera, D, Miranda Jofre, L, Jimenez, D, Edser, C and Martens, W 2015, A facility for simulating
room acoustics, employing a high density hemispherical array of loudspeakers, Acoustics
Australia vol. 43, pp. 7781. doi:10.1007/s40857-015-0010-y
Cox T and DAntonio P 2009, Acoustic Absorbers and Diffusers: Theory, Design and Application,
Taylor & Francis, London.
Kolarevic B 2005, Computing the Performative, in Kolarevic, B and Malkawi, A (eds).
Performative Architecture: Beyond Instrumentality Spon Press, New York, p. 201.
Oxman R 2008, Performance-based Design: Current Practices and Research, Issues, international
journal of architectural computing, vol. 06, no. 1. http://tx.technion.ac.il/*rivkao/topics/
publications/performance%20based%20design%20IJAC_2008.pdf. Accessed 10 June 2015
Peters, B, Burry, J 2011, Responsive Acoustical Surfacing Cluster at Smart Geometry 2011,
Copenhagen, access date 20.08.2012, http://www.responsive-a-s-c.com/.
Peters B and Olesen T 2010, Integrating Sound Scattering Measurements in the Design of
Complex Architectural Surfaces -Informing a parametric design strategy with acoustic
measurements from rapid prototype scale models, ECAADE 28 Proceedings.
Reinhardt D, Cabrera D, Jung A, Ulacco G, Niemela, M 2014, TriVoc- Robotic Manufacturing
for Affecting Sound through Complex Curved Geometries, in McGee, W and Ponce de Leon,
M (eds.) Robotic Fabrication in Architecture, Art and Design 2014, Springer International
Publishing Switzerland, pp.163180
Reinhardt, D, Martens, W, Miranda, L (2013). Sonic Domes - Interfacing Generative Design,
Structural Engineering and Acoustic Behaviour. R. Stouffs, P. Janssen, S. Roudavski, B. Tuner
(eds.), Open Systems: Proceedings of the 18th International Conference on Computer-Aided
Architectural Design Research in Asia (CAADRIA 2013), 529538. ISBN: 978-988-19-0264-1.
Reinhardt, D, Martens, W, Miranda, L (2012). Acoustic Consequences of Performative
StructuresModelling Dependencies between Spatial Formation and Acoustic Behaviour,
Achten, Henri; Pavlicek, Jiri; Hulin, Jaroslav; Matejdan, Dana (eds.), Digital Physicality Proceedings of the 30th eCAADe ConferenceVolume 1/ISBN 978-9-4912070-2-0, Czech
Technical University in Prague, Faculty of Architecture (Czech Republic) 12-14 September
2012, pp. 577-586.ISBN:978-9-4912070-3-7.
Sabine, WC 1964, Collected Papers on Acoustics, Dover, New York.
Robotic Hot-Blade Cutting

An Industrial Approach to Cost-Effective
Production of Double Curved Concrete
Structures
Asbjrn Sndergaard, Jelle Feringa, Toke Nrbjerg,
Kasper Steenstrup, David Brander, Jens Graversen,
Steen Markvorsen, Andreas Brentzen, Kiril Petkov, Jesper Hattel,
Kenn Clausen, Kasper Jensen, Lars Knudsen and Jacob Kortbek
Abstract This paper presents a novel method for cost-effective, robotic production
of double curved formwork in Expanded Polystyrene (EPS) for in situ and prefabricated concrete construction. A rationalization and segmentation procedure is
developed, which allows for the transliteration of double curved NURBS surfaces to
Euler elastica surface segments, while respecting various constraints of production.
A. Sndergaard (&) J. Feringa
Odico Formwork Robotics Aps, Odense, Denmark
e-mail: asbjorn@odico.dk
J. Feringa
e-mail: jelle@odico.dk
T. Nrbjerg K. Steenstrup D. Brander J. Graversen S. Markvorsen A. Brentzen
Department of Applied Mathematics and Computer Science, Technical University
of Denmark, Kongens Lyngby, Denmark
e-mail: tono@dtu.dk
K. Steenstrup
e-mail: khor@dtu.dk
D. Brander
e-mail: dbra@dtu.dk
J. Graversen
e-mail: jgra@dtu.dk
S. Markvorsen
e-mail: stema@dtu.dk
A. Brentzen
e-mail: janba@dtu.dk
K. Petkov J. Hattel
Department of Mechanical Engineering, Technical University of Denmark,
Kongens Lyngby, Denmark
e-mail: kipekt@dtu.dk
151
152
A. Sndergaard et al.
An 18 axis, tri-robot system approximates double curved NURBS surfaces by means

of an elastically deformed and heated blade, mounted on the flanges of two manipulators. Re-orienting or translating either end of the blade dynamically deforms the
blades curvature. The blade follows the contours of the rationalized surface by
continuous change in position and orientation of the end-effectors. The concepts
potential is studied by a pilot production of a full-scale demonstrator panel assembly.
Keywords Robotic fabrication
Concrete structures
Hot-Blade
EPS-molds
Cost-efciency
1 Introduction
The vast majority of contemporary building designs are restrained to a formal
language of planar surfaces and derivative geometric constructs; a constraint that
stems from the practicalities of construction, which favors the use of mass-produced
semi-manufactures andfor concrete in particularmodular, reusable formwork
systems. An increasing number of high-prole project designs challenge the
dominant paradigm. The challenge is posed by advanced building design projects,
such as the Kagamigahara Crematorium (Toyo Ito Architects 2006) and Waalbridge
Extension (Zwart & Jansma, under construction), which utilize manual production
of formwork to achieve complex curvatures; and building projects which employ
large scale CNC-milling to realize advanced structures, such as the Museum
Foundation Louis Vuitton by Gehry & Associates (Paris 2014); the Nordpark cable
railway by Zaha Hadid Architects (Nordpark 2007), and the Metz Pompidou by
Shigeru Ban (Metz 2010).
However, neither manual formwork production nor large scale CNC-milling
provide a cost-effective option for general construction, and projects of this type
therefore require extraordinary budget frameworks for realization. Recent
J. Hattel
e-mail: jhat@dtu.dk
K. Clausen K. Jensen
GXN A/S, Copenhagen, Denmark
e-mail: kec@3xn.dk
K. Jensen
e-mail: kgj@3xn.dk
L. Knudsen J. Kortbek
Center for Robotics, Danish Technological Institute, Copenhagen, Denmark
e-mail: lak@teknologisk.dk
J. Kortbek
e-mail: jkk@teknologisk.dk
153
Fig. 1 Large scale RHWC production at Odico (left), and hotwire cut production sample (right)
developments in architectural robotics by authors of this paper have demonstrated

novel, cost-effective means of producing bespoke formwork with the constraint of
being limited to ruled surface. The Robotic Hotwire Cutting (RHWC) approach is
utilized to concrete casting in Expanded Polystyrene that has been developed to
industrial scale (Feringa and Sndergaard 2014, Fig. 1). Currently, Odico Aps is
putting forward RHWC in relation to a project design by the Danish artist Olafur
Eliasson, for the Kirk Kapital HQ in Vejle.1 Here, over 4000 m2 of formwork are
produced, achieving production speeds order of magnitudes faster than
CNC-milling through the principal mechanics of the method (McGee 2012). In
extension of these developments, experiments at Odico are performed in abrasive
wire-sawing. Through this technique, the same digital control proceduresfacilitated by the internally developed control software, PyRapidis applied to direct
processing of construction materials, such as industrial marble (Fig. 2, top). In
further maturation of the concept, the method is being adapted in partnership
development with Bumer AG for industrial machining. Prototype production have
revealed further signicant reductions in machining times, in which full scale
elements may be cut in matter of seconds (Fig. 2, bottom).
However, for a number of projects, the realization of general double curved
structures is imperative. Here, no effective methods currently exist for architectural
scale in industrial production. In 2012, Odico Aps. tendered as part of a consortium
for the realization of the aforementioned Extended Waalbridge project (Fig. 3).
Here, the double curvature of the columns of the bridge elegantly blending with the
bridge slab are dominated in a single direction. The considerable scale of the project
implied large local radii (between 1 and 2 m) of the surfaces. Since, for this scale,
CNC milling molds from EPS would have been a prohibitively ineffective method,
digital manufacturing would not be economically competitive with the more traditional approach that was chosen. While developing the tender documents, Odico
Aps. realized that the Hot-Blade cutting method discussed in this paper would
represent a competitive solution.
http://www.domusweb.it/en/news/2011/12/01/kirk-kapital-a-s-by-eliasson.html.
154
Fig. 2 Robotic abrasive wire-saw cutting of marble blocks at Carrara, Italy (top, left); cut samples
(top, right). Second abrasive prototype tool, developed in collaboration with Bumer AG (bottom,
right) Sample geometry cut in nonflammable acoustic foam under 16 s (bottom, left)
Fig. 3 The Waal Bridge Extension Impression of the artist (left) ( Zwart & Jansma) Ongoing
construction work using traditional formwork systems (right)
2 State-of-the-Art
Contemporary construction currently employs either manually produced, bespoke
formwork or CNC-milling of foam molds for the realization of complex concrete
structures. In addition to these techniques, actuated mold systems have been
explored by Danish Adapa and in the EU FP7 project TailorCrete (Jepsen et al.
2011; Hesse 2012). This technique employs actuation of a flexible membrane as a
casting surface; however, the method is limited to concrete prefabrication; by the
155
casting pressure the individual systems can take; and the need for multiple casting
aggregates for large volume production due the curing time for concrete elements.
In addition, dynamic slip-casting for column elements is being explored (Lloret
et al. 2014), as a variant of the additive manufacturing of concrete structures
(Khosnevic 2004; Lim et al. 2012).
These and related methods attempt avoiding the need for formwork altogether
however do so at the cost of signicant degrees of freedom, such as the capacity to
realize cantilevered designs. Finally, fabric formworks have been proposed and
experimentally applied as an alternative technique for the casting of advanced
designs (Veenendaal et al. 2011). This approach is challenged by the capacity of the
fabric to achieve desired designs, as well as the unpredictability of the fabric
behavior in combination with the required complexity of creating bespoke molds.
A common denominator of the described developments is the requirement of
shifting to entirely new modes of construction, which creates a high barrier for full
scale implementation; or limits the degrees of freedom achievable compared to
existing means of realization. In contrast, the method presented here proposes a
production cycle which is fully compatible with current in situ and prefabrication in
concrete construction, while achieving doubly curved formwork designs at
machining times more than a hundred times faster than comparable CNC-milling,
the most developed and applied strategy for industrial scale production.
Double curved surfaces with positive Gaussian curvature can in a vast majority
of cases be described via swept splines. The term splines nowadays refers to
piecewise polynomial or rational functions used in CAD systems to model curves
and surfaces. However, prior to the introduction of computers in the 1950s the term
was used for thin wooden rods the shapes of which were manipulated by the
placement of so-called ducks at various points to create a naturally smooth curve
for drawing designs. These were used in ship building and, later, in the aviation and
automotive industries. The placement of the ducks simulates the placement of ribs
in the hull of the ship, and hence the curve drawn by following the spline is an
accurate reflection of the natural shape adopted by the planks forming the ships
hull. The use of splines for the storage and transmission of a design goes back to the
Romans, in the form of physical templates for the ribs of ships (Farin et al. 2002).
Splines and ducks suitable for drawings of ship designs were developed later,
perhaps in Hull in the 1600s.
The mathematical shape of a physical spline can be described exactly, although
it requires the use of so-called elliptic functions, which are nonlinear in nature
(Fig. 4). The correct mathematical model for an elastic rod bent by a force at one
end with the other end xed was given by James Bernoulli in 1691 (Truesdell
1983). In his approximation of the solution for the case that the ends of the rod are
at right angles to each other, he recognized that the solutions would require
non-standard functions. Later, in 1743, Bernoullis nephew, Daniel, suggested the
problem to Euler, who then, in an appendix to his famous treatise on the calculus of
variations found all possible shapes for these so-called Euler elastica (Euler 1744;
Love 1906).
156
Fig. 4 Design of the

Concorde wing-section using
physical splines, 1964 (
Bristol Archives)
3 Geometry Rationalization
The presented geometry rationalization approximates the physical behavior of the
the Hot-Blade in order to convert arbitrary input surfaces to producible geometry.
The HotBlade is xed between two robot arms, which enable us to choose the
location and rotation of the blades ends. The shape of the blade is the curve that,
subject to the endpoint constraints, minimizes the elastic energy. These curves are
the above-mentioned Euler elastica or elastic curves. Before discussing the
approximation of a CAD surface, let us consider the class of surfaces dened by this
cutting process, namely the surfaces swept out by continuously varying families of
planar Euler elastica. A planar curve is geometrically determined by its curvature
function js h0 s, where \ is the angle function of the unit tangent.
One can show that the equation dening an elastica is the normalized pendulum
equation h00 s sinhs and the solution is the curve:
Ck s 2Es; k s; 2k1 cns; k;
k h0 0=2;
where cn(s,k) and E(s,k) are standard elliptic functions depending on a parameter k.
Applying all possible dilations, translations and rotations to Ck , one obtains all
possible elastic curve segments. Allowing all of these parameters to vary with time,
and then generating the time sweep so dened, one obtains all possible
elastica-swept surface patches. One can implement this numerically, to obtain
examples (Fig. 5).
When rationalizing a CAD surface to Euler-elastica for Hot-Blade cutting, the
surface is segmented into patches that can be approximated by surfaces of the type
exemplied in Fig. 5. We essentially do this simply by nding planar curves on the
original surface and then approximating these by segments of planar elastic curves.
157
Fig. 5 Elastica surfaces generated through implementation of the above formulation in Matlab
3.1
Curve Approximation
Given a parameterized planar curve segment we wish to nd a piece of an elastic

curve which has the same shape. We do this via an optimization algorithm that
minimizes the distance between two curves. By choosing a standard parameterization, we are able to describe any elastic curve segment by four control parameters,
which determine the length and shape of the segment. Three more parameters
determine the position and rotation of the curve in the plane. The distance between
the given curve and any elastic curve is thus a function of the seven control
parameters.
The approximation algorithm has two steps: rst, we analyze the geometry of the
given curve in order to nd control parameters for an elastic curve segment, which
has the same overall shape (Fig. 6). Then, starting from this initial guess, we tweak
the parameters, using the optimization tool IPOPT (Wchter and Biegler 2006),
until we get the closest t. We can do this either with or without endpoints xed.
3.2
Surface Approximation
We now consider a given CAD surface, and we want to approximate it by a surface

that can be obtained by moving elastic curves through space. From the CAD design
Fig. 6 Original spline curve (blue) and initial guess for approximating elastic curve (red dotted),
original spline curve (blue) and best approximating elastic curve (green, left). Input NURBS
surface (center); rationalized surface (right)
158
Fig. 7 A selection of planar curves (splines) on the original CAD surface (left); The original
surface with elastic curves that approximate the splines. Note that these curves do not lie exactly
on the surface (center); Rationalized surface swept by elastic curves (right)
we extract planar curves on the surface and approximate each of these by an elastic
curve. By interpolating the control parameters we obtain a rationalized designa
new surface, which is swept out by elastic curves moving through space. For larger
designs we need to segment the surface into pieces that can be cut individually.
Because we control the endpoints and directions of the blade, we can ensure smooth
transition from one piece to another (Fig. 7).
4 Surface Segmentation
A number of segmentation procedures are developed, targeting three production
constraints: (a) plane segmentation when exceeding the dimensions of the input
EPS work object; (b) instability of the blade due to multiple inflection points, or
(c) cutting the same area multiple times due to rotation of the blade prole. Figure 8
(left) illustrates an example of a surface with too many inflection points. An
inflection point is a point where the sign of the curvature changes; in other words
the tangent at the point will cut the curve in two. We use a subdivision scheme to
nd the inflections. Analysis of one of the curves shows six inflection points and
since many inflection points on the curve make the blade less stable, segmentation
is required.
Fig. 8 An input surface (left), Hot-Blade planar cuts with inflection points (center), and one of the
cuts close up (right)
159
Assuming the rationalization of each cut is curvature continuous, there will be

the same number of inflection points on the cut and the rationalization. Two
exceptions to this are inflection points near the edge of the cut that may disappear,
and pairwise inflections close to each other, which may cancel out, just like pushing
out a small dent.
Taking the above into account, we propose the following algorithm.
1.
2.
3.
4.
Find the planar curves on the surface.

Calculate inflection points for each curve.
Segment the surface into a grid of blocks.
For each block test if there are more than two inflection points; if so try to
a. Move the block if there are overlaps to improve.
b. Remove inflection points close to each other.
If there still too many inflection points continue to step 5.
5. Take two new blocks, each of the same size as the original block, and place
them so that they overlap both each other and the two adjacent blocks in the
row. Go to step 4.
In this algorithm we can control whether we keep the same number of blocks in all
rows or not. This affects the aesthetics of the segmentation. In the overlap of the
blocks we choose a cutting plane such that the segmentation follows the geometry.
An example of the output of this algorithm can be seen in Fig. 8 (right), showing
the surface subdivision.
The problem of cutting the same area multiple times arises when rotation of the
blade in the cutting direction is allowed (Fig. 9, column 2 from the left). We see
here that the curves intersect each other, and thus part of the surface will be cut
multiple times, which is undesirable. In most cases this problem can be solved by
segmenting the surface, as described above. We only need to add a test for intersecting curves in step 4.
Fig. 9 Segmentation schemes
160
Fig. 10 Deformation of the blade through orientation and positioning of the two end-effectors
5 Dataflow and Robotic System Conguration

The experimental setup consists of three robots. Robot 1 holds the EPS work object,
which is to be cut, and moves the block linearly through space, thus acting in
principle like a conveyer belt. Robots 2 and 3 control the ends of the HotBlade
thereby determining its shape and its position in relation to the EPS block (Fig. 10).
When the geometry rationalization is completed, we know a set of planar elastic
curves on the rationalized surface. The curve segments which lie on the surface are
shorter than the HotBlade cutting tool, but since we know not just the curve
segments, but the entire curves we can easily extend the curves to the required
length, i.e. the length of the HotBlade. These extended curves are the target shapes
for the HotBlade during the cutting. We extract the relevant data for the extended
curves, that is, we nd the coordinates for the endpoints and the tangents at the
endpoints. The endpoint coordinates determine the position of the tools of robots 2
and 3 relative to the EPS block. The tangents determine the rotation of the tools,
which in turn controls the shape of the blade.
For our experiments the robots were given 51 targets. That is, for each block that
was to be cut, we provided 51 sets of positions and rotations for the tools of robots 2
and 3. The robot program then interpolates between these targets to follow a smooth
path from the rst to the last target, thus moving the blade while changing its shape,
resulting in an EPS surface of the rationalized design.
6 Blade Mechanics and Cutting Experiments

The main cutting tool used in the process is a thin metal stripusually referred to as
a blademade of a nickel-chromium super alloy. The blade is pre-heated to a
temperature of 300400 C by means of Joule heating and then it is slowly brought
into contact with an EPS block to produce melting, and subsequently to form or cut
the block into a desired shape (also referred to as thermal cutting). At such high
operating temperatures, the blade has to be displaced (or bent) into an elastic shape
161
Fig. 11 Tri-robot hot-blade cutting conguration
with predened curvature and at the same time maintain its elastic and flexibility
properties. Using FEM simulations, the effect of mechanical properties on the target
geometry was investigated and a particular material was chosen to ensure smooth
cutting. The blade is attached to two robots, one at each end, by specially designed
sandwich based holders to ensure strong and safe supports during all cutting
operations. The physical displacement of the blade is achieved by moving the
robots into an appropriate position, at the same time maintaining the
elastica-strain-curvature relations. The temperature dependent variations of the
blade shape are to be incorporated in the computational algorithm to secure proper
shape representation (Fig. 11).
Two experiments were designed and performed in order to test the utility of the
setup. In the rst experiment a convex doubly curved surface was cut. The curvature of the blade was continually changing during cutting in order to test the limit
of complexity that can be achieved and ensure proper geometrical representation.
The presence of two inflection points on the discretized surface was considered as a
possible problem, but the experiments showed that it does not make the blade
unstable, since the robots compensate with the angles of the holders and the curvatures involved were moderate. Good surface quality was achieved at cutting with
an absolute speed of motion of 7 mm/s. The EPS block to be cut had the dimensions
of 600 600 600 mm.
The second test aimed to cut a number of EPS blocks and then assemble them
into a single structure that should represent a ready-made mold for concrete casting.
Different discretized pieces of doubly curved surfaces of both convex and concave
types, as well as hyperbolic surfaces (negative Gauss curvature), were successfully
cut with the setup. The size of each individual block was approx.
600 785 600 mm, resulting in an assembly of size 1800 2345 mm, comparable to the size of production frame molds. The cutting experiments are currently
continued for production of doubly curved concrete panels with expected completion December 2015.
162
Fig. 12 General production workflow diagram: segmentation (left); cut foam (center) and in situ
mold (right)
7 Formwork Systems and Production Workflow

The efforts described in the previous chapters outline the general method for the
cost-effective production of doubly curved formwork in Expanded Polystyrene.
From this, the following process is developed (Fig. 12):
The cyclical workflow links conventional CAD-modelling operations with the
robotic Hot-Blade fabrication and standard concrete casting techniques. This
requires the rationalization and segmentation of geometry types before rebuilding
the geometry to the constraints of the blade, robot work envelope, work object
dimensions and tolerances. After the input geometry has been translated to segments of swept Euler elastica surfaces and data deducted for tri-robot motion,
EPS-mold pieces are produced. The mold pieces are subsequently used in combination with existing pre-fabrication and in situ workflows. For element
pre-fabrication, molds are mounted on vibration tables and sides enclosed with
metal or wooden frames. For in situ applications, mold pieces are used in combination with standard scaffolding modules for casting pressure support. These
applications ensure a full compatibility of the end-products of the Hot-Blade with
established industry workflows, critically ensuring a low barrier to adoption.
8 Conclusion
A general purpose robotic fabrication method for producing doubly curved formwork has been presented. The efcacy of the method has been demonstrated
through geometry rationalization and pilot production of a sample formwork panel
design. The method is being implemented for industrial scale fabrication by one of
partners of the research consortium, and the identied challenges are being
addressed through this work.
163
Acknowledgments The work presented in this paper is part of the larger 3-year research effort,
BladeRunner established and generously supported under the program of the Innovation Fund
Denmark for advanced technology projects. The project is conducted by the partners Odico Aps
(project lead), the Technical University of Denmark, Department of Applied Mathematics and
Computer Science and Department of Mechanical Engineering, the Danish Institute of
Technology; GXN A/S and Confac A/S.
164
References
Euler, L 1744, Methodus inveniendi lineas curvas maximi minimive proprietate gaudentes;
Additamentum I: de curvis elasticis.
Farin, G 2002, A History of Curves and Surfaces in CAGD, in Handbook of Com-puter Aided
Geometric Design, North-Holland Publishers, Amsterdam, pp. 123.
Feringa, J and Sndergaard A 2014, Fabricating Architectural Volume in Kohler, M and
Gramazio, F (eds), Fabricate : negotiating design and making, gta-Verlag, Zrich, pp. 44-51.
Hesse, P 2012, TailorCrete, Flight Assembled Architecture in Architekturteilchen. Modulares
Bauen im Digitalen Zeitalter, Kln, pp. 126127, 164165.
Jepsen, C, Kristensen M, Kirkegaard, P 2011, Dynamic Double Curvature Mould System in
Gengnagel, C, Kilian, A, Palz, N and Scheurer, F (eds), Computational Design Modeling:
Proceedings of the Design Modeling Symposium, Springer, Berlin, pp. 291300.
Khosnehvis B, 2004, Automated Construction By Contour CraftingRelated Ro-botic and
Information Technologies, in Journal of Automation in Construction Special Issue: The Best
of ISARC 2002, vol.13, no.1, January 2004, pp 519.
Lim, S, Buswell, RA, Le, TT, Austin, SA, Gibb, AGF and Thorpe, A 2012, Devel-opment in
Construction-Scale Additive Manufacturing Processes, Automation in Construction, vol. 21,
no. 1, pp. 262268.
Lloret, E, Amir, R, Shahab, Mettler, L, Flatt, RJ, Gramazio, F, Kohler, M and Langenberg, S 2014,
Complex Concrete Structures: Merging Existing Casting Techniques with Digital
Fabrication, Computer-Aided Design, Elsevier, Amsterdam, NL, vol. 60, March, pp. 4049.
Love, A 1906, A Treatise on the Mathematical Theory of Elasticity, Cambridge University Press,
Cambridge, UK.
McGee, W., Feringa J., Sndergaard, A Processes for an Architecture of Volume: robotic hotwire
cutting. In: Brell, S., Braumann, J (eds) Robarch 2012: Robotic Fab-rication in Architecture,
Art & Design. Springer Verlag, Vienna, pp 6271.
Truesdell, C 1983, The Influence of Elasticity on Analysis: the Classic Heritage Bulletin of the
American Mathematical Society, vol. 9, no. 3, pp. 293310.
Veenendaal, D, West, M and Block, P 2011, History and Overview of Fabric Formwork: using
Fabrics for Concrete Casting, Structural Concrete, Ernst & Sohn, Berlin, vol.12, no 3.
Wchter,A., and Biegler L.T, 2006, On the implementation of a primal-dual interi-or point lter
line search algorithm for large-scale nonlinear programming. Mathematical Programming,
Springer Berlin, 106(1):2557.
Part III
Projects
Fabrication-Aware Design of Timber

Folded Plate Shells with Double Through
Tenon Joints
Christopher Robeller and Yves Weinand
Abstract Integral attachment, the joining of parts through their form rather than
additional connectors or adhesives, is a common technique in many industry sectors. Following a renaissance of integral joints for timber frame structures, recent
research investigates techniques for the attachment of timber plate structures. This
paper introduces double through tenon joints, which allow for the rapid, precise and
fully integral assembly of doubly-curved folded surface structures with two interconnected layers of cross-laminated engineered wood panels. The shape of the
plates and the assembly sequence allow for an attachment without additional
connectors or adhesives. The fabrication and assembly constraint based design is
achieved through algorithms, which automatically generate the geometry of the
parts and the G-Code for the fabrication. We present the fabrication and assembly
of prototypes fabricated with 3D CNC milling and laser cutting systems, comparing
and discussing the advantages and disadvantages of the individual techniques.
Keywords Folded surface structures
lumber Miura ori fold 5-axis CNC
Integral attachment Laminated

2D/3D laser cutting
veneer
1 Introduction
In the design of smooth segmented plate shells, methods such as the Tangent Plane
Intersection TPI (Troche 2008) can be used for the panelization of doubly-curved
freeform surfaces. Different methods are required for the design of irregular and
freeform folded surface structures. Previous techniques have been presented using
triangulations (Trautz and Buelow 2009) and Origami inspired techniques using
C. Robeller (&) Y. Weinand
Timber Construction Laboratory IBOIS, EPFL Lausanne, Lausanne, Switzerland
e-mail: christopher.robeller@epfl.ch
Y. Weinand
e-mail: yves.weinand@epfl.ch
167
168
C. Robeller and Y. Weinand
reflection planes and cross-section proles (Buri and Weinand 2006) or mathematical models (Tachi 2009).
In parallel, research is being undertaken for the construction of shells with
lightweight and sustainable timber plates using laminated veneer lumber (LVL).
The easy machining of wood combined with numerically controlled machines
allows for the integration of joints into the geometry of the plates, allowing for a
rapid and precise assembly, aesthetic and easy-to-recycle mono-material structures
(Robeller 2015).
This paper builds upon previous research in the eld and presents a new method
that integrates fabrication and assembly constraints specic to folded surface
structures built from LVL panels and assembled with integral attachment techniques.
2 2-Layer Assembly with Through Tenon Joints

Integral multiple-tab-and-slot joints (MTSJ) such as nger joints provide geometric
features for a fast and precise alignment and assembly of the plates, as well as a
high resistance to compression and shear (Roche et al. 2015b), which are the
primary forces in segmented and folded timber plate structures. However the joints
between the plates receive not only shear, but also traction and bending forces.
These forces are typically supported by metal connectors. Alternative solutions are
hybrid nger/screw joints, such as in the ICD/ITKE LaGa Shell (Krieg et al. 2014)
or prismatic integral joints such as dovetails, which provide additional features for
the assembly and a resistance to bending and traction forces.
A comparison of the bending moment resistance of different edgewise joints for
laminated veneer lumber (LVL) plates has recently been provided by the authors
(Roche et al. 2015a), including screwed-, nger-, dovetail-, nejiri arigata- and
through tenon joints. This comparison showed that the strength of the through tenon
joints was the highest, which comes at the cost of a short protrusion beyond the
jointed corners.
A design constraint of the through tenon joints is their restriction to connections
of plates in two planes. A connection of plates in one plane is impossible due to the
joint geometry. In consequence, these joints are not applicable to smooth manifolds,
however they can be used for the design of folded timber plate structures. In these
designs, plates are always connected in two planes, where an orthogonal dihedral
angle u 90 between the plates is benecial for the structural performance as
well as for the fabrication process (Fig. 1).
However, a deviation b from this orthogonal angle is required for the design of
curved and irregular shell structures. When using through tenon joints, b is
equivalent to the inclination of the cuts, which are required for the fabrication of the
joints. Such cuts can be fabricated with multi-axis cutting machines such as gantry
or robot routers or laser cutters, however the inclination bmax of these machines is
limited, which sets a hard fabrication-constraint that must not be exceeded anywhere in the design.
Fabrication-Aware Design of Timber Folded Plate
169
Fig. 1 Fully integral 2-layer assembly with double through tenon joints (left); fabrication
constraints (right)
Figure 1 shows the through tenon joints ability to connect to multiple adjacent
plates through intersection. An entirely integral attachment of four plates is possible
following the illustrated method. On a mountain fold (as illustrated) the lower plate
intersects both counterparts with a double through tenon joint. Then, the upper plate
is inserted onto the tenons on its counterparts like a splice plate, which we call
inverse assembly. On a valley fold, the upper plate is inserted with a double-tenon
and the lower plate is inserted inversely. Major benets of this connection include:
1. A direct connection of the lower layers to the upper layers, without transferring
the forces through additional elements such as connectors.
2. Integral spacing of the two layers, which are kept at the correct distance
3. Blocking of elements: In such an assembly, only the last segment (two plates)
can be removed. All other plates are blocked and rmly held in place by other
parts, which must be removed before. A disassembly is only possible in the
reversed piecewise order of assembly. Therefore, no additional connectors are
required to x the plates. This does not only bring aesthetic advantages, rapid
assembly and cost savings, but it also allows the use of thin plates, on which the
use of edgewise screwed joints may not be permitted (DIBt 2011).
3 Segmentation of Doubly-Curved Folded Plate Shells

For the construction of self-supporting, doubly-curved surface structures with
discrete plate elements, we must nd a segmentation that satises the previously
mentioned constraints. Figure 2 illustrates this procedure on a target surface with a
span of 10 m in the V-direction and a span-to-rise ratio of 3.
In a rst step we discretize this surface into quadratic quadrilateral polygon mesh
faces. The resulting value bmean 85:6 indicates that we cannot join the plates
with through tenon joints, because our 3D cutting techniques are limited to
bmax 45 .
Instead, we will use two folding patterns known from Japanese Origami paper
folding:
Pattern 1, Yoshimura Fold Pattern, is a triangulation of the previous quadrilateral
mesh. The deviation b is still very large, but can be reduced through a reduction of
170
Fig. 2 Constraint-based segmentation of a doubly-curved target surface
segments in the V-direction. This results in deformed thin triangles (Fig. 2, top
right) and large plates with Lmax being larger than 1/2 of the span of the structure.
The assembly of such large parts with integral joints is difcult, because the edges
of the plates must be kept parallel during the insertion of the joints.
Pattern 2, Miura-Ori Pattern, presents an alternative approach. While previous
Origami-related methods aim at flat-foldable designs, we plan to produce our
structure from discrete components and do not include a flat-foldability constraint.
Instead, we generate a pattern through the evaluation of a point grid on the parametric base surface, where every second vertex in the U-direction is shifted by a
half segment length along V, and every second vertex in the V-direction is raised by
the offset length h along the surface normal. With this method, we can reduce the
global deviation bmean to 11 at an offset height of h 0:75 m, which satises the
fabrication constraints of our 3D cutting methods. However the quadrilaterals
generated with this method are not fully planar. D denotes the closest distance
between face diagonals. We reduce Dmean to 0.004 mm in a second step using an
external optimization framework (Bouaziz et al. 2012), which flattens the faces
while it preserves the surface boundary. bmax increases slightly through this step,
which could be reduced through an integration of the dihedral angle constraint into
the external optimization framework. However, the solver cannot nd a fold pattern
that satises the dihedral angle constraints without an initialization mesh with the
correct mountain and valley folds. Therefore we have chosen our strategy in two
separate steps.
171
4 Algorithmic Joint Generation

From the Miura-Ori based segmentation we obtain 432 individually shaped
quadrilaterals with 1,728 edges with different dihedral angles that must be joined.
Numerically controlled fabrication technology allows for the rapid fabrication of
these individually shaped plates, however the generation of the joints, as well as the
machine code must be generated with an algorithm.
4.1
Assembly Order and Joint Conguration
We base our algorithm on a polygon mesh with a uniform sampling of the unit
circle (Fig. 3a). Figure 3b shows that for each plate, up to 2 edges must be joined
simultaneously. The common insertion direction v for these edges is found at their
bisector. This causes a deviation h from a line on the plane perpendicular to the
edge. It is assembly-constrained to hmax 30 . A piecewise assembly of our
structure is only possible as illustrated, with the x-direction changing in every
second row (due to the opposite obtuse angles of the faces changing in every second
row in the Miura Ori pattern). In Sect. 2 the tab-and-slot conguration was
described, as well as its inversion based on mountain and valley folds.
Fig. 3 a Assembly order, b blocking graph, c insertion vectors
172
Table 1 Joint processing and 2D cut adaptation for Fig. 2 and t = 3 mm, o = 4*t
Joint conguration
Plate Edge a
Joint cong.
28-1
28-1
28-1
28-1
28-2
28-2
28-2
28-2
1 double through
2 through tenons
2 through tenons
1 double through
None
1 double through
1 double through
None
0
1
2
3
0
1
2
3
0.96
1
0.94
0.99
0.96
1
0.94
0.99
Tenon rotations
2D cut adaptation
ltenon
wslot lslot
tenon
67.04
90.5
73.34
99.55
13.96
0.5
16.66
9.55
29.69
29.6
20.6
19.61
13.8
13.79
13.38
12.65
3.84
3.03
4.03
3.55
13.71
13.7
13.13
13.07
tenon
tenon
90.5
73.34
0.5
16.66
29.6
20.6
13.79
13.38
3.03
4.03
13.7
13.13
tenon
On the polygon mesh, we identify mountain and valley folds per edge, using the
normals ni , nj of the two adjacent faces and the vertices pk , pl of their shared edge:

pl . From this we obtain a [ 0 or a\0, indicating mountain
a ni nj pkd
or valley folds and therefore the correct joint conguration (Fig. 3c, Table 1).
4.2
Insertion Directions
In addition to the two simultaneously assembled edges with outgoing through

tenons (v0 and v1 in the positive x-direction and v0 and v3 in the negative
x-direction), the tenons on the other two edges connect to inversely inserted
incoming parts, such as plate 8u, Fig. 3c, which connects to the four
plates 3u,3l,7u,7l simultaneously. Here, the only possible insertion direction is
found at the intersection line between the plates 3 and 7. Generally, all insertion
directions for incoming, inversely assembled parts are determined through a cross
product with the face normals of the diagonal neighbours in the direction of
assembly: If a row is assembled in the positive x-direction, v3 n0 n4 and
v2 n0 n3 . In the negative x-direction the incoming through tenon directions are
v2 n0 n2 and v1 n0 n1 .
4.3
Connectivity and Blockings
Generally joints in timber constructions are semi-rigid, introducing a certain

weakness in the structure. Apart from improving the strength of the joints, it is
benecial to attach each plate to multiple adjacent plates. In a regular single-layer
assembly with quadrilaterals, each plate is connected to 4 adjacent plates, in the
2-layer folded structure with through tenon joints, each plate is attached to 8
adjacent plates. For example, plate 4u in Fig. 3b, c connects to 1u, 1l, 3u, 3l, 7u, 7l,
5u, 5l.
173
The mutual blocking of parts is illustrated in a so called Blocking Graph (Wilson

and Latombe 1994) in Fig. 3d. It shows that the pieces 4l and 4u are blocking all
other parts in the assembly. Due to the individual insertion directions (about 1,500
individual directions in Fig. 2), it is impossible to detach an entire row or column of
elements. The through tenon joints perform like diagonally, crosswise applied
screws.
5 Prototype Fabrication
5.1
Milling System
Figure 4 shows an arch prototype (based on a single-curvature target surface)

connected only with through tenon joints, built to test the fabrication, assembly and
the stiffness of the resulting structure system. The prototype with a span of
3.250 mm, a width of 295 mm and a weight of 82 kg was built from 9-layer birch
plywood panels with a thickness of t 12 mm. The offset between the plates along
the face normal was set to o 4 t. With a span-to-rise-ratio of 9, the arch
demonstrates the construction of a shell with a low curvature like in a typical roof
structure. The maximum tool inclination for the joint fabrication is reduced to
bmax 11:5 .
The parts were fabricated with a 5-axis gantry router equipped with a 10 kW
electro spindle, operated at 16.000 rpm and a feed rate of 5 m/min in 2 vertical
infeeds. The G-Code for the fabrication was generated with a custom script, based
on a Lofttype synchronization between upper and lower polygon outlines of the
Fig. 4 Prototype assembled only with through tenon joints
174
Fig. 5 Load test for Fig. 4
plates and a conversion of the 3D tool vector into cardan rotations using an arctangent function with two arguments. All joints and slots were cut without additional
gaps or tolerances. The tight tting pieces were inserted quickly and precisely, the
insertion force was applied with a rubber mallet.
Figure 5 shows an experimental load test that was performed on this prototype.
A load of up to 0.84 kN was rst applied and then removed in eight steps of 120 N.
This cycle was repeated two times. applying a load of 0.8 kN, which is equivalent
to the self-weight of the structure, a vertical deflection of 12.1 mm (Cycle 1) and
12.3 mm (Cycle 2) was measured at the center point.
5.2
Laser Cutting System
For the production of small-scale prototypes of doubly curved shells with medium
density berboard (MDF) and construction paper, laser cutting proved to be an
efcient technique. However, widely available 2D laser systems cannot cut angular
slots for our through tenon joints with two rotations b and h. In the automated
production of furniture with 3-axis milling machines, non-orthogonal joints are
often realized through an increased slot width, which allows the inserted part to
rotate to its predetermined rotated position. The contact between the two parts is
along the edges of the slot rather than its side faces. We have integrated this method
into the joint generation algorithm. Table 1 shows the joint processing for one of the
segments (from Fig. 2), 8 edges are being processed on the upper and lower plate.
The rst four columns show the joint conguration and the assignment of slots
to the adjacent plates, followed by the dihedral angle and the joint rotations. From
these rotations, as well the thickness t and the offset o, we can calculate corrections
175
Fig. 6 Doubly-curved 2-layer shell, fabricated with a 2D laser cutter
for the shortening of the tenon base ltenon , the extension of the slot width wslot , and
the extension of the slot length lslot .
ltenon
o
cos b
tan b
t
2
cos h
wslot
t
t tan b
cos b
lslot wtenon t tanh
This method allows for the rapid production of precise 3D models based on
doubly curved target surfaces. Figure 6 shows two of the smaller prototypes that we
have built with 3 mm MDF and 1 mm construction paper.
5.3
3D Laser System
Due to the contactless operation of laser cutting systems, there is a constant cut
quality, no tool wear and the multi-axis robot or gantry system is not exposed to
mechanical forces generated by the cutting. The resulting easeand rapidity of
production of this method raises the question whether it could also be used for
full-scale applications. Additionally, larger-scale laser systems in the automotive
industry are commonly used for the trimming of deep-drawn, curved sheets of
hardened steel and therefore capable of 3D simultaneous cutting (illustrated in
Fig. 1), similar to the 3D milling system used in Sect. 5.1. We have performed tests
with such a system, cutting through tenon joints and slots on structural grade spruce
LVL panels with a thickness of up to t 38 mm and a 3D rotation of up to b 45
using a gantry machine equipped with a 6 kW CO2 laser. Cutting at a feed rate of
11 m/min (in a single infeed) with N2 gas and 5 kW power, the accuracy of the
joints was high and independent from the rotation b. The cut width of only 0.6 mm
allows for thin cuts and small radii on corners in the cutting contours. However, a
disadvantage of the method is the charring and odor of the laser-cut edge surfaces.
This can be decreased through higher feed rates but remains noticeable.
176
6 Conclusion
Through tenon joints for LVL panels combine the shear strength of nger joints
with a high resistance to bending moments and out-of-plane traction. The prototypes presented in this paper demonstrate the additional possibility of using double
through tenon joints for the integral attachment and spacing of double-layered
timber plate structures.
The plate conguration based on the Miura-Ori pattern allows for the design of
fabrication- and assembly-aware doubly curved folded surface structures. While the
Yoshimura pattern is constrained to target surfaces with a high curvature and results
in large plate sizes, the Miura-Ori pattern can also be applied to surfaces with a low
curvature. However, the vertical elevation of the vertices in the Miura-Ori also
results in certain structural disadvantages (Stitic et al. 2015). Further research is
necessary to determine if the structurally advantageous shape of the Yoshimura
pattern outweighs the disadvantages in its fabrication, joining and assembly.
Clear advantages of the joint conguration and assembly sequence described in
Sect. 4 include the direct connection of each plate to 8 adjacent plates, as well as the
mutual blocking of the plates which only allows for a piecewise disassembly in the
reverse order of assembly. Therefore, a traction resistance of the joints is not
required and additional connectors such as screws, metal plates or adhesives are not
necessary.
The production of prototypes with 3D milling as well as 2D and 3D laser cutting
systems has shown advantages and disadvantages of the individual solutions. The
highest quality cuts on LVL plates can be achieved with saw blades, due to the
large diameter and the large number of blades. However the production of the
concave polygonal contours and slots of the through tenon joints is not possible
with such tools. Instead, we have used milling bits with a radius of 6 mm, which
allowed for the production of precise parts. The tight t and precision of the joints
was conrmed by a load-test of the arch prototype.
An alternative solution was presented for the fabrication of small-scale prototypes using a geometric adaptation of the joints for 2D laser systems. The method
allows for the rapid production of precise models, however the plates are only in
contact along lines, not surfaces. Further research is required analyzing the influence of this method on load-bearing joints. Finally, the advantages of the 3D
milling and the 2D laser cutting were combined using a 3D laser system for the
production of through tenon joints on structural grade LVL panels.
Acknowledgments The Authors would like to thank Franck Dal-zotto, Anders Holden Deleuran
and TRUMPF Laser Technology. This research was supported by the Swiss National Competence
Center in Research (NCCR) Digital Fabrication.
177
References
Bouaziz, S et al., 2012, Shape-up: Shaping Discrete Geometry with Projections, in Computer
Graphics Forum, vol. 31, pp. 16571667. Wiley Online Library.
Buri, H and Weinand, Y 2006, BSP Visionen -Faltwerkkonstruktionenaus BSP-Elementen in
Grazer Holzbau-Fachtage, Verlag der TechnischenUniversitt Graz, Graz.
DIBt 2011, Allgemeine bauaufsichtliche Zulassung Kerto-Q Z-9.1-100, Paragraph 4.2 and
Attachment No 7, Table 5. Deutsches Institut fr Bautechnik.
Krieg O, Schwinn T, Menges A, Li J, Knippers J, Schmitt A, Schwieger V 2014, Computational
integration of robotic fabrication, architectural geometry and structural design for biomimetic
lightweight timber plate shells, in Advances in Architectural Geometry 2014. Springer Verlag,
London
Robeller, C 2015, Integral Mechanical Attachment for Timber Folded Plate Structures, PhD thesis,
EPFL ENAC, Lausanne, Switzerland.
Roche, S, Mattoni, G, Weinand, Y 2015a, Rotational stiffness at ridges in folded plate structures,
in Elegance of Structures: IABSE-IASS Symposium 2015, Nara, Japan.
Roche, S, Robeller, C, Humbert, L, Weinand, Y 2015b, On the Semi-Rigidity of Dovetail Joint
for the Joinery of LVLPanels, European Journal of Wood and Wood Products, pp. 1-9.
Stitic, A, Robeller, C, Weinand, Y 2015, Form Exploration of Folded Plate Timber Structures
Based on Performance Criteria, in Elegance of Structures: IABSE-IASS Symposium, 2015,
Nara, Japan.
Tachi, T 2009, Simulation of Rigid Origami, Origami, no. 4, pp. 175-187.
Trautz, M and Buelow, P 2009, The Application of Folded Plate Principles on Spatial Structures
with Regular, Irregular and Free-Form Geometries, in IASS Evolution and Trends in Design,
Analysis and Construction of Shell and Spatial Structures, Valencia.
Troche, C 2008, Planar Hexagonal Meshes by Tangent Plane Intersection, in Advances in
Architectural Geometry 2008, Springer, Vienna.
Wilson, R and Latombe, J 1994, Geometric Reasoning about Mechanical Assembly, Articial
Intelligence, vol.71, no.2, pp. 371-396.
RBDM_Robodome: Complex Curved

Geometries with Robotically Fabricated
Joints
Alexander Jung, Dagmar Reinhardt and Rod Watt
Abstract This research reports on the robotic fabrication for the complex architectural geometries of three intersecting domes. The project explores systems for modules
through a tessellated skin (a) of hexagonal tile modules that produce a macro geometry
for a doubly curved, non-developable surface; and the smooth micro geometry of an
interpolating structural rib (b) that requires a customised manufacturing of modules
and their integrated joints (c). It outlines the computational workflow between geometrical conditions, structural requirements, toolpath development, and fabrication
process. The research concludes with a discussion of a new module and joint hybrid
informed by stereotomic and timber joint techniques, which takes advantage of the six
axis robotic fabrication for a standardized multiple face joint between modules of
varying sizes that enables a form and force tting connection.
Keywords Doubly curved geometry

Robotic milling
Joint connections
Project and practical application Computational design to production
1 Introduction
Engineers and architects have for centuries shared a passion for relationships
between form, force and structure. This is manifest in the complex curved
geometries of arches, vaults, hypars, and spheres that determine the mathematical,
structural, organizational and material rationale of shell roofs, bridges, cathedrals,
A. Jung (&) D. Reinhardt R. Watt
Sydney, NSW, Australia
e-mail: alexander.jung@sydney.edu.au
D. Reinhardt
R. Watt
179
180
A. Jung et al.
Fig. 1 a From skin to rib: robotic dome, b ribs in intersecting spheres, c module
or domes. In the legacy of precedents by Frei Otto (1990), Buckminster Fuller

(1999), and Felix Candela (1967), this logic of complexity has been set as
self-forming structures, and as rule-based geometries. These force active forms
differ radically in structural performance and organization of components, and
require a logic of parts for material processes that inform point, lines, surface planes
and solids. Industrial fabrication and serialization, craftsmanship, details and
intersections contribute to complexity here. The project discussed here develops a
primary geometrical logic that extends computational modeling and scripting
directly towards robotic fabrication of modules and joints. RBDM_Robodome uses
the structurally and organizationally efcient geometry of three intersecting spheres
to test systems of robotic fabrication for a tessellated tile skin intersected with a
series of modular ribs (Fig. 1).
In a context of current robotic fabrication, complex curved surfaces of domes
and vaults posit an interesting challenge for transfers from structural performance
towards fabrication. New material production techniques of modular elements and
the connection of parts through customized joint systems have been applied to
structurally compelling, form-dened or force active constructions that explore
robotic applications for complexity, namely; curved structures through robotic
deposition of standard modular systems (Gramazio and Kohler 2014; Trummer
2012); structural vaults based on optimized segments of RDM Vault (McGee et al.
2012); and discrete developable surface segments (Yuan et al. 2014). Solutions for
joint systems on the other hand have been designed as customized connections in
wood (ICD\ITKE 2011 Research Pavilion, Schwinn et al. 2012); in a curved folded
plate structures (Robeller 2014); or as interlocking modular joints (Explicit Bricks,
IAAC 2011, Gramazio and Kohler 2014). Yet two challenges remain; the robotic
prototyping of joints in material volumes, and the problem of connectivity for such
modules. This is interesting because an extended set of criteria needs to be considered: the material envelope, angle and length of toolbit, cutting path, and six axis
181
dened surfaces have to be synchronized in order to maximize the sum of calibrated, planar faces of modules that would stabilize both local connections, and
overall geometry.
2 Complex Curved Geometries: Towards Innity

In contrast to vaulting systems, RBDM_Robodome explores an exemplary series of
three intersecting domes with different sphere diameters, skin tessellations and
connective rib modules. Domes belong to a family of spherical truss systems, with a
diversity of forces distributed as hoop force, meridian force, crown force, edge
force, or radial force (Engel 1967). Domes are sphere segments and as such contain
levels of innite geometrical symmetry in the pattern repetition of a surface module,
and in the degree of sphere curvature for boundary arches that is always the same.
The mathematical logic of a dome is thus simple but smart as complexity becomes
affordable through repetition. In the following, the system geometry of spheres is
introduced, and further evaluated for the affordances of robotic fabrication for a
surface tessellation, and structural rib that organizes adjacent surface areas.
2.1
Geodesic Dome: Icosahedron to Tesselation
The project uses a Geodesic dome (Fuller 1999) as geometrical design model for
the tessellation into producible segments (Fig. 2).
As design model, an icosahdron is constructed using three planes in a golden
section, where the diagonal length of the planes equals the diameter of a dome
(Fig. 2a). The vertices of the planes dene the points for the triangles that will hit
the sphere with their four corners, thereby creating twelve equally sized triangles
(Fig. 2b). A recursive projection of the midpoints of each triangle side towards the
surface of the sphere creates the next frequency (Fig. 2c), resulting in smaller tiling
of the generic hexagon module that constitutes the overall surface when repeated.
Within this system, twelve pentagons appear on the tip of the planes. The frequency
of triangle divisions can be increased innitely, with the pentagons decreasing in
size, remaining in their original position in the surface, not as part of the rib
structure. This system is then extrapolated to the local intersection of two tesselation patterns arriving from two spheres (Fig. 2d, e).
182
A. Jung et al.
Fig. 2 Geodesic dome: a generic icosahedron, b triangulated tessellation, c geometry for spheres,
d 2 domes intersected, e rib section relative to 2 centers
Fig. 3 Comparison of robotic fabrication system. System 1: a double curved shell, b tesselation
formed over curvature, c robotic milling of form mold, d System 2: rib structure, e components
formed by two spheres, f robotic milling of rib module
2.2
Two Approaches for Robotic Fabrication: Skin

or Structure
The identical degree of curvature in domes allows the customization of segments

for the robotic fabrication of a dome. Yet in this particular project, this homogeneity
varies between domes. A further challenge came with differentiation of the base
geometry into the skin and ribs, which were tested in two system series (Fig. 3), and
evaluated for affordances of robotic fabrication. System 1 develops the skin/shells
183
based on tile sizes relative to sphere dimension and curvature, as a sheet that is
efciently formed over a milled plaster mold with varying radius (Fig. 3ac).
System 2 develops intersecting ribs that follow an intersecting curve between
spheres, and bridge two tiles arriving from each side, and with a focus on segmentation of the rib into modules that can be robotically milled from a volume
(Fig. 3df).
Both approaches were tested: several prototypical moulds of system 1 were
robotically milled into plaster with a KUKA KR 60-3 industrial robot (radius
R = 1700, R = 1950 and R = 2300 mm), using a 4KW milling spindle with 6 mm
toolbit. Onto these moulds, surface tiles of 5 mm perspex sheets were air-suction
formed, and assembled to matching faces into a compressive joint surface system.
System 2 required further revisions of robotic workflow towards structural
efciency.
2.3
Geometry Rules for Structural Ribs
In contrast to the skin, the structural ribs had to extend the pattern synchronization
between dome tessellations arriving from two sides, towards carrying a structural
load at the intersection. This required an increase in complexity for geometry, and
change in robotic fabrication method from sheet logic to subtractive process.
Consequently, a number of different scripts were modeled in McNeel Rhino and
GH Grasshopper (scripting plug-in) to link criteria for tessellation pattern, rib
curvature, modules and joints (Fig. 4). The intersecting geometry of two spheres
results in an inclined circle with a center point that anchors the geometry of the rib
Fig. 4 Mother geometry: a rib relative to 2 center points, b segmented, c weaverbird tiling,
d synchronization of two tile sizes in rib, e module divisions, f joint insert
184
A. Jung et al.
Fig. 5 Module with relation to angles of sphere mid point and joint geometry
(Fig. 4a). Modules are segmented relative to material size (Fig. 4b). Each side of the
rib follows the custom mother geometry (center point) of its sphere, with a tiling
run in Weaverbird (GH plugin for mesh associations) of 12 pentagons, and a
variable of hexagon-shapes (Fig. 4c).
All sides in the modules are planar, while connecting surfaces, bottom and top
surfaces are doubly curved. The resulting intersecting modules are used to generate
the outline for a rib that combines shape information of two different patterns from
each side (Fig. 4d). The rib is dened through the degree of surface curvature from
two spheres, and through planar sides that connect to the skin. Divisions between
the modules are generated through the orientation of intersecting faces of the edge
conditions of valley and ridge (Fig. 4e). Into this section plane, a second mother
geometry for the joint is inserted that is always the same, but varies position
(Fig. 4f). As a result of this geometry denition, each modular component features a
number of criteria that are continuous within the sphere; the sphere diameter; the
continuation of interior arch curvature; the length of side edges that connect to skin
patterns on each side; and the geometry of joint connection between a lower and
upper module (Fig. 5).
3 Robotic Fabrication of Structural Ribs and Joints

RBDM then developed the computational modeling of geometry rules towards
six-axis robotic milling, whereby the overall geometry is linked to a standardisation
of producible parts adequate to material properties and performance. The prototype
185
Fig. 6 Robotic milling for one rib segment: a simulation, b, c for rough SRF/3 grade surface and
rough and ne milling, and d milling insertion of canal
paralleled aspects of serialization and customisation; rstly, for a structurally

effective rib system able to lock into the complex doubly shaped surface and its
hexagonal tiling system; and secondly, a constructive detailing of compressive
joints and a tension canal that are integrated for construction.
3.1
Workflow and Robotic Fabrication of Rib Structure
The structural system for the exemplary module series was further explored in a
robotic simulation with KUKA|prc in order to adjust the size of the intersecting tiles
by rotation, so that enough material remains. This data set contains a series of
customized scripts for SRF rough and ne surface milling of top and bottom
surfaces for each particular surface angle, facets and nishes (Fig. 6a, b). For a rst
186
A. Jung et al.
material test (Fig. 6b, c), the robotic milling followed industry customs for volume
milling (as in sandstone or wood) with support of added feet that allow steady
positioning on the routing bed and precise turnover of the material sample.
Modules were then robotically milled with a KUKA KR 60-3 industrial robot,
using a standard flat headed 4KW milling spindle with 10 mm toolbit and 3 mm
stepover, in a series of robotic protocols that require multiple manual turnovers but
adequately present material behaviour of wood or stone, with fabrication axis
angles relative to robotic axis deployment. A canal is drilled through each module
at center of the joint to allow for insertion of a tension cable (Fig. 6d). Modular
components are unique along each rib and respond to force-flow changes where the
structure acts in compression, resulting in thicker sections (added to top of load
distributed along center line of structure) towards the ground.
3.2
Robotic Fabrication of Joint
Accompanying the robotic fabrication of skin plates and the structural ribs, a third
and novel part of robotic fabrication is nally discussed: each module also contains
a three-dimensional joint that is inserted into the section plane between segments
(Fig. 7). These joints are multifaceted elements that are developed as a modication
of traditional japanese wood connections, which go beyond simple nger joint
systems, such as the three-faced halved rabbeted oblique scarf splice (Sumiyoshi
and Matsui 1989). RBDM uses a similar variation of a male-and-female joint,
constructed here as a multiple of a 90 angle.
This angle serves multiple purposes: the joint is embedded as geometric information into both the scripting and robotic process, and capitalizes on an industrial
logic. Instead of many different connections, the same precise mother geometry is
maintained while each particular module can differ in direction and orientation
along the dividing surfaces. The triangulated connection also maximizes the surface
contact between two modules; prevents horizontal movement; and provides structural efciency.
3.3
From Geometry to Structural Behaviour
This allows the structure to work in compression as all vaults or domes, yet ensures
precise construction as modules are connected as a compression unit that transfers
loads to the ground, with the triangulated joints counteracting lateral shifts.
Through the combination of a geometric logic coupled with robotic fabrication,
structures can be produced as sequences of equals (skin) and individual parts (ribs)
that can be fabricated effectively, in series, to contribute to complexity (Fig. 8).
187
Fig. 7 Joint as hybrid between wood and stereotomy: a one single standard, b local position on
rib, c connecting faces, d robotic nal module/joint hybrid
In continuation, the research will develop the robotic protocol to integrate the
doubly curved geometries (Fig. 8a) for an optimization in subtractive rough milling
of stacked rib-modules (Fig. 8b), taking into account the shared surface degrees in
one sphere that allows elements to be nested into each other (Fig. 8c). These
butterfly modules can then be roughcut as stacked series in order to reduce material
waste. Through the combination of a geometric logic coupled with robotic fabrication, structures can be produced as sequences of equals such as the tessellated
skin, and individual parts such as the structural ribs that can be fabricated singularly, in series, and in true materials.
188
A. Jung et al.
Fig. 8 Efciency for robotic fabrication and structural performance: a sphere curvature in module,
b milling segent, c stacking system, d rib context under tension
4 Conclusion
This research has explored the robotic fabrication of three intersected domes: based
on a hexagonal pattern structure that is differentiated into serial surface elements, a
customised modular structure and integration of new joint system that combines
aspect of timber construction with a stereotomic process. In doing so, the research
project has demonstrated that a serial production of doubly curved surfaces both for
surface and solid elements was achievable through robotic fabrication. Robotic
fabrication was also applied to equip the material with a high level of joint detail,
thereby seamlessly bridging between design process and fabrication, and furthermore incorporating construction and structural performance. This research project is
currently developed for a 1:2 prototype.
Knowledge about geometrical logic, material and fabrication process can then
enable structures that are complex but geometrically smart, producible at affordable
189
cost, with low material waste, and with close references to industrial cutting processes. In sum, the application of robotic processes allowed us to reconsider
engineering precedents, and to reformulate this into a novel architectural system.
Acknowledgments This research is an ongoing initiative (2014-) that has been informed by the
SmartStructuresLab (20142015), and with parts of the research (Skin) developed in the robotic
elective CodeToProduction (2015). The authors would like to express thanks to Gabriele Ulacco
for the research elective, and the student team for their engagement. The research has been
generously supported by The Faculty of Architecture, Design and Planning, The University of
Sydney, through a SEED Grant, and produced at DMaF. The authors would further like to express
thanks to Marjo Niemel for continued support.
References
Engel, H 1967, Structure Systems, Hatje Cantz, Stuttgart.
Fuller, B 1999, Your Private Sky- R Buckminster Fuller Krausse, J and Lichtenstein, C (eds),
Design alsWissenschaft, Lars Mueller Publishers, Zurich Switzerland, pp. 268, 304.
Garlock, M and Billington, D 1967, Felix Candela: Engineer, Builder, Structural Artist, Princeton
University Art Museum, Princeton, N.J and Yale, New Haven.
Gramazio, F, Kohler, M 2014, The Robotic Touch- How Robots Change Architecture, Park
Books, Switzerland, pp 224237.
McGee, W, Feringa, J, and Sndergaard, A 2012 Processes for an Architecture of Volume
Robotic wire cutting. in Brell-Cokcan S, Braumann J (eds), Robotic Fabrication in
Architecture, Art, and Design, Springer Wien, New York, pp 6371.
Otto, F 1990, IL 25: Experiments - Form, Force, Mass. Institute for Lightweight Structures IL,
University of Stuttgart, Stuttgart.
Reinhardt, D, Jung, A, Cabrera, D, Ulacco, U and Niemela M 2014 TriVoc-Robotoc
Manufacturing for Affecting Sound Through Complex Curved Geometries in McGee, W,
Ponce de Leon, M (eds), Robotic Fabrication in Architecture, Art and Design. TriVoc-Robotoc
Manufacturing for Affecting Sound Through Complex Curved Geometries, pp. 181195.
Robeller, C, Nabaei, S and Weinand, Y, 2014, Design and Fabrication of Robot-Manufactured
Joints for a Curved-Folded Thin-Shell Structure Made from CLT in McGee, W and Ponce de
Leon, M (eds), Robotic Fabrication in Architecture, Art and Design 2014, Springer
International Publishing Switzerland, pp. 6781.
Schwinn, T, Krieg, O and Menges, A 2012, Robotically Fabricated Wood Plate Mor-phologies in
Brell-CokcanS and Braumann, J (eds), Robotic Fabrication in Architecture, Art, and Design,
Springer Wien, New York, pp.4961.
Sumiyoshi, T, Matsui, G, 1989/1991, Wood joints in Classical Japanese Architecture, Kajima
Institute, Tokyo, Japan.
Trummer, A, Amtsberg, F and Peters, S 2012, Mill to Fit - The Robarch, inBrell-Cokcan, S and
BraumannJ (eds) Robotic Fabrication in Architecture, Art, and Design, Springer Wien, New
York, pp. 6371.
Yuan, F, Meng, H and Devadass P 2014, Performative Tectonics- Robotic Fabrication
Methodology Towards Complexity in McGee, W and Ponce de Leon, M (eds), Robotic
Fabrication in Architecture, Art and Design 2014, Springer International Publishing
Swit-zerland, pp. 181195.
Topology Optimization and Robotic

Fabrication of Advanced Timber
Space-Frame Structures
Asbjrn Sndergaard, Oded Amir, Phillip Eversmann,
Luka Piskorec, Florin Stan, Fabio Gramazio and Matthias Kohler
Abstract This paper presents a novel method for integrated topology optimization
and fabrication of advanced timber space-frame structures. The method, developed
in research collaboration between ETH Zrich, Aarhus School of Architecture and
Israel Institute of Technology, entails the coupling of truss-based topology optimization with digital procedures for rationalization and robotic assembly of
bespoke timber members, through a procedural, cross-application workflow.
Through this, a direct chaining of optimization and robotic fabrication is established, in which optimization data is driving subsequent processes solving timber
joint intersections, robotically controlling member prefabrication, and spatial
robotic assembly of the optimized timber structures. The implication of this concept
is studied through pilot fabrication and load-testing of a full scale prototype
structure.
A. Sndergaard (&)
Aarhus School of Architecture, Aarhus C, Denmark
e-mail: asbs@aarch.dk
O. Amir
Israel Institute of Technology, Haifa, Israel
e-mail: odedamir@cv.technion.ac.il
P. Eversmann L. Piskorec F. Gramazio M. Kohler
ETH Zrich, Zrich, Switzerland
e-mail: eversmann@dfab.ch
L. Piskorec
e-mail: piskorec@dfab.ch
F. Gramazio
M. Kohler
F. Stan
Odico Formwork Robotics Aps, Odense, Denmark
e-mail: florin@odico.dk
191
192
Keywords Topology optimization

Advanced timber structures
Digital fabrication Architectural robotics
1 Introduction
Topology optimization (Bendse and Sigmund 1999) may broadly be dened as a
family of procedures aimed at creating efcient structural layouts. In the form of
continuum representations, this equates the redistribution of material within a Finite
Element-discretized design space. In the form of truss-based representations, it
equates the determining of the topological connections and cross section sizes from
a pre-dened set of possible members.
In a preceding research work, continuum optimization was explored for architectural concrete structures (Dombernowsky and Sndergaard 2012; Feringa and
Sondergaard 2015; see also Fig. 1). While these studies successfully indicated
signicant potentials for design innovation and reduction of material consumptions
compared to commonly found standard structures (Dombernowsky 2011), the work
also found an inherent complexity in translation from optimization result to construction design. Furthermore, current continuum procedures are not directly
applicable to the majority of construction projects, which are realized through
assembly of prefabricated semi-manufactures and components, necessitating alternative modes of optimization. These limitations can be conceptually addressed
through the application of truss-based topology optimization (Sndergaard 2013).
This approach enables the optimization of pre-dened memberand connection
types within predetermined ranges of cross-sections, hereby facilitating the generation of optimized designs, which align closer with current building and construction practice. However, the topological complexity of the optimization results
derived from such processes necessitates digital new means of pre-fabrication and
assembly to become practical to full scale building implementation. As of today, no
Fig. 1 The unikabeton prototype ( Dombernowsky and Sndergaard, Copenhagen 2010, left).
The opticut prototype structure ( Sndergaard and Feringa, under construction, right)
Topology Optimization and Robotic Fabrication
193
method exists for direct realization of optimization result, in which the complex
challenges in prefabrication and assembly arising from the complexity of the
optimized topologies are handled in an integrated, digital process. The collaborative
research presented in this paper addresses these challenges targeted at the special
application area of digital manufacturing of timber structures.
2 State-of-the-Art
Recent developments within architectural robotics have presented novel procedures
for digital fabrication of advanced timber structures. In the seminal experiments
conducted at ETH Zrich (Gramazio and Kohler Research), integrated robotic fabrication and assembly have been demonstrated within layered assembly of timber
structures (Willmann et al. 2016; Gramazio et al. 2015). This process is currently
being applied for large scale production of the 80 22.5 m Arc-Tech-Lab roof
structure under construction at the ETH Hnggerberg university campus (Apolinarska
et al. 2016). Furthermore, recent experiments at the ICD Stuttgart (Knippers and
Menges 2013) have demonstrated full scale fabrication of plated timber structures in
combination with manual assembly through robotic CNC-milling of bespoke elements for the Research Pavilion 2011 while long-threaded developments at the EPFL
Lausanne are investigating digital structural design and fabrication of new timber
structures (Weinand 2009).
Most recently, research was undertaken at ETH Zrich to explore the potential of
robotic assembly of single-joint spatial structures in combination with application
of fast-curing, 2-component chemical binder (Helm et al. 2016). The collaborative
work presented in this paper builds on this development, while process parameters
have been extended to enable the realization of topology optimized structures.
3 Optimization of Timber Structures

The structure fabricated within the current study is essentially a rationalization of a
result of a topology and sizing optimization procedure. We rely on well-established
formulations from the eld of structural optimization, where the purpose is to nd
the optimal structural layout of a truss (locations of existing members) as well as the
optimal cross-section areas. Such layouts have been investigated since the early
20th century stemming from Michells classical work on least-weight grid-like
continua (Michell 1904).
194
3.1
Implementation of Truss Optimization Using Optimality

Criteria Approach
The current implementation uses the so-called ground structure approach (Dorn
et al. 1964), where for example the design domain is discretized using a xed set of
nodal points, which are then connected by a set of potential truss bars. The only
requirement in setting the ground structure is that it should be able to transfer the
loads to the points of supports without forming a mechanism. Then, the objective of
topology and sizing optimization is to determine the optimal topology and
cross-section areas of all potential bars, including eliminating unnecessary bars by
assigning them a zero cross-section area. In its most basic form, the topology
optimization procedure aims at nding the stiffest truss. This can be stated as
follows: nd the structural topology and cross-section areas, so that stiffness is
maximized (i.e. external work is minimized), subject to an upper limit on the
volume of material used, and provided that structural equilibrium can be satised.
The corresponding mathematical statement is:
min f T u
a
s:t: : Kau f
NBARS
X
ai l i V 0
i1
ai 0
i 1; . . .; NBARS
Here a is a vector of non-negative cross-section areas; f is the external load vector;

u is the displacements vector; K is the stiffness matrix, depending on a; li represents
the length of the i-th member; and V is an upper limit on the volume of the structure.
The solution is found by an optimality criteria approach where the over-stressed
cross-section areas are gradually increased while the under-stressed ones are
decreased.
This implementation was applied in the development of prototype designs for
the purpose of testing rationalization and fabrication methods (Fig. 2). The prototype design was reached by optimization from 1711 possible connections in an
irregular, trapezoid ground structure with 555 subdivisions in the XYZ directions. The conguration was fully supported on three nodes at points (1,1,1; 1,5,1
and 5,1,1) and eccentrically loaded with 5 kN single point load at (3,2,5).
Optimized for a minimization of compliance under a volume constraint of 0.14 m3,
the processes resulted in a geometrically complex 34 bar structure.
195
Fig. 2 Optimization result of a 4-point supported single-load conguration (left). Early result of
the prototype optimization (right)
4 Geometry Rationalization
The optimization procedure described in the previous sections accounts for the
structural load-capacity of the topological conguration, and the dimensioning of
member cross-sections. However, the output does not solve the geometrical intersection of members at node levels, which must be processed in a secondary step. To
accommodate for this, a rationalization procedure is developed and implemented in
GhPython (Fig. 3). The objective of this procedure is to, given any topological
structure, solve the necessary cutting sequence and orientation for bars in each
Fig. 3 Workflow diagram of the rationalization procedure
196
node, based on the limitations of the robotically controlled sawing process. The
outcome is a discrete list of revised bar geometries, which avoid in-node overlaps,
while (a) ensuring structural continuity from load-points to points of support, and
(b) minimizing the number of necessary intersections while maximizing the contact
surface area at member joints.
Members are discretized into ranges of pre-determined cross-section dimensions,
and joints rationalized according to the member cross-section dimensions included
in the joint: joints containing members of only one dimension type are trimmed
against a shared plane derived from the bi-sector of the center axis of the intersecting members. For joints containing members of several dimension types, largest
dimension types are trimmed per bisector as previously described, while smaller
dimension types are trimmed against the cross-section proles of larger members.
The result of this operation is that lower level bars will share a surface only with
one higher level member, leading to simplied joining faces in this situation.
5 Assembly Procedure
A predominant challenge for robotic fabrication of topology optimized
space-frames is the auto-generation of valid assembly motion sequences, which
must determine the chronological order of member insertion while avoiding collision with the structure under construction. Addressing this challenge an assembly
processor is developed, which computes the assembly sequence and the respective
trajectories directly from the node geometries, hereby avoiding simultaneously
in-node collision at joint level and global collision at the structural level (Fig. 4).
This is conducted in an operation, in which members and nodes are sorted
according to their distance to the robot base, and the bar with the smallest angle
relative to the base-plane is selected. Once the rst valid bar is found, insertion
Fig. 4 Assembly procedure workflow diagram
197
Fig. 5 Assembly sequence example. From the list of possible node connections (left), the inserted
member is selected based on connectivity and priority level, and insertion trajectory computed
from the node geometry (right)
trajectories are computed by the sum vector of the normals of the contact faces of
neighboring members within the same joint, dening collision avoidance as any
trajectory which is >90 to any normals of the neighboring contact faces. If collision is found, and cannot be solved through incremental search for alternative
trajectories in the trajectory solution space, a combinatorial search is performed for
the insertion sequence with least collisions; the obstacle member is retracted and
validity is re-checked after every insertion of a new node.
For every insertion operation, a connectivity check is performed at the end-node
of each inserted bar member (opposite of the joint node). If found, the connecting
bar will be inserted ensuring, where possible, a build-up through triangulation,
which help to ensure physical stability during assembly (Fig. 5).
6 Robotic Operations
The fabrication setup at ETH Zrich consisted of a KR 150 L110-2 KUKA robot on
a 7 m linear axis (Gramazio et al. 2015) and a Mafell Erika 85 circular table saw.
A custom positioning table for material feeding was added to the saw. The robot is
equipped with a custom parallel gripper, which is capable of holding the beams stiff
enough during the cutting process. Within the robotic process, the following steps
were repeated for every beam: rst, a wooden beam was gripped, and then positioned in 5 axes for cutting. Then, the positioning and cutting procedures were then
repeated for all cutting planes. Finally, the robot could reach the nal assembly
position, where multiple beams can be prepared for gluing. Each step of the geometric constraints, the robotic movement and the assembly is explained in more
detail in the following paragraphs.
198
6.1
Geometric Constraints
The idea of the cutting operations was to perform every cut with a vertical sawblade, while the robot could perform all complex geometric orientations. The orientations can be computed through a transformation matrix, which aligns the
trimming plane vertically, while keeping on of the edges of the beam horizontal
(Fig. 6).
This approach allows for a wide range of possible cutting angles (Fig. 7, left).
The positioning angle (in the XY plane) depends on the specic length of each
beam, the distance to the tool and the saws maximal pulling distance. The cut angle
(in the vertical plane) depends on the size of the gripper and parts of the end
effector, which could cause collisions. A threshold angle of 60 was identied for
the current setup.
Fig. 6 Vertical cutting orientation
Fig. 7 Maximum angles (left), and assembly surface (right)
199
Building on previous research (NRP-66 research projects, 201217), joints

where connected using a fast-curing, two-component adhesive with curing times of
510 s (Zock et al. 2014). Each connecting face was perforated to allow for the
adhesive to permeate deeper into the structure, hereby increasing resistance to
tensile stress. Due to the high viscosity of the adhesive, gaps between connecting
faces were sealed with tape during injection, to avoid leakage during the short-term
curing. This approach allowed for accommodating tolerances between 210 mm,
while ensuring strong connections.
6.2
Robotic Movement
A number of challenges were encountered in the programming of the toolpaths for

robotic movements. The geometric operations and toolpath data was computed in
python. For the simulation and post-processing the software HAL was used.
For angles beyond the threshold angle in Sect. 6.1, a regripping procedure was
created. During this procedure, the robot places the beam on the table vertically and
regrips it afterwards at a 90 angle. Due to safety reasons, beams were rotated
horizontally in a safety-plane above the sawblade. Therefore, the robots a6 joint had
to perform most of the movements. The rotational limits were easily reached. This
issue was solved with using joint-movements to a custom unwinding position
between each of the cuts, where the joint of the 6th axis can rotate in interpolated
movements to a zero value. This created some additional movements, but provided
a safe position for the motion-planning. It helped in avoiding collisions between
very long beams and the robot during the cutting and positioning. Custom positioning of the robot base was used to allow the robotic rotations of the beams,
which outer corners reached beyond the linear axis (Fig. 8).
For the positioning of the robot base three cases needed to be considered.
Ideally, the base moves along a position normal to the current target plane (a). This
works only for rotations along the maximal offset domain of the robot. Therefore
repositioning of the base can be anticipated and performed during a safe position as
the unwinding position (b). Inward rotations can be performed with the maximum
offset depending of the reach of the robot (c). Therefore a pattern for all base
positions had to be calculated in advance for all toolpath targets.
Fig. 8 Custom robot base positioning
200
6.3
Assembly Trajectories
For the assembly motion planning a slightly curved reference surface was created,
slightly hovering over the planned structure. The assembly direction planes derived
from the assembly processor described in Chap. 6 were then pulled in the normal
direction on the surface. This allowed for an easy control of the trajectories since
the upper paths are projected above the structure, whereas the lower paths remain in
ideal safe regions around the structure. Figure 9 shows the nal prototype, which
was constructed using the above described processes.
7 Analysis and Load-Test

The fabricated timber truss was nally tested under point loading in order to validate the effectiveness of the overall design and fabrication process (Fig. 10, left).
Since the performance of timber structures is determined by the capacity of the
joints, a key question was the structural capacity of the glued connections. In the
numerical simulations, both in MATLAB (within the optimization procedure) as
well as in RSTAB, very small displacements were predicted under a load of 5 kN.
Furthermore, the difference in displacement between the optimized design (with
variable cross-sections) and the fabricated design (with three bar types only) was
under 10 %. At the moment of writing, load-testing went to 13.6 kN, but failed to
Fig. 9 Final prototypical structure ( Michael Lyrenmann)
201
Fig. 10 Load testing setup (left), and displacement and force-diagram of the physical load-test
(right)
proceed to collapse due to rupture of the connecting metal braces (Fig. 10, right).
While this limits the measurement of the actual stiffness of the structure, it nevertheless indicates that jointsdespite geometrical complexity and variability of
gluing conditionsperform overall within expected range and that the prototype
indeed is very stiff due to its optimized conguration.
8 Conclusions and Outlook

This paper has presented a process that facilitates integrated optimization, production rationalization, robotic fabrication and assembly of topology optimized
space-frame structures. The method discussed presents a solution for the production
of spatial structures with a high concentration of bars at individual nodes, the
implication of this is demonstrated through optimization, fabrication and
load-testing of a full scale structure. Analysis and tests show general consistency
between predicted capacities in optimization, the analysis of the rationalized
geometry and the performance during physical testing.
While the presented work demonstrates the feasibility of the proposed process, a
number of challenges were identied for further work. The high level of complexity
of all steps of the described process necessitates either fully automated or highly
automated construction processes to remain feasible in full scale architectural
applications. This implies in particular (a) that custom adaptive/feedback-based
processes are needed in the future for handling un-modeled material effects (in
particular gravitational sagging of the structure during assembly and tolerances
stemming from member warping during cutting) and (b) therefore the development
of novel, fully integrated design and fabrication workflows/tools are required.
Finally, automation of the sealing process as presented is challenged by the high
degree of joint complexity. While manually solvable, robotic automation would be
key to improving the industrial applicability of the process.
202
Acknowledgements The research presented in this paper was performed within a research
exchange between ETH Zrich and Aarhus School of Architecture in collaboration with the NCCR
Digital Fabrication MAS Programme and Israel Institute of Technology, Haifa. This research was
supported by the NCCR Digital Fabrication, funded by the Swiss National Science Foundation
(NCCR Digital Fabrication Agreement # 51NF40-141853). The contribution of Aarhus School of
Architecture was enabled through the generous nancial support of the Danish Ministry of Higher
Education and Science under the Elite Research Travel Grant program.
The presented research builds directly on the research ndings and developments from the
ongoing SNSF research project NRP-66, established in collaboration between ETH Zrich, Bern
University of Applied Science and Nolax AG. The primary constituent ndings for the presented
work are (a) ongoing joining experiments involving two-component, super-curing adhesives and
(b) the principal process of robotic pre-sawing and spatial assembly of timber members.
In particular, the authors would warmly like to thank: Dr. Volker Helm and Dr. Jan Willmann
for their helpful organizational support and discussion of research and paper content; NRP-66
collaborators Dr. Thomas Kohlhammer, Aleksandra Apolinarska and Peter Zock for fruitful discussions of analytical and structural approaches, knowledge transfer and help regarding the
adhesive process; Student assistants Micha Ringer and Lazlo Blaser for their involvement in the
fabrication of the prototype structure; Michael Lyrenmann for excellent photographic documentation; and Dominik Werne and the ETH HIF-Halle staff for their tireless involvement and support
in the load-testing of the structure.
References
Apolinarska, A, Knauss,M, Gramazio, F, Kohler, Mc. 2016, The Sequential Roof, in Menges, A
(ed), Advancing Wood Architecture, Routledge,Abingdon, UK ; New York, USA.
Bendse, MP and Sigmund, O 1999, Material Interpolation Schemes in Topology Optimization,
Archive of Applied Mechanics, vol. 69, no.9-10, pp. 635654.
Dorn, W, Gomory, R and Greenberg, M 1964, Automatic Design of Optimal Structures, J de
Mecanique,vol. 3, pp. 2552.
Dombernowsky, P and Sndergaard, A 2011, Unikabeton prototype, in Sheil, R and Glynn, R
(eds), Fabricate: Making Digital Architecture, Riverside Architectural Press, Waterloo, CA.
Dombernowsky, P and Sndergaard, A 2012, Design, Analysis and Realiza-tion of Topology
Optimized Concrete Structures, International Association for Shell and Spatial Structures vol.
53, pp. 209216.
Feringa, J and Osterhuis, K 2011, Investigations in Design & Fabrication at Hyperbody, in
Glynn, R and Sheil, B (eds.), Fabricate: Making Digital Architecture Riverside Architectural
Press, Waterloo, CA.
Gramazio, F, Kohler, M and Willmann, J 2015,The Robotic Touch How Robots Change
Architecture, Park Books, Zrich, pp. 466-467.
Helm, VM, Gramazio, F and Kohler, Mc. 2016, Additive Robotic Fabrication of Complex Timber
Structures,in Menges, A (ed), Advancing Wood Architecture, Routledge, Abingdon, UK /
New York, USA.
Knippers, J and Menges, A 2013, ICD/ITKE Research Pavilion 2011, in Hu, C (ed),
Architectural Material and Texture I, pp. 266273.
Michell, AGM 1904, The Limit of Economy of Material in Frame Structures, Philosophical
Magazine, vol. 8, no. 6, pp. 589597.
203
Sndergaard, A, Amir, O, and Knauss, M 2013, Topology Optimization and Digital Assembly of
Advanced Space-Frame structures, in Beesley, P, and Kahn, O (eds), Adaptive Architecture,
Proceedings of the 33rd ACADIA Confe-rence 2013, Riverside Architectural Press, Waterloo.
CA.
Weinand, Y 2009, Innovative Timber Constructions, in Journal of the International Association for
Shell and Spatial Structures, vol. 50, no. 2, pp. 111120.
Willmann, J, Gramazio, F and Kohler, M c. 2016, New Paradigms of the Automatic Robotic
Timber Construction in Architecture, in Menges, A c. 2016, Advancing Wood Architecture,
Abingdon, UK / New York, USA.
Zock, P, Bachmann, E, Gramazio, F, Kohler, M, Kohlhammer, T, Knauss, M, Sigrist, C and
Sitzmann, S, 2014, Additive robotergesttzteHerstellungkomplexerHolzstrukturen, 46,
Tagungsband FortbildungskursHolzverbindungenmitKlebstoffenfr die Bauanwendung,
Swiss Wood Innovation Network (S-WIN), Weinfelden, pp. 197-208.
Mobile Robotic Brickwork

Automation of a Discrete Robotic Fabrication
Process Using an Autonomous Mobile Robot
Kathrin Drfler, Timothy Sandy, Markus Giftthaler,
Fabio Gramazio, Matthias Kohler and Jonas Buchli
Abstract This paper describes the implementation of a discrete in situ construction
process using a location-aware mobile robot. An undulating dry brick wall is
semi-autonomously fabricated in a laboratory environment set up to mimic a construction site. On the basis of this experiment, the following generic functionalities
of the mobile robot and its developed software for mobile in situ robotic construction
are presented: (1) its localization capabilities using solely on-board sensor equipment and computing, (2) its capability to assemble building components accurately
in space, including the ability to align the structure with existing components on site,
and (3) the adaptability of computational models to dimensional tolerances as well as
to process-related uncertainties during construction. As such, this research advances
additive non-standard fabrication technology and fosters new forms of flexible,
adaptable and robust building strategies for the nal assembly of building components directly on construction sites. While this paper highlights the challenges of the
current state of research and experimentation, it also provides an outlook to the
implications for future robotic construction and the new possibilities the proposed
approaches open up: the high-accuracy fabrication of large-scale building structures
outside of structured factory settings, which could radically expand the application
space of automated building construction in architecture.
K. Drfler (&) F. Gramazio M. Kohler
ETH Zurich, Chair of Architecture and Digital Fabrication, Zurich, Switzerland
e-mail: doerfler@arch.ethz.ch
F. Gramazio
M. Kohler
T. Sandy M. Giftthaler J. Buchli
ETH Zurich, Agile & Dexterous Robotics Lab, Zurich, Switzerland
e-mail: tsandy@mavt.ethz.ch
M. Giftthaler
e-mail: mgiftthaler@ethz.ch
J. Buchli
e-mail: buchlij@ethz.ch
205
206
K. Drfler et al.
Keywords In situ robotic construction

fabrication Robot localization
Mobile robotic fabrication Adaptive
1 Introduction
The degree of automation in the construction industry is constantly rising, particularly in the area of pre-fabrication. On construction sites, however, the level of
automation is rather low, and nal assembly tasks of building components predominantly imply the use of manual labor (Gambao and Balaguer 2002). This is a
fundamental difference to other industries (e.g. the automotive industry), where the
entire process from production of single parts to nal assembly is often fully
automated (Balaguer and Abderrahim 2008). Therefore, robotic in situ fabrication
performed directly on the construction siteholds the potential to nally close
the digital process chain between design and making (Helm 2014; Helm et al. 2014)
and to leverage novel aesthetic and functional potentials in the eld of non-standard
architectural construction Gramazio et al. (2014).
However, the inherent characteristics of construction sites substantially differ from
those in factory environments, which makes the implementation of in situ fabrication
tasks signicantly more difcult. Building sites are generally considered unstructured1 (DeSouza et al. 2002) because they are gradually evolving and continuously
changing shape during construction, floors are not necessarily flat and there is no
guarantee for regular structures in the surroundings, as opposed to prevalent constant
conditions in industrial production. Additionally, robots for pre-fabrication are
commonly employed at an anchored position within a work cell and work pieces are
brought to the stationary unit. Yet, to enable the fabrication of large-scale building
structure that exceed the workspace of a xed robot, the employment of robots on
constructions sites requires them to be mobile (Fig. 1). Robots need to be able to travel
to the place of production and to move during construction, while still being able to
localize themselves with respect to the working environment and fabricate structures
accurately in space (Seward 2002; Feng et al. 2014).
To take on these challenges, the two ETH Zurich groups Gramazio Kohler
Research2 and the Agile & Dexterous Robotics Lab3 are developing an autonomous
area-aware mobile robot, called the In situ Fabricator (IF).
Within this paper, the term unstructured is used to describe the environment of building sites,
although, in most cases they can be dened as semi-structured, due to partially con-strained and
dened conditions.
2
http://www.gramaziokohler.arch.ethz.ch
3
http://www.adrl.ethz.ch
1
207
Fig. 1 Mobile in situ fabrication entails that instead of bringing the work piece to the xed robotic
work cell and referencing the work piece in relation to the robot (left), the robot travels to the
worksite and needs to localize itself in relation to the working environment (right)
Following its predecessor dimRob (Helm et al. 2012), described in the next
section, IF consists of an industrial robotic arm mounted on a base driven by
hydraulic crawler tracks. It is intended as a generic mobile fabrication robot for the
future employment on construction sites. This paper presents a rst physical construction experiment using IF: the fabrication of an undulated dry-stacked brick wall,
made up of discrete production steps, in a laboratory environment set up to mimic a
construction site. The experiment serves to demonstrate the robots generic functionalities and system architecture, as well as its integrated digital design and control
software framework. In this context, objects of detailed investigation are (a) the
automated tting of the geometric description of key features of building site
components (e.g. floor, walls, pillars) to captured laser range measurements made by
the robot, (b) the precise robot localization using point cloud registration, and (c) the
adaptability of a parameterized brick walls geometric description and its corresponding assembly sequences to process-related parameters during construction.
2 Context
Concepts and exploratory setups to employ industrial robotic units for automated
in situ fabrication tasks have been explored since the 1980s and 1990s, the most
advanced of them being the mobile bricklaying robots ROCCO (Andres et al. 1994)
and BRONCO (Pritschow et al. 1996). These early concepts, however, are characterized by heavy duty machinery and rigidly planned production routines. As a
result, assembly procedures largely depend upon uniform, standardized building
elements, standardized connections, strictly organized fabrication routines and well
208
K. Drfler et al.
controlled environments. In the last decade, however, robots have evolved through
new developments in sensing, real-time computation and communication, within
which inflexible top-down organization principles are replaced by flexible and
adaptive bottom-up approaches. These advancements allow also for their customization as advanced design and construction tools.
In 2010, the Gramazio Kohler Research group, together with the industrial
partner Bachmann Engineering AG, developed and built IFs predecessor, the
mobile platform dimRob (Helm et al. 2012). It consisted of an ABB IRB 4600
industrial robot arm mounted on a tracked mobile base. Its hydraulic drive system
was powered by a diesel engine and the system was steered manually using
hydraulic levers. While dimRob already successfully demonstrated core concepts
for in situ fabrication on the basis of a variety of experiments its applicability was
limited by a few key aspects. The original design of dimRob lacked the sensing
required to allow the robot to build with high accuracy without being anchored to
the ground using fold-out legs. This made it infeasible to build structures that would
require the robot to move many times during construction. Also, dimRob had to be
repositioned manually. This not only required substantial human intervention, but
also placed a limit on the precision with which the robot could be repositioned.
Finally, the robot arm was powered and controlled by a control box, which was not
integrated into the robotic system, which signicantly limited the autonomous
capabilities of the overall setup. This motivated a major revision to drastically
extend its capabilities. The result is IF, the In Situ Fabricator, whose main features
are described in the following sections.
3 IF Setup
3.1
In Situ Fabricator System Architecture
IF is designed such that it can autonomously complete building tasks directly on a

construction site. The level of autonomy intended for the robot is dened to contain
all of the facilities required for precise manipulation of building materials. In this
way, human interaction with the robot is narrowed down to the specication of
building tasks through high-level planning environments and dedicated interfaces.
In order to achieve this, the robot is designed to be self-contained, with all components needed for construction on-board: mainly sensing, control hardware, and
computing systems. A dependence on excessive setup of the construction site for
building is also avoided. For this reason, the robot is designed such that it should
not depend on external referencing systems (e.g. Nikon iGPS, Vicon, etc.).
209
Fig. 2 Two separate processes control the robots arm and base respectively. Coordination of the
two is performed within Grasshopper Rhinoceros
3.2
Hardware Overview
IF features the same robot arm as its predecessor, but additionally it carries a
complete, retro-tted ABB IRC5 industrial controller. The whole system is electrically powered by lithium-ion batteries, which enable it to operate for 34 h
without being plugged in. The robots hydraulically driven tracks can still be
controlled manually, but are predominantly operated in automatic mode where the
tracks are steered precisely using an on-board control system. The drive system can
achieve a maximum speed of 5 km/h on flat terrain at a total robot weight of 1.4
tons. IFs on board computer runs a real-time enabled version of Linux (Xenomai),4
which allows for hard real-time data acquisition and processing, along with the
robot operating system (ROS)5. For the experiment described in this paper, IF was
equipped with a Hokuyo UTM-30LX-EW laser range nder mounted on the arms
end-effector and an Xsens inertial measurement unit, attached to the robots base
frame. Additionally, the robot was equipped with a vacuum gripper to pick and
https://xenomai.org/
http://www.ros.org/
210
K. Drfler et al.
place bricks, and a brick feeder on its back, which can carry 6 bricks at a time and
has to be manually lled.
3.3
Computer Architecture and Communication
The high level planning of fabrication tasks, such as the sequencing of the robots
positions and brick laying procedures, and computing the arm and gripper commands, is implemented within the architectural planning tool Grasshopper
Rhinoceros (Fig. 2).
A custom TCP/IP implementation allows the online control of the robots arm
and base. Commands are sent through a Python interface within Grasshopper to the
robots ROS nodes for base movement, as well as to the ABB Robot Control
Software for arm manipulation procedures. In return, all state and sensor data needed
within the high-level planning tool before and during construction is received within
Grasshopper. Generally speaking, the robots setup allows for feedback loops at
multiple levels of the system. All time-sensitive tasks are executed by control loops
running on the robots low-level computer and the ABB controller, to control base
and arm motion, respectively. The control of the overall building process, which is
much less time-sensitive, is closed via the architectural planning tool.
4 Experiment
This section details an initial experiment performed with IF, in which a dry stacked
double-leaf brick wall is constructed in between two pillars. The material system
consisting of discrete building elements and simple assembly logicsis specically
chosen in order to be able to solve basic problems of adaptive control strategies,
construction sequencing and repositioning operations of IF, while still being able to
subdivide the sequential building process into discrete production steps. (Note that
while a more elaborate hardware setup could have been employed to use adhesive in
between the bricks, or also to avoid the manual placement of bricks in the feeder on the
robot, this was not done because these tasks did not t the main goals of the experiment.)
4.1
Adaptive Building Process
The building process begins once IF is moved to the construction site. (Note that in
this initial experiment, the robot was positioned manually via a remote controller.
While the robot has all of the sensing and computing capabilities for autonomous
navigation on-board, the development of the autonomous navigation capabilities
211
Fig. 3 Workspace geometry matching functionality of IF: the geometric description of existing
structures within the building space (here: blue pillars and brown floor) is t to a point cloud,
captured by the robot when moved to the construction site. Notice how the brown plane, signifying
the floor, initially lies above the scan points on the ground (left), but ts into the points after
matching (right)
Fig. 4 A brick walls geometric description is adjusted to the real-world sensor measurements of
the robot. A mesh relaxation algorithm is used to align the individual building blocks orientation
and position with respect to the true location of the pillar, as well as to level the spacing between
the single bricks
required is left as future work.) At this time, it takes a 3D scan of its surroundings,
which serves as a reference scan for the robots localization in space (Sect. 4.2).
Additionally, this scan is used to locate the true positions of key features of the
working environment. These key features identify the interfaces to which the
structure being built must attach (Fig. 3). This information is then fed back to the
architectural planning tool (Grasshopper), and is integrated as a parameter into the
generation of the walls geometric description (Fig. 4). Since the true dimensions of
the construction site generally deviate signicantly from the ideal dimensions of
building plans, it is important to consider these inaccuracies before starting the
construction.
212
K. Drfler et al.
Fig. 5 The wall is stored as a graph, within which each brick is a node. Single patches of bricks to
be assembled are generated from this graph, according to the location of the robot and the
reachability constraints of the robots arm. After the construction of each patch of bricks, the robot
needs to change its position, localize itself and continue the construction from its new location
As soon as the building environment is properly identied and the walls precise
geometry is dened, IF is moved to the rst position required for building. When in
the desired building position, it needs to localize itself (Sect. 4.2) and communicate
its precise position to the high-level planning tool within Grasshopper. Within the
planning environment, the location information, along with the robots reachability
constraints, is used to determine a patch of bricks to be built (Fig. 5). At this point,
the robot can begin with fabrication.
As soon as IF has placed all bricks within its reach, it is moved to a new position.
There it scans, localizes itself, and builds another patch of the structure. This
process then continues iteratively until the structure is completed (Fig. 6).
4.2
On-Board Pose Estimation
In order to build with high accuracy on the construction site, the robot needs to be
aware of its position with respect to its work-piece. Because one goal of IF is to
avoid dependence on external sensing systems, this means that the robot must be
able to localize itself in its surroundings using on-board sensing and computing.
For this experiment, the primary sensor used for localization is a
laser-range-nder, mounted on the end-effector of the robots arm. By executing
sweeping motions with the arm, 3D scans of the robots environment are generated.
213
Fig. 6 Simulation of one possible building sequence of the walls construction: The double-leaf
brick wall with the dimensions of 6.5 m length and 2 m height, consisting of 1600 bricks, is
subdivided to be fabricated from 15 different positions of the mobile robot
Fig. 7 Robot localization: Point clouds before (1) and after registration (2). Point cloud
registration is used to nd the relative transformation from the current robot position to the
reference robot position
Point cloud registration is then used to nd the relative transformation from the
current robot position to the reference robot position (Fig. 7). Non-linear least
squares optimization performed using Googles Ceres Solver (Sameer 2015) is used
to nd the relative transformation required to minimize a measure of point cloud
quality between the measurement and reference point clouds (Maddern et al. 2012).
This registration method requires no reference markers to be placed on the construction site a priori, makes no assumptions about the structure of the robots
surroundings, and is not severely impacted by objects that move within the site
during building. For these reasons, this method should generalize to a wide variety
of construction environments.
214
K. Drfler et al.
Fig. 8 Images taken from a time lapse video of the building process. https://youtu.be/
loFSmJO3Hhk
Initial experiments showed that, with only static loading, the robots base tilts up
to 2 when the arm reaches far from the base. This can result in end effector
positioning errors of up to 70 mm. To compensate for this base tilting, an inertial
measurement unit is used to continuously measure the orientation of the base. This
information is then used to adjust the target end effector position while the arm is
reaching to place a brick.
4.3
Experimental Results and Validation
The designed double-leaf brick wall was successfully constructed using the process
described herein, requiring the robot to be repositioned 14 times. An average
duration of 40 s for the picking and placing of a single brick was observed. While in
the scope of the experiment it was not possible to directly measure the localization
accuracy and the consistency between the nal built structure and the initial design,
the accuracy of the system was evaluated by measuring the position of placed bricks
relative to previously placed neighboring bricks. These relative measurements were
consistently within 3 mm of the value expected from the CAD model. It was also
observed that the bricks placed closest to the pillar opposite from where the robot
started building were within 7 mm of their expected position relative to that pillar.
This indicated that global localization and brick placement errors did not accumulate over the course of the building process, since rst, the pillar locations were
only determined and considered before any bricks were placed, and second, every
point cloud captured from a new location was always registered against the same
initial reference scan. IF was therefore successful in building a structure which was
aligned to existing features of the construction site using solely on-board sensing
and computation. It is important to note that, in the experiment presented, a specic
production sequence was not dened beforehand, but derived from the resulting
215
locations of the robot during construction: While the human operator navigated the
robot to an arbitrary location, the machine then identied and reacted to the
resulting location and continued building with no further human interaction (Fig. 8).
While the automated navigation and optimized production sequencing of the robot
is left for future development, the chosen strategy demonstrated a successful integration of human intervention and automated construction.
5 Conclusion and Future Challenges

The experiment presented in this paper demonstrates a signicant step towards
enabling the robotic construction of complex structures directly on the construction
site with minimal human intervention. As mentioned in the previous sections, the
continuous exchange of information between true measurements and the underlying
computational model allows for the compensation of material and process related
inaccuracies during fabrication (Sha et al. 2009). With respect to the mobility of the
machinery, production sequences can radically be redened, which allow for the
construction of continuous structures. These structures dont have to be discretized
into separate building components due to constraints prevalent in pre-fabrication,
but rather have to be redened in accordance with the fabrication logics of the
chosen material system, the mobile machinery and conditions on site. Eventually,
this will demand novel mobile robotic building strategies, not only to realize
complex design propositions directly on construction sites, but also to enable design
processes, whose formal language and constructive details comply with the fabrication logic of the respective machinery used.
Future research into in situ construction methodologies using IF will be focused
on moving towards a more fully integrated and continuous construction process,
aiming at simultaneous arm and track maneuvers and continuous location-aware
manipulation procedures. In this experiment, the robot base was driven and repositioned manually while the industrial robot arm was controlled from within
Grasshopper. In a next step, these separate processes need to be unied in a
whole-body control framework that allows to plan optimal, simultaneous base- and
arm motions. This will then open up the possibility to address open questions like
optimal building sequencing in terms of required energy or overall building time.
Finally, formal influences in the design vocabulary through structural and
process-related boundary conditions by using a mobile robot for fabrication need to
be investigatednot only in the context of their functional, but also in their aesthetic capacities.
216
K. Drfler et al.
Acknowledgements This research was supported by Swiss National Science Foundation through
the NCCR Digital Fabrication (NCCR Digital Fabrication Agreement #51NF40-141853) and a
Professorship Award to Jonas Buchli (Agreement #PP00P2_138920). The building ma-terial was
sponsored by Keller AG Ziegeleien. Special thanks also go to the lead technician of NCCR Digital
Fabrication and photographer Michael Lyrenmann, as well as the project leader of IFs predecessor
dimRob, Dr. Volker Helm.
References
Andres, J, Bock, T, Gebhart, F, and Steck, W 1994, First Results of the Development of the
Masonry Robot System ROCCO: a Fault Tolerant Assembly Tool, in Chamberlain, DA (ed),
Automation and Robotics in Construction Xi, Elsevier, Oxford, pp. 87-93.
Balaguer, C, and Abderrahim, M 2008, Trends in Robotics and Automation in Construction,
INTECH Open Access Publisher.
DeSouza, GN and Avinash, KC 2002, Vision for mobile robot navigation: a survey, 24 IEEE
Transactions on Pattern Analysis and Machine Intelligence.
Feng, C, Xiao, Y, Willette, A, McGee, W, and Kamat, V 2014, Towards Autonomous Robotic
In-Situ Assembly, Sydney: ISARC - The 31st International Symposium on Automation and
Robotics in Construction and Mining.
Gambao, E and Balaguer, C 2002, Robotics and Automation in Construction, IEEE Robotics and
Automation Magazine, vol.9, pp. 46.
Gramazio, F, Kohler, M, and Willmann, J 2014, The Robotic Touch. Park Books. Zurich.
Helm, V 2014, In-situ-Fabrikation: Neue Potentiale roboterbasierter Bauprozesse auf der
Baustelle. Dissertation. Kln: Kunsthochschule fr Medien Kln.
Helm, V, Ercan, S, Gramazio, F and Kohler, M 2012, Mobile Robotic Fabrication on Construction
Sites: dimRob. IEEE/RSJ International Conference on Intelligent Robots and Systems,
Vilamoura, Algarve.
217
Helm, V, Willmann, J, Gramazio, F, and Kohler, M 2014, 'In-Situ Robotic Fabrication: Advanced
Digital Manufacturing Beyond the Laboratory' in Rohrbein, F, Veiga, G and Natale, C (eds),
Gearing Up and Accelerating Cross-fertilization between Academic and Industrial Robotics
Research in Europe, Springer International Publishing. vol. 94, pp. 6383.
Maddern, W, Harrison, A and Newman, P 2012, Lost in Translation (and Rotation): Fast Extrinsic
Calibration for 2D and 3D LIDARs. Minnesota: IEEE International Conference on Robotics
and Automation (ICRA).
Pritschow, G, Dalacker, M, Kurz, J, and Gaenssle, M 1996, Technological aspects in the
development of a mobile bricklaying robot. Elsevier, Warsaw, Poland, vol.5, no. 1, pp. 313.
Sameer Agarwal, KM 2015, Ceres Solver, http://ceres-solver.org Accessed June 12, 2015.
Seward, DW 2002, Automating the Construction Workplace: Positioning and Navigational,
Construction Innovation, vol.2, no.3, pp-167189.
Sha, L, Gopalakrishnan, S, Liu, X, and Wang, Q 2009, 'Cyber-Physical Systems: A New Frontier',
in Yu, P and Tsai, J (ed), Machine Learning in Cyber Trust: Security, Privacy and Reliability,
Springer, New York, pp. 3-13.
Closeness: On the Relationship

of Multi-agent Algorithms and Robotic
Fabrication
Abstract This paper demonstrates the effect of feedback between algorithmic,

robotic and material behaviors on the emergent formal character of several recent
design projects. These projects demonstrate a progression from single step linear
feedback between fabrication and simulation constraints to the attribution of new
material agency through real-time and recursive feedback between multi-agent
behaviors and physical material. We present a prototype robotic control system and
methodology that allows design to take place in and on an object rather than in its
anticipation, and we speculate on the implications for generative design and robotic
fabrication.
Stigmergic fabrication
Multi-agent algorithm Autonomous robotics
1 Introduction
The development of robotic fabrication strategies within architecture has matured
sufciently to enable the discourse to shift from predetermined operations and tool
paths to an emerging interest in real-time feedback and rule-based autonomous
operations. While a signicant body of literature documents the value and role of
autonomy in robotic machining processes, the design agency of robotic behaviors
has experienced relatively limited practical investigation. We posit a strategy for
encoding architectural design intention within robotic behaviors as an extension of
multi-agent generative design processes. Intrinsic to this position is an argument for
R. Snooks (&)
RMIT University/Kokkugia/Studio Roland Snooks, Melbourne, Australia
e-mail: roland@kokkugia.com
G. Jahn
RMIT University/Elseware Collective, Melbourne, Australia
e-mail: gwyllim.jahn@rmit.edu.au
219
220
R. Snooks and G. Jahn
an ontological closeness between physical and digital material, robot and computational agent, design and fabrication. Our research explores the emergent characteristics of form and articulation generated through varying degrees of feedback
between robotic fabrication and multi-agent generative algorithms in several
recently completed projects.
The conceptual domain of autonomous fabrication and distribution of design
authorship has perhaps been most vividly established by Francois Roche through
projects such as Ive Heard About or the FRAC Orleans proposal (Corbellini 2009).
In these speculative projects a simulated feedback between robotic behaviors and
material results in the compression of design and fabrication decisions into a single
process without an a priori model. Within industry focused or more applied
domains, pragmatic real-time sensing and feedback is enabling the necessary
control and accuracy to deal with material and fabrication tolerances exemplied in
the Stratications project of Gramazio and Kohler et al. (2014).
2 Material Agency and Feedback

The shift away from top-down hierarchical approaches in favor of attributing design
agency to feedback between material behavior and robotic operations is embodied
in the recent work of architects such as Del Campo (2014). Del Campo (2014)
describes his research as moving away from optimization and efciency as the
primary drivers of digital fabrication in pursuit of a model where materials assume
maximum agency in the fabrication process. Menges (2011) has argued that
embedding material characteristics, manufacturing constraints and assembly logics
allows a design to be driven through intrinsic performative capacities rather than
through hierarchical relationships that prioritize form over materialization. While
the discussion of the agency of human, material, digital or robotic design behaviors
is perhaps becoming increasingly common within this broader context, our concern
for agency and the nature in which it is dened in this paper has developed out of
the computational design processes that draw on the logic of swarm intelligence and
operate through multi-agent algorithms.
The approaches posited in the following three projects range from single step
linear feedback between fabrication and simulation constraints, to recursive feedback between goals and behaviors within multi-agent systems and material phenomena. The pattern of silicon inlay in the Composite Wing project is generated
through a multi-agent algorithm that is conditioned by structural performance prior
to being robotically extruded. The limitations of robotic rod bending are encoded
directly as agent behaviors within the Brass Swarms generative algorithm.
Through the development of real-time computer vision and robotic control systems,
the Feedback Deposition Studies explore the notion of an ontological closeness
between robotic and multi-agent behaviors.
Closeness: On the Relationship of Multi-agent
221
3 AgentBodies
Multi-agent algorithms such as Craig Reynolds (1987) Boid algorithm consider the
agent to be a point within Cartesian space. The multi-agent algorithms developed
for the Composite Wing and Brass Swarm projects draw on the self-organizing
logic of swarm intelligence and embed the agents within hierarchical structures
described as agentBodies. The conceptualization of the agentBody resembles the
logic of ant-bridges, where it is the interconnected geometry of the ants bodies that
forms structural or architectural matter (Fig. 1). Such connections emerge through
feedback between the behaviors of neighboring agentBodies within the generative
algorithm. When these behaviors are conditioned by structural forces, such as
deflection and bending moments on a ber-composite surface, the patterns generated by the agentBodies are a negotiation between behaviors designed to generate
emergent patterns and those resisting structural load.
4 Project 1: Composite Wing

Composite Wing is part of an ongoing series of prototypes that explore the design of
ber-composite surfaces through the use of multi-agent algorithms. Composite
Wings translucent ber-composite structure is embedded with an intricate inlay of
vein-like bodies that grasp, intertwine and disperse (Fig. 2) in a complex interplay
of structural and expressive concerns.
Each agentBody is a unique shape that emerges from repeated interaction with
other agentBodies and local structural conditions, and as a result the bodies inlayed
into the ber-composite were fabricated at two scales. Large-scale bodies are
robotically milled from high-density foam and provide the primary structure.
Fig. 1 AgentBodies are dened by strings of agents that respond to their context
222
Fig. 2 Composite Wingrobotically fabricated inlay within a composite surface
Small-scale bodies are robotically extruded from silicon and provide local
stiffening by increasing the structural depth and imparting a corrugation, or
micro-beam, in the surface (Fig. 3). We have explored two approaches to robotically extruding inlay within composite surfaces: extruding rigid thermoplastic onto
molds, and extruding flexible silicon onto malleable meshes.
The thermoplastic is extruded using a plastic extruder originally developed for
welding plastics. The proprietary motorized screw based end-effector enables a
consistent and reliable inlay, however, the rigid nature of the plastic requires that it
is extruded directly on a mold. This was prohibitive for the Composite Wing project
as the inlay was being extruded concurrently with the mold being used in a separate
factory for laminating the berglass. Consequently a pneumatic silicon extruder
was developed to extrude onto a flat mesh, which could then be laid into the curved
mold. The extruder was mounted to a KUKA KR150 robot on a 4 m linear track
enabling large surfaces of the agentBody pattern to be extruded. The speed and
Fig. 3 Left Plastic extrusion onto a mould. Right Silicon extrusion onto a mesh
223
precision of the robotic extrusion process allowed the highly differentiated pattern
of the agentBodies to be realized. However the relationship between these two
operations is linearone enables the other. Without feedback between the computational and robotic these two realms do little to expand the space of possibility of
the other.
5 Project 2: Brass Swarm

As opposed to the linear relationship established in the Composite Wing, Brass
Swarm encodes constraints developed through collaborative robotic rod-bending
techniques to construct the complex and highly volatile geometries of a multi-agent
systemthus establishing a feedback between fabrication limitations or behaviors
and generative design procedures (Fig. 4).
The rod-bending technique, which utilizes two KUKA industrial robotic arms, is
constrained by factors such as a minimum length between bends, a maximum bend
angle and in some cases the relative direction of subsequent bends. By encoding
these limitations in an existing set of multi-agent behaviors, robotic fabrication
constraints interacted with a broader set of design concerns intended to drive the
formation of pattern and geometry within the project.
The generative design behaviors of the Brass Swarm are primarily concerned
with the formation of a coherent manifold surface topology, the interaction of the
limbs of agentBodies, and the associated emergent characteristics of their form. The
more pragmatic constraints of the rod-bending process are intended to condition
Fig. 4 Brass Swarm
224
this emergent outcome rather than drive its formation. These constraints can be
classied as those derived from the empirically tested constraints of the robotic
bending process, and those, which are informed by the topology of the assemblage.
The robotic constraints relate to the entire rod, but operate computationally on each
agent-based control-point of the rod. The brass rods all have a standard length, so
while the bodies digitally stretch and deform to interact with their neighbors, a
pragmatic behavior operates to maintain the length of individual rods. The size and
shape of the robot grippers determine the minimum distance between bends as well
as maximum angle of any given bend. These constraints are translated into
behaviors, which influence the interaction of the control-point agents.
6 Collaborative Rod Bending

Several methods for rod bending were developed using two KUKAAgilus KR10
R1100 SIXX robots. The interaction of the robots was programmed in KUKA|prc
with the use of KUKA RoboTeam synchronization functions linking the master and
slave robots. The rst method rotates the rod to a consistent bending axis enabling it
to bend in any plane. The second method, which we describe as shear-bending,
creates two bends in a single action (Fig. 5).
These bends can be non-planar without a separate rotation operation. This is an
efcient and fast method avoiding the constant rotation of the rod for each bend,
typical of many bending procedures. This technique, however, is limited to making
pairs of bends in opposite directions. Thus a Z shape can be bent, but not a C shape.
The process is fully-automated, with one of the robots picking up a rod from a
pre-labeled array of rods.
The constraints on the geometry imposed by the limitations of each robotic
bending technique were incorporated within the design of the agentBody as well as
its algorithmic behaviorcreating a single step feedback between robotic operations and the generative algorithm. While fabricating all of the rods with unique
Fig. 5 Robotic shear-bending of brass rods with two KukaAgilus robots
225
bends is efcient (all 300 rods were bent in one day), the manual assembly of these
parts is labor intensive and difcult. The advantage of using industrial robots for rod
bending (as opposed to standard CNC rod bending machines) is not in the bending
operations themselves, but in the potential to combine fabrication and assembly.
The robotic positioning and welding of rods aided by sensor feedback has been
demonstrated by Dave Pigram and Wes Mcgee (2014), however the application of
this to highly complex assemblies such as the Brass Swarm will require a
sophisticated vision system and complex assembly planning and approach paths.
7 Constraints and Character

The topology or connectivity of the network of agents is important in understanding
their stability and structural integrity. Topological descriptions of the network
(Fig. 6) provide a tool for global analysis of what is an inherently globally ignorant
generative system.
A graph is used to describe and interrogate the connectivity of the bodies, which
enables structural analysis as well as an understanding of connectivity (which
influences stability) to be passed back to the individual bodies and influence their
behaviors through a set of structural heuristic behaviors. Thus the character of the
Brass Swarm is conditioned by both fabrication constraints that are localized to a
single agentBody, and global constraints that are an emergent property of the form.
By exploring alternative algorithmic behaviors capable of satisfying these local and
global constraints the emergent character of the resulting designs can exhibit radical
and unexpected formal novelty.
The drawings in Fig. 7 demonstrate the effect of an alternative set of robotic and
material constraints upon a similar self-organizing agent system. An arbitrary initial
distribution of square-proled rods tends towards woven and knotted congurations
when rods are allowed to twist and are conditioned to form pinned in-plane joints.
This conditioning of generative behavior is a form of design agency. As a
Fig. 6 Brass Swarm robotic behavioral constraints
226
Fig. 7 Knotted rods generated through the constraint of a pinned in-plane joint
consequence the robot influences the formation and contributes to the character of
the project. This difference is evident through a comparison of the Brass Swarm,
which privileges alternating bends (capable of being shear-bent), and the
rod-twisting experiments. Brass Swarm maintains a strong directionality and relatively even eld, while the rod-twisting assemblies rotate and knot to enable planar
connections (Fig. 7). So while these behaviors may be intended as pragmatic
conditions, they are instrumental in establishing the highly expressive and often
unanticipated characteristics of the project.
8 Agent Based Robotic Control Paradigms

The third approach posited in this paper to the relationship of the agent and the
robot is explored through a workflow that integrates feedback between material,
sensor, agent and robot. This workflow is tested through a series of short experiments. Feedback collapses the sequential relationship between these processes and
establishes a closeness of behaviors. What is signicant is that the robotic, material,
and computational processes can now run concurrently, enabling feedback to
become intrinsic to the ontology of the computational modelthe robot and the
agent are polymorphic, as are the physical and digital models of materiality.
A typical workflow for programming KUKA robots is a linear sequence of translations from a desired 3D model, to toolpaths, to robotic instructions written as a
linear sequence of KRL commands. To establish real-time feedback we have
developed a workflow in which the KUKA robots send commands at 4 ms intervals
through a custom server that operates between the design software
(Processing/Java) and the robot. This server handles path-planning through
KUKA RSI based on design responses from Processing and relays the robots
position and orientation back to Processing.
227
9 Project 3: Feedback Deposition Studies

To demonstrate a generative approach to real-time feedback between material, robot
and computational agents, we have undertaken an initial pilot project. This workflow links a vision system (Microsoft Kinect), agent design behaviors
(Java/Processing), real-time robotic control (RSI/RSI Server), and a volatile
material deposition end-effector (polyurethane foam). A computational agent is
used to navigate a scanned point-cloud of the current state of the deposited foam
(Fig. 8). A series of design behaviors influence the path of the agent, which is
referenced as the target for the robot and its further deposition of foam.
To explore feedback between agent behaviors and robotic material deposition
several algorithms were developed to nd peaks or valleys within a snapshot of the
point cloud scan. This iterative re-forming of the surface of the foam generated a
stigmergic interaction in which the volatility and behavior of both the polyurethane
foam and computational agent created a negotiated form (Fig. 9). The closeness of
Fig. 8 Processing screenshots of Kinect point clouds showing robotic deposition
Fig. 9 Feedback Deposition Studies. Polyurethane foam is extruded in response to the existing
deposition of foam based on real-time agent behaviors
228
the behavior of an agent within a digital environment, and the behavior of the robot
in a physical environment frees the digital from any concern for modeling the
physical. The computational agent simply responds to the physical and doesnt
require any encoded knowledge of material. Inherent within this process is a
principle of anti-simulation, whereby physical and material phenomena are
observed rather than anticipated by the digital model.
10
Conclusion
This series of projects and experiments trace what began as an attempt to use
robotics to construct geometry generated through multi-agent algorithms, through
to positing the robot as the agent in a compression of design and fabrication. This
progression parallels a shift in emergent character from that of the algorithm
expressed within the Composite Wingto a negotiated character of the material and
computational agentevidenced by the Feedback Deposition Studies.
The experiments with stigmergic depositions embody an ontological shift from a
closeness of agent behaviors operating on simulated material to the anti-simulation
of robotic behaviors operating on physical material. Within this anti-simulation,
design takes place in and on the object rather than in anticipation of the object and
conceptually attributes algorithms, robots and material with equal design agency.
Explicitly engaging with material performance and behavior as design drivers is a
prevalent theme within the discourse of generative design. By contrast,
anti-simulation is not an attempt to digitally model material behavior in order to
anticipate known structural, material or formal constraints, but is instead predicated
on the attribution of new agency to material through real-time feedback between
digital and material agents. We demonstrate that such feedback gives rise to
material behavior within digital models, and emergent character within stigmergic
material depositions, without encoding the epistemology of these behaviors by a
human designer.
The ambition of continuing this research in the future, as an extension of the
trajectory of feedback approaches outlined above, is to explore further design
implications arising from the closeness of computation and material agency. We
speculate that this will open a space of experimentation around an engagement with
error, inaccuracy and unpredictability within the design. A situation where fabrication precision becomes irrelevant and instead the precision of sensing is critical to
the closeness of the digital and physical. This approach requires replacing empirical
testing and calibration with feedback loops that self-correct and self-stabilize over
time.
229
References
Corbellini,G 2009, Bioreboot: The Architecture of R&Sie(n), Princeton Architectural Press, New
York.
Del Campo, M 2014, Autonomous Tectonics 2, in Gerber, DJ and Ibanez, M (eds), Paradigms in
Computing: Making, Machines and Models for Design Agency in Architecture, eVolo Press,
New York.
Helm, V, Willmann, J, Gramazio, F and Kohler, M 2014, In-Situ Robotic Fabrication: Advanced
Digital Manufacturing Beyond the Laboratory, in Rhrbein, F, Veiga, G and Natale, C (eds),
Gearing Up and Accelerating Crossfertilization between Academic and Industrial Robotics
Research in Europe, Springer International Publishing, Cham.
Menges, A 2011, Polymorphism,Time+Architecture, no.06/2011, pp. 118123.
Reynolds, CW 1987, Flocks, herds and schools: A distributed behavioral model, ACM Siggraph
Computer Graphics, vol. 21, pp. 2534.
The SPIDERobot: A Cable-Robot System

for On-site Construction in Architecture
Abstract The use of robots in architectural construction has been a research eld
since the 1980s. Driven by both productive and creative concerns, different systems
have been devised based on large-scale robotic structures, mobile robotic units or
flying robotic vehicles. By analyzing these approaches and discussing their
advantages and limitations, this paper presents an alternative strategy to automate
the building construction processes in on-site scenarios. The SPIDERobot is a
cable-robot system developed to perform assembly operations, which is driven by a
The original version of this chapter was revised: The missing author names were added.
The erratum to this chapter is available at DOI 10.1007/978-3-319-26378-6_40
J.P. Sousa (&) C.G. Palop
Faculty of Architecture, University of Porto + CEAU/DFL, Porto, Portugal
e-mail: jsousa@arq.up.pt
C.G. Palop
e-mail: cpalop@arq.up.pt
E. Moreira A.M. Pinto G. Veiga
INESC TEC, Porto, Portugal
e-mail: eduardo.j.moreira@inesctec.pt
A.M. Pinto
e-mail: andry.pinto@fe.up.pt
G. Veiga
e-mail: germano.veiga@inescporto.pt
J. Lima
INESC TEC + Polytechnic Institute of Bragana, Bragana, Portugal
e-mail: jlima@ipb.pt
P. Costa P. Costa A. Paulo Moreira
INESC TEC + Faculty of Engineering, University of Porto, Porto, Portugal
e-mail: paco@fe.up.pt
P. Costa
e-mail: pedrogc@fe.up.pt
A. Paulo Moreira
e-mail: amoreira@fe.up.pt
231
232
J.P. Sousa et al.
specic Feedback Dynamic Control System (FDCS) based on a vision system. By

describing and illustrating this research work, the authors argue about the advantages of this cable robot system to deal with the complexity and the scale of
building construction in architecture.
Keywords Cable-Robot Spiderobot

cation Non-standard architecture
Automated construction Digital fabri-
1 Introduction
The use of robots in architectural construction can be traced back to the 1980s. By
then, robotic technologies were employed in Japan to introduce a high level of
automation not only in the factory but also in the construction site. However, the
efciency of such approaches still required a lot of manual work and design standardization (Cousineau and Miura 1998; Kolarevic 2001). As a consequence, such
on-site robotic systems didnt prove to be satisfactory and lost importance overtime.
In 2005 at the ETHZ, Gramazio and Kohler (2008) recovered the interest in
robotics with a stronger focus in enhancing creativity. When digital tools assisted an
unprecedented design freedom, it was crucial to nd the appropriate flexible
manufacturing technologies to materialize novel tectonic strategies. The architectural interest in robotics has then spread to other schools and research groups, and
motivated the creation in 2010 of the Association for Robots in Architecture.
Despite its success in other industries, the use of industrial robots still presents
some limitations when facing the scale and complexity of the building construction
industry. Its limited range of action and movement makes its application more
suited to prefabrication than on-site construction. Furthermore, while the factory
space provides a controlled and safe environment to work with robots, the accidental and weather-exposed conditions in the construction site sets a highly
unstable scenario to work with such machines. The adaptation of the industrial
robot for on-site construction is thus a complex challenge, so other robotic
approaches may be explored. By considering the move from the fabrication of
components to the construction of buildings, the next chapter surveys some of the
current trends facing automation and robotics in construction.
2 Robotic Systems for On-site Construction

Looking to introduce robotic technologies into the construction site, architects and
engineers have explored several strategies. This paper proposes to resume them
according to the following categories: large scale robotic structures; mobile robotic
units; or flying robots (Fig. 1).
233
Fig. 1 Three research approaches to on-site robotic construction: large scale robotic structures
(left); mobile robotic units (center); flying robots (right)
2.1
Large Scale Robotic Structures
This section comprises those approaches looking for augmenting the scale of
conventional structures and robotic setups to t them to the scale of the architectural
buildings. The robotic construction initiatives in the 1980s and 1990s in Japan were
similar to a big scaffolding structure, integrating robotic systems to perform different operations. Bock and Langerberg (2014, p. 98) refer to this concept as
Integrated Automated Construction Sites and describe them as partly automated,
vertically moving on-site factories providing shelter for an on-site assembly.
The WASCOR (WASeda COnstruction Robot) group and the Shimizu Corporation
were among the rst initiatives to promote this trend.
In a different way, Behrokh Koshnevis devised a large-scale gantry bridge
structure moving horizontally along two parallel lanes to support his Contour
Crafting technology (Koshnevis and Bekey 2002). At the University of Southern
California, the team conceived an automated system for carrying a material deposition nozzle to 3D print architectural buildings in a single-run. Resembling a big
CNC router, this type of structure concept was followed in other similar strategies,
like the D-Shape technology1 developed by Enrico Dini.
Despite the robustness and high operational capacities proposed by these systems, their real application in the construction sites is problematic. According to
(Gambao et al. 1999, p. 600), these big and heavy robots are difcult to transport
to the construction site, have some unsolved scientic and technical problems, and
need a very high investment.
2.2
Mobile Robotic Units
To overcome the stationary condition of industrial robots and cope with the large
size of building constructions, the placement of robots over mobile platforms has
been another research avenue. Started in 1992, the ROCCO (Robot Assembly
1
http://www.d-shape.com.
234
J.P. Sousa et al.
System for Computer Integrated Construction) project departed from the understanding that an articulated robot placed over a mobile platform (a lorry, a towable
platform or an autonomous mobile robot) results very appropriate for the assembly
tasks on a construction site (Gambao et al. 1999, p. 600). Besides introducing this
mobile condition, two large robotic arms with a higher payload and range of action
than conventional ones were developed to allow using larger and heavier material
blocks (e.g., in concrete and stone). The focus in automating construction tasks
based in repetitive operations had led to other similar initiatives, like the robotic
bricklayer S.A.M. (Semi-Automated Mason) developed by Construction Robotics
(Petters and Belden 2014).
With a deeper interest in addressing creative issues, Gramazio and Kohler initiated in 2011 a research line on In Situ Robotic Fabrication at the ETHZ. To assure
the adaptation to the continuous changing conditions, unpredictable events,
obstacles, and the activities and movements of people working on-site, they devised
a robotic arm mounted on a mobile unit integrating additional systems, like sensor
and scanning technologies and different end-effectors (Helm et al. 2012, p. 169).
This mobile robotic strategy is interesting to avoid the complex setup of heavy
large-scale structures. However, ground mobility still have to solve some technological challenges to overcome the unstructured constraints of the construction site
environments.
2.3
Flying Robotic Vehicles
The exploration of aerial modes of robotic construction is a recent research avenue

in the eld. Launched by Gramazio and Kohler in collaboration with Raffaello
DAndrea at the ETHZ in 2011, this strategy employs flying vehicles to manipulate
building components in the air, thus avoiding the problems of ground-based
mobility and the need for scaffolding or cranes (Wilmann et al. 2012). This
approach also considers the cooperation of several aerial robotic units to allow the
execution of different and synchronized building construction tasks.
This team rst demonstrated this approached in the Flight Assembled
Architecture installation at the FRAC Centre in rleans (France). A set of four
quadcopters lifted, transported and assembled a tower structure made out of 1500
lightweight foam modules (Wilmann et al. 2012). The research on Aerial
Constructions continued in other experiments, by testing the assembly of space
frame structures and also the erection of tensile structures (Mirjan et al. 2014).
The advantages of the aerial robotic construction applications promise an
unprecedented freedom in building construction, which can stimulate new ways of
thinking and designing architecture. However, this research eld is in an early stage
of development. The automation and cooperative control technologies, the energy
autonomy or the payload capabilities are some of the technology challenges to face
in the near future.
235
3 The Cable-Robot SystemSPIDERobot

Facing this trend of developing automated and flexible modes of on-site construction, this paper presents the research of an alternative technology based on a
cable-driven robot system (i.e., referred in this paper as cable-robots)the
SPIDERobot. Under development since 2013 at the INESC TEC and the FAUP
(Faculty of Architecture of the University of Porto), this approach is based on
similar systems developed in other elds, like the sports and entertainment industries, like the Skycam2 or the CableCam3. The transfer of such systems to a building
construction eld has the potential to overcome some of the problems found in the
examples discussed in the previous chapter. The next sections describe the principles of the SPIDERobot, as well as the rst functional prototype that was built and
tested to evaluate the concepts. This research is still in an early stage, with rst
results being discussed in the conclusion.
3.1
System Description
Cable-driven robots are automated systems where multiple cables are attached to a
mobile platform or end-effector. A positioning system controls the cables by
actuating the motors for extending or retracting the cables (Bosscher et al. 2007).
The overall system is thus relatively simple, which opens some interesting
advantages to use cable-robots for on-site construction applications.
When compared with other robotic construction systems like those presented in
Sect. 2, cable-robots are easy and inexpensive to transport, assemble and dissemble
on-site, due to the lightness of the cable-based system. Furthermore, the conguration of the cable-based structure allows the denition of larger translational
working spaces, which is decisive to face the scale of architectural constructions.
Unlike flying robotic vehicles, cable-robots can have much higher payloads and
work continuously by means of constant energy supply (i.e., avoiding the use of
batteries), while complying with the safety requirements. Despite these advantages,
cable-robots also present some critical features. The number and movement of the
cables can cause interference within the working space, and their force in the
downward direction is limited. The cable system also faces specic technological
challenges regarding the control of the precision due to the tension forces and some
elasticity of the cables. In this context, the SPIDERobot is a low-cost prototype of a
cable-driven robot developed to perform assembly operations in on-site construction scenarios (Fig. 2).
http://www.skycam.tv.
http://www.cablecam.com.
236
J.P. Sousa et al.
Fig. 2 SPIDERobot system: 1 frame; 2 vision system; 3 motors; 4 cables; 5 mobile platform with
the gripper (left), and the built prototype (right)
Its structure consists in 4 actuated cables, which are xed on the top corners of a
frame with 120 60 135 cm, and connected to a central mobile platform equipped
with a rotating gripper. The system conguration presents 4 degrees of freedom
(DOF), which comprise the xyz movements and the rotation angle around z-axis.
Regarding other cable-robot systems, the SPIDERobot presents some combined
distinct features. By using only 4 cables, it reduces the possibility for cable interference with obstacles and leaves more useful working space than systems with
more cables, like the contour crafting robot system with 12 cables presented in
(Bosscher et al. 2007). However, because this option leaves the kinematics of the
robot under-constrained, the gravity force affects the cables tension and consequently the precision of the whole system (Moreira et al. 2015). To deal with this
situation, the large majority of cable robots use tensor-feasible controlling systems
for positioning the mobile robot in the workspace. In a different way, the
SPIDERobot presents a specic Feedback Dynamic Control System (FDCS) that
does not require sensors for measuring the cables tension. Instead, the proposed
FDCS control is based on a vision-based system, which can be something similar to
a differential GPS or laser measurement system on the construction site. By using
the information available in the environment, the FDCS controls the positioning of
the robot while assuring that the length of the cables is always within safe values.
3.2
237
Practical Experiment
The SPIDERobot prototype was tested in the assembly of an irregular structure

made out of 18 foam blocks with 120 60 30 cm. The design of the structure was
modeled in Rhinoceros with the goal of dening a geometry that could challenge
conventional modes of construction. Then, the different spatial coordinates and
orientation of each block (i.e., dened by the coordinates of two points) were listed
in an Excel le with the help of Grasshopper. This information was used to inform
the SPIDERobot about the position of the blocks in the structure (Fig. 3).
For picking them from the feeder site, the SPIDERobot took advantage of its
FDCS based on a vision system to automatically detect and recognize them in the
working space. With this feedback, the robot adjusted its height and orientation to
pick the blocks correctly (Fig. 4). With this kind of intelligent behavior, the
placement of the blocks in the construction feeder site does not have to be rigorous.
In the experiment, the blocks were placed in the feeder site in stacks up to 5 units.
The assembly of the 18 blocks was completed in around 16 min. The whole
process was slow, but revealed to be accurate (Fig. 5).
Fig. 3 Design model of the structure used in the experiment (left), and spatial positioning
information of the blocks (right)
Fig. 4 Sequence showing the automatic rotation of the gripper with the help of the FDCS, when
picking the blocks from the feeder site
238
J.P. Sousa et al.
Fig. 5 Sequence of photos showing the SPIDERobot assembling the structure
4 Conclusion
This paper presented a cable-robot system as an alternative strategy for automating
the on-site construction in architecture. Moved by design creativity concerns, the
authors tried to overcome some of the technological, physical and economical
limitations presented in other research approaches based on large-scale robotic
structures, mobile robotic units or flying robotic vehicles. The current stage of the
research work converged in the production of the SPIDERobot prototype, and it can
open the discussion both on a technological and on an architectural level.
On the one hand, in analytical studies conducted by the authors (Moreira et al.
2015), the results demonstrated that the topology of the FDCS implemented in the
SPIDERobot revealed to be more accurate than the traditional force-feasible
approach to the kinematics of cable-robots when performing pick-and-place operations. Therefore, the FDCS proved to be a promising system to be scaled to the
size of real construction environments (Moreira et al. 2015). Currently, the research
work is already centered in further exploring the vision system of the FDCS to
expand the autonomous capabilities of the system, like in the automatic detection
and avoidance of obstacles. Future research directions will be focused in scaling-up
the prototype, rening the vision-system and improve the robustness and speed of
the motors.
On the other hand, the SPIDERobot has the potential to challenge the traditional
concepts of designing and building in architecture. By considering its 4 DOF and
the geometric conguration of the 4 cables, architects can incorporate such
parameters in the creative process to drive design customization possibilities
towards aesthetically convincing and functionally efcient buildings. Indeed, unlike
other robotic approaches, the cable-robot system is not limited to assist the prefabrication of building parts (e.g., brick walls). One of its great promises is in the
on-site construction of buildings through the assembly of pre-fabricated building
parts or the stacking of building units (e.g., prefabricated housing modules), like in
239
high-rise building solutions developed in the FCL design studio in Singapore

(Budig et al. 2014). Furthermore, the simplicity and flexibility of the system also
facilitate its integration in both empty construction sites (e.g., with the help of
cranes) and in highly dense urban scenarios (e.g., by taking advantage of existing
buildings to set up the cable system).
In conclusion, the exploration of cable-driven robots can be an effective solution
for stimulating design creativity and expanding digital fabrication processes to the
realm of digital construction in architecture. Its application in practice can also
foster the vision of different and complementary robotic construction technologies
cooperating in the on-site construction of architectural buildings.
Acknowledgments This work was developed in the scope of the Research Project with the
reference PTDC/ATP-AQI/5124/2012, funded by FEDER funds through the Operational
Competitiveness ProgrammeCOMPETE and by national funds through the FCT (Foundation for
the Science and Technology).
References
Bock, T and Langenberg, S 2014, Changing building sites. Industrialization and Automation of
the Building Process, in Gramazio, F and Kohler, M (eds), Made by Robots - Architectural
Design, John Wiley & Sons. no.229, pp. 8899.
Bosscher, P, Williams, RL, Bryson, LS, Castro-Lacouture, D 2007, Cable-suspended robotic
contour crafting system, in Automation in Construction, May 2008, vol.17, no.1, pp. 4555.
Budig, M, Laurer, WV, Petrovice, R and Lim, J 2014, Design of robotic fabricated high rises, in
McGee, W and Ponce de Leon, M (eds), Robotic Fabrication in Architecture, Art and Design
2014, Springer, Michigan USA, pp. 111130.
Cousineau, L and Miura, N 1998, Construction Robots: The Search for New Building Technology
in Japan, ASCE Press.
Gambao, E, Balaguer, C and Gebhart, F 1999, A Robotic System for Automated Masonry, in
Automation and Robotics in Construction XVI, pp. 509602.
Gramazio, F and Kohler, M 2008, Digital Materiality, Lars Muller Publishers, Basel.
Helm, V, Ercan, S, Gramazio, F, Kohler, M 2012, In-Situ Robotic Construction: Extending the
Digital Fabrication Chain in Architecture, in Synthetic Digital Ecologies: Proceedings of the
32nd Annual Conference of the Association for Computer Aided Design in Architecture
(ACADIA) 2012, San Francisco, pp. 169176.
Khoshnevis, B and Bekey, G 2002, Automated Construction using Contour Crafting Applications
on Earth and Beyond, Proceedings of the 19th ISARC, Washington, USA, pp. 489495.
Kolarevic, B 2001, Digital Fabrication: Manufacturing Architecture in the Information Age, in
Jabi, W (ed), Reinventing the Discourse, Proceedings of the ACADIA 2001 Conference,
Washington DC, pp. 268277.
Mirjan, A, Gramazio, F and Kohler, M 2014, Building with flying robots, in Gramazio, F, Kohler, M
and Langenberg, S (eds), FABRICATE: Negotiating Design and Making, Zurich, pp. 266271.
Moreira, E, Pinto, AM, Costa, P, Moreira, AP,Veiga, G, Lima, J, Sousa, JP and Costa, P 2015,
Cable Robot for Non-Standard Architecture and Construction: A Dynamic Positioning
System, in Industrial Technology (ICIT), 2015 IEEE International Conference, pp. 31843189.
Petters, S and Belden, R 2014, SAM, the robotic bricklayer, in SMART / Dynamics of Masonry,
vol.1, no.4, pp. 1014.
Wilmann, J, Augugliaro, F, Cadalbert, F, DAndrea, R, Gramazio, F and Kohler, M 2012, Aerial
robotic construction towards a new eld of architectural research, in International Journal of
Architectural Computing, vol. 10, no.3, pp. 439459.
Developing Architectural Geometry

Through Robotic Assembly and Material
Sensing
Kaicong Wu and Axel Kilian
Abstract Advances in robotic fabrication and computational geometry have opened

up new possibilities for including robotic assembly and material selection into the
loop. We introduce a method for computing and constructing architectural geometry
through the negotiation between the design intention and the constraints of assembly
and materials. A small scale experimental structure has been modeled and partially
built from EPS foam sheets, using an industrial robotic arm to pick, cut and subsequently assemble the components of the structure. To reduce waste, a sensing procedure was developed to generate component based on the form of the found material
piece and t it in the existing structure, similarly to how the Caddisfly Larvae builds
its cocoon exclusively with found material. We aim to investigate how the sensor
enabled waste control can potentially adjust the form of the assembled structure.
Keywords Robotic assembling prototyping

Robotic fabrication
sensing Waste control Architectural geometry
Material
1 Introduction
One of the major trends of contemporary architecture is about free forms, which
triggers many geometric problems that are collectively called Architectural
Geometry (Pottmann et al. 2014). According to Pottmann (2014), the discussion of
the related problems focuses on two main areas: rationalization- and
fabrication-aware design, which are also referred to as post-rationalization and
pre-rationalization (Kolarevic 2003). Fabrication-aware design as digital modeling
which automatically generates buildable formal solutions, poses more unsolved
problems (Pottmann et al. 2014). Robotic fabrication as one of the advanced
K. Wu (&) A. Kilian
Princeton University, Princeton, USA
e-mail: kaicongw@princeton.edu
A. Kilian
e-mail: akilian@princeton.edu
241
242
K. Wu and A. Kilian
prototyping methods provides potentials for nding formal solutions in this

research area. It has been used to demonstrate the advances in performing custom
fabrication such as wire-cutting (McGee et al. 2012), milling (Menges 2012) or
incremental-forming (Kalo and Newsum 2014).
Recently, increasing attention is being paid to robotic assembly research. For
instance, designers have used robotic arms to assemble custom brick walls
(Gramazio et al. 2010) and assemble on-site constructions1 or prototype tower
models (Budig et al. 2014; Gramazio et al. 2012). More assembly based research
projects have automated the construction of complex timber structures2 and roofs.3
This paper presents a robotic assembling prototyping method, in which fabricating
and assembling irregular components are controlled by sensor enabled material
selection. The form of the structure is modeled and constructed by iteratively
computing feedback from the negotiation between the design intention and the
constraints of robotic assembly and found materials.
2 Form, Assembly and Material

The form of a constructed structure is intrinsically linked with the assembling
process and the material it is built with. Construction by a Caddisfly larvae serves as
a precedent where a cocoon is assembled on shape recognition in found material
elements, and t into local context of the cocoon (Stuart 2000) (Fig. 1a).
A study of human assembly of laser cut parts varies this to precomputed and
custom cut pieces instead of found ones (Fig. 1b). Whereas constraints of assembly
and materials are critical to design and construction of architecture, here, a
Caddisfly larvae cannot customize found materials but instead develop an ability to
assemble the fragments by nding a tting position in context. In a context of
robotic fabrication, this approach can act as a valuable framework for formal
feedback in robotic construction and assembly.
3 Methodology: Robotic Assembly Prototyping

with Sensor-Enabled Material Selection
This methodology is used for a project with assembly based robotic fabrication
setup using EPS foam sheets with sensor-enabled material selection. The core
challenge is to achieve the integration of picking, cutting, and tting the components of a structure. In this research, a 6-axis ABB IRB 7600 industrial robot arm
was used to run the prototyping. The components are cut by hot wire from
1
www.gramaziokohler.arch.ethz.ch/web/e/forschung/273.html.
3
2
243
Fig. 1 a Diverse array of case morphologies among families of Caddisfly ( Alison Elizabeth Stuart
2000), b Human assembly based on pre-computed and custom cut pieces instead of found pieces
24 48 1 in. Expanded Polystyrene (EPS) Foam Sheets. Special end-effectors

with small diameters were designed to avoid collisions with the cutting tool and to
hold components of different sizes. Hot glue is used to quickly attach the components into tting positions.
3.1
Geometry and Material Constraints
Based on the constraints of the chosen EPS foam sheets, all the component
geometries need to be flat. The Tangent Plane Intersection Mesh developed by
Troche is used to generate a planar hexagonal mesh from a double curved surface
(Troche 2008). To develop the thickness of the components, all the components are
offset outwards with the thickness of the foam sheets (1 in.). The resulting offset
mesh of constant thickness cannot be a polygon mesh of equal valence (Pottmann
et al. 2007). To maintain the valence of the offset polymesh, the edge surfaces of
each component become twisted (Fig. 2).
Component sizes are based on the curvature of the design surface and assembly,
so that curvature details are maintained while the total amount of cutting lines is
optimized. In addition, size was further evaluated for pick-up limits of the
end-effector, and the fabrication constraints of the foam sheets.
3.2
Sensor Enabled Picking
Without sensing, the EPS foam sheet has to be prepared carefully and positioned in
the modeled orientation. As a result, scraps are created by cutting off undesired
foam. Sometimes the placement inaccuracies result in partial cuts. The total waste
consists of the scraps and the partial cut sheets. Reusing these oddly shaped pieces
makes positioning them manually much harder. Thus, Kinect sensing was used to
identify component orientation in the found material piece (Fig. 3). So a procedure
244
K. Wu and A. Kilian
Fig. 2 Curvature and components (left 2), polygon mesh offset outwards (center), project
geometry (right)
Fig. 3 Kinect sensing for manually identifying components orientation and size
was developed using computer vision to detect the shape of a scrap piece and
automatically determine the correct pickup position for a to-be-cut component. For
this, a Kinect sensor delivered data into Processing as two separated point clouds
using the SimpleOpenNI library4 and coding reference (Borenstein 2012).
For the end effector, the center of the point cloud is averaged to be at the end effector
center. For the foam, the point cloud of the central area of the top surface is averaged (to
reduce noise) to be the height of the pickup surface. Both the center of the end effector
and the projected point cloud of the foam are read by Grasshopper (GH) with the add-on
gHowl5 into a digital model with their coordinate system origin being the Kinect
position. Given the absolute coordinates of the end effector center in the digital model,
the material point cloud can be calibrated into the model space. From the point cloud, a
Laplacian Mesh is created by GH as the geometric domain of the found piece. Finally,
the component is parametrically oriented into the piece by aligning the longest edges.
3.3
Hot-Wire Cutting
For robotic fabrication, the edge surfaces of a component are sorted by segment order
of the inner polyline. Each edge surface is divided into several section planes and the
surface is oriented to align with the cutting tool by the planes to cut off the unwanted
4
www.code.google.com/p/simple-openni/.
www.grasshopper3d.com/group/ghowl.
245
part (Figs. 4 and 5). The project adopts here the Mussel add-on for Grasshopper
developed by Johns6 to generate Rapid Code to control the tool paths of robot arm.
3.4
Fitting to Existing Structure
The component remains attached to the end effector and is tted into the assembly
position until human operator lls hot glue into the gaps to attach it to the neighbors. A non-trivial problem is determining the collision free assembly sequence of
parts. The components are sorted by the height of their area centers to establish the
assembly sequence and to ensure that the arm will always approach the already
installed components safely from above. The cutting and assembly tool paths are
compiled into a program simulated in ABB Robot Studio.7 The robotic arm
effectively acts as a temporary scaffold to secure the new component in its correct
position. Once the new component is glued to the existing ones, the new component
becomes part of the structure and the arm can be removed (Fig. 6).
4 Formal Feedbacks from Waste Control

Fitting a found material piece least altered into the existing assembly is achieved by
orienting the geometry measured by Kinect in the parametric model. The current top
edge of the structure is modeled by isolating the boundary of the inner polygon of the
structure model. The lowest point on the top edge is located and the predicted
material is oriented by referencing to the two attached edges of the lowest point. The
oriented material is shifted to cover the lowest segment of the edge. It is reoriented to
be tangent to the design surface and is trimmed off by the top edge. The left over
geometry is extruded to the material thickness by referring to the vertex normal of
the polygon of the existing structure (Fig. 7). The generated component denes the
new tool paths and updates the existing model as the input of next generation.
4.1
Linking Formal Adjustment and Waste Control
Fitting the found material piece determines the generation of components and how
they are cut and t into the existing structure. By changing the tting strategy and
the form of the found piece, the structure is remodeled by the assembling of
components that are made within the material constraints. A relationship between
the formal adjustment and the sensor enabled waste control can be roughly
approximated (Fig. 8).
6
www.grasshopper3d.com/group/mussel.
www.new.abb.com/products/robotics/robotstudio.
246
K. Wu and A. Kilian
Fig. 4 Tool paths sequenced by moving foamsheets through hot wire
Fig. 5 Cutting components with a hot wire in different angels and the twisted edge surfaces
Fig. 6 Assembling components with the end effector and gluing for attachment
247
Fig. 7 The new components are generated from the geometric detection of found piece
Fig. 8 The amount of waste and area difference between the designed and modeled geometry. In
the waste control test, some scraps are reused so the waste is reduced by sensing
5 Conclusion and Future Work

This project has conceptualized and prototyped a robotic assembly method based on the
combination of design intention, the constraints of assembly and the constraints of
found materials, with several prototype studies. By robotically fabricating and
assembling irregular components from EPS foam sheets, the project built a link between
computing programs, fabricating tools and sensors with formal feedbacks. By detecting
found materials, fabrication was measured and modeled within the material constraints
in digital space to reduce fabrication waste and control assembling tool paths.
Future research includes several new steps: Firstly, the sensing technique can be
developed to track differences between the physical and the digital models, and
allowing for compensation. Secondly, assembly has been restricted to manual xing
components, which should be replaced by robotic xing. Thirdly, the sensing
procedure could check potential structural failures while the object is assembled.
Finally, a more robust and optimized computation is required to model the relationship between waste control and adjustment. Similar to the cocoon of Caddisfly
larvae, the assembly outcome will be different based on the material context in
248
K. Wu and A. Kilian
which it is built. Further material prototyping will deliver feedback for adjusting the
form of a structure. Yet the rst results discussed here on geometric detection of
found material piece, robotic assembly and minimizing waste are contributing to the
design of new methods for freeform architecture.
References
Borenstein, G 2012, Making things see: 3D vision with kinect, processing, Arduino, and
MakerBot: OReilly Media, Inc, OReilly Media, Maker Media, Inc, Sebastopol.
Budig, M, Lauer, WV, Petrovic, R and Lim, J, 2014,Design of Robotic Fabricated High Rises, in
McGee W and Ponce de Leon (eds), Robotic Fabrication in Architecture, Art and Design 2014,
Springer, Cham, pp. 111129.
Gramazio, F, Kohler, M and DAndrea, R 2012, Flight Assembled Architecture, Editions HYX,
Orlans.
Gramazio, F, Kohler, M and Oesterle, S 2010, Encoding material, Architectural Design, vol.80,
no.4, pp. 108115.
Kalo, A, and Newsum, MJ 2014, An Investigation of Robotic Incremental Sheet Metal Forming
as Method for Prototyping Parametric Architectural Skins, in McGee W and Ponce de Leon
(eds) Robotic Fabrication in Architecture, Art and Design 2014, Springer, Cham, pp. 3349.
Kolarevic, B 2003, Architecture in the Digital Age: Design and Manufacturing, Spoon Press, New
York, pp. 127140.
249
McGee W and Ponce de Leon (eds) Robotic Fabrication in Architecture, Art and Design 2014,
Springer, Cham, pp. 3349.
Menges, A 2012, Morphospaces of Robotic Fabrication, in Rob| Arch 2012, Springer,
Springer-Verlag, Vienna, pp. 2847.
Pottmann, H, Asperl, A, Hofer, M and Kilian, A 2007, Architectural Geometry. Bentley Institute
Press, Exton.
Pottmann, H, Eigensatz, M, Vaxman, A and Wallner, J 2014, Architectural geometry, Computers
and Graphics, vol. 47, pp. 145164.
Stuart, AE 2000, The Utility of Behaviour in Macroevolutionary Studies,Analyses of Caddisfly
(Trichoptera) Case Building Behaviour, Publisher, Location.
Troche, C, 2008, Planar Hexagonal Meshes by Tangent Plane Intersection, Advances in
Architectural Geometry vol.1, pp. 5760.
BotBar: A Platform for Multi-disciplinary

Design Education
Marjo Niemel, Samantha Horlyck, Susana Alarcon-Licona,
Dylan Wozniak-OConnor, Gabriele Ulacco, Rodney Watt
and Rob Saunders
Abstract The BotBar has been developed to respond to the signicant challenge of
integrating smart technologies and sensor loops with industrial robot arms. The
process has focused on the robot as an open design platform, utilized as a nexus for
education and collaboration between the disciplines of Architecture and Interaction
Design. This paper discusses the success and challenges that have emerged from
this project, while also documenting an interaction design studio that prototyped
sensor-based integrations with the BotBar.
Keywords Robotic education
Creative robotics
Multi-disciplinary design Smart technologies
M. Niemel (&) S. Horlyck S. Alarcon-Licona D. Wozniak-OConnor

R. Watt R. Saunders
Faculty of Architecture, Design and Planning, University of Sydney, Sydney,
Australia
e-mail: marjo.niemela@sydney.edu.au
S. Horlyck
e-mail: samantha.horlyck@sydney.edu.au
S. Alarcon-Licona
e-mail: susana.alarconlicona@sydney.edu.au
D. Wozniak-OConnor
e-mail: dylan.wozniak-oconnor@sydney.edu.au
R. Watt
R. Saunders
e-mail: rob.saunders@sydney.edu.au
G. Ulacco
AR-MA, Sydney, Australia
e-mail: gabriele.ulacco@ar-ma.net
251
252
M. Niemel et al.
1 Introduction
The BotBar is a flexible design project that aims at exploring the use of robots,
human-robot interaction, sensors and supplementary technologies by employing a
robotic bartender as a versatile and engaging teaching and training tool. Robotic
bartenders are not a novel concept nor is the use of technology, and this project
does not situate the creation of the BotBar itself as the end result. Instead, it
positions the BotBar as a useful platform with which to foster collaboration and
skill sharing between disciplines with a mutual interest in robotics. There are
diverse examples of the development and use of creative robotics throughout
industry and education; however, the introduction of industrial robot arms into an
education setting presents challenges regarding accessibility, integration and
knowledge propagation. The BotBar was developed to provide students and
researchers with the training and tooling to successfully overcome the challenge of
incorporating industrial robots in their projects. However, the BotBar is not only a
facilitation tool but a boundary object for design disciplines that can be repurposed
and expanded in a variety of directions (Fig. 1).
Within the broad disciplines of architecture and interaction design, robots are
increasingly being used in design education (Brell-Cokcan and Braumann 2013).
The BotBar aims to advance our understanding of the bridge between sensor loops
and the possibilities these offer to robotic fabrication and human-robot interaction.
Additionally, the BotBar aims to alter the customary view of an industrial robot arm
Fig. 1 BotBar 1.0 following a pre-programmed toolpath
BotBar: A Platform for Multi-disciplinary Design Education
253
from that of a labour machine to that of an autonomous social robot, which

responds to dynamic inputs, gives feedback, and interacts with humans. In their
paper The Framed Pavilion, Dank and Frieissling (2012) explore some of the
challenges of working on non-standard architecture and design projects involving
robotics, caused in part by the wide range of collaborators often involved. The
BotBar posits an alternative method of collaborating whereby all information is not
necessarily input to a parametric model. Instead, it aims to embrace the diversity of
collaboration and draws on varied inputs to create a series of intertwined feedback
loops: from loops between sensor input and real-time robot adaptive technology to
the larger conceptual loops of enabling creative collaboration and skill sharing. The
skills of those designers who work on a project such as the BotBar act as catalysts
which feed back into the project as improvements or variations upon the design and
function of the BotBar.
2 Function as Boundary Object in a Social Context

Robots are a prime example of boundary objects capable of connecting groups with
different backgrounds in the construction of new knowledge and new approaches.
As Pickering (1992) states, boundary objects aim to t local needs, to promote
shared coherent actions and knowledge and to allow parties to resist evident
translation and reconstruct methods. The concept of the bar as a complementary
boundary object makes this a successful exercise with potential to be expanded
but is not without its limitations. This project began as a training platform and
method to understand how to foster and manage shared work across diverse skill
sets which exist within many design and education contexts. Since then it has been
developed into a platform able to engage people from diverse technical backgrounds collaborating on a singular, approachable output which serves as a functioning interactive system to bring robotics to the wider community and highlight
human-robot interaction (Power 2014 Sydney Morning Herald.)
As a boundary object, the BotBar engages distinctive groups such as designers,
programmers, and architects with each others disciplines and skill sets while also
dening important reference points and tasks between each group. At the boundary
of these elds the robotic arm can function as a point of crossover between robotic
fabrication and the many potential uses within Interaction Design. The creation of
the BotBar has involved tasks ranging from design, fabrication, parametric modeling, motion programming and hardware integration; a critical part of the process
involves the input of skills and knowledge sharing from each discipline involved, as
well as negotiating the best approach from such a vast skill base. Advances in
technology, such as the use of industrial robot arms, present changes in the nature
of contemporary societies and consequently it is relevant for students and
researchers working in areas of designing, improving and constructing environments to experiment with and generate meaning from these technologies.
254
M. Niemel et al.
3 BotBar 1.0 and 2.0

This project has enabled students, academics and researchers from multiple disciplines to work hands-on in a safe environment with a KUKA KR6. The rst
iteration of the BotBar engaged architectural and interaction design researchers
through a division of tasks whereby the physical design and construction of the bar
was largely handled by the architecture discipline while the control interface (utilizing Python scripts to activate programs on the robot controller) was developed by
the interaction design discipline. However, motion programming with Grasshopper
and KUKA|PRCan integral part of the BotBar 1.0was a task both disciplines
could collaborate on.
The rst BotBar iteration followed a pre-programmed toolpath, the second
iteration has focused on exploring robot-human interaction, sensor input, and
social/experimental interfaces for real-time and pre-programmed processes in
addition to the use of industrial robot arms in challenging environments such as an
unstructured and crowded room. This social or public environment can present
difculties in using standard, accessible input sources such as Kinect motion sensors which may struggle with detecting multiple users. It also raises issues in
regards to safe operation where the robot may be more accessible than in a factory
setting. Particular attention was paid to designing the bar for safety without needing
to include barriers, and as such the second iteration of the bar ensures a safe
distance from the robot to the user by separating them with the structure of the bar
surface, which is greater than the reach of the robotic arm. It also incorporates
rotating bar sections controlled by Arduino to further the physical separation
between the work envelope of the robot and the users (Fig. 2).
The flexibility of the BotBar platform is in part that each of the areas explored in
each version can be developed, reiterated, and rethought by a design team who can
use their existing skill sets (such as parametric modelling, Arduino, Python, and
traditional fabrication) in an environment which fosters the cross-pollination of a
Fig. 2 Plan view of BotBar 2.0 with component locations and motion sequence of robotic arm
255
Fig. 3 Design flow of the BotBar 2.0 as intersection between physical and virtual organizations,
shared between Interaction Design and Architecture
broad range of skills, ideas, and knowledge between different disciplines (Fig. 3). In
addition, each area can scale from relatively simple designs to a level of complexity
that incorporates multi-process, multi-robot interactions. The BotBar is intended to
be used as an accessible system; by utilizing well documented and easily controllable components, such as Arduino powered motors and Raspberry Pi controlled
pumps, different elds can provide input to the project. Instead of controlling these
various components and the robot with a relatively unfamiliar language for many
users, such as KUKA Robot Language, control of the BotBar is handled through
the computer with the simplest tools available. This allows motion programming
with Grasshopper and KUKA|PRC to be used side by side with interfaces built with
Processing, and allows other disciplines or elds to contribute to the BotBar using
familiar computer or hardware based processes.
4 Challenges
Robotic arms, while increasingly commonplace in architecture and design faculties,
still present challenges for design educators as their inherent complexity of operation often renders them inaccessible to researchers and students who tend to work
within short project cycles (Schwartz 2012). Following a social constructivist
philosophy, where technology is seen as an integral part of society and properties
and effects are usually dened in a particular social context (Boudourides 2003),
the training and teaching model was shaped to accommodate the needs and skills of
our community and explore three signicant challenges.
256
4.1
M. Niemel et al.
Accessibility
Students and researchers have the opportunity to work in the further development of a
real application with areas of interest ranging from motion path programming to the
exploration of different interaction paradigms. The development of flexible training
platforms is one approach that has enabled students and researchers to engage with
robots on research-relevant projects by providing them with enough knowledge to
start using robotics as soon as possible in a safe and creative environment.
4.2
Knowledge
Experimental robot projects, such as the BotBar, aim to create, promote, and deploy
the training modules, teaching skills, and specialist knowledge required to successfully use industrial robotic arms. They do so by providing a space to explore
different approaches to hardware conguration, motion programming, programming languages, controller familiarity, and other complex tasks that can create
opportunities for innovation such as the implementation of custom end-effectors or
of communication between robots and external devices.
4.3
Integration
Different modes of operation are being explored to allow students to work on projects without requiring a comprehensive understanding of the complete system. We
have implemented the communication with different programming environments
(i.e., Python and Processing) as gateways to integrating supplementary tools that
support the background and skills (e.g. Arduino and Grasshopper) of the students.
5 Training and Teaching Programs Utilizing the Botbar

Utilizing the BotBar initially for a training program and then a teaching program
has seen the creation of several successful projects by research teams. The program
is structured in a way that introduces the fundamentals of working with a six axis
industrial robotic arm but quickly focuses on integrating individual research
interests, aiming to use industrial robots as open design interfaces with adaptable
tools to achieve a goal (Brell-Cokcan and Braumann 2013). This program begins
with an interactive hands-on tutorial with the robot that focuses on safe operation,
basic movement, six-axis familiarization, and an awareness of physical restrictions
such as axis limits. The program then moves to manual path programming, such as
a simple drawing program, before introducing software tools which enable complex
parametric control of the robot.
257
An understanding of the fundamental concepts of robot motion and control

empowers students and researchers to explore and partake in a wide range of robotics
projects. In order to encourage collaboration, propagate knowledge, and share skill-sets,
robotic researchers have formed three broad and overlapping research clusters:
Subtractive, Additive, and Smart Technologies. These clusters include robotics projects
and units of study across multiple disciplines and provide the opportunity for students
and researchers to interact with robots in a creative environment.
6 Botbar as a Concept to Explore Interaction Design

BotBar 1.0 raised attention and highlighted opportunities among different disciplines including design computing and interaction design. The second version of
the project was envisioned as a platform for student participation as Robots provide an embodiment and the ability to add social interaction to the learning context
(Mubin et al. 2013). By working with a standard robotic arm, interaction design
students were able to gain knowledge and inspiration from previous work in education, industry, and society.
With the aim of exploring human-to-robot interaction in a specic social setting,
students from the Master in Interaction Design and Electronic Arts were immersed
into a studio-based unit around the BotBar project (Fig. 4). This platform aimed to
develop possible scenarios where robots and humans interact in a harmonic and
collaborative way. These ideas were nourished by lectures encouraging user
research and context analysis, as well as tutorials adapted from the training
Fig. 4 Three interfaces developed to explore interaction with the BotBar in a social environment:
Pubpop QR code reading (left), Bounce Kinect-Processing code (top right), Faces tactile
interactive menu (bottom right)
258
M. Niemel et al.
programs. On the one hand, modules on Rhinoceros, Grasshopper, and KukaPRC

determine the basics of robot control and set up, as well as 3D modelling and
fabrication methods. On the other hand, sessions on Arduino and Processing
demonstrate the opportunities of external hardware and software integration.
In this study, where the initial objective was adapting the resources for interaction design students, a custom XML interface was implemented. This Processing
interface bridges the communication between robot and computer and allows users
to activate movements without the need of using the robot teach pendant. Students
were encouraged to explore their concepts through experimentation with a broad
range of technology and processes, generating interesting and diverse results:
Team 1/Pubpop implemented a webcam mounted to the robot arm as part of the
end-effector. Custom software and hardware were developed to allow QR codes
to be used to select drinks and to construct a tactile skill game that brings
together multiple participants. The focus of this project was on creating an
amusing social experience, where the robot acts as a facilitator and provides
feedback to the participants through its movements based on both the scanned
QR codes and the output of the tactile skill game.
Team 2/Faces developed a tangible tabletop interface as an integral functional
and aesthetic element of the BotBar design. The interactive menu uses tokens of
different colours to represent different ingredients and triggers corresponding
programs from the robot controller. In this case, the groups attention was
directed towards interaction innovation within the specic context.
Team 3/Bounce aimed to engage users in an interaction with the robot through a
dance battle. A Kinect sensor was used to facilitate the interaction, to capture
users movements and to provide information that generated feedback from the
robot. An important part of this exploration was the analysis of human to robot
movements and the implications of robot motion as a creative communicative
method. The design of the environment, including the lights and designated
interaction spaces were also incorporated into the project.
From a tactile interface triggering pre-programmed movement routines to a
Kinect sensor based interaction and custom robot motion study; the flexibility of the
BotBar has been as an open platform for exploration and knowledge acquisition in
different areas which complement the students education. The direct result has
been to highlight several areas for further development. In particular, the complexity of sensor integration in the BotBar will to continue to expand as it is
integrated into a wider range of future units of study.
7 Success and Limitations

The aim of this project was not merely to create a robotic bartender: this incarnation
of the project has been versatile enough to transition from a training program into a
teaching platform. However, increasing complexity will begin to reveal limitations.
259
The BotBar has been particularly successful in familiarizing fabrication staff,

students, and researchers with KUKA|PRC, end-effector design, and knowledge
such as path planning and singularity identication. In one semester the BotBar has
enabled the above documented studio of Master of Interaction Design students to
work with robotic motion programming and sensor technologies as applied to the
robot arm. It is anticipated that with more open tools to interact with the robot and
generate results quickly the BotBar will be used within undergraduate studios next.
The success of this program has also been in cultivating interest in robotics with
a human-centered scenario that is broad and relatable while still able to showcase
real outputs from multiple disciplines collaborating around a common object of
interest. The limitations of the project in its current form as a bartender are limitations of scale which affect architectural output. A future incarnation has been
proposed which would use the robot to self-assemble on-site, however with a 6 kg
payload and limited reach this may not be a feasible scenario. It may be necessary
to move away from the experience-focused output of a cocktail to a more formal
exploration of spatial interaction. The BotBar has proved itself to be an ideal
opening to this conversation, but in order to have professionals and students
working together on integrated projects that are of equal interest to architects and
interaction designers another evolution of the project is likely needed.
8 Conclusion
The breadth of engagement with a robotics program that used the BotBar as a catalyst
is encouraging. However, the future of the platform will be to develop an interface to
increase the accessibility of robot arms for students and researchers who lack programming skills. An API which uses a library of sensor inputs alongside a range of
end-effectors will foster further cross discipline engagement while allowing more
advanced interactions and programs to be explored intuitively and prototyped quickly.
This multi-disciplinary engagement has the potential to evolve further with the
robot continuing to act as a boundary object, located at the intersections between the
disciplines of interaction design and architecture, that can foster collaboration and the
intermingling of skill setsdemonstrating that since meanings are not embodied in
boundary objects, divergent uses, interpretations, and reconstructions are likely
(Pickering 1992, p. 174). In BotBar 1.0 and 2.0 the work was divided so that the
architecture discipline handled the design of the environment, the physical bar fabrication, the end-effector fabrication, and the automation processes while the interaction design discipline developed the interfaces and software bridges. In future
iterations of the BotBar, this divide needs to become more apparent and thus more
efcient or to be dissolved so that the tasks overlap and better foster research, training,
and education in creative robotics and fabrication in architecture and design.
260
M. Niemel et al.
261
Acknowledgments This research has been supported by the Faculty of Architecture, Design and
Planning, and has been developed at DMaF Lab. The authors would like to thank Interaction
Design students Duane Allam, Angela Graf and Guilherme de Macedo from Faces project;
Abhiruchi Chhikara, Qingwei Kong and Yonghan Ji from Bounce; as well as Yu Guan, Yan
Song and Dan Zhang from Pubpop, for their inspiring projects and their participation in this
research.
References
Boudourides, MA 2003, Constructivism, Education, Science, and Technology in Canadian
Journal of Learning and Technology/La revue canadienne de lapprentissage et de la
technologie, vol. 29. no. 3 Retrieved from http://www.cjlt.ca/index.php/cjlt/article/view/83/77.
Brell-Cokcan, S and Braumann, J 2013, Industrial Robots for Design Education: Robots as Open
Interfaces beyond Fabrication, in Zhang J and Sun C (eds), Global Design and Local
Materialization, Springer Berlin Heidelberg.
Dank, R and Freissling, C 2012, The Framed Pavilion, in Brell-Cokcan S and Braumann, J (eds),
Rob|Arch Robotic Fabrication in Architecture, Art, and Design, Springer-Verlag, Vienna.
Mubin O, Stevens CJ, Shahid S, Al Mahmud A and Dong, JJ 2013, A Review of the Applicability
of Robots in Education, in Journal of Technology for Education and Learning. vol. 1, no.1,
pp. 17
Pickering A 1992, Science as Practice and Culture, University of Chicago Press, Illinois,
pp. 168-176.
Power, M 2014, Architecture Students Toast their Designs on Robots, The Sydney Morning
Herald, 24 August, accessed May 14, 2015, www.smh.com.au/technology/technology-news/
architecture-students-toast-their-designs-on-robots-20140823-107lku.html.
Schwartz, T 2012, HAL: Extension of a Visual Programming Language to Support Teaching and
Research on Robotics Applied to Construction, in Brell-Cokcan S and Braumann, J
(eds) 2012, Rob|Arch Robotic Fabrication in Architecture, Art, and Design, Springer-Verlag,
Vienna, pp. 92101.
RECONstruction
A Survey of Surface Scanning Techniques
for Architectural Substrates in Robotic Assembly
Joshua Bard, Richard Tursky and Michael Jeffers
Abstract While there has been substantial development in the use of industrial
robots for the tool pathing and assembly of fabrication components for architecture,
there exists a scope for improving a methodology for the mapping of material
substrate in architectural construction settings. Construction tolerances posit a
problem since they vary widely from rough to nish applications and are often at
odds with the demanding precision required in robotic fabrication processes. This
paper discusses a series of tests of scanning techniques on three example substrates
typical to wood construction, including: lath for plastering and stucco, spaced
sheathing for cedar shingles, and traditional stick framing. Scanning substrates
accounts for the gaps in tolerance that emerge from rough to nish construction
such as variation in as-built dimensions, misalignment of members, and the
adaptive behavior of materials as they adjust to new conditions. From a comparison
of scanning techniques, a cost benet matrix is developed to aid in evaluating the
appropriate application of scanning techniques for various robotic applications.
Keywords Reality capture
fabrication Photogrammetry
Motion capture
Robotic sensing
Robotic
1 Motivation
Robotic technologies have been incorporated in nearly every aspect of building
construction from excavating sites to erecting structural frames and applying
material nishes. While the application domains of robotic technologies in
J. Bard (&) R. Tursky M. Jeffers
Carnegie Mellon University, Pittsburgh, USA
e-mail: jdbard@cmu.edu
R. Tursky
e-mail: rtursky@cmu.edu
M. Jeffers
e-mail: mjeffers@cmu.edu
263
264
J. Bard et al.
architecture may span the entire construction process, the typical implementation of
industrial robots on construction sites is not seamless. Instead, robots are typically
deployed for discrete tasks in the construction process with varying degrees of
autonomy, often working in concert with a complex array of human activities (Lee
and Moon 2014). This stands in contrast to factory settings where automation and
strict control of manufacturing tolerances can currently be coordinated across the
entirety of complex product workflows.
On-site building construction poses particular challenges to industrial robotic
applications. In addition to issues of mobility in changing environments (Feng et al.
2014), the need to interface with multiple material systems installed with varying
acceptable tolerances by the building trades calls for improved methodologies for
the mapping of material substrates in construction settings. Construction tolerances
vary widely from rough to nish applications and are often at odds with the precision required in robotic fabrication processes. For example ASTM C926 stipulates a plane tolerance for veneer plaster of in (6 mm) in 10 ft (3050 mm)
(Ballast 1994). In order to generate safe and accurate motion planning for fabrication processes, the translation from planned virtual models to as-built realities
must account for variation in dimensions, misalignment of members, and the
adaptive behavior of materials as they adjust to new conditions (Vasey et al. 2014).
This paper discusses a series of tests with common scanning techniques by using
three typical wood construction systems (Fig. 1) where a nish material can be
applied to an architectural substrate:
1. Lathfor plastering and stucco (Lath)
2. Rib Assemblyfor cedar shingles (Digitally Fabricated Substrates, DFS)
3. Stick Framingfor interior and exterior sheathing (Stud Wall)
Fig. 1 Three test substrates related to wood construction precedent
RECONstruction
265
Surface descriptions generated by these scanning techniques were benchmarked

against robotically probing each substrate with an end-of-arm-tool (EOAT). From
this comparison, a cost benet matrix was developed to aid in evaluating the
appropriate application of scanning techniques for various robotic applications.
2 Development of Cost Benet Matrix

The technology space of real-time sensing and reality capture is rapidly expanding,
thus providing designers with ready access to autonomous workflows, adaptive
feedback, and environmental awareness in robotic fabrication. Understanding the
breadth of sensory devices and techniques is the rst step to isolating what
approaches are best for specic applications. This paper reviews three common
scanning approaches and develops a matrix that describes the various costs and
benets of each system (Table 1). These will be discussed in the following. The
Table 1 Cost benet matrix
Sensor type
Photogrammetry
Motion capture
1D depth sensor
EOA Probe
(Benchmark)
Specication
Nikon D3100
OptiTrack
Flex13
Sharp
GP2Y0A02YK0F
Aluminum
shaft with
soldering iron
tip
10 min; real
time
External, xed
camera array
10 min; real time
>3 h; real time
Robots
reach + Sensor
range
15100 cm
Robots
reach + probe
length
Robot work
envelope
$9000
Proprietary
$55
User-developed
$100
N/A
Yes, with
intensive
integration
Yes, with
moderate
integration
Yes, with
minimal
integration
[x,y,z, rx,ry,rz]
for all detected
rigid bodies
single value(float)
2.667 mm
44.0131 mm
[x, y, z] [q1,
q2, q3, q4]
(plane
w/origin)
N/A
Use constraints
Scan Time
10 min; 824 h
(live); (post)
Space
Robots
reach + camera
FOV
Workable
Min
Range
Dist >100 cm
for best image
Integration constraints
Cost
$400
Software
Cloud-based
service
Mountable
Yes, with
to EOA?
minimal
integration
Output and communication
Data (size);
Large output,
(type)
Produces Mesh
Mean
Accuracy
4.148 mm
Volume of
cameras FOV
266
J. Bard et al.
factors evaluated cover issues or advantages in use, integration, and output (Vh
and Pentti 2013).
2.1
Use Constraints
Scan Time describes how much information can be acquired, interpreted, and
brought back to the end-user in a meaningful format on which to base next actions.
This may be a single scan cycle or many scan-cycles arrayed over a given
search-space. Information processing time also weighs heavily on some strategies.
Scalability refers to the ability for the same device(s) to jump scale of
search-spaces. This is largely tied to workable ranges. Workable Range describes
the devices optimal or allowable range of (most) accurately detecting or acquiring
real-space data.
2.2
Integration Constraints
Cost of System include all associated (nancial) expenses of the device, peripherals,
and software. Required Software discusses all necessary interfacing technology to
communicate with, extract, and interpret data from sensor device. Lower-end
devices may involve user-developed communication strategies and communication
techniques to successfully integrate data to an applied robotic workflow.
EOA-Integration may also be referred to as Eye-in-Hand. Devices that can be
hosted EOA have the advantage of a mobile frame in space, and constant data of the
location of that frame (robotic kinematic model), thus allowing a given sensor to be
in multiple locations, orientations, and distances in relatively short periods of time,
and tied to a continuous process. However, spatial constraints of the end-effector,
tooling, task-conditions, data transmission, and power requirements must be successfully integrated or solvable for EOA.
2.3
Output and Communication
Accuracy refers to the margin of error, or delity of resolutions available on device.

Latency is the factor of lag time present within a single scan cycle and communication. This does not include processing time to compute higher levels of information and data types (this is a factor of Job Time). Data Type is the end-output
data type(s) available to the end-user. If user-developed software, often this will be
primitive data types. Proprietary software will likely provide multiple options,
formats, les, for output. Provided SDKs for such software may further extend this
category.
RECONstruction
2.4
267
Benchmark
Manually jogging a robot tted with a probe is the most effective way to directly
acquire, with very tight tolerance, a point in space with respect to the robots
coordinate system. For the purposes of establishing a benchmark, and demonstrating a high-accuracy method of obtaining substrate data, this technique was
applied on all three substrates with upwards of 100 sampling points each. Note that
manually operating the robot in direct contact with work surfaces requires human
proximity, which increases risk of damage, human injury, and increases scan time.
Probing is best utilized for nding key, or dening, elements of a known shape
in an unknown location to register digital model data to real world probed data for
subsequent robotic actions. Both human operated and automated probing can be
performed with high levels of accuracy, at the cost of invested time. All probe data
is immediately understood relative to the robot-space model without the need for
further transformation from another coordinate system. Drawbacks from probing
include the fact that input is still point-based, a pixel in an image, and therefore
slow to enumerate an array of information. The act itself requires slow and careful
manual motion, especially with probes of higher precision and sensitivity.
3 Three Scanning Approaches

3.1
Photogrammetry
Using an EOA-mounted camera (Nikon D3100 with AF-S Nikkor Zoom Lens set at
24 mm, f/18 aperture priority), planar toolpaths were prepared to quickly guide and
aim the camera at many sections of the target surface area (Fig. 2). The advantage
of utilizing the robot arm was primarily an issue of reach. A signicant factor is a
high degree of coverage, since missing data cannot be translated into the output
mesh.
This allowed issues of occlusion, or hard to reach areas of the surfaces to be shot
from a number of angles, producing the best results. Images were processed using
Autodesks Memento Beta. Since this software is cloud-based, there is signicant
lag time (test have shown typically 824 h) between uploading the images and
getting a workable mesh le. The input data required are simple JPEG images that
have framed signicant shared features among the other images. These shared
features are then used to correlate and compute a change in distance relative to other
features in the image plane. This software is designed to deliver 3D depth information as a factor of correlating the image data across all other images in the set.
High face-count meshes were generated for each substrate, with very little noise in
the target areas.
268
J. Bard et al.
Fig. 2 (Clockwise) Robot probe, DSLR Camera, 1D depth sensor, hand-held MOCAP probe
3.2
Infrared Motion Capture
A six camera OptiTrack Flex 13 camera array was ceiling mounted above the test
substrates. Each camera had a resolution of 1280 1024 and a frame rate of 120
fps. OptiTracks proprietary software, Motive Tracker, was used for internal calibration of cameras and streaming camera output. A custom streaming component in
Grasshopper (plug-in for McNeel Rhino and visual scripting environment) brings
rigid body tracking information directly into the CAD environment (Schwartz et al.
2014). A custom calibration script automatically aligns all tracking with the lab
robots base coordinate system. Once calibrated, the location and orientation of
custom tools can be tracked relative to the robots coordinate system. A custom
probe was created with interchangeable tips for various types of scanning. Users
positioned the handheld probe at each of the benchmark probe points for comparison and also traced along key features of each substrate for direct surface
generation from scanned guide curves. There exist benets and drawbacks within
this approach: Benets include the user directed flexibility in determining how to
scan specic features in an open-ended environment; a high level of accuracy and
precision; and the capacity to track plane orientation in addition to point location.
Drawbacks to this scanning approach include the need for human input during the
scanning process; high monetary cost of system, size limitations; and potential
occlusions of capture volume.
RECONstruction
3.3
269
1D Depth Sensor
A simple proximity sensor (Sharp GP2Y0A02YK0F) wired into a microcontroller

with a digital radio platform allows one to point-and-shoot for depth measurement.
Using the tcp normal vector and origin can generate the resultant pinged location as
a factor of its distance along the normal vector: In this case study, three separate
toolpaths were prepared based on the design-documentation of three surface
geometries and rough locations. Scan-lines were estimated as valuable areas to
search in hopes of capturing an edge condition to then justify the 3D geometry
against. However, with 200 m between a read of the sensor and overall scope of
toolpath distance, resolution was sacriced for the sake of speed. Scan paths
operated with <20 mm steps along lines that swept the extent of the 4 8
substrates. This then generated upwards of 300 data points per surface. Many
readings also fell off, or through, openings in the more sparse substructures,
introducing clear biases in the data. using design-geometry data, most readings
could be culled as invalid readings. However, those readings that were not culled
still exhibited ambiguity from sensor noise or registered points on parts of the
structure that were not strictly the desired top surface.
4 Comparison of Results
Each of the following graphs explains the performance of a given scanning process
against the three different substrates (Fig. 3). These histograms show the frequency
of a given scan point being within some millimeter tolerance range of the benchmark data model. MOCAP comparisons could be done point-to-point, whereas
Momento and 1D comparisons had to perform nearest-point to a mesh constructed
from the benchmark data. Therefore, error takes into account our closest approximation to the probe data, and not actual surface geometry, as is discussed here:
1. The 1D sensing has the worst range, as the scanline technique was designed to
capture the difference in depths that could be understood as edge. Even culling
complete misses does not take into account scan points that hit near, or just past
the top surface, where the probing was targeted.
2. Photogrammetry delivers a mesh object and the most error is introduced in
simple transformations to calibrate it back against the real object, as the mesh
output has no dimensional register to real-space. A best-t genetic algorithm
was used to obtain the best transform for this output to reposition the mesh in the
working-model.
3. IR Tracking was used with a hand-held stylus that could be placed on any
surface given the tool was not occluded from the cameras. For this data, the
points were directly obtained on top of the actual physical locations of the
270
J. Bard et al.
Fig. 3 Histograms of photogrammetry, motion tracking and 1-D depth sensor
probing benchmark data. This represents a clear depiction of device-to-device

tolerances, whereas the other techniques required additional geometric relationships to be established to build an equitable metric.
The following graphs (Fig. 4) also show more precisely the relationship of a
scanning technique to the three substrates, and how surface geometry and area
covered introduce additional errors that affect some sensing approaches more than
others.
RECONstruction
271
Fig. 4 Graphs mapping error by substrate paired with tabulation of standard deviation results.
(Top to Bottom) Photogrammetry, MOCAP, 1D Depth
272
J. Bard et al.
5 Conclusion
After test scanning three material substrates using three common reality capture
techniques a signicant amount of information can be leveraged to better select the
appropriate scanning technique for a desired operation. Signicant criteria to be
considered include the price point, ease of use, accuracy, and robust performance
under onsite environment constraints. For overall accuracy, hand-held MOCAP
probing provided the most reliable results. This technique also favors flexible
denition of environment features based on user input. If autonomous scanning
using EOAT is desired, then the photogrammetry approach provided accurate
results with automatic mesh generation. If evenly distributed, dense information is
desired across the scene then Photogrammetry provides the most expedient
approach, excluding cloud based post-processing of meshes. If select features are of
interest, MOCAP provides the best flexibility while scanning. If targeted reference
points are of interest, 1D depth sensing can be quickly deployed using motion
planning.
Future research building up on these ndings aims to rene each scanning
technique. It is suggested that for more robust 1D depth sensing, a custom triangulated laser range nder can be developed. For MOCAP scanning, better automatic calibration scripts are being developed to align live streaming with robot
coordinate systems. Through the development of an on-robot MOCAP camera
mount issues of limited capture volumes with xed, ceiling mount systems could be
overcome.
The survey of scanning techniques for surface description of architectural substrates, as has been discussed here, offers a scope for further investigations in
applied fabrication techniques for each substrate system, and can be extended
towards robotic plastering, robotic shingling, and autonomous wood frame
assembly.
Acknowledgments Special thanks to research assistants Brian Smith, David Blackwood, Nidhi
Sekhar. Autodesks beta version of Memento was used to for the photogrammetry section of this
research. Optitrack Cameras were used for all motion capture. HAL Robot Programming and
Control was used to stream MOCAP sessions into Grasshopper.1
References
Ballast D, 1994, Handbook of Construction Tolerances, McGraw-Hill, New York.
Feng, C et al., 2014, Towards Autonomous Robotic In-Situ Assembly on Unstructured
Construction Sites Using Monocular Vision, Automation and Robotics in Construction and
Mining (ISARC), 31st Symposium, pp.163170.
Hal.thibaultschwartz.com.
RECONstruction
273
Lee, S and Moon, J 2014, Introduction of Human-Robot Cooperation Technology at Construction

Sites, Automation and Robotics in Construction and Mining (ISARC), 31st Symposium,
pp. 978983.
Schwartz, T, Bard, J, Ganon, M, Jacobson-Weaver, Z, Jeffers, M and Tursky, R 2014, All Bent
Out, in McGee, W and Ponce de Leon M (eds), Robotic Fabrication in Architecture, Art and
Design, pp. 305317.
Vh, P, Heikkila, T, Kilpelainen, P, Jarviluoma, M, Gambao, E, 2013, Extending Automation Of
Building ConstructionSurvey On Potential Sensor Technologies And Robotic Applications,
Automation in Construction, Vol.36, pp.168178.
Vasey, L, Maxwell, I and Pigram, D 2014, Adaptive Part Variation, Robotic Fabrication in
Architecture, Art and Design, pp. 291304.
Robotics-Based Prefabrication
in Architecture
Prefabrication the Shinsa-Town Free-Form
Ceiling Structures
Xun Li, DongHan Shin, JinHo Park and HyungUk Ahn
Abstract Existing methods for the production and installation of free-form ceiling
structures were not suitable with respect to construction cost and time period that
were assigned to a Shinsa-town project. Hence, we selected the Robotic-based
Digital Fabrication Method that was being tested at that time. Considering the
construction cost and period, we selected expandable polystyrene (EPS) as
the material of the ceiling structure, and we developed and utilized BAT
(a Grasshopper plug-in), to process the work as a free-form production method. We
also invented a new cutting method to implement the specic types of components
that were otherwise unlikely to be implemented due to the limitation of the straight
hot-wire. This paper describes a transport system for the components of the
framework, and a robotics-based on-site installation method that is required for the
utilization of a robot in the fabrication of these structures.
Keywords Curved ceiling shapes
Industrial robot
Hot-wire cutting
Robotic fabrication
X. Li (&) J. Park
Department of Architectural Engineering, Inha University, Incheon, Korea
e-mail: leehoon126@inha.edu
J. Park
e-mail: jinhopark@inha.ac.kr
D. Shin H. Ahn
B.A.T Architecture, Incheon, Korea
e-mail: donghan.shin@b-at.kr
H. Ahn
e-mail: parisene@b-at.kr
275
276
X. Li et al.
1 Introduction
Industrial robots can be widely applied in various areas such as constructions,
interiors, and industrial design because of their accuracy, flexibility, and immediacy
(Saidi et al. 2008). Particularly, they can bring many advantages in terms of time
and cost, especially for producing curved surfaces by processing the expandable
polystyrene (EPS) selected in the Shinsa-town project. Indeed, the reduction in
processing time of the robot-based hot-wire foam cutting method is outstanding
when compared with the conventional CNC machining or manual laboring that is
normally used (McGee et al. 2012). Consequently, the reduction in processing time
implies a reduction in the construction period and in construction cost.
Because the cost and time available in the case of small-scale constructions such
as interiors and remodeling of outer walls are limited, both time and nance are
insufcient for performing research on the methods and materials to implement
free-form shapes. If a designer or a building owner requests to re-work due to errors
pertaining to accuracy, the responsibility shifts completely to the manager of the
construction. In order to prevent such problems in advance, commonly manual
labor of highly skilled workers is required for free-form architecture, resulting in
high production costs and extended construction time.
The Shinsa-town project is the rst project where an industrial robot is utilized
on a Korean construction site. For this reason, there is no precedent with regard to
the system required to establish the robot-based digital fabrication method. Thus,
thorough recording and investigation of the details of Shinsa-Town project should
be considered as the most crucial tasks that enable the robot to be more widely used
in the Korean constructions.
This paper presents all the issues that occurred during a series of processes from
the selection of free-form production method to the production of components and
the installation at the site. Based on the analytical results of the types proposed by
Chohelo A+U, an autonomous analysis was conducted by comparing the data of
traditional manufacturing methods with the data obtained through research on the
robotics-based digital fabrication method to identify the advantages and disadvantages of each method (Fig. 1). In addition, by classifying the phases of the
Fig. 1 Shinsa-Town Project. Before installation (left), and render proposed by Chohelo A+U,
CEO, ShinHyung Cho (right)
Robotics-Based Prefabrication in Architecture
277
Shinsa-town project into production and installation phases, and establishing

detailed sub-phases, the problems that occurred in each phase and the solutions for
improvement are discussed.
2 Methodology
2.1
Context
To implement this design after receiving a request for this project, conventional
production methods were investigated and classied into three types: Galvalume
Pipe Bending Method (GPBM), CNC Machining (Table 1).
Next, we investigated the robotics-based digital fabrication method from a
production time, materials, and cost perspective, based on the experience we gained
previously by using the ABB robot in the Hot-wire Cutting Research Project
(HCRP) (Fig. 2).
Table 1 Comparison of traditional methods (source HyungUk Ahn, construction site manager,
BAT Architects)
UC
MC
LC
GP
$18*500
MT
FT
IT
CC
M
$200/p*2
10d
10d
$20,360
I
$200/p*4
CNC
MDF
$18*400
$90/h
I
$180/p*4
7d
15d
$19,000
(M Material, GP Gavalume Pipe, UC Unit Cost, MC Manufacturing Cost, LC Labor Cost,
M Manufacturing, I Installation, MT Manufacturing Time, FT Fabrication Time, IT Installation
Time, CC Construction Cost, h hour, p person, d day)
GPBM
Fig. 2 Hot-wire Cutting Research Project (HCRP)
278
X. Li et al.
Table 2 Predictable information, using a robotic based pre-fabrication method

M
UC
MC
LC
FT
IT
RC
CC
EPS
$63*40
$90/1 m
I
$180/p*4
10d
5d
$3,600
$14,350
(M Material, UC Unit Cost, MC Manufacturing Cost, LC Labor Cost, I Installation, FT Fabrication
Time, IT Installation Time, RC Rental Cost, CC Construction Cost, m month, p person, d day)
Through this project, we collected data for estimating the time and cost required
for cutting the foam by hot-wires (Table 2). Based on the results presented, it was
concluded that the robotic-based digital fabrication method is the best choice in
terms of efciency and economic feasibility; and based on this nding, production
was begun.
2.2
Controlling Material Distribution
First, the curved surface was divided into a regular grid. The size of the grid was
determined by considering robot size, module weight, and minimization of material
loss. Next, a volume mass that is intended according to a standard plane was
determined; it was divided into 3 sections, with thicknesses of 600, 900, and
1560 mm, such that the curved surface is included in the volume occupied by these
three sections (Fig. 3). Based on the grid set above, the entire volume was divided
into approximately 200 components, and each component was labeled.
The basic material processing method is solid-based cutting, which produces a
basic component by cutting the EPS (Expandable Polystyrene) into a xed size, and
then cuts them as per the desired shape. However, the three types of BC for the
framework were congured with consideration for minimizing material loss, since
the construction period would have been extended indenitely if the production of
each of the 200 BCs was optimized to size. In order to simply minimize the loss of
material, each component was categorized as one of three basic types according to
size.
Fig. 3 The entire volume divided into approximately 200 components
279
Fig. 4 Robotic fabrication sequence: 1 basic component, 2 guide curves, 3 straight lines, 4 planes,
5 hotwire cut, 6 nished component
2.3
Tool Path Development
For hot-wire applications, two curves or guidelines are required to create a path.
The tool path along which a robot moves is made based on these guidelines. The
tool-path of hot-wire is a set of coordinate planes that continuously moves between
two guidelines, which are determined by a series of straight lines and a given point.
The x-axis is located on the straight line and the z-axis is determined by the center
of the straight line and the given point. Thus, the hot-wire is always located in a
series of coordinates and the center of hot-wires always corresponds to the origin of
the coordinate plane. The process of generating the tool-path is as follows. First,
two curves that serve as guidelines are extracted relative to the shape of the
component. Once the curves are dened, they are divided into a number of points.
A straight line connects the uniformly divided points in a sequence, and the
coordinate system is based on the straight line and a given point (Fig. 4).
As the number of points increases, the difference between the designed and
produced curve surfaces decreases, but the time for processing increases. The time
for processing affects the material loss because it is related to the time that hot-wires
stay on the EPS (Ahn et al. 2002). In addition, the optimized dividing of guidelines
is crucial because of the processing time affecting the construction time as well.
Thus, the guidelines were divided into 30 mm gaps in this project.
In the Xun Cutting Method only the ruled surface can be cut by hot-wires
(Yuan et al. 2014). In the case of most components, a method where hot-wire cut
the component along with the guideline considering two separated curves as the
guide curve is applied, however the shapes shown below cannot be cut by
280
X. Li et al.
Fig. 5 Experiment 2: for a curved surface where only one side is a straight line and other sides are
all curved, there are several options as given
designating the two guide curves. In order to solve this problem, several experiments were performed through computational modeling, with a rst set that tested
methods for cutting a simple plane, followed by tests for curved surface cutting.
While none matched the original curved surface precisely, method alt. 4 leads to the
closest approximation (Fig. 5).
Considering the results, a new method was developed to cut curved shapes.
While selecting the two guide curves, these were not selected both as curves, but
select one side for curve and the other for multi-curves (Fig. 6). The number of
separation between the curves is important in order to achieve the shape closest to
the original shape during the guide curve conguration process.
Fig. 6 1 3D model of irregular component, 2 tool-path of component, 3 Wire Cut Object
281
Fig. 7 Comparison between 3D model (left), and wire cut (center), and site context (right)
If a curve is exactly divided at the point where two curves meet, the boundary of
the side in the component that was cut appears clearly, and four lines will be
recognized as one. This results in a smoother shape when dividing overall lines, and
yields a result close to the actual shape. Furthermore, since one guiding line is much
shorter than the other, the amount of foam loss increases because the time that
hot-wires stay at the shorter curve is relatively long. We reduced the required
strength of hot-wires, as well as the moving speed of the robot by considering this
point (Ahn et al. 2003). While delivering a code to the robot, if two guide curves
coincide, the starting or ending point is not recognized. Thus, the closed curve
within the error rate should be cut, so that it can be recognized as two independent
curves. These tests were further evalutad for accuracy between 3D modeling and
robotic fabrication (Fig. 7).
3 Assembly and Onsite Construction

After cutting all components, these were assembled in groups of three and cut one
more time after tting them in a 900 mm 900 mm frame to minimize inconvenience of additionally cutting the components due to difference in their sizes.
A special EPS suitable adhesive was used here, which combined foam bond and
cement. Subsequently, the module was moved to the site, where further assembly
into larger areas was performed (Fig. 8).
Fig. 8 Module Assembly: 1 tting in a frame, 2 combination of three per group, 3 additional
xing
282
X. Li et al.
Fig. 9 Onsite installation of assembled modules
Most shapes can be implemented by a general cutting method that sets two
guidelines, but this method is not suitable for some shapes. In order to solve this
problem, we intended to use the Xun Cutting Method previously introduced. This
method implements the shape required for the second processing at once within
acceptable error, by giving guidelines for changes in conguration. The components that were delivered in accordance with the grid marked at the site were xed
with bolts at the precise location based on the name marked on the component.
Then, the nishing work was additionally done, followed by flame resistant treatment and plaster coating (Fig. 9).
4 Discussion
The research faced a number of issues during production and installation processes.
First, additional types besides the basic type were necessary, while making the unit
box because of irregular grid gaps. Thus, we should make the tool-path that cuts the
box. If we had considered production issues more deeply in the design process,
those problems would have been solved.
Secondly, we used the BAT plug-in (which is still under development) to design
a tool-path, convert it into a code, and deliver it to the robot; and because the
real-time upload module was not developed, it was delivered via a portable device.
We spent a lot of time in this process. Especially, it was inconvenient to proceed
when modifying the tool-path was required. Consequently, a real-time communication will be necessary to increase the efciency in robotics-based fabrication.
Thirdly, although the robotics-based fabrication is a highly accurate method,
errors still occur. The small errors combined together result in a large error in the
relatively large-scale construction. In this project, if there is a small error on a small
component, then a major problem can happen when 200 small components are
assembled together. In order to minimize this error, we cut the unit box to a larger
size and expanded the starting and ending point a little more while generating the
tool-path that cuts the component.
283
5 Conclusion
This paper reported on the process of production and installation of a curved surface
ceiling structure by a robotics-based digital fabrication method. We successfully
implemented the proposed design via this method with less cost and in shorter time
than other conventional methods. We experienced the efcient, economic, and
flexible features of robotic fabrication and could verify the vision as an actual
interior constructions in this process. It is also expected that the robotic fabrication
will be widely applied in a shorter time, if several problems can be solved. Firstly,
the robotics-based fabrication has the advantage that proposed designs can be
executed effectively. In order to apply it more efciently in the design phase, a
concept of robotic fabrication needs to be integrated at the beginning phase of the
design process. Additionally, we should identify the movement of the robot and get
information about the tool installed on the robotic arm. Secondly, we should
investigate the specication and performance of the materials that are used for
production. In this project, we could have reduced the time for production if we had
designed with such knowledge. If we had considered the hot-wire cutting method in
designing the curved surface of this range, we would have predicted the situation
and invented a new production method such as the Xun Cutting Method and there
would not be the need for an experiment during the production. Thirdly, further
studies of on-site installations using the robot should be performed. If our study
proceeds to the phase where the robot has information about the location of each
component and installs it in the exact location, it can eliminate errors that can occur
while installing manually, and thus further reduce the time for installation.
Acknowledgments The authors would like to thank Daeyeon Kim (Robotics department), ABB
Korea and Youngmin Park, and KUKA Robotics Korea for their continuous and generous support.
References
Ahn, DG, Lee, SH, Yang, DY 2002, Investigation into Thermal Characteristics of Linear Hotwire
Cutting System for Variable Lamination Manufactureing(VLM) Process by using Expandable
Polystyrene Foam, in International Journal of Machine Tools and Manufacture March, vol. 42,
no.4, pp. 427538.
Ahn, DG, Lee, SH, Yang, DY 2003 A Study on the Influence of the Sloped Cutting Angle on
Kerfwidth and Part Quality in the Hotwire Cutting of EPS Foam for the VLM-s Rapid
Prototyping Process, in International Journal of Machine Tools and Manufacture, November,
vol. 43, no.14, pp. 14471464.
McGee, M, Feringa, J and Sondergaard, A 2012, Processes for an Architecture of Volume, in
Brell-Cokcan, S and Braumann, J (eds), Robotic Fabrication in Architecture, Art, and Design
2012, Springer, New York.
Saidi, KS, O'Brien, JB and Lytle, AM 2008, Robotics in Construction, in Siciliano, B and Khatib,
O (eds), Springer Handbook of Robotics, 1st ed, Springer-Verlag, Berlin, pp.10791099.
Yuan, PF, Meng, H and Devadass, P 2014, Performative Tectonics: Robotic Fabrication
Methodology Towards Complexity, in McGee, W and Ponce de Leon, M (eds), Robotic
Fabrication in Architecture, Art and Design 2014, Springer International Publishing, Switzerland.
Stereotomy of Wave Jointed Blocks

Toward a Wave-Jointed Stone Construction
Using Wire Cutter Toolpath Generation
Simon Weir, Dion Moult and Shayani Fernando
Abstract This paper focuses on developing a system for using 6-axis robotic arms
to cut interlocking blocks with wire. Tracing the trajectory of stereotomy through
millennia of practice, an extrapolation is presented that stereotomy will serve
increased formal and structural complexity. The addition of robotic carving to
stereotomy also removes the ethical-aesthetic connection between the carvers effort
and the visual attention given to the object. This leads to the design of a wave
jointed block capable of an extended structural ability, concealing the majority of
the cutting effort inside the joined blocks. The proposed fabrication system uses a
wire cutter end effector following a toolpath generated from quad based mesh
topologies. This single tool cutting system maximises the efciency of the cutting
process and returns the once technical aspects of robotic construction back to the
designer.
Keywords Stereotomy
Wave joint
Toolpath automation
Digital fabrication
Vaults
1 Original Values
The art of stereotomy has its origins in the prehistory of architecture. The earliest
unambiguous records of the Greek word , the origin of the English word
architect, are Herodotus two and half thousand year old accounts of architects
directing the careful cutting and assembly of natural materials into massive objects
S. Weir (&) D. Moult S. Fernando
Faculty of Architecture, Design and Planning, Sydney University, Sydney, Australia
e-mail: simon.weir@sydney.edu.au
D. Moult
e-mail: dmou8237@uni.sydney.edu.au
S. Fernando
e-mail: shayani.fernando@sydney.edu.au
285
286
S. Weir et al.
(Herodotus, 440BC). Whether using timber or stone, the architect was a type of
sculptor, taking the earths materials and rearranging them; rst hewing them from
their origin in nature, fashioning them in a precise manner, and nally joining these
pieces together. The deliberate step-by-step process of taking irregularly shaped,
roughly hewn materialstypically stoneand fashioning them into precisely
predetermined forms is the art of stereotomy.
Procuring very large slabs of stone was a difcult task. This was recognized by
the ancient communities that encountered them. Herodotus described an impressively large shrine carved from a single piece of stone that was dragged whole from
the quarry by two and half thousand men and their beasts for three years.
Michelangelos over-life-sized gure of David was renowned as a masterwork of
stereotomy and comprised of an extraordinarily large block of Carrara marble that
was quarried in the 1460s. Sculptor Agostino di Duccio roughed out the gure in
the quarryto reduce the blocks weightbefore moving it to Florence (Wallace
1994). Apparently overwhelmed by the responsibility of the block, Duccio abandoned the block and it remained in Florence until Michelangelo took on the task in
1501. To reduce his labour and to honour this rare behemoth slab, Michelangelo
removed as little as possible. It was then the size of the gure cut from this block
that impressed his peers.
A second principle for considering stereotomy is displaying the nesse of the
carver. In sculpture, as in architecture, eyes are drawn toward some areas more than
others. The wise architect typically directs the most energy to these details: whether
they be a gures eyes or a columns capital. The joints within the walls, never seen,
are cursorily cut. Directing the carvers time to the visible details, the architect
enacts an ethic of aesthetics by allowing the carver to see their labour and knowing
that others can too. This second principle dissolves when a robotic arm wields the
saws and les. Robotic masons change this fundamental aspect of stereotomy.
The third principle recognizes that across the history of stereotomy, increasing
structural efciency aligned with new developments such as vaults evolved from
corbeled surfaces smoothed in situ; pre-cut voussoir vaults with joint surfaces at
normals to geometrically estimated thrust lines; and Antonio Gaudis hanging
catenary models tracing asymmetrical thrusts. With the arrival of steel and concrete
Modernism, stereotomy diminished into an academic focus on descriptive mathematics. This research ultimately led to the CAD and CAM systems that are
accelerating structural analysis options such as Philippe Blocks Thrust Network
Analysis tool (Block 2009).
2 Wave Jointed Blocks

Extrapolating the trajectory of stereotomy points towards two conditions. First is that
continuing to increase cutting complexity for greater structural capacity by using the
tensile strength of typically compression-only vaulting materials. While adding a
second material to provide tensile strength can produce remarkable results, as
287
Fig. 1 Wave jointed block: diagrammatic elevation and axonometric of curved blocks
demonstrated by Fallacara (2012), the different thermal expansion rates ultimately

cause damage, reducing lifespan; consequently we seek single material construction.
The second observation is that with robotic fabrication, the ethical relationship
between effort and effect is attenuated. This opens the possibility that more effort
might be used on an unseen surface. These trajectories led to the development of the
wave jointed block, whose edges interlock to transfer bending loads.
Construction blocks with interlocking sinusoidal edges have appeared several
times, e.g. patented by Robson (1976) and developed by Estrin (2014). These
blocks assist the construction process since they slip into place. Consequently,
Estrin and Dykin et al. (2005) proposed these for robotic construction in
extra-terrestrial environments. These blocks use a single wavelength along their
long sides and half a wavelength on the short side; and have been tested for axial
compression strength (Khor et al. 2004; Yong 2012).
Differing from these precedents, this project uses blocks that are designed with
multiple wavelengths of relatively high amplitude that transfer bending loads
between the blocks. The amplitude of the wave can be altered for different amounts
of forceshallow waves transfer little force, while deeper waves can transfer more.
Secondarily, while Robsons and Estrins blocks appear to be a cast material with
sinusoidal curves on both the long and short edges of the blocks, these blocks are
designed with flat edges on the short sides so the whole block can be cut using only
a wire cutter.
The design of the wave joint includes a repeated horizontal section which adds
robustness to the otherwise narrow and fragile edge. Then, in order for the blocks to
interlock, the horizontal tab is also included in the middle of the block (Fig. 1).
3 Toolpath Generation Using Quad Meshes

The cutting process for the wave joint blocks has been prototyped using a hot wire
to cut through foam blocks and an innovative approach to toolpath generation.
Descriptions of 3D objects in a computational environment are usually done with
NURBS based formats or mesh based formats. NURBS are an equation-based
288
S. Weir et al.
format, whereas meshes construct shapes using vertices, edges and faces. Each face
is made out of 3 or more edges. A 3 edged face is known as a tri, a 4 edged face is
known as a quad, and an n-edged face, where n > 4 is known as an n-gon. Despite
the typical usage of NURBS for digital fabrication, our approach towards designing, prototyping, and digital fabrication employs a mesh based topology using
quads.
For organic, sculptural forms, mesh modelling is generally much faster than
modelling with NURBS. This speed increase gives the opportunity for the designer
to prototype and experiment with iterative forms more efciently. Once the design
has been nalised, the mesh retains valuable information about the relative resolution of different parts of the mesh that has further impacts on simulation and
fabrication processes.
For example, during extruded print prototyping, most extrusion toolpaths are
generated from a triangulated meshed STL format. The resolution differences allow
the designer to specify different print resolutions on different parts of the mesh. In
the wave jointed block, a higher resolution was applied to interlocking surfaces than
to external surfaces. This allows the interlock to be accurately studied while the
external non-interlocking surfaces are less precisely and more quickly extruded.
The meshes of the blocks have been built using quads, as they can produce
organic or natural contours around the meshs surface which may be manipulated
by the designer. These contours may then be converted into a cutting toolpath that
can be adapted for different end-effectors. In the wave jointed block, it is envisaged
as being cut out of stone with a diamond wire cutter. To prototype this process, a
hot-wire cutter was used to cut through foam. The wire cutter end-effector implies
that the geometry follows a ruled surface.
There are also some potential benets to using this quad-meshed approach for
CNC milling. Typically, mills inevitably leave striations on the object. These
striations are then sanded down into the nal product. Using quad meshes allows
the designer to determine (or influence) the directions of the toolpaths, and hence
the directions of these striations. These striations can then be positioned as a
desirable effect, rather than an unwanted after-effect of the cutting operation.
4 Edge Loops
Meshes built with quads are able to describe the directionality of contours
throughout the surface of the mesh using the concept of edge loops. An edge loop is
determined using the algorithm below:
selects or deselects edges that:
if edges has 2 faces:
has vertices with valence of 4
not shares face with previous edge
289
if edge has 1 face:

has vertices with valence 4
not shares face with previous edge
but also only 1 face
if edge has no face:
has vertices with valence 2
As shown, the directional path dened by an edge loop is terminated when a vertex
with 3 or more than 4 valence is reached. For this reason, meshes built with tris or
n-gons are inappropriate for deriving natural toolpaths from edge loops (Fig. 2).
The two directions of a quad meshs edge loops are used to determine the initial
cutting path, as well as which neighboring edge loop to cut next. By moving from
one edge loop to another, the shape can be completely described (Fig. 3).
The KUKA robots KRL instructions determine the robots kinematics using 6
parameters: X, Y, Z, A, B, and C. X, Y, and Z refer to the translational coordinates
of the tools centre point while A, B, and C refer to the rotations around the XYZ
axes respectively for the tool. To determine the translational coordinates of the
toolpath the midpoints of each edge in the edge ring are used to provide a base
solution. Depending on the length of the wire, and the available buffer space along
the length of the wire, the midpoint may be moved along the edge. To determine
Fig. 2 Quad mesh and triangular mesh indicating their inherent directionality
Fig. 3 Quad mesh showing edge loops (left), edge rings (centre), vertex and face normals (right)
290
S. Weir et al.
rotational values the wire cutter can be assumed to approach co-planar to the quads
of the edge ring, or to approach normal to the quads of the edge ring. Of the three
rotational axes, one axis always runs along each edge of the edge ring whereas the
other two depend on the assumption made. The other two axes may be derived from
the cross-product of the quads vertex or face normals and the cutting edge. Using
these axes produces workable toolpaths, but they may be altered if collisions are
detected between the object and the end-effector.
As the KRLs rotational values are intrinsically rotatedi.e. the rotational axes
are performed in sequence and change after each rotationa mapping needs to be
done to covert the global 3D rotational axes to the values required by the KRL. This
mapping is shown in the following equations, where R represents the transformed
axes:
0 1
1
B C
^x @ 0 A
0
0 1
0
B C
^y @ 1 A
0
0 1
0
B C
^z @ 0 A
1
0
R11
R ^xt ; ^yt ; ^zt @ R21
R31
R12
R22
R33
1
R13
R23 A
R33

q
hy atan2 R13 ; R211 R212
when hy 6 90
hz atan2R12 ; R11
hx atan2R23 ; R33
when hy 90
hz hx sin hz atan2R21 ; R22
Once the mandatory parameters are determined these can be converted into different
types of continuous paths, such as linear or spline paths (if interpolation is desired
due to a lower resolution mesh). Further manual adjustments of the robots kinematic calculations can be introduced but is generally unnecessary.
291
5 Prototypes for Wave Joint Stereotomy

The toolpaths for the wave jointed block were staged as a 2-pass sequence
beginning with a rectangular block. First, the block is trimmed into an arced block
similar to a traditional voussoir block using a low resolution mesh. Next, the wave
joints are cut from both ends. Through the wave joints, the resolution of the mesh is
increased, as indicated by abundance of the orange lines indicating surface normals
around the wave joint (Fig. 4).
This process was prototyped in foam using a KUKA robotic arm mounted with a
hot wire end effector to produce approximately 1:2 scale blocks, 60 30 20 cm
(Fig. 5).
When identical prototype blocks are stacked into a tower 5 blocks high the
bending load is easily transferred to the base producing a rigid bent column
Fig. 4 Diagram of 2 pass cutting sequence for wave jointed block
Fig. 5 The 2 pass cutting sequence in foam for wave jointed block
292
S. Weir et al.
Fig. 6 Wave jointed blocks assembled into bent column
Fig. 7 5 Wave jointed blocks assembled as vault surface
(Fig. 6). The regular alignment of the waves in this design also allows for the
rearrangement of the blocks into the surface of a vault (Fig. 7). Further analysis of
the structural loads carried across the joints will later determine the size of the joints
needed to bear the weight of a stone material used in a similar manner.
6 Conclusion
These models demonstrate the practicality of using wire cutters to produce wave
jointed blocks that can be easily assembled into arches and vaults. Although the
arch and vault surfaces are presented here as prototypes, the system can be adapted
into an unlimited number of architecture elements and structural solutions. In a
293
complete vault the joints do not need to carry their bending load between blocks,
and the wave joint will weaken the blocks compressively. Therefore a design for an
arch using this joint would need to be more massive relative to its structural load.
The bent column, like the arch surface, is not a form that is impossible from solid
stone. It is only impossible to produce with 5 traditionally joined columns sections.
The possibility that this joint provides would be best expressed in very large
structures where the fabrication options, transport logistics and lifting capacities
limit the maximum size of pieces assembled. The wave joint design can take many
variations. Constructed as a ruled surface between two waves, there are two frequencies and amplitudes that can be modied, as well as their relative positions.
These variations constitute a range of structural capacities.
After millennia of stereotomic innovation the interactions between descriptive
geometry and construction robotics are enabling new opportunities in stereotomy to
emerge. The progress toward greater structural effect has always been produced by
the judicious joining of blocks. The mass production capabilities of construction
robotics in stereotomy can be directed toward precision joint cutting that can
achieve new forms in architecture.
Acknowledgments This research has been supported by The Faculty of Architecture, Design and
Planning, The University of Sydney, and was produced at DMaF.
References
Block, P, 2009, Thrust Network Analysis: Exploring Three Dimensional Equilibium. http://web.
mit.edu/masonry/papers/Block_PhD_2009.pdf. Accessed 2015.
Dyskin, AV, Estrin, Y, Pasternak, E, Khor, HC and Kanel-Belov, AJ 2005, The Principle of
Topological Interlocking in Extraterrestrial Construction, Acta Astronautica, vol. 57, no. 1,
pp. 1021.
Estrin Y 2014, Topological Interlocking and Osteomorphic Blocks, in G. Fallacara and V.
Minenna (eds), Stereotomic DesignCatalogo della mostra omonima per Inside Marmomacc
and Abitare il Tempo Veronaere, Verona 2014, Edizioni Gioffreda, Maglie.
Fallacara, G, 2012, Stereotomy: Stone Architecture and New Research. France: Presses des Ponts.
http://www.lithicdesign.it/portfolio/building-the-stereotomic-design/. Accessed 2015.
Herodotus, The History of Herodotus, 440BC, 2.175.5; 3.60.3,4; 4.87.1, 4.88.1; 7.36.1. Available
from: http://classics.mit.edu/Herodotus/history.html. Accessed 21 June 2015.
Khor, HC, Dyskin, AV, Estrin, Y and Pasternak, E 2004, Mechanisms Of Fracturing In Structures
Built From Topologically Interlocked Blocks, in Atrens, A, Boland, JN, Clegg, R and
Grifths, JR (eds), Structural Integrity and Fracture International Conference (SIF'04).
Robson, D 1976, Verbundstein, German Patent DE2554516 (A1), British Patent GB1533980.
Brisbane, Australia, 26-29 September 2004, pp. 189194.
Wallace, WE 1994, Michelangelo at San Lorenzo: The Genius as Entrepreneur, Cambridge
University Press, Cambridge.
Yong, HTD 2012, Utilisation of Topologically-Interlocking Osteomorphic Blocks for
Multi-Purpose Civil Construction, PhD, University of Western Australia, Perth.
Crafting Robustness: Rapidly Fabricating

Ruled Surface Acoustic Panels
Nicholas Williams and John Cherrey
Abstract This paper reports on the extension of a simple design concept into a
technique for the rapid fabrication of customized components of acoustic panels with
ruled surfaces. Recent proposals for the robotic fabrication of construction components include examples of techniques for cutting ruled surface geometries through
the pairing of an industrial robot arm with a linear blade. While these demonstrate
the fabrication of curved and complex geometry, they do not resolve many technical
issues around speed, accuracy and material nish, critical to a robust process
demanded by the manufacturing industry. To address these, the research presented
here pursued a detailed investigation into of the history of bandsaw cutting technology. Key knowledge of material crafts and obsolete applications of ruled
geometries both offer signicant insights. Using these in an iterative development, a
rapidly improved robotic design and fabrication process is demonstrated here.
Robust manufacturing
Digital fabrication
Mass-customization.
1 Motivation
The potential integration of robotics as an active element in the design and fabrication of customized, non-standard architecture has been intensely explored in
recent years. Many research projects have demonstrated that through the consideration of fabrication constraints in early stage design and the iterative production of
material prototypes, signicant opportunities exist to enhance design exploration
(Gramazio et al. 2014). Enhancements have been described and documented for a
potential to conceptualize novel material effects, and process efcacy, closely
aligning design proposals with means of fabrication.
N. Williams (&) J. Cherrey
RMIT University, Melbourne, Australia
e-mail: nicholas.williams@rmit.edu.au
J. Cherrey
e-mail: john.cherrey@rmit.edu.au
295
296
N. Williams and J. Cherrey
The research project discussed is equally interested in robotic fabrication,

applied here towards the design context of a music teaching space (Music Space,
2015-). The project brief demands the retrotting of an existing classrooma
demountable room within an existing space. This constrained the design to a rectilinear space of approximately 12 m2. Within these constrains, the design seeks a
novel solution for a panel system that lines the interior walls as the critical element
for aesthetic and acoustic room criteria. An iterative research process was undertaken to develop a design system, and to develop a robust outcome which could be
applied at larger scales of production. Materials demanded durability and a high
quality of nish appropriate to the classroom environment. The robotic fabrication
process warranted a rapid and reliable solution that could be comparable in efciency to existing processes of product manufacture.
2 Background
2.1
Acoustic Imperatives and Standard Solutions
This design of a music practice room presents challenges in terms of architectural

acoustics, and specically sound attenuation, which centers on the minimization of
sound transfer between the proposed space and those surrounding it. A Sound
Transmission Class (STC) of 60 rating was identied as a target performance level.
To achieve this, a modular system of massive sheet materials mounted on timber
frames was designed and constructed. This was calculated through variations on
guides for acoustic attenuation with commercial wall systems (The Red Book 2011).
Key issues and strategies for the acoustic design of small music spaces have
been outlined by Osman (2003, 2010), and include diffusivity to reduce occurrences
of flutter echoes and specular reflections. Standard solutions are the application of
absorptive materials and the use of quadratic diffusers. Furthermore, previous
research into architectural acoustics includes the shapes of wall surface articulation
applied to the design of acoustically diffuse spaces by Peters (2010); or acoustic
behavior of ruled-surfaces of hyperboloids (Burry et al. 2011); or shaping spatial
acoustics through robotic fabrication (Reinhardt et al. 2014). These studies attest to
a rich eld of acoustic research through empirical studies on complex forms.
2.2
Fabricating Ruled Surfaces
Ruled surfaces have a long tradition in architecture, and have been manufactured
through both additive and subtractive processes. More recently, studies have been
produced through exploring the flexibility offered by industrial robotic arms to cut
volumetric materials using a range of linear blades. Other applications of robotic
endeffectors include the use of hot wire to cut foam (Feringa 2012; McGee et al.
2012; Brell-Cokcan and Braumann 2013); wire diamond blades for stone cutting
297
(Feringa 2014); or bandsaw blades for cutting timber (Johns and Foley 2014).
Specically these latter studies explore variations of blade or material mounting,
from a blade mounted on the end of the robot arm, to alternately a xed position of
the blade and the material held by the robot. Here, a robot is programmed to move
the tool in a trajectory that causes the blank to move through the blade in a given
sequence in order to achieve a desired shape.
3 A Means or an End?
While some of these examples are explicitly experimental, others take current
industry processes and increase flexibility of geometry through robotics. Of particular relevance to this research is Johns and Foleys timber cutting (John 2014).
But as this paper argues, while this suggests exciting industrial opportunities, many
barriers remain in scaling up the process.
This research contents that robustness is at the heart of scaling both physical size
and volume. While the term is commonly used by designers to describe material and
construction durability, in manufacturing it relates more commonly to the resilience of
processes. Here, robustness describes a high volume manufacturing process in which
speed and reliability are vital. In a contemporary manufacturing context, flexibility is
added through robotic fabrication to enable customization of parts. Furthermore, the
need for processes to remain robustthrough minimization of complexity, and
uncertainty in other aspects of fabricationhas been noted (Piller 2013).
While current research has prioritized the design of novel formal and aesthetic
effects, the research presented here privileges considerations of process robustness,
and aims at a fabrication process that is reliable and rapid. In doing so, constraints
to formal outcomes may become apparent, challenging researchers to identify
productive trade-offs in negotiating the constraints.
4 Developing a Design System

4.1
Design Concept
Engaging these criteria, this research centers on robotic fabrication through a

combination of a robot arm that feeds material to a floor-standing bandsaw in order
to cut laminated blanks. It was proposed that long, rectilinear blanks be cut in a
single pass with each of the resulting pieces on either side of the cut then applied as
wall panels. The application of both pieces produces minimal material waste.
A number of considerations are embedded within this approach. Firstly, the
resulting curved, non-parallel wall surfaces are good for the acoustic diffusivity.
The blanks are a custom lamination of differing materialsprimarily timbers but
also foam and acrylic. These are cut with a surface shape dened by two planar
curves running along either side of the piece. The cutting process reveals these
298
Fig. 1 Diagrams showing a rectilinear blank with ruled surface for cutting (left); resulting pair of
component panels (right)
layers, cut at acute angles across their boundaries. By varying the shape of cutting
the surface not only is the form of panels varied but also the prole shapes of the cut
layers of laminated materials. This has implications for acoustics, with the differing
materials effecting reverberation time. There are also aesthetic implications through
the patterns created.
For the design of the room, a generative process was developed to explore
patterns. This centered on the translation of amplitude curves of sound waves to
create cutting geometry for blanks. The selected pattern applied to the room is based
on a set of tuning frequencies, with time steps mapped across the surface (Fig. 1).
4.2
Failing Fast
A rst mock-up panel was rapidly and successfully produced by cutting a 600 mm
foam blank with a hot wire. The blank was held by vacuum on the end of the robot
arm and moved through the wire. This mimicked precedent research, albeit with a
part composed of several foam layers glued together. A wire bandsaw blade was
sourced to enable a larger scope of material cutting; a spiral blade of a compact
1.5 mm diameter typically used to cut metals. A second material palette was prepared with 600 mm long blanks that were fabricated including a mixture of timbers,
Dupont Corian and Echopanel (a felted PET sheet with acoustic absorptive properties). These comprised the desirable palette of materials for panels.
First cuts proved to be highly problematic in terms of fabrication. The materials
cut with only varying levels of success: the Corian was clean; the timbers, both
hardwood and softwood, tore and burnt; while the Echopanel melted producing bad
surface smears (Fig. 2). Both the burning and melting was the result of sawdust not
clearing from the blade and hence becoming clogged.
Substitute materials were briefly considered but the larger fabrication problem
remained; the easy deflection of the wire blade restricted cutting to very low speeds.
The rst 600 mm blanks required 20 min to be cut without blade deflection. Such a
slow process was not only laborious but also contradicts aspirations for a robust
process. It was decided that the use of a wire blade presented insurmountable
barriers and so this approach was abandoned.
299
Fig. 2 Mockups in foam (left) and a combination of timber and Echopanel cut with a wire blade
but heavily burnt (right)
4.3
Seeking Older Precedents for Progress
A further process of investigating craft precedents was undertaken to better

understand historic applications of band saws. Two discoveries proved important.
The rst was the application of a similar technique to that used to fabricate curved
beams for boats. Through adding a rotary axis to the bandsaw, the blade could be
tilted as a timber member is passed through. Ruled surfaces could be accurately
robotically fabricated through an indexed relationship between the linear movement
of a workpiece and the rotation of the bandsaw structure itself (Fig. 3).
In parallel to identifying this potential, the design of bandsaw blades was also
investigated. Blade manufacturers commonly specify a minimum radius of cut that
can be achieved in a plane perpendicular to blade. This is proportional to the
thickness of the blade. There is, however, scant information published as to the
rotation in a perpendicular plane, twisting against the plane of the blade. The planar
depth of a standard blade means that twisting of a work piece commonly causes
distortion of the blade. While previous research from Johns (2014) has mentioned
the need for such twist, rotation has been constrained by conventions of the blade.
To compensate for this, cutting speeds and path have been compromised.
Through understanding blade manufacture, it was identied that a signicant twist
could indeed be achieved through modifying the blade to increase the kerf. With a
wider channel around the cut, the blade misses the twisting workpiece. With a precise
control of twist enabled through a robot arm, a detailed process of blade modication
and testing was undertaken to establish the geometric limits of such an approach.
4.4
Material Prototyping
With the adoption of a planar bandsaw blade with modied teeth and prole, a
sequence of example cuts were undertaken stepping through radius and twist in the
two planes. This blade was iteratively modied in order to increase kerf width.
A series of short 300 mm blanks of cheap timbers served as generic blocks to test
geometric limits. Further tests with Echopanel and other selected hardwoods was
also undertaken to identify limits at which a good cut nish could be demonstrated.
300
Fig. 3 Robot and bandsaw pairing to cut laminated blanks
The technique was also applied to the production of further full-scale prototypes to
address a combination of shape, custom laminated blanks, and the control of pattern
across multiple panels. Large prototypes include those fabricated with geometries
taken from small section of the larger wall design (Fig. 4). They conrmed the
geometric limits identied through test cuts to shorter blanks. A level of control was
achieved to demonstrate continuous shapes and selective wash boarding in key areas
of the designed surface. Final fabrication of the Music Room is currently being
undertaken with cutting of 50 blanks scheduled to be completed within one day.
In exploring the geometric limits of the system, a wash boarding effect was
also identied which occurs where the blade is periodically but minimally deflected
(Fig. 5). The geometric conditions at which this occurs are readily identied and
this effect can be introduced to accentuate key parts of the larger surface pattern.
5 Results
Through close attention to craft precedents, the technique presented has developed
to be more robust in terms of cutting speed and reliability of nish. A conventional
bandsaw blade has been adopted allowing a material feed rate of over 500 mm/min
across material up to 200 mm wide to be cut. With two parts produced simultaneously, this equates to a rate of 0.2 m2/min.
Through modication of this blade, good rates of curvature in the two key planes of
rotation are accommodated. Using a 12 mm blade, turns of as little as 50 mm radius
can be achieved. Twisting is also accommodated at up to 24 per 100 mm linear travel.
301
Fig. 4 Image has dropped out?Large mockups showing example sections of the proposed panels
Fig. 5 Protoype panel showing wash boarding from controlled deflection of the blade
Most importantly, these rates of curvature can be achieved without deformation of the
blade or reducing speed. This results in a nish which is consistent with existing
saw-cut materials on the market. Compared to early tests with wire blades with spiral
cutting surfaces this is a signicant improvement in speed (close to 20 times faster)
and reliability of nish (no burning or tearing of the material).
Despite modication, the conventional blade does not provide the geometric
freedom of a wire blade. Rather, tradeoffs between geometric freedom and the
robustness of process have been carefully considered. We have been able to extend the
geometric possibilities of a conventional bandsaw blade to a point where a 180 twist
can be achieved over the relatively short length of 750 mm. This is sufciently flexible
for application in the Music Room project. It would also be geometrically applicable to
other freeform timber components such as curved glue-lam beams. Importantly, cutting speeds and surface nish have not been compromised through this.
6 Discussion and Conclusion

This paper has presented key iterations in developing a technique for robotic fabricating of customized panels with ruled surfaces. With the improvements through
integrated knowledge from material craft and manufacturing, a robust technique
302
was developed and applied to the production of a wall panel system for a music
classroom.
This project demonstrates improvements to process achieved by extending a
known technology and integrating the deep knowledge around this. This is an
important shift from early research focus. Early attempts to improve a novel but
untested blade and technique were abandoned in favor of maximizing the geometric
freedom offered by a more conventional tool. This has implications across much
research engaging digital fabrication, through demonstrating benets of close
interactions between robotics and practices of material craft.
Future research will explore the application of this technique to other design
applications, next targeting the prototyping of other building components and
products. Many of these components are currently milled and this technique offers
potential reductions in volumes of dust as well as machine time. Furthermore, the
research will be further expanded through the application of additional end effectors
such as the Tungsten Carbide and bimetal bandsaw blades, so as to enable cutting a
wider range of materials, and to enhance renements of surface nish.
303
Acknowledgments The Music Room project was supported through an Innovation Voucher
awarded by the Business R&D Voucher Program through the Department of Business and
Innovation, the State Government of Victoria, awarded in partnership with Deutsche Schule
Melbourne. Professor Xiaojun Qui from RMIT University contributed expertise and research in
acoustics. Fabrication was undertaken in the Architecture and Design workshops at RMIT University.
References
Brell-Cokcan, S and Braumann, J 2013, Industrial Robots for Design Education: Robots as Open
Interfaces beyond Fabrication in Zhang, J and Chengyu, S (eds.), Global Design and Local
Materialization (Proceedings of the 15th International Con-ference on Computer Aided
Architectural Design Futures, Shanghai, China, pp. 109117.
Burry, J, Davis, D, Peters, B, Ayres, P, Klein, J, Pena de Leon, A, and Burry, M 2011, Modelling
Hyperboloid Sound Scattering: The Challenge of Simulating, Fabri-cating and Measuring, in
Gengnagel et al., Computational Design Modelling, Pro-ceedings of the Design Modelling
Symposium, Springer, Berlin, pp. 8996.
Feringa, J and Sonegaard, A 2012, An Integral Approach to Structural Optimiza-tion and
Fabrication, in ACADIA 12: Synthetic Digital Ecologies, Proceedings of the 32nd Annual
Conference of the Association for Computer Aided Design in Architec-ture, ACADIA, San
Francisco, pp. 491497.
Feringa, J and Sondergaard, A 2014, Fabricating Architectural Volume: Casting concrete, cutting
stone, building with foam, in Gramazio, F Kohler, M and Langen-berg, S (eds), Fabricate
2014: Negotiating Design and Making,Verlag, Switzerland.
Gramazio, F, Kohler, M and Willman, J2014, The Robotic Touch: How Robots Change
Architecture, Park Books, Switzerland.
Johns, R and Foley, N2014,Bandsaw Bands, in McGee, Wand Ponce de Leon, M(eds), Robotic
Fabrication in Architecture, Art and Design, Sprin-ger,Switzerland, pp.1732.
McGee, W, Feringa, J, Sondergaard, A 2012, Processes for an Architecture of Vo-lume, Robotic
Wire Cutting, in Bell-Cokcan, S and Braumann, J (eds.), Robotic Fa-brication in Architecture,
Art, and Design, Springer, Wien, pp. 6271.
Osman, R2010, Designing Small Music Practice Rooms for Sound Quality, Pro-ceedings of 20th
International Congress on Acoustics, ICA, Australia.
Osman, R and Fricke, R2003, Perceived Acoustic Quality of Small Rooms for Mu-sic,
Proceedings of 8th WESPAC, Melbourne, Australia.
Peters B, Olesen T (2010) Integrating sound scattering measurements in the design of complex
architectural surfaces: informing a parametric design strategy with acoustic measurements
from rapid prototype scale models, future cities. In: 28th eCAADe conference. Switzerland,
pp 481491
Piller, F2013, Three Capabilities that Make Mass-Customisation Work, in Pirooz-far, P
andPiller, FT(eds), Mass Customisation and Personalisation in Architecture and Construction,
Routledge, New York, pp. 1730.
Reinhardt, D, Cabrera, D, Niemel, M, Ulacco, G and Jung, A2014,TriVoc: Robotic
Manufacturing for Affecting Sound Through Complex Curved Geometries, in McGee, W&
Ponce de Leon, M(eds), Robotic Fabrication in Architecture, Art and Design, Springer,
Locaton, pp.163180.
The Red Book: Fire and Acoustic Design Guide, 2011, CSR, NSW Australia.
From Analysis to Production and Back

Attempts and Results of Reusable Adaptive
Freeform Production Strategies for Double
Curved Concrete Construction Elements
Felix Amtsberg, Gernot Parmann, Andreas Trummer
and Stefan Peters
Abstract The submitted paper presents the results of a 3-year research project in the
eld of adaptive forming technologies for freeform structures made of UHPC
(Ultra-High Performance Concrete). The focus of the research is found in the analysis
and comparison of the developed robotic-driven formwork. During the research it
was observed that the typical concept of process creation of a direct feedforward
material formation can be combined or even replaced by feedback-based production
processes. The different time points of material analysis not only allow for greater
control but also enable completely new production methods.

Sensory feedback
Material analysis Adaptive formwork
Formation strategies
1 Introduction
1.1
Relevance in Architecture
The efcient production of freeform building elements is still of essential importance to architecture. Concrete has been used as a construction material to realise
inspiring double curved structures during the 20th century by Nervi, Saarinen, Isler
and others, but has become less important to this kind of architecture in the present
F. Amtsberg (&) G. Parmann A. Trummer S. Peters
Institute of Structural Design, Graz University of Technology, Graz, Austria
e-mail: felix.amtsberg@tugraz.at
G. Parmann
e-mail: gernot.parmann@tugraz.at
A. Trummer
e-mail: andreas.trummer@tugraz.at
S. Peters
e-mail: stefan.peters@tugraz.at
305
306
F. Amtsberg et al.
day. Its castability in combination with reinforcement materials like steel, glass, and
carbon bre, and its further development to high and ultra-high performance concrete can result in a material, which is ideal for the fabrication of freeform structures. But double curved concrete structures remain at the fringes of the building
industry even as the development of new digital design tools has enabled more
freeform designs.
This is due to the relatively high expenses of the form works necessary for
production, as noted by Jrg Schlaich (Sobek 1987) in the foreword of Werner
Sobeks dissertation, which aimed to reduce formwork expenses with innovative
approaches to moulding systems. Time is still the essential factor in production
costs and waste reduction is an additional factor in each production cycle with new
production techniques analysed according to machining time (McGee et al. 2012)
or material waste (Jones and Foley 2014).
1.2
Research Motivation (Adaptive Formation)
Over the past decades many research projects have focussed on automated,
digitally-controlled processes for generating form. Several researches have investigated the realisation of recongurable tools for a sheet metal formation process
(Walczyk et al. 1998) from within the building industry, as well was within the
aerospace industry. But as of 2003 hardly any of the solutions patented during the
last 140 years have been applied commercially (Munro and Walczyk 2007).
Especially during the last ten years, digital fabrication has started to step into the
environment of architecture, art and design. Newly-developed digital tools have
enabled a direct link between design and fabrication; developed new machines and
systems; and deeply changed the way architects, artists, and designers work. The
concept of adaptive formworks was rethought by several research groups in the
environment of architecture: Adaptive Moulds developed by ADAPA (Raun et al.
2011) and the Flexible Mould Project (Michel and Knaack 2014) are just two
examples in this research eld. However, big innovative construction companies
and specialized companies still predominantly use subtractive CNC milling of foam
to produce double curved structures. Concluding, it can be said that the adaptive
form generation in the eld of double curved concrete structures is still not state of
the art.
2 Research Project
In response to this the research project Shell Structures made of Ultra High performance Concrete (UHPC)Thin Walled Double curved Construction Elements
Made of High-Performance Concrete for Innovative Shell Fabrication was started
in 2011 to develop a resource efcient production process for double curved
307
Fig. 1 Digital model to fabrication: analysis and subdivision of the geometry (left), Production:
formwork concepts, casting techniques and post fabrication processes e.g. formatting and dry
joining (right)
thin-walled concrete elements using just one industrial robot as the central unit for
all essential production steps. The implementation of industrial robots was the
logical choice for production as they combine versatility, accuracy, and controllability while also allowing for a direct transfer between digital design and
fabrication.
2.1
Digital Process Chain/New Developed Fabrication

Processes
The designed production process (Fig. 1) starts with the creation of a digital form.
This form is analysed and subdivided into elements that can be prefabricated. This
geometrical data is then used for the following production steps: adjusting the robot
driven adaptive formwork; generating toolpaths for different casting concepts; and
for formatting and sanding down the contact faces using wet machining.
3 Adaptive FormworkConcepts and Strategies

3.1
Adaptive Formwork General Concept and Requirements
The geometric analysis of a pavilion designed for the research project dene the
parameters for the construction of the formwork system. The project had the following requirements:
Realizable bending radius of 1.5 m as derived from research about the curvature of shell structures and designed freeform geometries.
Form accuracy of <1.0 mm as dened by the pre-stressed dry joint technology.
308
F. Amtsberg et al.
Robustness and simplicity as dened by the adjustment as done by an industrial

robot where complex machined elements should be avoided within the
formwork.
Three different formwork concepts where designed and chosen during the research:
the Pixeleld, the Pineld, and the Clay Mold.
3.2
Pixeleld
The concept of the Pixeleld is based on a dense high-resolution package of

rectangular plastic bars with a spherical head and a detached membrane. The rst
conceptual model was planned and tested with an arrangement of 15 30 pixels at
10 10 60 mm. For the real scale prototype, the rst and last pixel were halved
every second line to create a bracing. The resolution is 30/31 30 pixels; the pixel
size 30 30 400 mm and the mold area 900 900 mm. The formation process is
done serially by pushing each pin separately. Earlier tests used a wheel adjustment
(Fig. 2 left). When a row is adjusted, the position is xed by clamping. A membrane
surface is laid on the actuation pins after adjusting the whole area, and the form is
then imprinted with a lling material.
3.3
Pineld
The concept of the Pineld consists of a uniformly spaced eld of pins with ball
joint heads that attach to the mold surface. The rst conceptual model was planned
and tested with an arrangement of 7 7 pins with a distance of 40 mm while the
full-scale model was designed with 5 5 pins and a distance of 250 mm. The
formation process is done by a serial screwing of every pin (Fig. 2 center), which
Fig. 2 Different concepts of adaptive moulds: Adjusting by displacement of a dense packed pixel
eld (resolution 15 30) (left). Spaced Pineld attached to a membrane by screwing (7 7)
(center), and Iterative plastic formation of clay as a mold material (variable resolution) (right)
309
requires cyclic adjustment of the pins to avoid local overstretching of the mold
surface.
3.4
Clay Mold
The concept of the Clay Mold differs in terms of production logic. Instead of
developing a formwork using a complex adjustable mechanism, the intention here
is to plastically deform a raw mold material using different actuation techniques.
The tool affects the material directly, and tool development for specic tasks is
relatively easy. The material used for the research is an oil-based clay that doesnt
dry and can be reused. The displacement of the material is done by hammering and
pushing with a simple end-effector (Fig. 2 right). The formation process can be
described as material-neutral because there is no subtraction or addition of material
during the production process.
4 Information and Feedback/Development

and Production
An important distinction in this research is between information and feedback. The
relation of information, analysis, and approximation is immanent in the development of fabrication processes. Work steps, actuation, and the response of the
effected material are observed and adapted to the specic requirements.
4.1
Optimization
The optimization is based on design criteria, material information and production

methodology. The process itself is redeveloped, optimized and adjusted, until the
requested parameters can be achieved. The production itself is not under observation anymore. The process is planned, production is started, and the nal result is
expected (Fig. 3).
4.2
Information as Feedback
Information is dened as feedback if it takes a direct influence on the fabrication

process. This characterization is difcult, as the determination of a stock predenes
the milling process, but does not affect the nal result, just the fabrication process
310
F. Amtsberg et al.
Fig. 3 Information based process development
that in this case is the toolpath. It is the acquisition of specic information during
the creation of a process, and its use for analysis that allows the implementation of
the information before and/or while the fabrication process is running.
The simplest example of information-based process (Fig. 4, green cycle) is in the
analysis of different scrap pieces and their use for the production of a desired shape
(Amtsberg et al. 2015) or in the analysis of a specic log and the use of geometrical
information for resource-efcient production according the analysed shape (Jones
and Foley 2014).
The next category is the multiple cycle feedbackthe continuous gain of
information during the production process. Between two actuation cycles the fabrication machine stops, the current state of the work piece is then scanned, analysed,
and the resulting the information used for the next fabrication step (Fig. 4, blue
cycle). The extraction of sensory information and actuation are separate processes
occurring at different times. This strategy enables reaction on appearing indeterminablenesses, and thus enables iterative production cycles. An example is the
iterative formation of clay as a mold material, as presented in this paper.
Real time feedback is the continuous gain of information while the production
process is running. The production process is permanently under observation.
Fig. 4 Feedback based production development, input (green), cyclic (blue), real time (red)
311
It enables reaction while actuating (Fig. 4, red cycle). This process requires a direct
link between actuation machine and sensor. An example is the sensory controlled
bending of timber elements (Schwartz et al. 2014).
5 Sensory Information and Feedback in Development

and Process Generation of Adaptive Formworks
The implementation of sensors has catalysed recent discourse on robotic fabrication
in architectural design and introduced a new emphasis on the value of information
in the production process. Results within production cycles can be hard to predict,
so information describing current production states can be used to improve the
process (Raspall et al. 2014). All three formworks were designed independently and
with strongly differing actuation strategies, material use, and expected deformation
behavior, yet each process was remarkably simple. While moving, screwing, and
displacing are processes that are easy to realize by using industrial robots, their
detailed effects are hard to control and hard to predict. The successful use of each
formwork depended on the implementation of visual feedback during the fabrication process to ensure the generated form satised the dened requirements.
5.1
Pixeleld: Uncertainty
While the adaptive mold was planned meticulously and constructed precisely, the
process still left space for uncertainty. The clamp and release of every line resulted
in slightly different bearing situations, where deviations of the pixel size in the
range of less than a 1/10 of a mm could add up. The increased dead loads in every
line and varying friction between the pixels were hard to predict and impossible to
simulate.
Almost all of the adjusted pixels remain within the requested tolerances. Close to
70 % are within an accuracy of 0.25 mm; 20 % within 0.50 mm; 10 % within an
accuracy of 0.50 mm; and the number of pixels out of range below 1 %. After
adjustment a scanning cycle was implemented to detect the pixels out of range, and
they were then readjusted. The next control scan shows the successful replacing and
the continuance in position of the neighbors (Fig. 5).
5.2
Pineld: Indirectness
The Pineld is the rst concept that was tested during the research. It was simple,
fast, and using screwing as a formation process proved reliable. The actuation
influences only singular points on the surface. Most of the area is deformed
312
F. Amtsberg et al.
Fig. 5 Adjustment of the Pixeleld, extraction of the adjusted information, detection of the
inaccurate pixel and scanning after readjustment
indirectly by local deflection of the adjacent eccentric pins, or generated under the
increasing tensile force that appears when the mold surface is stretched. The same
adjustment results in different geometrical output curvature, depending on different
aspects such as the position or orientation of the desired shape on the mold surface
or mold material changes (Fig. 6). The high number of varying influences, and the
fact that a readjustment of a single pin effects on a whole area invites further
research on iterative adjustment process with continuous visual information.
5.3
Clay Mold: Iteration
The Clay Mold uses a process of plastic material formation where the amount of
material is not subtracted during the production process, but rather displaced with
Fig. 6 Adjusted Pineld (same curvature, different material response)
313
Fig. 7 Start analysis of material sufciency (red insufcient/green sufcient, bottom) and robot
production cycle, convergence of the surface after each press cycle
any deformation affecting the surrounding area. This reaction to the actuation
changes from centre to side, from ridge to recess, while the the implementation of
visual sensors enables an iterative production process. Cycle-by-cycle the stock is
scanned, analysed and the desired form is approximated (Fig. 7).
6 Conclusion
At the start of the research project, there was no intention to employ sensory
information as feedback within the production process. The production methods
were developed at different stages of the research project and seemed to be reasonable concepts, easy to generate and to control.
Sensory information was an essential tool for the development and improvement
of the concepts, but as feedback it has shown new possibilities within digital
controlled fabrication.
All of the formworks presented need not just information to generate the production process, but also feedback on their current state in order to control, readjust
or approximate the form in order to achieve the desired shape. The implementation
of feedback enables new perspectives on project-relevant production techniques
whereby continuous control of formation is possible and feed-forward processes are
improved across multiple cycles or even in real-time.
314
F. Amtsberg et al.
Acknowledgments The presented project is part of the FFG-Bridge Project 836524 Shell
structures made of Ultra High Performance Concrete (uhpc). The concept of sensory information was developed in collaboration with Felix Raspall and Martin Bechthold and part of the
workshop Digital Material Formation at the Rob|Arch 2014. The adaptive clay mold is part of
Florian Landsteiners Master Thesis written at the Institute of Structural Design in 2014-15, Graz
University of Technology.
References
Amtsberg, F, Raspall, F and Trummer, A 2015, Digital-Material Feedback in Architectural
Design, in Proceedings of the CAADRIA Conference, Daegu, South Korea, pp 631640.
Johns, R.L and Foley, N 2014, Feature-based Design and Fabrication of Nested Freeform
Surfaces in Wood, Proceedings of RobArch 2014, University of Michigan Taubman College
of Architecture and Urban Planning, Ann Arbor, Michigan
McGee, W, Feringa, J, Sndergaard, A 2013, Process for an Architecture of Volume: Robotic
Wire Cutting, in Brell-okcan, S and Braumann, J (eds), Rob|Arch 2012: Robotic Fabrication
in Art Architecture and Industrial Design, Springer, Vienna, pp 2847.
Michel, MandKnaack, U 2014, Grundlagen zur Entwicklung adaptiver Schalungssysteme fr
freigeformte Betonschalen und Wnde in Bautechnik, 91, Ernst & Sohn Verlag fr
Architektur und technische Wissenschaften GmbH & Co KG, Berlin, pp 845853.
Munro, C and Walczyk, D 2007, Recongurable Pin-Type Tooling: A Survey of Prior Art and
Reduction to Practice, Journal of Manufacturing Science and Engineering, vol. 129, pp 287296.
Raspall, F, Amtsberg, F and Peters, S 2014, Material Feedback in Robotic Production, in McGee,
W and Ponce de Leon, M (eds), Robotic Fabrication in Architecture, Art and Design 2014,
Springer Cham, Heidelberg, New York, Dordrecht, London, pp 333345.
Raun, C, Kristensen, MK and Kirkegaard, PH, 2011,Dynamic Double curved Mould System, in
Gengnagel, C et al.(eds), Computational Design Modelling, Proceedings of the Design
Modelling Symposium Berlin 2011,pp 291300.
315
Schwartz, T, Bard, J, Ganon, M, Jacobson-Weaver, Z, Jeffers, M and Tursky, R 2014, All Bent
Out, in McGee, W and Ponce de Leon, M (eds), Robotic Fabrication in Architecture, Art and
Design 2014, Springer Cham, Heidelberg, New York, Dordrecht, London, 2014, pp 305317.
Sobek, W, 1987, Auf pneumatisch gesttzten Schalungen hergestellte Betonschalen, Verlag
Ursula Sobek, Stuttgart.
Walczyk, F, Laksmikanthan, J and Kirk, D 1998, Development of Recongurable Tool for
Forming Aircraft Body Panels, Journal of Manufacturing Systems, vol. 17 no. 4, pp 551565.
Free Form Clay Deposition in Custom

Generated Molds
Producing Sustainable Fabrication Processes
Kate Dunn, Dylan Wozniak OConnor, Marjo Niemel
and Gabriele Ulacco
Abstract In a context of free fab printing, this research explores a series of
investigations into the potential of 3D printing with clay that address the problems
of viscosity, tool paths and setting times. The material of clay is explored here in
order to simulate architectural building processes that use both subtractive and
additive methods of construction that cannot be performed by a gantry style model
of robotics. The use of clay deposition on a robotic tooling path enables a continuous and sustainable adaptation process due to the fact that clay is reusable, can
mimic other materials in viscosity and is compatible with a range of sustainable
aggregates through to ring stages. This paper describes ongoing research into a
two-step robotic fabrication of free form clay printing; namely, as (a) the robotic
milling of a sustainable formwork; and (b) as controlled deposition of liquid clay
into a form or mold.
Keywords Robotic free form
processes
3D printing techniques Sustainable tool path
K. Dunn (&)
University of NSW, Sydney, Australia
e-mail: k.dunn@unsw.edu.au
K. Dunn D.W. OConnor M. Niemel
University of Sydney, Sydney, Australia
M. Niemel
G. Ulacco
AR-MA, Sydney, Australia
e-mail: gabriele.ulacco@ar-ma.net
317
318
K. Dunn et al.
1 Introduction
3D printing processes are an area of robotic fabrication that enables architectural
building processes through additive manufacturing techniques. Free Form
Fabrication (FFF) or Extrusion 3D printing processes rely on the extrusion of a
material in a pattern determined by an STL le. In FFF, material is extruded
successively onto a work bed or plate, whereas in 3D printed robotic processes,
only the robotic reach, and the work bed limitations limit the object size. While
studies have been conducted in a diverse range of material from polymer to plastics
and clay, specically the area of robotic clay printing has signicance since the
material characteristics of clay such as viscosity, tool paths and setting times present
a challenge to robotic manufacturing, but also offer the benet of sustainable
adaptation processes.
The research builds on ceramic material and delivery methods as an ideal
approach for large-scale 3D printing, due to the properties and material specications. The unique property held by clay is that it combines the strength of a solid
with the fluidity of a liquid, with plasticity in terms of slide or shear as initial
response to pressure, and strength attained by a mix of particle sizes and thixotropy,
which holds ne particles in a strong network (Hamer 2004). Ceramics or clay in a
raw state change consistency, and can mimic other materials through the addition or
reduction of water, defloculant, sodium silicate, alcohol and a range of up-cycled
industry waste aggregates (crushed materials added to a body to deliver different
properties of strength, moisture absorption and refractory qualities). Aspects of
dissolvability allow for the use of clay specically in robotic manufacturing where
machinery and equipment are sensitive to material blockage. Furthermore, robotic
3D clay printing in a customizable mold support new work methods that are highly
sustainable because the subtractive continued shaping of molds enables reuse and
renement.
This paper describes preliminary results from an ongoing research into a
two-step robotic fabrication of free form clay printing; namely, as (a) the robotic
milling of a sustainable formwork; and (b) as controlled deposition of clay composites into a mold. The research investigates a series of robotic precedents and
commercialized 3D printing techniques in order to develop a framework for
material and fabrication techniques. It focuses specically on multi-functional
processes including robotic arms while also addressing material sustainability in the
3D printing process, and the potential of up scaling for free form printing.
In the following, a background overview is presented. A second part describes
the present framework for the research (including technical components of pump,
setup of robotic arm and work envelope, limitations of aggregate clay body, and
customized end effectors). The paper concludes with a report on ongoing exemplary
studies into the custom manufacturing of a mold and slip cast in clay, free form
deposition of clay, and robotic deposition of clay within a mold.
Free Form Clay Deposition in Custom Generated Molds
319
2 From Commercialized 3D Printing to Robotic

Applications
The medium of ceramics has a long tradition of forming techniques such as wheel
throwing, hand building or casting, but has seen some recent research into robotically controlled investigations into free from deposition and scale adaptations that
test the material limits. Ceramics is used in industry and design in a range of modes
from slip casting injection molding and wheel forming to 3D printing, which differ
in terms of techniques and strategies, use of material body, and deposition techniques. 3D ceramic printing is currently commercialized on a small scale by
companies such as Figulos and Shapeways, production costs are high, and there
exist quality and stability issues with material shrinkage or cracking that is typical
for ceramic or porcelaineous materials.
In a context of free form robotic fabrication, research has been undertaken into
three- dimensional depositing of plaster within the work envelope of an industrial
robot of Morpheaux (Bard et al. 2012). Robosculpt (Schwartz and Prasad 2012)
shapes molds for bre-glas chairs by robotically subtracting from a manual packed
clay solid. Objects of Rotation (Dickey et al. 2014) uses different clay shaping
tools in a collet chuck attached to a robot arm that are used to mark or shape
columns of clay secured on a clay throwing wheel, and which enables digital
automation and up-scaling for clay modeling techniques. A signicant research for
combination of subtractive and additive combinations of formwork has been
undertaken by Woven Clay (Friedman 2014). Here, a styrene formwork is created
using a router subtracting material, and additive robotic fabrication is applied. The
porcelainous clay is deposited by robotic extrusion of a paste style of clay into a
woven pattern onto an undulating foam bed, so that the material is set to dry within
the formwork. This set an interesting precedent for the combination of different sets
of robotic applications within one work process. Furthermore, material processes of
ceramics can be exploited through robotic manufacturing for serial tests of surrogate or compounds, such as waste or up-cycled aggregates that address sustainability in large scale 3D printing.
3 Framework for Subtractive Robotic Milling

and 3D Printing
Using a robot to rout out plaster molds and subsequently to deposit material
selectively into the mold can be signicant because this reduces several steps from a
traditional ceramic fabrication workflow, and advances the potential of free form
ceramic deposition by utilizing the full range of movement available with a 6 axis
robot. A continuous trajectory from the subtractive cutting of molds towards the
additive deposition of material allows for a simplied workflow, and more
importantly demonstrates a two-fold robotic process where one single machine is
320
K. Dunn et al.
Fig. 1 Plaster Milling for Slip Cast mold (left), and clay deposition (right)
deployed for multiple fabrication steps (Keating and Oxman 2013). The research
developed these considerations into a series of material and robotic path studies into
free form clay deposit. Several aspects were included: designing with time; viscosity of material, placement, and hardening; three-dimensional deposit of material
(in space); velocity as aspect between tool path and material sedimentation; 3D
printing as gestural tracing in material, equivalent to drawing, a comparison of
different 3D printing materials and techniques; studying differences between
material bodies using clay as a composite material combined with different
aggregates. These considerations were tested against the project brief that explores
robotic fabrication of a mold, and consecutive clay deposition as a two-step processes that combines additive and subtractive techniques (Fig. 1).
In order to derive a framework, the research tested these criteria through a series
of prototyping experiments, including:
1. Subtracting material to shape details by testing the fabrication of molds for slip
casting that are fabricated using a spindle router from a block of pottery plaster.
While the creation of a mold traditionally requires a master object to cast from,
here, sequences of subtractive cutting of forms of the mold can be pre-determined
digitally, and continuously removed with completed stages of fabrication.
2. Exploring different aggregate conditions of a material body in order to understand requirements for robotic deposition process of liquid material bodies.
3. Using a robotically controlled extruder end effector to lay material rst in a
repeating, incrementally stepped up 2D pattern, then at angles, speed and precision unachievable by hand slip trailing or by a simpler gantry frame setup.
321
4 Workflow and Robotic Free Form Clay Printing

Several technical elements further contributed to the framework, such as the
ne-tuning of the pump setup, hoses and deposition bed; the limitations of aggregate
body of the clay material; and the development of a customized end-effector, which
are discussed in the following.
4.1
Setup and Work Envelope of Robotic Arm
The 3D clay depositing process is prototyped here on a KUKA KR10-R1100 with a

working envelope of approximately 1 m3, using the robot as a 3D clay printer.
A remote pump draws material from a reservoir, attached to the robot arm with a
6 m hose to utilize the full range of robot reach and axis movement. The extrusion
system is congured to work with a volume of material vastly beyond a traditional
robotic payload. The KUKA KR10-R1100 is stationed at a standard workbench
height of 900 mm on a portable, custom fabricated steel base. Initially a series of
intuitive tech pendant movements to test appropriate speed and height for tool-paths
was conducted, followed by extrusions for various material clay aggregates with
high control in KUKA|prc (Fig. 2).
4.2
Technical Components of Pump, Hose

and Depositing Bed
The research used a Moineau or Progressive cavity pump with a marine grade
stainless steel core or driver for deposition. It draws from a plastic reservoir
(20200 L) connected by a flexible hose that feeds the printing material into the
Fig. 2 Milling (left) and Testing tool path for robotic deposition in sections on custom bench
(right)
322
K. Dunn et al.
Fig. 3 Robotic work envelope, material storage and accessibility, flow of clay depending on
viscosity and speed of the pump and raw and red test samples
chamber of the pump, and into a hose (diameter 25 mm) attached to the robot end
effector for extrusion (Fig. 3). The extruder setup utilizes an aluminum coupling
which holds several off the shelf irrigation components to act as the extrusion
nozzle during the material testing phase. The current conguration enables
researchers to experiment with a range of materials of various viscosities as well as
aggregates in an accurate, repeatable and documentable way. In addition, there are a
series of control valves to ensure safety of operation, maintenance and cleaning of
the equipment, including one valve at reservoir outpoint, one in the pump cavity,
and a third attached to the nozzle or end effector. At a later stage, this setup will be
recongured to incorporate an air powered piston valve capable of being synchronized with remote control of pump velocity.
4.3
323
Limitations of Aggregate Body of the Clay

Material-Scaling up
As a specic focus of the research, different materials for robotic extrusion were
tested based on white earthenware casting slip (commonly used for casting) composed of clay particles suspended in a liquid composite body, and with additives
including kaolinite, crystalline silica, water, sodium silicate and polyacrylate dispersant (dispex). Furthermore, waste materials were included as aggregates to
address sustainability, including recycled paper pulp and softwood sawdust.
Maltodextrin and ne sugar were added to the mix to aid adhesion of the layers
during deposition, while cellulose bre and bentonite are added to provide structural integrity. Alcohol (methylated spirit) was added to decrease setting times
through evaporation. Initial tests were successfully prototyped by extrusion in the
closed form mold (Table 1). This is signicant because it presents an alternative to
common slip casting. In slip casting, liquid porcelain or clay is lled into a mold,
and water is absorbed by the mold so that the material solidies, and excess liquid
can be emptied. While slipcasting can include negative aspects such as overspill,
material wastage or mistimed absorption, the controlled robotic deposition can
improve these aspects by selective delivery of material.
Table 1 Material mixes for robotic deposition

Material mixes
Robotic deposition consistency
(1) Earthenware casting slip

(2) Cellulose bre
Cellulose bre stays in suspension in the

material, but tends to clump in
extrusion/deposition
Paper pulp is aerated through fabrication
process and when added to the base creates an
aerated even, lumpy texture. Tool path lines,
once layered, tend to slump on corners as the
tool path changes direction
Texture similar to mix 2 but sugar and
maltodextin result in thickness, facilitating the
layered toolpath deposits adhering to each other
and the deposition bed

(2) Paper Pulp, 30 % by volume water

(2) Maltodextrin & ne sugar added to
base mix 16 % each before water
(3) Paper Pulp 30 % by wet volume
(2) Maltodextrin & ne sugar dry weight
added to base mix 16 % each before water
(3) Paper Pulp 30 % by volume
(4) Methylated spirits 3 % wet
(2) Softwood particles 20 % by volume
(3) Bentonite 5 %
Texture changes with the addition of the

methylated spirits, becomes thicker and holds its
shape better. Once set, the material goes hard
and is reasonably strong even in an unred state
Textured with some particles being up to 4 mm

long and 2 mm wide. Bentonite absorbs
moisture and acts to hold the materials in
suspension
324
K. Dunn et al.
Fig. 4 Successive development of prototypes for clay deposition end effectors
4.4
Development of Customized End Effectors
In conjunction with the material composite studies, several iterations of

end-effectors were produced to explore strategies for clay deposition. An electrically triggered solenoid valve with a roll plate coupling resulted in limitations of
working with multiple viscosities, because maintaining a consistent pressure with
water to correctly operate the valve during movement proofed difcult. Another
aspect included the need for different off-the-shelf irrigation ttings to enable trials
of nozzles with varying diameter. As a result, a modular end effector was developed
so that individual components can be swapped for different material experiments,
programs, or in the event of passages becoming clogged or damaged (Fig. 4).
In addition, once experiments moved from 2D toolpaths that either spiraled or
stepped up in the Z direction to depositions within 3D forms, a nozzle 150 mm in
length was added to enable placement of materials with less movement of the
robotic arm, and less danger of collision with the work-piece, allowing increased
velocity of movement and a more versatile placement angle.
5 Conclusion
This paper has discussed the initial stages of an ongoing research in the context of
free fab printing as a series of investigations of the potential of 3D printing with
clay. These initial tests that have been undertaken enable us to better understand the
problem of relationships for robotic printing of clay in terms of mold, viscosity, tool
path and multiple times (depositing times, liquid run times, setting times). The
research discusses a robotic tooling for a two-step robotic fabrication of free form
clay printing; as the robotic milling of a sustainable formwork; and as controlled
deposition of liquid clay into a form or mold. The research successfully prototyped
rst results with the robotic milling of formmolds that have minimal waste in the
325
production, and the consecutive deposition with different clay mixtures. The robotic
deposition demonstrated an improvement of common slipcast processes and
reduction of spillage and material waste, which is part of traditional practices.
Furthermore, the use of waste substances from industry as a component of the
deposition material was successfully tested, and thus aspects of sustainability could
be associated within the context of 3D printing.
As a future research trajectory, the process could be adapted to specic site
conditions by the including locally sourced materials as aggregates, and furthermore be upscaled towards architecture applications.
References
Bard S, Mankouche S and Schulte M 2012, Morphfaux-Recovering Architectural Plaster by
Developing Custom Robotic Tools, in Brell-Cokcan, S and Braumann, J(eds),
RobArch2012-Robotic Fabrication in Architecture, Art, and Design, Springer, Wien, New
York, pp.139141.
Dickey R, Huang, J and Mhatre S 2014, Objects of Rotation in McGee, W and Ponce de Leon, M
(eds), Robotic Fabricationin Architecture, Art and Design, Springer International Publishing,
Switzerland.
Friedman, J, Kim, H and Mesa, O 2014, Experimentsin Additive Clay Depositions-Woven Clay,
in McGee, W and Ponce de Leon, M (eds),Robotic Fabrication in Architecture, Art and Design
2014, Springer International Publishing, Switzerland, pp.261272.
Hamer, F and J 2004, The Potters Dictionary of Materials and Techniques, 5th edition, A&C
Black Publishers, Ltd London.
Keating, S and Oxman, N 2013, Compound Fabrication: A Multi-Functional Robotic Platform for
Digital Design and Fabrication. Robotics and Computer-Integrated Manufacturing, vol.29,
no.6, pp.439448.
Schwartz, M and Prasad, J 2012, RoboSculpt, in Brell-Cokcan, S and Braumann, J (eds), Rob
Arch 2012 Robotic Fabrication in Architecture, Art, and Design, Springer Wien New York,
pp-230237.
http://oliviervanherpt.com/functional-3d-printed-ceramics//accessed 10/06/2015.
Solar Bytes Pavilion

3D Printed Interlocking Modules
Brian Peters
Abstract The Solar Bytes Pavilion is a temporary structure that highlights a

potential for architecture, where buildings are fabricated using new techniques
(robot arm, 3D printing), incorporate smart technologies (light sensors) and are
powered by renewable energy sources (solar power). Taking advantage of a robot
arms strength and range of movement, the pavilion was 3D printed with an
experimental extruder and the result is a structure comprised of ninety four unique
modules that charge during the day and glow at night.
Keywords Robotic fabrication Parametric design

manipulation Renewable energy 3D printing
Digital fabrication Robotic
1 Background: 3D Printed Architecture

Currently, several architects and engineers (Khoshnevis et al. 2006; Buswell et al.
2006; Fixsen 2015) around the world are creating large-scale 3D printers in order to
produce large-scale structures. They are utilizing various 3D printing techniques
and exploring different materials, such as concrete, adobe, articial stone and
plastics. At the same time, there is an emerging, alternative approach to 3D printing
architecture that is inspired by an ancient building component; brick. Smaller
machines are utilized to fabricate smaller building blocks that assemble to form
larger structures (Peters 2014), which was a technique that helped drive this
research project.
There are two approaches currently being utilized to 3D print plastic with a robot
arm. The rst is printing with a standard fused deposition modeling
(FDM) technique, where successive layers are printed on top of each other and rely
on the layer below for structural stability. Dirk Vander Kooij was one of the
B. Peters (&)
College of Architecture and Environmental Design, Kent State University, Kent, USA
e-mail: bpeter27@kent.edu
327
328
B. Peters
pioneers of the FDM style technique, printing functional furniture in recycled

plastic. The second approach relies on the plastic quickly drying and hardening in
midair during the printing process, allowing one to print without support below.
This technique was rst introduced with the use of resin in the Mataerial project
(Laarman et al. 2014) developed in collaboration between Joris Laarman Studio and
the Institute for Advanced Architecture of Catalonia (IAAC). Subsequently, there
have been many projects investigating this technique, such as the Mesh-Mold
(Hack and Lauer 2014) by Gramazio and Kohler, as well as many others. While
these recent projects highlight the geometric possibilities of utilizing plastics and
3D printing with a robot arm, they are not creating structural enclosures.
2 Project Goals
The Solar Bytes Pavilion is the result of research with multiple goals and is the rst
of two iterations of a full scale, 3D printed pavilion. The rst overarching goal for
these pavilions is to 3D print a full-scale structure using a relatively small machine.
This approach requires individual modules to be printed and joined together to
create the full form. It was decided early in the project that interlocking joints would
be developed to eliminate the need for mechanical fasteners between the modules
and allow the structure to be self-supporting. A second goal for this project was for
it to serve as a test of the structural capability of the 3D printed plastic, as there
would not be a separate structural framework. Finally, the intent was for the
pavilion to glow at night, while not being connected to the grid, and therefore solar
powered lights needed to be integrated into the design. The Solar Bytes Pavilion
serves as a prototype for a second structure and was analyzed both during and after
its fabrication to determine its performance and consider improvements to the
system before embarking on the design of the second structure.
The overall form of the pavilion follows the form of a catenary arch that is then
extruded to create a barrel vault. The vault is then slightly skewed from north to
south to help with the overall stability of the pavilion. The vault was then positioned
to follow the path of the sun, spanning from east to west to maximize solar exposure
for the solar powered lights (Fig. 1). Once the form was established, the structure
was subdivided into modules. Several subdivision patterns were experimented with,
such as triangular, Voronoi, and hexagonal (Fig. 2).
The hexagon was determined to be the best option for two reasons: (1) the
smaller surfaces of the sides/faces of the modules limited warping during printing,
and (2) the shape offered efciencies in both material and fabrication time. For
example, initial tests were performed with triangular patterns, however the prints
often warped because the three faces of the module had large, flat surface areas.
While the hexagon proved the most reliable to print, it is less resistant to compression loads. Therefore an additional interior wall was incorporated into the
design of the base to resist those loads and produce a very stiff and resilient module.
329
Fig. 1 The vault spans from the east to the west, following the path of the sun
Fig. 2 Subdivision pattern explorations
The nal pavilion is constructed out of 94 of these hexagonal modules that each
have an integrated solar powered LED at their peak. All of the modules are unique,
ranging in size from 35 to 48 cm in both width and length, and with a median height
of approximately 300 cm. The modules were 3D printed with translucent plastic,
allowing the structure to lter sunlight during the day and create a uniform glow at
night. The light effect is enhanced by the use of interlocking, dove-tail joints that
reduce the visual division between each module, which ultimately creates a lightweight structure, both visually and structurally.
The arch form of the pavilion highlights the structural capabilities of the modules, which were designed to carry the load of the structure in compression with
minimal reinforcements. Two thin-gauge steel arcs on each end of the pavilion help
resist wind loads and prevent the pavilion from moving laterally. The pavilion is
secured to the ground with a series of stakes that attach to the steel arcs and assist in
both resisting tipping and sustaining the compression vault.
330
B. Peters
Fig. 3 Compression load test
Compression strength testing was performed on the modules to determine the

weight a module could withstand before structural failure (Fig. 3). The tests revealed
that a single module can withstand approximately 450 kg of pressure. Since the
modules were printed in plastic, they were susceptible to bending before the failure,
but compared to more brittle building materials, such as ceramics or concrete, plastic
offers a great strength to weight ratio as the each module weighs approximately 1.3 kg.
3 Parametric Model
A parametric model (GrasshopperTM denition) was developed for this project,
which was essential from design to fabrication. There were several key design
parameters that were controlled by the denition, such as the subdivision pattern,
size and shape of the interlocking joint, and the height of the peaks that contain the
solar powered lights. Fabrication parameters could also be manipulated in the
parametric model, including the printing speed, layer height, and fabrication code.
The prototyping process was expedited by creating a direct link between the parametric model and physical tests, which minimized the time spent redesigning and
modifying the module. This was essential for the research, since the 3D printing
material and technique were highly experimental. The nal part of the denition was
a custom script to contour each module into a series of polylines stacked in the
Z-coordinate, and then subdivide those contour lines into xyz-coordinate that were
then used as the G-code for the robot arm. The KUKA|prc plugin was used to
translate those xyz-points into a programming language specic for KUKA robots.
4 Fabrication: Material
One of the main design objectives of the project was to create a glowing pavilion at
night, which required the use of a translucent material that can be 3D printed. There
are only two materials that t these criteria: glass and plastic. Since 3D printing
331
with glass is still difcult at this scale, plastic was used. There were several
obstacles that needed to be understood about printing with plastic that signicantly
influenced the design of the project. The rst was which material could be used with
the extruder. At the outset, the project aimed to use either polylactic acid (PLA) or
recycled plastic. There would have been two main advantages to using PLA versus
polypropylene (PP), polyethylene (PE), or polyvinyl chlorine (PVC). First, PLA is a
bio-based plastic, meaning that it can be created from plant-based starches. And
second, PLA warps less because the rate of shrinkage is minimal (www.stelray.
com/reference-tables.html). However, at the time of nal production it was not
viable because it could not be consistently extruded during the prototyping phase.
A translucent natural PP lament, with no added color, was ultimately used to
achieve the design goal of light transmission at night and warped far less than
polyethylene (Fig. 4).
4.1
Fabrication: 3D Printing Extrusion
The modules were 3D printed using a DOHLE hand welding extruder, the Mini CS
that was attached to a KUKA Agilus and utilized a FDM style printing process
(Fig. 5). The extruder was capable of extruding half a kilogram of plastic per hour
by accepting 4 mm thick PP plastic lament that was fed from a spool suspended
above the robot arm. The modules were printed with a continuous extrusion;
material flowed from the print head without stopping and starting, following a series
of vertically stacked printing paths composed of polylines (Fig. 6).
The goal for each polyline was to create an unbroken path that was both
structurally stable and time efcient. The size of the extrusion resulted in rigid and
strong print walls that were roughly 1 cm thick. The fabrication time for each
module ranged from 3 to 4 h depending on their size. Since the modules required
absolute precision during the fabrication process to allow for the interlocking joints
to work, the robot arm provided an extremely precise tool, as well as a very stable
base to mount the extruder.
Fig. 4 Difference in plastic warping: polyethelene (left) versus polypropylene (right)
332
B. Peters
Fig. 5 3D printing process with plastic extruder mounted on robotic arm
Fig. 6 3D printing process showing stages of completion (left to right)
4.2
Fabrication: Delamination
One of the biggest challenges with FDM style plastic 3D printing is the possibility
of delamination between layers either during or after printing. In this printing
process, material is extruded onto the layer below and therefore relies on the bond
between those layers to produce a structurally sound object. Delamination is often
the reason for an objects failure and can be the result of multiple variables. First,
each type of plastic has a unique rate of shrinkage when undergoing the phase
change from liquid to solid, which can affect the surface area between layers.
Second, the temperature of the printed material, printing environment, and the
cooling method are all factors. If the printing temperature is too low, then the
material will not be hot enough to bond with the layer below, and if the temperature
is too high the material will burn. Cooling the printed plastic effectively is also one
of the subtle intricacies of the printing system, since if the material is cooled too
333
Fig. 7 Delamination between

layers
quickly it will delaminate, and if it is cooled too slowly the print will begin to warp
signicantly.
The initial prototypes often failed due to delamination, so the layer height and
3D printing speed were both reduced, while the printing temperature was increased
(Fig. 7). The initial layer height was set at 2.5 cm, but was ultimately adjusted to
1.8 cm, while the printing speed was reduced from *2010 mm/s. The printing
temperature, however, was increased from 104 to 121 C. To cool the plastic during
the printing process, an air nozzle was integrated into the design of the extruder.
During the prototyping stage, several variables were tested, such as temperature and
direction of airflow, and it was determined that the nozzle should be directed
directly after the point of extrusion and set at a temperature of 21 C.
To better understand the lamination process, an infrared camera was used to
examine the thermal properties of the plastic and study the bonding strength
between layers (Fig. 8). The thermal imaging camera was able to visualize the
complex thermal properties of the 3D print during the printing process, and was
integral to identifying which settings led to a stronger bond between layers.
Fig. 8 Infrared image of 3D
printing process
334
B. Peters
Fig. 9 Detail of the module design (left). Exploded axon of the assembly (right)
5 Interlocking Joint Connection

A primary design feature of the pavilion was the development of interlocking joints
between the modules. It was inspired by a traditional method of construction,
stereotomy, which has experienced a rebirth with the rise of multi-axis digital
fabrication tools (Diles 2014; Clifford 2012). However, it utilizes an additive, rather
than subtractive, process where complex interlocking blocks are 3D printed rather
than CNC milled out of a solid material. Since the modules are unique, each one
was designed and fabricated for a specic location in the assembly (Fig. 9).
The interlocking connection is based on a sliding joint, where each of the six
sides of the hexagonal modules has either a male or female connection. The unique
tails and pins of the isosceles trapezoid shape are reminiscent of a dove-tail joint,
which works to prevent the modules from moving and is critical to the pavilion
holding its form. Since the underside of the pavilion is exposed to reveal the
structural geometry, the design of the interlocking joint became a design feature
rather than merely a structural detail (Fig. 10).
Fig. 10 Interior view of the pavilion revealing the interlocking joint detail
335
Fig. 11 Integrated solar powered light shown during the day (left) and night (right)
Fig. 12 Time-lapse during a 12 h period (4 pm4 am) showing the transition from day to night
The joints also allow for easy assembly and disassembly of the pavilion, which
is aided by unique numbers embedded into the base of each module that indicate its
position within the overall form. Several rounds of prototyping were needed to
336
B. Peters
perfect the interlocking connection. The joints needed to be loose enough so that the
modules can easily snap together, but not so loose that they ruin the structural
integrity of the pavilion.
6 Solar Power
The module design includes an integrated solar powered light (Fig. 11) that comprises a small photovoltaic panel, a rechargeable battery, an LED light, and a light
sensor. Each solar cell acts independently, capturing and storing energy as well as
sensing light levels for each individual LED, leading to an autonomous system that
reacts to its isolated solar conditions (Fig. 12).
This creates a recording of the suns exposure over the course of the day that is
then reflected at night as the LEDs light up. For example, if it is cloudy in the
morning and clear in the afternoon, the pavilions east side will be illuminated for a
shorter period of time than the west side, reflecting those environmental conditions.
The solar powered LEDs are installed at the central peak of each module to evenly
disperse light within the interior of the module and add to the glowing effect of the
pavilion.
7 Conclusion
This pavilion is an achievement in 3D printing a full-scale structure; nevertheless
there are multiple design improvements that will inform future fabrication. For
example, a subdivision pattern that better highlights that each module is unique
should be utilized, since it was hard to perceive that each module is different.
Additionally, the depth of the structure could vary to use less material on the top
than bottom and to improve material efciency and structural integrity. Thirdly, a
double curved surface could aid structural stability and perhaps further eliminate
secondary reinforcements. During the fabrication process, connecting and controlling the extruder through the robot arm programming language could permit
stopping and starting during printing without stringing between the openings.
Another fabrication improvement we are currently experimenting with is using an
extruder that can accept plastic pellets, which could lead to the utilization of
material mixtures that include recycled plastics, bio-based materials and composites. Finally, a larger extruder could decrease the relatively slow fabrication time.
As a future research, the Solar Bytes Pavilion will proceed to take full advantage
of the geometric freedom that a robot arm offers. In a second iteration, the pavilion
will be recycled by 3D printing a completely different structure using the same
plastic. Modules will be shredded into small pellets and directly feed into an
337
extruder that prints with recycled plastic. We plan on fabricating a larger structure
with a similar set of goals, but apply knowledge gained during the design and
fabrication of the Solar Bytes Pavilion.
References
Buswell, R, Gibb, A, Soar, R, Thorpe, A 2006, Freeform Construction Appli-cation Research.
Advances in Engineering Structures, Mechanics & Construction Solid Mechanics and Its
Applications, vol. 140, pp. 773780.
Clifford, B 2012, Volume: Bringing Surface into Question. Blurb Publica-tions.
Diles, J 2014, Intricate Stereotomic Assemblies: Hollow Masonry From Buckled Surfaces
ACADIA 14: Design Agency Proceedings of the 34th Annual Conference of the Association
for Computer Aided Design in Architecture, Los Angeles, October, pp. 257266.
Fixsen, A 2015, The Fine Print: Three design teams employ three different methods to arrive at
3-D printed structures Architectural Record, April, pp. 3436.
Hack, N and Lauer, W 2014, Mesh-Mould: Robotically Fabricated Spatial Meshes as Reinforced
Concrete Formwork Architectural Design: Made by Ro-bots: Challenging Architecture at a
Larger Scale. May/June, vol. 84, no. 3, pp. 4453.
Khoshnevis, B, Hwang, D, Yao, KT and Yeh, Z 2006, Mega-scale fabrication by contour
crafting International Journal of Industrial and Systems Engineering, vol. 1, no. 3, pp. 301
320.
Laarman, J, Jokic, S, Novikov, P, Fraguada, LE and Markopoulou, A 2014, Anti-gravity additive
manufacturing in Gramazio F, Kohler M and Langenberg S (eds), Fabricate: Negotiating,
Design and Making, gta-Verlag, ETH Zurich, Swit-zerland, pp. 192197.
Peters, B 2014, Building Bytes: 3D Printed Bricks FABRICATE Conference and Publication,
ETH Zurich, Switzerland, February, pp. 112119.
Materially Informed Design to Robotic

Production: A Robotic 3D Printing System
for Informed Material Deposition
Sina Mostafavi and Henriette Bier
Abstract This paper presents and discusses the development of a materially

informed Design-to-Robotic-Production (D2RP) process for additive manufacturing aiming to achieve performative porosity in architecture at various scales. An
extended series of experiments on materiality employing robotic fabrication techniques were implemented in order to nally produce a prototype on one-to-one
scale. In this context, design materiality has been approached from both digital and
physical perspectives. At a digital materiality level, a customized computational
design framework has been implemented for form nding of compression only
structures combined with a material distribution optimization method. Moreover,
the chained connection between the parametric design model and the robotic production setup has enabled a systematic study of specic aspects of physicality that
cannot be fully simulated in the digital medium. This established a feedback loop
not only for understanding material behaviours and properties but also for robotically depositing material in order to create an informed material architecture.
Keywords Informed design
tecture Material behavior
Robotic 3D printing Porosity Material archi-
1 Introduction
Materially informed Design-to-Robotic-Production (D2RP) systems explore the
extents to which rapid and flexible robotic fabrication methods can inform and
enhance established generative design to materialization and production practices.
In the case study of this paper, the focus is on experimentation with optimized
S. Mostafavi (&) H. Bier
TU Delft, Hyperbody, Robotic Building, Delft, The Netherlands
e-mail: s.mostafavi@tudelft.nl
H. Bier
e-mail: H.H.Bier@tudelft.nl
339
340
S. Mostafavi and H. Bier
material deposition for a compression-only computationally derived topology. The

study has explored the possibilities of designing and fabricating material architectures with various levels of porosities, ranging from architectural (macro) to
material (micro) scales. By employing performative and generative design methods,
industrial robotic production techniques and material science experiments, the
D2RP aims to close the loop from design to 1:1 scale fabrication. With this goal, the
main research components of the presented case study deal with specic aspects of
materiality in relation to design computation and robotic 3D printing. In this context, the chosen fragment of a computationally designed pavilion required translation of the optimization results from a nite geometry into a continuous robotic
path for material deposition in order to create an applicable material architecture.
The integration of physical material properties into design by means of digital
design interfaces and computational design methods has been explored in both
practice and academia (Borden and Meredith 2011; Kolarevic and Klinger 2008;
Gramazio and Kohler 2008). The historical survey with respective related cases and
paradigms is not within the scope of this paper but relevant to the goals of the
presented case. In order to position this project in this larger eld of research two
major types of approaches have been identied. One presents cases in which, in
order to study design materiality, the design system relies only on virtual modelling,
simulation, analysis, and abstraction of physicality through implementation of
certain computation methods such as Finite Element Method (FEM),
Computational Fluid Dynamics (CFD), Particle Systems, etc. The other one presents material experimentations and the design system focuses mostly on constraints and potentialities of certain material and/or fabrication method that is
integrated into digital modelling platforms, i.e. parametric design models. The
proposed D2RP system establishes a feedback loop between the two approaches. In
order to achieve this goal, at digital materiality level, a systematic and chained
strategy for design information exchange is established by designing and implementing a customized parametric form nding system for compression-only
structures combined with topology optimization. At physical level, the direct
connection to the robotic production system, in addition to improving the production method has led to the direct study of certain aspects of physicality that
cannot be fully modelled inside the digital design platform. Therefore, the production system becomes not only a means of fabrication but also simulation.
Recent advances in both robotics and 3D printing have introduced new
approaches towards architectural materialization and production. Considering
materiality and architecture at multiple scales, there are a few projects that successfully bring the two together. In some examples a scaled up printing machine is
employed to horizontally deposit layer-by-layer building material (Khoshnevis
2004; Khoshnevis et al. 2006; Kestelier 2012; Dini et al. 2006). The explored and
presented robotic 3D printing project proposes an alternative method of material
deposition aiming to create a multi-dimensional material architecture (Fig. 1). This
is achieved while taking the behaviors and properties of the implemented material
into consideration, which in this case is ceramics, as well as by integrating material
optimization routines in the D2RP system.
Materially Informed Design to Robotic Production
341
Fig. 1 3D model continues robotic single robotic path and emergent material architecture
Fig. 2 D2RP explores multi-scalar porosity at building, component, and material level
D2RP consists of four main research components: Design computation,

tooling/production set-up, robotics, and materiality. Each set of experiments and
design exercises presented in the following section, explores possibilities of integration by establishing feedback loops between the four components. Parallel to the
lab-based explorations for the development of the D2RP a studio design project was
considered as a pilot case study. n this project architectural and material porosity at
various scales is considered as the main design driver (Fig. 2).
2 D2RP Development
The D2RP proposes a roadmap for development and improvement of a robotic 3D
printing technology for fabricating 1:1 building components. The roadmap includes
three initial case studies, concluding with creating a direct link between design and
production: multi-colored light robotic 3D printing, robotic pattern studies, and
ceramic robotic printing.
Multi-colored light robotic 3D printing involves mounting a color changing light
source on the robotic arm. This project addresses the connection established
342
between motion and information extracted from the virtual 3D model. Being able to
study the three-dimensionality of robotic motion contributed to developing a new
approach to 3D printing, different from the slicing-in-layers printing technique. This
provided possible directions for dening a 3D printing method, in tune with the
structural characteristics of the nal prototype.
The study of robotic motion denes the boundaries of the digital design-space in
relation to the physical solution-space. This informs the parametric setup with
ranges of reachability and optimized orientations. It also contributes to being able to
maximize the overall space used. In addition, by numerically controlling the on-off
light pattern and light colors by means of an Arduino microcontroller, the team
reached the goal of further extending design possibilities in such a way that multiple
materials can be deposited at certain coordination based on the information
extracted from the virtual 3D geometry. As the rst step, any given curve, in digital,
is reproduced, in physical, with multi-colored light curves captured by means of
long exposure-time photography. Later this approach is tested on the
compression-only designed pavilion represented by a network of curves (Fig. 3).
As part of the second set of preliminary studies, the robotic pattern project
focuses on drawing geometric patterns that explore variation in densities and resolutions to reach the desired porosity. This informed the design of robotically
controlled routines for material deposition to reach a functionally graded structure
(Oxman et al. 2011). The established parametric system, derived from these
experiments, involved size of the overall shape, thickness of nozzle for material
deposition, number of targets to describe the robotic motion and the method of
approaching dened targets. As a consequence the team formulated two categories
of material deposition: Continuous flow and on/off numerically controlled flow
patterns (Fig. 4).
Fig. 3 Robotic motion: Multi-colored Light, 3D printing studies
343
Fig. 4 Pattern and material-architecture studies: on-off material architecture tests (left),
differentiated porosity tests at material micro scale (right)
The ceramic robotic printing explores possibilities of production of 3D printed

building parts and establishes a production method where all parameters are calibrated for the developed physical set-up. The team designed an extruder connected
to an end-effector mounted on the head of a robotic arm, where the material source
was exterior to the robotic arm in order to maximize the freedom of movement, in
order to achieve an optimum multi-dimensional material-architecture. Initial
experiments ranges from simple layer-by-layer material deposition to study material
flow to 3D dimensional printing on doubly curved surfaces (Fig. 5).
Fig. 5 Robotic 3D printing: Nozzle/resolution customization, tests on curved surfaces
344
Considering the fact that natural materials are not fully predictable several
material properties like plasticity, viscosity, flow rate and short-term material
behavior were investigated and documented at different robot-motion speeds in
order to provide complete information sets for the next prototyping phase.
3 Design and Prototype

In order to develop a coherent computational design system specic to this project,
the rst step was to implement methods for form nding of compression-only
structures, derived from the innate characteristics of the material. In addition to
eliminating tension forces in the derived topology, this part of the design system was
implemented as a parametric strategy to dene the porosity at macro or architectural
scale to fulll certain functional and locational requirements. Furthermore, in order
to achieve the micro porosity level, a nite element method for material distribution
optimization was implemented on a part of the designed pavilion. The optimization
also considered local and global load and support conditions. To implement a
generic and repeatable method on other parts of the topology, the challenge was to be
able to parametrically change the method of nite-mode geometric representation
like point cloud and mesh to a vector based or NURBs (non-uniform rational
B-spline) geometry. This was achieved by applying a segmental system in the very
initial topology, retrievable at different stages of form nding and parametric geometric transformation (Mostafavi et al. 2013; Mostafavi and Tanti 2014). By
applying the computational design system several congurations were generated, in
each distributing the compression only material where needed and as needed, while
taking the structural performance at both macro and micro scales into consideration.
The challenge of the next step was to materialize these differentiated densities by
creating unied topologies that express structural loads consistent to the design
approach and robotic fabrication potentialities and constraints. At this stage, various
algorithmic form nding and optimization techniques, mostly in the
Rhino-Grasshopper platform and Python scripting-language, were applied. This
allowed the systematic exploration and evaluation of design alternatives in the
design-solution space, eventually providing the required information for production, path generation and kinematics simulation with the ABB-Robotstudio.
Simultaneously, the initial material experiments and information sets informed the
design process, design materiality and robotics. This was achieved through
step-by-step documentation of a series of purposeful design-to-robotic production
experiments with xed and variable parameters. Specic to this project, the
resulting dataset provides information on the possibilities of the developed D2RP
system for robotic ceramic 3D printing, such as maximum angle of cantilevering,
maximum length of bridging material without supports, minimum and maximum
size of the nozzle, material flow, motion speed, etc.
For production purposes, the topology of the pavilion was sub-divided into unique
components. As the research progressed it became apparent that due to the signicant
345
Fig. 6 Left to right fragment chosen for 1:1 fabrication, informed point cloud chosen fragment,
continuous curve as toolpath
variety of custom building components featured in the design, the robot manufacturers software functionality needed further customization. For this purpose a link
between the design and the simulation environments (Rhinoceros platform and its
add-ons) and the rapid code interpreter of Robotstudio has been implemented. This
direct link between the design model and robot controller enabled the implementation
of a greater range of unique, longer, continuous tool paths (Fig. 6).
As a construction material, clay-ceramics is commonly used for compressiononly structures. The structures based on compression perform through stability due
to signicant mass and specic geometry. What the study aimed to prove was that
by controlling the geometry and the material deposition, compression structures
could become lighter, and signicantly improve their material cost and their thermal
insulation performance. A way of achieving material deposition optimization is by
controlling the parameters of the production setup. This is briefly described as
follows: The extruder system designed and built by the D2RP team manages a
plunger-based mechanical extruder of a paste of ceramic-clay, water and a specic
water based color-pigment that increases gluiness. The numeric control of clay
extrusion was experimented and valuable results for dynamic extrusion were
recorded, while implementing a discontinuous porous pattern. But due to shifting
research objectives, only continuous clay extrusion was used for the fabricated
prototype. Therefore, a custom design routine was developed in order to extract a
continuous motion path to generate the designed material architecture.
To achieve continuous material deposition, similar to the challenge of translating
mesh to NURBs in macro scale, in micro level a generic parametric system is
developed to translate the discrete result of optimization to continuous curves
(Fig. 7). In brief, the method involved picking a starting point and recursively
searching in the extracted point cloud to generate a continuous spline. From
topological and computational point of view, this helped the system to directly and
efciently provide an applicable tool path, considering material properties and
behaviors and hardware-software specications of the developed D2RP system.
Throughout the process the extrusion speed was adjusted empirically according
to observed structural and aesthetic considerations. Extrusion parameters were
346
Fig. 7 Deposition process on one of the driven continuous curve from the discrete result
controlled through line-size and nozzle customization at the tip of the robot
end-effector. We experimented with nozzles of various prole sizes and shapes. For
the fabricated prototype, a nozzle featuring a square 1 cm2 aperture was used.
Finally, within the studys agenda of 1:1 fabrication and architectural performance
aims, it can be concluded that the prototype achieves both improved 3D printing
speed and reliability.
4 Prototype
According to the design brief, the architectural object was relating to the surrounding environment via pores of varying in size according to functional, and
structural requirements. The fabricated fragment explores these connections,
materializing a piece of structurally optimized compression only urban furniture at
1:1 scale. While developing a customised design-to-production setup, the team
achieved optimization in motion path generation. Common 3D printing techniques
employ non-differentiated routines for slicing and ordering material layers into
motion paths. The prototype was produced embedding fabrication potentialities and
constraints into the design. It must be noted that, although the computational 3D
model comes close to the actual prototype, the two entities remain different mainly
due to emergent material properties. Differences between virtual and material
exemplify emergent aesthetics inherent to the material behaviour. The emergent
347
Fig. 8 Left: Test of the material deposition method (left), robotic 3D printing of the one-to-one
prototype (right)
aesthetics inherent to the prototype are as much due to the 3D layering technique as
to how material extrusion varies along the path (Fig. 8).
5 Conclusion and Discussion

Advancements in robotic building as presented in this paper indicate that future
building systems are customizable and increasingly robotically produced and operated. The presented D2RP system demonstrates that informed porosity in additive
manufacturing is relevant for the development of materially informed architecture.
Porosity at macro (building), meso (skin), and micro (material) scales implies optimization of spatial congurations and material distribution. Using this approach we
strive not only to control mass-void ratios but also to achieve an integrated design,
from overall building conguration to the architectural material. In the context of the
third and fourth industrial revolutions (Anderson 2012), the flexibility of such D2RP
system can be understood with respect to the interaction between designers, users, and
NC systems aiming to produce highly customizable and on-demand building components. Robotic Building (RB) eliminates the current problem of missed optimization opportunities due to a fragmented and sequential process of architecture
engineeringmanufacturing. In a larger context, the additive D2RP approach
348
presented in this paper is part of the Robotic Building (RB) project, which focuses on
linking design to materialization by integrating multiple functionalities (from functional requirements to structural strength, thermal insulation, and climate control) in
the design (Bier 2013, 2014) of building components.
Scaling up the technology of 3D printing from object to building was the specic
goal of the presented case study. This was achieved by integrating the technology in
an informed, chained design-to-production system, in which the 3D printing and
robotics are not only ways of manufacturing but also methods and tools for simulation and testing of certain aspects of materiality, which lead to new opportunities
for design exploration and creation. For the authors, it was important to develop the
technology not as an isolated node but as an integrated working-operating module
connected to a real-life design problem. The main consideration in architecture and
building construction is that the factory of the future will employ building materials
and components that can be robotically processed and assembled. This requires the
development of multi-materials, -tools, and -robots for D2RP processes that will be
implemented incrementally in the next phases of this research.
349
Acknowledgments This paper has proted from the contribution of the Robotic Building team
(authors, Ana Anton and Serban Bodea) and Hyperbody MSc 3 students (Fall semester 2014). The
project presented has been sponsored by 3TU. Bouw Center of Excellence for the Built
Environment, Delft Robotic Institute, 100 % Research and ABB Benelux.
References
Anderson, C 2012, Makers the New Industrial Revolution, Random House, London, pp. 185191.
Bier, H 2013, Robotics in Architecture, in Oosterhuis, K and Bier, H (eds), Robotics in
Architecture, JSB, Heijningen, pp. 68.
Bier, H 2014, Robotic Building(s). In Oosterhuis K (ed), Next Generation Building 1(1) (doi:10.
7564/14-NGBJ8), p. 8392
Borden, GP and Meredith, M (eds) 2011, Matter: Material Processes in Architectural Production,
Routledge, New York.
Dini, E, Nannini, R and Chiarugi, M 2006, Methods and Device for Building Automatically
Conglomerate Structure,WOPatent WO2006100556.
Gramazio, F and Kohler, M 2008, Digital Materiality in Architecture, in Gramazio and Kohler,
Lars Muller Publishers, Boden.
Kestelier, XD 2012, Design Potential for Large Scale Additive Fabrication, Free Form
Construction, in Fabricate Making Digital Architecture 2nd ed, Riverside Architectural Press,
Cambridge, pp. 244-249.
Khoshnevis, B 2004, Automated Construction by Contour Crafting-Related Robotics and
Information Technologies, Autom Constr, vol. 13, no. 1, pp. 519.
Khoshnevis, B, Hwang, D, Yao, K and Yeh, Z 2006, Mega-Scale Fabrication by Contour Crafting,
Int J Syst Eng vol. 1, no. 3, pp. 301-320.
Kolarevic, B and Klinger, K (eds) 2008, Manufacturing Material Effects: Rethinking Design and
Making in Architecture, Routledge, New York.
Mostafavi, S, Morales Beltran, MG and Biloria, NM 2013, Performance Driven Design and
Design Information Exchange, in Stouffs, R and Sariyildiz, S (eds), Proceedings of the
Education and Research in Computer Aided Architectural Design in Europe (eCAADe) 2013
conference, Delft, The Netherlands, vol. 2, pp. 117-126.
Mostafavi, S and Tanti, M 2014, Design to Fabrication Integration and Material Craftsmanship',
Thompson, Emine Mine (ed.), Fusion - Proceedings of the 32nd eCAADe Conference,
Newcastle, UK, 10-12 September 2014, vol.1, pp. 445-454.
Oxman, N, Keating, S and Tsai, E 2011, Functionally Graded Rapid Prototyping, Proceedings of
Innovative Developments in Virtual and Physical Prototyping, The 5th International
Conference on Advanced Research in Virtual and Rapid Prototyping, Leiria, Portugal, Sept.
28-Oct. 1, 2011.
Robotics-Enabled Stress Line

Additive Manufacturing
Kam-Ming Mark Tam, James R. Coleman, Nicholas W. Fine
and Caitlin T. Mueller
Abstract This paper presents a new robotic additive manufacturing (AM) framework
for fabricating 2.5D surface designs to add material explicitly along principal stress
trajectories. AM technologies, such as fused deposition modelling (FDM), are typically
based on processes that lead to anisotropic products with strength behaviour that varies
according to lament orientation; this limits their application in both design prototypes
and end-use parts and products. Since stress lines are curves that indicate the optimal
paths of material continuity for a given design boundary, the proposed stress-line based
oriented material deposition opens new possibilities for structurally-performative and
geometrically-complex AM, which is supported here by fabrication and structural load
testing results. Called stress line additive manufacturing (SLAM), the proposed method
achieves an integrated workflow that synthesizes parametric design, structural optimization, robotic computation, and fabrication.

Keywords Robotic fabrication Additive manufacturing Principal stress line

Oriented material deposition Fused deposition modelling Topology optimization
1 Introduction
Contemporary additive manufacturing (AM) technologies, such as fused deposition
modelling (FDM), compliment earlier CAD advances to enable complex geometric
exploration. Current 3D-printing platforms, however, conform to a traditional
K.-M.M. Tam (&) J.R. Coleman N.W. Fine C.T. Mueller
Massachusetts Institute of Technology, Cambridge, USA
e-mail: kmmt@mit.edu
J.R. Coleman
e-mail: colemajr@mit.edu
N.W. Fine
e-mail: nwne@mit.edu
C.T. Mueller
e-mail: caitlinm@mit.edu
351
352
K.-M.M. Tam et al.
CNC-based paradigm (Schwartz 2013), in which machine operation is isolated from

design. With a limited interface between conception and materialization, designers
have few opportunities to affect the qualities that the fabrication embeds in the nal
artefacts. This paper presents a new robotic-enabled FDM technique that is
structure-based and material-centric, to enable the systematic reproduction of
high-quality, and performance based printed structures for a given 2.5D (surface)
design boundary using common software interfaces and electronic components.
Called Stress Line Additive Manufacturing (SLAM), the proposed method is contextualized within a broader conceptual structural design methodology; the integrated
software-hardware framework addresses FDMtodays most common AM method
(Gibson et al. 2010)and seeks to facilitate flexible design-space exploration and
fabrication standardization. From standardized 3D printing platforms, such as the
hobbyist and professional options developed by MakerBot Industries and Stratasys, to
full scale AM systems like D-Shape1 and Contour Crafting,2 layer-based conventions
reduce aesthetic quality, material efciency, and geometrical accuracy. Particularly
problematic for FDM-based techniques, layer-based conventions produce anisotropic
material behaviour with strength and ductility properties that vary signicantly
depending on the lament orientation: the tensile capacity of specimens loaded perpendicular to lament orientation can be up to 50 % weaker than specimens loaded in
parallel, because the weak fusion between horizontal layers provides a natural weak
point for breakage (Mueller et al. 2014). These problems limit both the durability of the
printed specimen, and the end-use application of AM. This paper focuses on FDMs
most fundamental mechanism: robotic flexibility is used to directly achieve lament
depositions that conform to desired material behaviourprinted along trajectories
revealed computationally to enhance assembly strength. Recognizing the directionally
dependent performance characteristics of FDM processes, laments are aligned to
major axial networks of curvatures called principal stress lines.
2 Stress Lines: Theory, Suitability and Additive

Manufacturing
The emergence of structural analysis tools within common design platforms, such
as Karamba 3D3 and other similar plug-ins, has created an environment favourable
to stress-line-inspired fabrication, yet often structural patterns are employed without
being substantiated by structural logic. Precedents combining robotic fabrication
with stress lines include the ICDs Leichtbau BW Installation,4 and the GSDs
Robotic Beat Rolling (Friedman et al. 2014). In contrast, this paper focuses on the
production of stress-line-based surfaces with enhanced structural performance.
1
http://www.d-shape.com,.
http://www.contourcrafting.org/.
3
http://www.karamba3d.com/.
4
http://icd.uni-stuttgart.de/?p=10941.
2
Robotics-Enabled Stress Line Additive Manufacturing
353
Fig. 1 Stress line properties and suitability in design and additive manufacturing
Stress lines are numeric integrations of principal stress directions over each
innitesimal element that comprises an investigated structural body. Designers are
interested in principal stress lines because they provide a visualization of the natural
force flow in a structure, which shows the lines of desirable material continuity for a
given design domain (Michalatos and Kaijima 2014). This characteristic is evidenced by the typical convergence of results obtained from analytical and
numerical optimization procedures and from the principal stress lines generated for
the same design boundary (Fig. 1.1).
Theoretical properties of stress lines support their application in FDM: Firstly,
conventional numerical optimization methods, such as ground structure and
homogenization methods, tend to produce complex results that are computationally
exhaustive to manipulate (Brackett et al. 2011), and difcult to summarize as
line-based paths. Stress lines offer computational ease for fabrication and optimization information (Fig. 1.1). Secondly, design abstraction is enabled as stress
line methods are highly suitable for geometric exploration (Fig. 1.2), so that results
can reflect design boundaries, regardless of the scaling of material properties,
applied force, or the objects dimensions (Li and Chen 2010). Thirdly, geometric
compatibility is signicant (Fig. 1.3). Since stress distribution for elastic continuum
bodies are also continuous, stress lines create contour-like elds with curvatures
that typically traverse from one design boundary to another: these properties lend to
their consistently-clean robotic deposition and efcient linkages as printing paths.
Thus, a framework becomes available from stress line computation to load testing
to robotic fabrication (Fig. 2).
3 Methodology
To capitalize on the critical relationship between architectural geometry and
structural behaviour in order to generate innovation in each, the project combines
structural design space exploration, topology optimization, and robotic fabrication,
and introduces a physical testing protocol in order to deliver structurally informed
354
K.-M.M. Tam et al.
Fig. 2 Robotics-enabled SLAM-based design framework
geometric feedback. It builds on previous research into stress lines where computation is broken down into initialization, generation, and processing (Tam and
Mueller 2015). The design scope is here a 4-support grid shell with application for
design space exploration with 2.5D membrane structures. The fabrication environment contains a custom extrusion module that is mounted to a KUKA KR6
R900 small robot that is located centrally inside a contained envelope (Fig. 3).
For the prototyping, the project used 3D modelling software McNeel Rhinoceros
and the scripting plugin Grasshopper, with structural analysis conducted using the
plug-in Karamba on a surface design form-found using Kangaroo Physics. Robot
programming followed a precise set of criteria for the KRL code, with robot
instructions generated in RobotMaster,5 a plugin to MasterCAM.6 The print surface
(milled from a laminated MDF block) is positioned at an eccentricity from the robot
to minimize joint issues. PLA plastic is used as lament material, for its adhesion
capacity to MDF and printability on an unheated surface without warping.
Adapted to SLAM, a multi-objective processing method was developed to
balance aesthetic, fabrication, and performance objectives. Essentially, the optimization procedure assembles a new 3D-frame structure in each iteration with stress
lines that are heuristically selected by a genetic algorithm. A FEA is conducted to
determine the strain energy, normalized by material volume, of the new frame
structure with each iteration; the combination of stress lines minimizing total strain
energy is selected for materialization (Fig. 4).
Specically, each stress line within the base eld is assigned a binary value that
determines its inclusion or exclusion in or from the assembled frame structure for
each iteration. The list of binary values for each of the two principal planes is in
turn generated according to a ratio of on- to off-valuesa ratio that acts as proxy
for stress line spacing corresponding to various global stress line densities. Thus, an
innite design space is characterized only by two design variables in the procedure.
http://www.robotmaster.com/.
https://www.mastercam.com/.
355
Fig. 3 Robotic operation in fabrication shop (left), and path generation in MasterCAM (right)
Fig. 4 Stress line processing by strain energy optimization
With the density of the stress lines optimally calibrated, a series of rules-based
corrections are iteratively applied to the resulting stress line topology in order to
achieve additional improvements in the results (Fig. 5).
A rule-base system can codify and implement existing rules, and can expand to
accommodate additional rules as the experimentations continue to develop new
knowledge. Rules that were designed specically for the SLAM framework to
improve lament extrusion ease include the removal of line segments in areas with
signicant stress line overlap, and the realignment of otherwise converging stress
lines in highly-stressed areas. General structural rules include modication of stress
line curvatures to facilitate force transfers at intersection nodes, and the insertion of
bracing members. Although the initial implementation of the rule-based system was
largely dependent on manual operations, the process is to be automated in future
developments.
356
K.-M.M. Tam et al.
Fig. 5 Stress line processing by rule based corrections
4 Robotic Tool and Workflow Design

4.1
End Effector Design
Referencing current extrusion devices in consumer-grade applications, the custom

extruder is composed of a waterjet-cut aluminium frame that is mechanically
coupled to the robot via a pneumatic tool changer (Fig. 6.1). The aluminium frame
holds a commercially available extruder and control electronics. The Signstek
extruder accepts 1.75 mm PLA (polyactic acid) plastic and is composed of a 1.8
degree stepper motor, heating element, thermistor, and cooling fan (Fig. 6.2). Using
an Arduino Uno microcontroller and a N-Type MOSEF, the closed loop temperature control (PWM) of the extruder nozzle was utilized. The KUKAs 24 V signal
outputs were monitored by the on-board Arduino and were used to start and stop the
stepper motor/extrusion. Control of the stepper motor was achieved with an
EasyDriver board that uses an Allegro A3967 motor driver chip.
4.2
Robotic Programming Interface and Parameters
The typical workflow consists of the following iterative and trial-and-error procedure:
(1) Import of print surface and stress line data from CAD into CAM software;
(2) Specimen positioning in the workspace; (3) Print surface calibration in the CAM
model space; (4) Clustering of paths based on robot work volume and reach limitations; (5) Assigning stress-line-based geometry for toolpath generation; (6) Iteration
of possible congurations in CAM and RobotMaster interface; (7) Optimization and
simulation of robot trajectory; and (8) Export KRL code and run program in KUKA.
Due to robot work volume and reach limitations, stress-line-based paths are clustered
into a series of separate print programs. Reasonable estimation of the KUKA arms
limitation guides the global clustering of stress lines corresponding to different print
surface orientation (Fig. 7.1), whereas geometric similarities guide the clustering of
stress-line-based path internal to each surface orientation (Fig. 7.2). Next, the stress
357
Fig. 6 Various components in custom extrusion module with 1 tool changer and 2 extruder
Fig. 7 Procedure to create robot print programs from stress lines
lines assigned to each cluster are linked, and sequenced in a way that minimizes the
KUKAs total travel distance. (Fig. 7.3).
Joint congurations and toolpath generation parameters may be altered to affect
the lament quality. These quality determinants include:
1. Offset. To ensure a consistent quality in the printed lament and improved
adhesion between lament and print surface, the offset between the nozzle and
the printed surface is set at a level to allow the tip of the nozzle to drag along the
top of the extruded lament uniformly across a stress curvature (doff in Fig. 8.1).
2. Extrusion activation timing. To mitigate the loss of lament due to a residual
pressure gradient across the nozzle and to ensure the production of normalized
flow at the start of new stress-line-based paths, the lament is retracted
immediately after the end of a tool path to break connection with the previous
lament, and extruded again just prior to recommencing print for the following
stress line (lst and lend in Fig. 8.1).
3. Travel and move rate. The robots movement is measured by its move rate, which
differs from the actual travel rate of the extrusion tip over the print surface due to
rotational requirement of the joints that are specic to the geometries of each
stress-line-based path. Typically, decrease in the travel rate corresponds to the
thickening of the deposited material, whereas increases may cause poor adhesion.
4. Tool axis orientation. 5-axis settings achieve the greatest consistency in the
laments cross section prole, thus leading to the best aesthetic and structural
358
K.-M.M. Tam et al.
Fig. 8 Robot programming control parameters
performance. However, 4-axis settings can be used when joint rotational and
collusion issues prohibit stress-line-based paths to be printed at complete normalcy (Fig. 8.3).
5 Results: Quality and Performance

Several surface topologies corresponding to common loading cases were produced
for the 4-support grid shell case geometry. The superimposition of several layers of
stress lines based on different loading conditions to induce additional structural
resiliency was also explored. The SLAM methodology is relatively successful at
achieving complex, structurally-meaningful geometries and topologies. To validate
that SLAM-produced specimens perform better than conventionally 3D-printed
parts, a comparative load test was completed on a number of specimens: three
printed using the SLAM method, and four printed using a conventional layer-based
3D printer. The load test consists of a single centralized vertical point load that was
applied until a peak load was reached (Fig. 10.1). MakerBot was selected as the
technology to be compared to the SLAM method, as it is one of the most popular
consumer-grade 3D printing platform available to designers that uses PLA plastic.
An effort was made to ensure that all specimens have similar total material volume
(Fig. 9).
The MakerBot MB prints included both a solid constant-thickness shell (labelled
M-C) and three variously-discretised shells, which include a random-generated
(M-D.R), a stress-line-inspired (M-D.DL), and a grid-based topology (M-D.G).
While the number of tests conducted is not high enough to be statistically conclusive, these preliminary results suggest that the SLAM method does lead to
359
Fig. 9 Fabrication results of artefacts printed using the SLAM method
improved structural performance, as indicated by the increase in ultimate load and

improved ductility after initial failure, as shown in the normalized
load-displacement plot (Fig. 10.2). The potential gain is particularly evident in the
MB specimen with uniform thickness, where failures occurred in the shell where
tensile stresses were predicted to occur (Fig. 10.3). More signicant advantages for
the SLAM method are expected in complex geometries and loadings that induce
more tension.
6 Discussion
There are several important directions for future work in SLAM that can be broadly
categorized into four objectives: (1) Continue to investigate the material and
strength behaviour of artefacts printed using the SLAM method; (2) Standardize
and improve the SLAM procedures to achieve better precision; (3) Develop a
computational framework to automate the generating stress-line-based structures
and robot paths; and (4) Continue to explore and expand the design applications of
SLAM.
Considering that SLAMs development was motivated by the recognition of
anisotropy in artefacts printed using conventional FDM methods, future SLAM
360
K.-M.M. Tam et al.
Fig. 10 Load tests and results
research will continue to seek validation on the merits of oriented-lament deposition. Particularly, research is underway to conrm the tension capacity of prints
using the SLAM method, since the form-found 4-support grid shell geometry used
in the initial implementation is not expected to experience signicant tensile
stresses. Modication on the boundary conditions of the grid shell, such as its
supports and loading positions can signicantly induce greater tension in the system. The exploration of more complex surface geometries that are expected to
experience both tension and compression can also provide a better understanding
on the tensile strength properties of artefact printed using the SLAM method.
Beyond the immediate time frame, signicant future milestones include the computational development of 3D-solid stress line computation, the elimination of the
support structure, and the expansion of the extrusion modules hardware capabilities, such as the incorporation of sensors to allow extrusion parameters to vary
intelligently according to emergent conditions of the printed surface. These
advances will open possibilities for free-form and real-time additive manufacturing,
thus allowing technique to be implemented for more complexly curve 3D-surface
design, and for full-scale construction.
361
7 Conclusion
The research pursued in this paper constitutes a promising rst step in validating a
new approach to AM that synthesizes multi-axis machining capability and
structural-led computation to enable the production of structurally-performative and
geometrically-compelling 2.5D surface designs. Signicantly, structural load testing provided initial verication that the proposed method outperforms methods
using the conventional layer-based paradigm. The research also presented new
strategies that specically addressed the challenges of robotic manipulation at the
prototyping scale. The most important contribution, however, is the demonstration
of a new consolidated methodology encompassing parametric design, form-nding,
structural optimization, robotic computation and digital fabrication, which uses
robotic-integration to achieve a structurally-informed method of fabrication that
provides designers with an opportunity to explore a fuller design space that considers both geometry and performance.
Acknowledgments The authors wish to thank the following students who assisted with various
aspects of this research: Jonathan Mackaman, Akshat Bubna, Elizabeth Bianchini, Katie Gertz,
Colin Poler, Xinyi Ma. Additionally, the authors acknowledge MIT fabrication lab coordinators
Justin Lavallee, Chris Dewart, Jen OBrien, shop monitors Ins Ariza and James Addison, and
testing lab technician Stephen Rudolph.
References
Brackett, D, Ashcroft, I, and Hague, R 2011, Topology optimization for additive manufacturing,
Proceedings of the Solid Freeform Fabrication Symposium, Austin, Texas, United States of
America, pp. 348362.
Friedman, J, Hosny, A and Lee, A 2014, Robotic Bead Rolling: Exploring Structural Capacities in
Metal Sheet Forming in McGee, W and Ponce de Leon, M (eds), Robotic Fabrication in
Architecture, Art and Design 2014, Springer Science & Business Media, Cham, pp. 8398.
Gibson, I, Rosen, DW and Stucker, B 2010, Additive Manufacturing Technologies: Rapid
Prototyping to Direct Digital Manufacturing, Springer, New York.
Li, Y and Chen, Y 2010, Beam structure optimization for additive manufacturing based on
principal stress lines, Solid Freeform Fabrication Proceedings, pp. 666678.
Michalatos, P and Kaijima, S 2014, Eigenshells: Structural patterns on modal forms in
Adriaenssens, S, Block, P, Veenendaal, D and Williams, C (eds), Shell Structures for
Architecture: Form Finding and Optimization, Routledge, London, pp. 195210.
Mueller, C, Irani, A and Jenett, B 2013, Additive Manufacturing of Structural Prototypes for
Conceptual Design, Proceedings of the International Association for Shell and Spatial
Structures (IASS) Symposium, Brasilia, Brazil.
Schwartz, T 2013, HAL: Extension of a visual programming language to support teaching and
research on robotics applied to construction in Brell-okcan, S and Braumann, J (eds), Rob|
Arch 2012 Robotic Fabrication in Architecture, Art and Design, Springer-Verlag, Vienna,
pp. 92101.
Tam, KMM, Mueller, C 2015, Stress Line Generation for Structurally Performative Architectural
Design, Proceedings of the Association for Computer Aided Design in Architecture
(ACADIA) Conference, Cincinnati, Ohio, US.
BUILD-ing the MASS Lo-Fab Pavilion

Dynamo-Driven Collaborative Robotic
Fabrication Workflows for the Construction
of Spatial Structures
Nathan King, Nathan Melenbrink, Nick Cote and Gustav Fagerstrm
Abstract This project-based paper describes the iterative design, structural optimization, and fabrication of the experimental grid shell structure developed for the
MASS Lo-Fab pavilion. In this case, formal complexity is resolved through
functional complexity that emerges in both elements of the structural systemthe
node and the strutthat each maintain a level of simplicity appropriate to
respective manufacturing processes and material properties. The structure was
fabricated using state-of-the art collaborative robotic fabrication techniques and a
combination of traditional craftsmanship and computationally driven manufacturing
processes. In order to move from the computational design environment to one of
material, the team worked in collaboration with AutodeskTM to develop a novel
design-to-robotic fabrication workflow using the emerging visual scripting interface
Dynamo. A custom robotically assisted welding process was developed to assemble
1880 steel parts making up 376 nodes that saved over 3 weeks of labor when
compared to traditional processes.
Keywords Robotic fabrication Collaborative robotics

Robotic welding Dynamo Grid shell Design robotics
Digital fabrication
N. King
MASS Design Group, Center for Design Research-Virginia Tech School
of Architecture + Design, Blacksburg, Virginia, USA
e-mail: natlking@vt.edu
N. King N. Melenbrink
The United Nathans, Boston, Massachusetts, USA
e-mail: nathanmelenbrink@gmail.com
N. King N. Cote (&)
Autodesk, Boston, Massachusetts, USA
e-mail: nick.cote@autodesk.com
G. Fagerstrm
BuroHappold, Center for Design Research-Virginia Tech School
of Architecture + Design, New York, USA
e-mail: gustav.fagerstrom@gmail.com
363
364
N. King et al.
1 Introduction
Formal complexity often has implications for structural typology and conversely
structural typology impacts formal complexity (Tomlow 1989). Grid shell structures offer the potential for large unsupported spans; can achieve structural stability
through double curvature; are typically composed of nodes and struts; and have
historically had particularly stringent construction criteria that aim for the formation
of identical elements (Bechthold 2008; Wallner and Pottman 2011). Today, computational design and analysis tools have eliminated many of the design restrictions
that governed previous examples (Knippers and Helbig 2009).
In many cases formal complexity results in custom components that exhibit a high
level of functional complexity (Fagerstrm et al. 2014). In this particular case, individualized steel nodes were fabricated, each having four custom tabs and a single
uniform elementa central cylinderthat provided registration for robotic tooling
and welded assembly. Individualized struts were uniform in section but varied in
length. Each strut had a custom dado in either end to receive the steel tabs from the
welded assembly. On-site, a uniform bolted assembly process was utilized thus
segregating the highly custom to the automated or high craft environments of the shop
and simplifying onsite assembly such that unskilled labor could participate in the
construction with limited training using a codied brand on each strut for guidance.
The MASS Lo-Fab Pavilion was developed during a series of collaborative
charrettes held between MASS Design Group, The Virginia Tech Center for Design
Research (CDR), students from VT Architecture and Industrial Design programs,
and the computational design consultant, The United Nathans (Fig. 1).
Over a period of several weeks students from the School of Architecture and
Design produced building components at the Virginia Tech Research and
Fig. 1 Detail of the MASS Lo-Fab pavilion installed on the Rose Kennedy Greenway in Boston,
MA for the 2015 Design Boston Biennial ( Virginia Tech Center for Design Research)
365
Demonstration facility in Blacksburg, Virginia. Working with designers from

MASS Design Group, a Boston-based architecture rm, and Rudabega, a
Blacksburg-based furniture design and construction rm, the students deployed the
experimental structure on the Rose Kennedy Greenway in Boston Massachusetts as
part of the Design Boston Biennial in July of 2015. This paper describes the
project-based development of the gridshell structure from conception to fabrication
with a particular focus on the newly developed collaborative robotic fabrication
environment and associated Dynamo-based robot motion control plugin that were
developed as part of this research.
2 Structural Form Finding, Optimization and Analysis

The initial toroid form was developed by MASS Design Group, the project
architects, and the projects student collaborators based on programmatic desires.
The guidelines for the global form stipulated that it be a single surface structure in
order to simplify construction and to unify structural integrity as a shell. Further
design constraints included site-specic pedestrian axes to which the form should
be oriented as well as a height restriction of 12 feet. The primary tools for
form-nding and surface discretization were Grasshopper, Kangaroo, and
Millipede. Rather than following a conventional workflow of rst generating a
NURBS surface and then discretizing it, a low-polygon mesh approximating the
topology was rst generated and then subdivided into a diagrid. This diagrid
structure was then deformed according to specic forces using the Kangaroo
physics solver for Grasshopper (Fig. 2).
Fig. 2 Left mesh topology used as input to Grasshopper model. Right real-time visual feedback
from Grasshopper, showing discretization and Finite Element analysis the rainbow gradient
indicates surface deflection where blue is 0 and red is 0.75
366
N. King et al.
While the forces entered into the Kangaroo simulation were guided by an
expected global output, the nal curvatures and discretization patterns were
unknown and left to the described form-nding procedure to determine. The
Kangaroo physics solver was set to include a force of upward thrust, forming the
mesh into a catenoid shell, and as an objective to seek planarity between adjacent
struts, thus leading to higher density and shorter strut lengths at areas of greater
curvature, specically the top of the structure (Pottman et al. 2007).
The workflow was arranged such that the designer could edit the control points
of the input mesh with a corresponding real-time visualization of the physics
simulation, complete with diagrid surface discretization and a Finite Element model
generated with Millipede. It should be noted that the metric from Millipede (in this
case a visualization of deflection) was used as visual feedback, but was not
explicitly entered into any kind of topological optimization. In other words, it
afforded the designer a real-time visual evaluation of the structural behavior in order
to better inform intuitive decisions involved in manipulating the global form of the
pavilion. Once a nal form was proposed, further analyses for deflection, bending,
and torsion were also conducted to ensure the design passed baseline standards. The
nal form and construction detailing was evaluated though non-linear analysis
based on centerline geometry.
2.1
Structural Analysis
3D non-linear analysis was carried out using centerline geometry of the grid shell.
The base conditions of this unique geometry, with inner ring and outer ring, provide
efciency in structural performance. Both inner and outer rings would be anchored
to the ground using 3/16 in. rebar irons, which were placed adjacent to joints to
minimize bending in the ring beam steel plates. No further foundation work was
deemed necessary as there was minimal risk of uplift or overturning due to the high
ratio of self-weight to wind load (this structure is over 70 % permeable). Due to the
hyperstatic nature of the structure, a decision was made to size all members and
joint steel according to the highest load occurring throughout the structure (Kilian
and Ochsendorf 2005). This decision was further reinforced by the high probability
of visitors climbing on the structure. Because of the high efciency of the gridshell
layout, where there was very little bending or shear, joint fastening could be
designed based predominantly on keeping members in tension and compression.
For this reason, two bolts per connection, placed perpendicularly to the member
direction, proved sufcient. Bolts were tted with washers on both sides to alleviate
stress in the wood immediately surrounding the bolt holes. Members were oriented
normal to the notional design surface throughout to ensure maximum cross sectional efciency in the plane where any bending would occur. This layout of the
members also brings the additional benet of minimizing the rotation of the dado
slot receiving each joint plate.
367
3 Dynamo-Driven Robotic Motion Control

Recent developments in the use of graphical editors for the development of robot
control have involved closed, unidirectional processes and only a few rigid, albeit
simple, tools. In the past 5 years a number of integrated design-to-robotic fabrication workflows have provided many designers and researchers with ready access
to robotic control and reduced the need for programming to the point where it is no
longer required for machine access (King et al. 2014). These tools (HAL, Robots.io,
DRG-PRG, and others)almost entirely relegated to a Rhino-based design environmenthave begun to exhibit complications that rival even the proprietary
motion control packages provided by the robot manufacturers. As tool developers,
we are left searching for a middle groundmaximum functionality with minimum
complexity.
Dynamo is an open source, graphical algorithm editor that enables users to
construct generative tools according to their own specications. Dynamo offers a
robust selection of nodes that may be arranged into custom graphs. These terms
respectively refer to elemental (low-order) and composite (high-order) operations in
a Dynamo workspace with the graph being an arrangement of nodes connected by
wires relating input and output. Nodes tend to be functionally agnostic; they are
specic, basic, and versatile such that their combination avoids predened outcomes. While these notes would suggest that a Dynamo workspace demands
somewhat greater interaction than comparable softwaresuch as Grasshopperit
exposes advanced functionality while avoiding totalizing simplicity.
To take advantage of this functionality the robotic control workflow integrates
with the advanced functionality of the Dynamo core in order to control the production of Rapid code: the language used for the ABB robotic manipulators (or
other robot-specic syntax). The workflow uses a custom library of zero touch
nodes developed in C# as a plug-in for Dynamo whose names, inputs, outputs, and
methods correspond with Rapid methods. The nodes are organized into four classes
target data, program data, constants, and instructionsthat, when wired properly
as a graph, output PRG les to a workstation, virtual controller, or external drive.
The workflow allows users to:
1. Arrange data types, instructions, and functions as in Rapid using the Dynamo
core, utilities, and le-writers.
2. Interface with Dynamo, the virtual controller, and workstation to communicate,
edit, and enter the following: target data, position data, tool data, work-object
data, and program les.
3. Manage multiples of this data per program le.
4. Conduct Run Modes, Number Sliders, and While Loops in the Dynamo
workspace to populate a destination with PRG les for production.
5. Parametrically generate program les in Rapid syntax using Dynamo.
6. Interface with the virtual controller in RobotStudio for simulation and data
entry/retrieval.
368
3.1
N. King et al.
Robotic Programming Procedure for Collaborative

Welding
The purpose of each tool path was to weld four steel flanges onto a steel cylinder
assuming that each must rotate in the same direction; to clear the table while
rotating; to not over-rotate; and to be time efcient. Numeric data, which specied
four flange angles for each of the 376 steel cylinders, was imported from a CSV le
and referenced in Dynamo. With each of the four values the sixth variable of a
known joint conguration was replaced indicating the position of the sixth axis of
the robot. At each of these joint congurations, the cylinder attached to the end
effector would be in position to weld a flange. Before and after each of these
positions was another joint conguration assigned such that each flange would
correspond with three key operations: raising the assembly above the work surface,
rotating the sixth axis into position, and then lowering the assembly to the work
surface. The rotations were cumulative and totaled to *270 per assembly. A joint
conguration which positioned the sixth axis at 0 and the initial assembly beneath
the table was inserted at the beginning of the list to avoid cumulative over rotation.
These congurations were wired to a joint target node (JointTargetAtVals) and
subsequently to a movement instruction node (MoveAbsJ). The movement
instruction node also received speed-data, zone-data, set name, tool name, and
work-object name. Two nodes for program data (ToolAtVals and WobjAtVals)
were dened with tool and work-object conguration data gathered from the
workstation. The outputs of these nodes all connected to a PRG writing node
(CreateRapid) which, when the graph ran, produced a le for production. Each
flange group was indexed to a value on a number slider that interacted with the
automatic run-mode in Dynamo to populate the destination drive with all 1,504
unique positions in a matter of seconds.
4 Fabrication
4.1
Codication and Part Organization
An organizational system was developed that provided all construction data within
the individual parts so that much of the fabrication and assembly could be conducted
without any shop drawings (Fig. 3). For each node a series of four tabs were custom
laser-cut and etched with an index number denoting (node#_flange#) and the process
was executed from flange zero (node#-0) to three (node#-3) in order to maintain
consistent organization of the nearly 2,000 steel parts throughout fabrication and
assembly. A uniform brand was burned into each rough-cut wood member to
organize strut data and to communicate relational data corresponding with each steel
flange. The H brand organized the following data that drove the nal stages of the
wood fabrication and ultimately the assembly on site. All fabrication data was
369
Fig. 3 Photograph and diagram showing the labeling system used to guide all manual fabrication
and assembly ( Virginia Tech Center for Design Research)
gathered and organized through a model-linked CSV le that fed a cloud-based

spreadsheet. First, the strut number was notated, followed by the strut length
rounded to the third decimal place, and on either side of the marking, a flange
association and dado angle was placed. Once codied, the rough-cut lengths contained all the manufacturing and assembly data needed to complete the structure.
4.2
Computer Assisted Craft-Based Strut Production
Each strut had a uniform 2.5 2.5 section but varied in length and end condition.
Drawing from the part details embedded in each brand, an automated compound
miter saw was used to cross-cut precise lengths and the dimensioned members were
then drilled uniformly to accommodate the bolted assembly. Using the data
embedded in the part, each 3/8 dado was located and cut to 3-1/8 depth using a
single table saw pass. These parts were then nished and organized into groups for
shipping and assembly.
4.3
Collaborative Robotic Node Fabrication
To facilitate the robotically assisted welding operation, a multi-robot

human-machine collaborative work cell was developed consisting of two inverted
ABB IRB 120 s mounted beneath the work surface (Fig. 4). A custom end effector
was developed consisting of a 3-jaw chuck capable of gripping the inner or outer
wall of the 2 steel pipe that remained consistent throughout all nodes. The end
effector was isolated from the robot using a phenolic extension to ensure no electrical interference from the grounded welding table. In this case the gripper was
manually engaged to account for deviation in pipe dimension but in other
370
N. King et al.
Fig. 4 Collaborative robotic work cell developed at the Virginia Tech Center for Design Research
for robot-assisted welding of the steel grid shell nodes. ( Virginia Tech Center for Design
Research)
applications, precision materials or force sensing could adapt to a fully automated

production strategy.
The end effector was positioned to pass through the work-plane to facilitate
welding while protecting the inverted industrial robots mounted beneath. Here, the
Fig. 5 Robotically assisted welding. Located beneath the work surface, the end effector holds the
steel pipe section. ( Virginia Tech Center for Design Research)
371
robotic arm is completely separated from the user, enabling safe collaborative
interaction (Fig. 5).
To amplify safety and to achieve enhanced quality control a machine operator, in
addition to the welder, managed individual node les and maintained manual
enabling operation. Each team of two fabricatorswelder and operatordeveloped
a call and response cadence to verify that the program pointer, the manually placed
steel tab, and the robot position all correspondeda strategy that eliminated all but
a few placement errors. In a comparative analysis it was determined that the collaborative robotic fabrication of the steel nodes reduced manufacturing time from
39 to 3 min per nodea total savings of 225 labor hours at a sustained production
rate. This savings ultimately accounts for an overall reduced labor time of
approximately 3 weeks when accounting for teams of two fabricators.
4.4
Assembly Logic
Once fabricated, all parts were packaged in groupings based on their respective
position on the structure. Onsite, these packages were catalogued and organized
about the perimeter of the structures base for use during assembly. The assembly
sequence was established to maintain a geometry that was continuously
self-supporting, moving from the outer ring of the structure inward (Fig. 6). This
strategy allowed for concentric assembly without the need for sintering or formworkonly limited shoring was used for added safety. The bolted connection
allowed zero tolerance between the bolt (pin) and the wooden strut, but holes in the
Fig. 6 Aerial view of the completed pavilion installed on the Rose Kennedy Greenway in
Boston MA, July 2015 ( MASS Design Group)
372
N. King et al.
steel allowed for 1/16 tolerance to resolve any site variation or t issues between
parts.
5 Conclusion
The MASS Lo-Fab pavilion enabled close collaboration between design practice
(MASS), academia (Virginia Tech), and industry (Autodesk and Rudabega). This
type of collaboration is increasingly important to realize innovation in the
Architecture, Engineering, and Construction (AEC) industries. From a practical
perspective the grid shell represents a structural typology that can be used in
resource-limited settings where long span members are not available (Otto et al.
1996; Ahlquist and Flesichmann 2009). By positioning a need that emerges from
design practice in the context of applied academic research new opportunities
emerge. This project served as the platform for advanced research in robotic fabrication, the development of novel computational workflows, and established new
partnerships that will help bring emerging Design Robotic automation strategies
closer to application in AEC industries.
The construction resulted in the fabrication of 376 custom nodes consisting of
1880 parts and 720 individual struts consisting of 1440 individual end conditions
through a combination of high-craft and high-tech fabrication strategies. On-site,
strategic assembly instruction embedded in each component allowed for construction by volunteers, students, and staff having limited to no construction
experience. During the project two new software tools were developed including
the discussed Dynamo-to-robotic fabrication workflow and an emerging Autodesk
FusionTM-based kinematic simulation environment. These tools will be further
developed and tested during the Dynamo-BUILD workshop at the 2016
International conference on Robotic Fabrication in Architecture, Art, and Design
conference in Sydney, Australia.
Acknowledgments This project-based research was conducted with the support of Autodesk, the
Autodesk BUILD Space; Virginia Tech College of Architecture and Urban Studies, Virginia Tech
School of Archtiecture + Design, Center for Design Research and The Institute for Creativity,
Arts, and Technology; the Rose Fitzgerald Kennedy Greenway Conservancy; The United Nathans;
and Rudabega. Project collaborators include Adam Allard, Alan Ricks, Ashleigh Otto, Brendan
Kellogg, Chip Clark, Cole Smith, Conor Byrne, David Barrett, David Scurry, Ed Coe, Giorgia
Cannici, Gustav Fagerstrom, Jason Zawitkowski, Jeff Snyder, Jonatan Anders, Jonathan Rugh,
Justin Lavallee, Kyle Barker, Mark Leach, Martin Philipp Angst, Michael Murphy, Mike Steehler,
Nathan King, Nathan Melenbrink, Nick Cote, Nikki King, Paul King, Robert Dunay, Steve
Bickley, and Victoria Smith. Mike Dewberry, Matt Jezyk, Ian Keough and others who contributed
to the development of the Dynamo-to-robot workflow.
373
References
Ahlquist, S and Fleischmann, M 2009, Cylindrical Mesh Morphologies: Study on Computational
Meshes Based on Parameters of Force, Material, and Space for the Design of Tension-Active
Structures. Computation: The New Realm of Architectural Design, Proceedings of Education
and research in Computer Aided Architectural Design in Europe (eCAADe) Conference,
Istanbul, vol. 27, pp. 39-46.
Bechthold, M 2008, Innovative Surface Structures: Technologies and Applications. vol. 2, Taylor
& Francis, New York.
Fagerstrm, G, Verboon, E and Robert A 2014, Topo-faade: Envelope design and fabrication
planning using topological mesh representations, in Gramazio, F, Kohler, M and Langenberg,
S (eds), Fabricate: Negotiating Design and Making, gta Verlag, Zurich.
Kilian, A and Ochsendorf, J 2005, Particle-Spring Systems for Structural Form-Finding, Journal
of IASS, vol. 46, no.147.
King, N, Pender, S, Grinham, J, Vroman, R and Clark, D 2014, Beyond Digital Steroids: a
Pedagogical Approach to Foundation Design Through Design Robotics, Proceedings of the
2014 International ACSA Conference, Seoul Korea.
Knippers, J and Thorsten H 2009, Recent developments in the design of glazed grid shells, in
International Journal of Space Structures vol. 24 no.2, pp.111-126.
Otto, F, Schanz, S, and Rasch, B, 1996, Finding Form: Toward an Architecture of the Minimal,
Axel Menges, Fellbach.
Pottmann, H, Asperl, A, Hofer, M and Kilian, A 2007, Architectural Geometry, Bentley Institute
Press, Exton, PA.
Tomlow, J 1989, The Model, Information of the Institute for Lightweight Structures, Stuttgart,
vol.34.
Wallner, J and Pottman, H 2011, Geometric Computing for Freeform Architecture, in Journal of
Mathematics in Industry vol.1, no.4.
Part IV
Workshops
Robot UI
User Interfaces for Live Robotic Control
Curime Batliner, Michael Jake Newsum and M. Casey Rehm
Abstract Through the development of user interfaces that leverage real-time

control, the robot emerges as a design platform where programming, simulation and
execution collapse into a singular act in time. This reduction of the typical robot
workflow allows design processes to continuously engage with adaptive contexts
whether they be deformations of material, nuanced data or instantaneous design
input. The case studies, presented in this paper, demonstrate design potentials for
developing new interfaces, where the digital and physical are mutable, letting
designers intuitively engage with matter and representations in flux through robotic
interactivity and autonomous agency.
Keywords Real-Time
User interface
Design research
Autonomous agency SCI-Arc Robot house
Interactivity
1 Context
1.1
Physical Interface
SCI-Arcs conceptual approach to design cannot be separated from the development of software. Since the beginning of the Robot House, students and faculty
have been working on custom design interfaces by integrating robotic motion
control into existing design tools that architects and designers use every day in
school and practice. Bespoke plugins for Maya and Grasshopper have reduced the
C. Batliner (&) M.J. Newsum M. Casey Rehm
Southern California Institute of Architecture (SCI-Arc), Los Angeles, USA
e-mail: curime_batliner@sciarc.edu
M.J. Newsum
e-mail: jake_newsum@sciarc.edu
M. Casey Rehm
e-mail: michael_rehm@sciarc.edu
377
378
C. Batliner et al.
learning curve of the robot as a technology and have allowed designers to integrate
the robot into their design process. Treating the robot as a physical design interface,
where real world feedback is introduced into this digital process, the digital model
serves as a point of departure rather than an idealistic goal (front image).
Eventually, the model as a design tool can be removed altogether (Atwood 2012).
As a consequence, the limits of where design stops and production starts are
becoming increasingly intertwined.
While signicant efforts have been made for narrowing the gap between digital
and physical, the research is inherently limited by the hierarchical structure of
offline programming where the robot is being constrained to a predened timeline
or sequence. With this model, synchronization with other machines and digital
devices happens at predetermined positions in the sequence. Feedback is limited to
observation of the particular process. Errors in the underlying logic, the sequencing
of the robot, are fatal as once programmed it cannot be altered while the program is
running. Since motion planning and runtime are separated in time, many experiments need to be reset resulting in a linear way of working. While in this scenario,
the design process can be iteratively rened and calibrated, yet design processes
ultimately remain constrained by the sequence.
The ability to alter a sequence, during runtime, narrows the gap between digital
and physical workflows which signicantly increases speed in project development.
More important than speed is the potential of opening up new research trajectories,
which require alternative modes of control that break away from a static sequence.
The ability to alter the robots logic in runtime can change how we design, since it
establishes an intuitive interactive bridge between the digital and physical. This
platform, where information can be transferred at any time in both directions,
allows the user to reshuffle, reverse, replay, layer and combine sequences. The
digital environment gets access to the physical as the physical gets access to the
digital, and new design processes, which require simultaneous data flow in both
directions, are possible.
1.2
Software Mediation
Intelligent real-time software has saturated both contemporary design and, more
importantly, society. Our methods of interfacing our world are increasingly mediated by software, which interpret, analyze and sometimes adapt to our intentions.
The sophistication of software, along with its integration with hardware, allows it to
slip seamlessly into our daily tasks and produce an intuitive extension of intentionality into ever expanding territory.
Robot UI
379
The mouse, joystick and keyboard have dominated how we convey intentionality to computers. For designers, especially digital artists, this is not always
intuitive. Interfaces, such as the Wacom tablet and touch screens, include drivers
that have been tuned to effectively translate perceived human motion into useful
movements within digital environments for vector and pixel graphics. Through
additional software packages, the user can further dene the translation of these
moves. This model of the interface extends beyond purely simplifying the translation of human inputs into effective outputs. It allows the user to manipulate these
translations towards amplied effects.
Without physical registers, touch inputs are imprecise, thus technology must be
forgiving to create a pleasing user experience. Software, such as Swype, adapts and
learns the users tendencies by creating bespoke gesture libraries for the user.1 This
signicantly increases speed and accuracy while enhancing user experience. The
speed with which it adapts and performs sophisticated tasks allows for it to
effectively disappear in the user experience.
Mediating software applications have recently gained more popular attention
through the release of the Google Deep Dream to the public. The underlying
software was originally developed to explore the use of deep learning neural
network architectures to identify specic elements in images to improve
image-based search.2 Google Deep Dream exposes the generative potentials of
malappropriation of these tools. Digitally generated images, similar to experiences
of pareidolia, are made when the image search is applied to images that lack the
specied elements. The software probabilistically transforms the image by amplifying what is most similar to those absent elements.
2 Live
Live, a real time motion control platform developed for industrial robots paired with
a series of novel interfaces, aims to open up the robot for nonlinear processes, by
coupling sensing devices and advanced programing with robotic technologies.
This entails rethinking the directionality of the standard unidirectional workflow
and encourages hybrid modes of digital practices where simultaneous visualization
and material manipulation inform the design process (Bard et al. 2014).
The Live platform is an expedient for the exploration of interfaces that intuitively translate human inputs into robotic motion, the interplay between human
inputs and analysis driven robotic interpretations, and interfaces for developing
robotic agency.
https://gigaom.com/2012/06/20/nuance-swype-living-keyboard-predicts-learns/ Accessed August

8, 2015.
2
http://googleresearch.blogspot.com/2015/06/inceptionism-going-deeper-into-neural.html
Accessed August 8, 2015.
380
C. Batliner et al.
Live utilizes the robots ability to send and receive information, through timed
synchronous tasks, which gives the robot the capacity to continuously engage with
current context available to the designers programming logic. Incoming information is processed into motion, tooling and conguration commands. The designer is
given the ability to add information for later use as well to intervene or alter the
robotic operations at any time. This programming method has alternative goals
where the robot is no longer programmed to be optimized for repetition and precision. Instead, the robot is programmed to be versatile, nuanced and interactive
(Batliner et al. 2015b).
This removes the constraints of offline-programmed operations that need to
anticipate the contextual information through simulation and contextual restrictions.
Online Programming encourages design processes to move away from the idealized
static CAD model to a context aware model where a shifting environment is
anticipated. Inputs can be dynamically adapted through external sensors, and
events. At any time the design processes are described through a model of adaptable
recongurable relationships.
3 Case Studies
3.1
Tango
Tango is a dance between two performers, a human and a robot. Using a Microsoft
Kinect, the humans movements are tracked and translated. Whenever the human
moves, the information is sent to the robot in real-time. The robots movement is
instantly activated by the control platform Live. While the program runs real-time,
the performance itself gets out of sync. This gap never fades completely. With some
training from the human and the possibility to adjust to how the system processes
and translates information in real-time, the gap becomes an intuitive part of the
interaction. Human movement varies between individuals, therefore the user
interface needs to be easy to learn, while being recongurable, on the fly. While the
robot gives feedback to its human partner with its body, a visual interface mapped
onto the walls gives the human additional information about the internal workings.
The visualization includes the real-time representations of the human and robot
bodies, so the participant can understand how each step will impact on the performance. Additional information, such as speeds, position and ltering, communicate the robots limitations to the human, helping them understand if they need to
slow down or if the robot responses can be adjusted (Fig. 1).
Robot UI
381
Fig. 1 Tango: interactive spatial interface
Fig. 2 Tango: interactive network
All lters and damping functions are adjustable on the fly by the operator, and
some variables, such as speed and pausing, are also made available to the performer
through body gestures (Fig. 2). The choreography, from a programing perspective,
is set up as a master-slave condition. Yet, the performance appears as a real dance
between equal partners, since the robots body influences the humans behavior,
382
C. Batliner et al.
which closes the interactive loop. Additionally, the robots motion can be recorded,
replayed and modied letting the user choreograph the robot in real-time for later
use. The flexible interface allows the human to tune the relationship granting them
mediating agency in this spatial interface.
3.2
Automaton
Utilizing Henri (Batliner et al. 2015a) in conjunction with Live, an application was
developed in which the platform attempts to perceive potential features in a blank
mass and then iteratively amplify those features (Fig. 3). In this project, the human
intervention occurred at the level of designing the intelligence of the platform to
analyze its context and then effectively manipulate it. This type of application has
uses in creating production methods for dynamic material or for complex contexts
in which direct human intervention or analysis is prohibitively difcult.
Unlike in Augmented Materiality (Johns 2014), a project which also establishes
a live link between sensor material and robot with a totalized goal of a rationalized
physical output, in the above example the robot analyzes recursively the wax
monolith for potentially conical regions, and then creates tool paths to amplify these
features until a maximum limit on material porosity is achieved to avoid structural
collapse (Fig. 4). In these conditions, it is necessary to generate design through the
codifying of intention at the level of local behaviors (Batliner et al. 2015b).
Fig. 3 Automaton: autonomous cone sculpting behavior
Robot UI
Fig. 4 Automaton: interactive network
Fig. 5 Draughtsman: interactive drawing interface
383
384
C. Batliner et al.
Fig. 6 Draughtsman:
interactive network
3.3
Draughtsman
The drawing series executed with Henri and Live synthesizes the approaches
explored in Tango and Automaton. Multiple levels of interactivity, between
interaction and production, are quickly explored in this 2D setting. Three basic
versions of the application have been developed that parallel ranges of involvement
between the human operator and the intelligence of the interface discussed in the
Real-Time section of the text. The rst iteration of the application utilizes a tablet
interface to translate human pen marks into robot painting behaviors (Fig. 5). In this
version, the underlying algorithms for translating inputs into motion are tuned to
make the real-time control of the robot, through the pen, completely seamless and
intuitive. Filters and analytic algorithms are used to interpret and rene potentially
noisy inputs into velocities and paths that have a clear scalar relationship and
linearity towards the designers intention.
The second iteration of the application begins to push the agency of the softwares intelligence forward hierarchically. Generative algorithms create a collaborative relationship between the software and the designer. In this case, the operator
makes a mark on the page (Fig. 6). The robot platform utilizes a web camera to
analyze the marks for brightness levels. Mapping the pixel information of the
camera to a grid, the robot manipulates the captured mark using a growth algorithm.
It then adds its interpretive marks to the painting. In this way, the nal composition
becomes a negotiation between the user as artist and the designed machine
intelligence.
Robot UI
385
Fig. 7 Mediated sculpting interface
In a similar approach to the misuse of the image analysis in the Google Deep
Dream series, a third iteration of the app was developed.3 It expands the analytics
package in the above example, to adapt its brightness and contrast settings until it
falsely perceives shaded areas on a blank page of a similar quantity to the human
marks in the second example. With this interface, the machine takes on greater
autonomy like in Automaton, and the users intentionality must be coded into their
design of the platforms agency.
4 Conclusions
Real-Time robotic interfaces as design methodologies are still new territories at
SCI-Arcs Robot House. The use and appropriation of input sensors and output
tooling are still in their infancy. Live interfaces continue to be prototyped for a wide
range of working modes where designers can approach robotics as an intuitive
extension of their own design processes.
http://googleresearch.blogspot.com/2015/06/inceptionism-going-deeper-into-neural.html
Accessed August 8, 2015.
386
C. Batliner et al.
Using the Live platform, digital and analog methods converge through real-time
inputs and outputs between both worlds where the robot is the mediating device.
These new interfaces allow human skills to be transferred directly into informed
robotic motion controls. This workflow moves designers and craftspeople away
from offline programming through the integration of gestures and sensor embedded
smart tools that interact with robot collaborators in high skill applications (Bard
et al. 2014). Moving to trajectory based models, hands can drive the robot intuitively, through vision inputs like low delity devices such as the Leap Motion or
the Xbox Kinect. These digital input devices can be tuned through the interfaces to
perform as physical emulations of brushes that are found in design software such as
Photoshop or ZBrush. The robot can then be situated as a physical modeling
program where the designer can work directly with real material properties, so
design and fabrication emerge as one continuous act (Fig. 7).
When the designer wants to carve or draw with their hands, they can engage the
material directly. Conversely, when the designer needs parametric patterns or
machine precision, the designer can leverage parametric controls. This allows each
component of the design process to inform the other from the onset. Recognizing
the co-dependency of these elements, the process cannot proceed without the
simultaneous cooperation of its four players: the human designer, the robotic
manipulation, the computer simulation and the material reaction (Johns 2014).
In this context, digital designers do not need to simulate material affects to try to
anticipate the fabricated results. Instead, their assumptions can be tested and calibrated simultaneously. The time distance between output, evaluation and modication are reduced to milliseconds, so that design intentions can be expressed
without compromise. Vision inputs can be used to monitor material in real-time,
thus feeding information into algorithms that analyze the material properties to
output updated paths according to codied design intentions. This opens the possibility of correcting processes that the robot can layer on top of an existing
operation, either autonomously or after notifying the user.
This approach lets digital designers engage with less deterministic processes.
These novel interfaces aim to layer intelligence, mediated through software, with
manual intuition. Equally, it invites skilled craftsmen to digital making, through
sophisticated human-robot collaboration models based on sensory inputs. This
builds a common base for exchange and collaboration amongst designers of different skill sets and machines. Here, design and making is not a one directional path
with an initially determined end. It becomes a multidirectional open-ended process.
Robot UI
387
References
Atwood, A 2012, SCI-Arc Presents Film, Symposium for Advances in Architectural Geometry,
Centre Pompidou, Paris.
Bard, J, Gannon, M, Jacobson-Weaver, Z, Jeffers, M, Smith, B and Contreras, M 2014, Seeing is
Doing: Synthetic Tools for Robotically Augmented Fabrication in High-Skill Domains,
ACADIA 14: Design Agency, Los Angeles, California, pp. 409-416.
Batliner, C, Newsum, MJ and Rehm, MC 2015a, Real-Time Platform for Robot Design
Interfaces, Proceedings of the eCAADe Conference, Vienna, Austria.
Batliner, C, Newsum, MJ and Rehm, MC 2015b, A Creative Platform, Proceedings of the
Robotic Futures Shanghai Biennale, Shanghai, China.
Johns, RL 2014, Augmented Materiality: Modelling with Material Indeterminacy, Fabricate
Conference Proceedings, Zurich, Switzerland.
Towards On-site Collaborative Robotics

Voice Control, Co-speech Gesture and
Context-Specic Object Recognition
via Ad-hoc Communication
Thibault Schwartz, Sebastian Andraos, Jonathan Nelson,
Chris Knapp and Bertrand Arnold
Abstract This work presents a novel set of accessible and unied hardware and
software solutions that facilitate the implementation of natural human-machine
interactions, as required by collaborative robotics in both indoor and outdoor
environments. This extensible framework supports vocal control, co-speech gestures, and object recognition with feature tracking and adaptive resolution. The
interactions are based on a new network messaging protocol that allows any device
using TCP/IP to share variables with the full abstraction of the original machine
software platform and can therefore be used synchronously by a vast array of
equipment including CNC machines, industrial robots, construction equipment,
mobile devices and PLCs. We conclude with the description of a testing scenario to
be deployed during the conference workshop.
Keywords On-site worker assistance

Collaborative robotics
communication Advanced optical sensors
Multimodal
T. Schwartz (&) S. Andraos

HAL Robotics, London, United Kingdom
e-mail: t.schwartz@hal-robotics.com
S. Andraos
e-mail: s.andraos@hal-robotics.com
J. Nelson C. Knapp
Bond University, The Abedian School of Architecture, Gold Coast, Australia
e-mail: jnelson@bond.edu.au
C. Knapp
e-mail: cknapp@bond.edu.au
B. Arnold
Soundisplay, London, United Kingdom
e-mail: arny@soundisplay.com
389
390
T. Schwartz et al.
1 Introduction
1.1
Collaboration Context
The present prevalence of new tools and techniques provides an opportunity for the
paradigm of the architect to evolve. At Bond University the robotic curriculum is
aimed at exploring and nurturing this possibility, with particular focus on collaborative robotics as an evaluative territory in which to conceive new ways of
working and thinking. HAL Robotics, with their HAL software suite (Schwartz
2012), are aiming to create an accessible vendor-agnostic solution for the simulation, programming, and real-time control of robots and peripheral devices, while
Soundisplay develops high-performance, acoustic and optical sensing technologies
for entertainment and various industrial applications, including robotics.
It is our shared belief that collaborative robotics represents the next phase of the
intersection between design and constructionan undoubtedly fertile space where
many architects and designers have the distinct opportunity to take more control over
not only their work but also the methods and processes with which it is made.
Exercising this control is seen to facilitate a fuller understanding of design processes
and the decisions that help drive these processes from both ends: how construction
works backwards to affect design and how design moves forward to affect construction.
1.2
Technical Context
In-situ applications of roboticswhether applied to fully automated construction

scenarios, or on the contrary envisioned as intelligent, service-oriented, distributed
manufacturing systemshave remained an ongoing research topic since the early
1980s. Beyond the various sociological and economic challenges surrounding the
application of such innovative processes, major technical issues still need to be
addressed in order to ensure the compatibility of past and future research results
with the requirements of the construction industry. The authors have identied three
of these bottlenecks which, amongst the large scope covered by this research eld,
appear to directly impact the progress of efforts led by actors from the Architecture,
Engineering & Construction (AEC) industry:
The absence of generic communication protocols allowing the exchange of
structured data between industrial machines and the highly limited abstraction of
the alternatives currently proposed by research consortiums (Schwartz 2015).
The inability (and lack of interest) of robot manufacturers to provide intuitive
control interfaces for the full range of their products even though the quickly
growing market of new generation, low-payload robots has clearly proved
customers enthusiasm for such improvements.
The incompatibility of most sensing technologies with intensive outdoor
applications such as construction sites, which are often radically different to the
391
traditional industrial environments which these technologies have been primarily developed for.
Instead of attempting to solve these problems individually, the authors decided
to tackle them simultaneously, thus enabling various levels of interaction to be
natively linked to each other and minimizing the risk of subsequent technical
limitations resulting from design omissions.
This work presents a novel set of accessible, unied hardware, and software
solutions facilitating the implementation of natural human-machine interactions, as
required by collaborative robotics in both indoor and outdoor environments. This
extensible framework supports vocal control, co-speech gestures, and object recognition with feature tracking and adaptive resolution. The interactions are based on a
new network messaging protocol allowing any device using TCP/IP to share variables
with the full abstraction of the original machine software platform and can therefore be
used synchronously by a vast array of equipment including CNC machines, industrial
robots, construction equipment, mobile devices and PLCs. We conclude with the
description of a testing scenario to be deployed during the conference workshop.
2 Generic Robot Communication Protocol

In order to harmoniously account for context and operator information, robots need
the ability to manage and exchange large amounts of data with additional devices.
Traditionally, such a machine is often linked to a cell equipped with a few sensors
to take simple events into account, such as the intrusion of an operator into the robot
workspace during the execution of a task triggering an alternative behaviour (reduced motion speed, immediate halting of program execution etc.).
The content shared by systemic components in these situations is extremely
basic: switches return digital, electric signals while other sensors return digits in the
form of an electrical current. Ultimately this means that any data is tightly linked to
its electrical reproduction. Collaborative robotics requires much higher abstraction
and density of data: the most basic vision application will already deal with messages 3 orders of magnitude larger but smart sensors can necessitate even bigger
amounts of data to be processed.
It is obvious that there is a need for radical change in industrial
machine-to-machine communication systems that can move beyond simple signal
sharing via PLC towards peer-to-peer communication on robust eldbuses. Such
scenarios have been anticipated since the mid-1980s, but the very late and
restrictive standardization of these protocols has left industrial networks a jungle of
low-level, low-bandwidth networks. A few modern solutions, mostly based on the
Ethernet technology stack, are now slowly replacing these old infrastructures and
are a very interesting vector of modernization of both networks and endpoints as
they support various protocols already in use for internet communications such as
TCP/IP. Allowing a virtually unlimited number of connected devices (millions, in
comparisons to dozens or hundreds at best for older systems in use), they enable the
392
T. Schwartz et al.
development of technologies required by specialized machine networks, such as

those that could animate building sites in a not-so-distant future.
Considering this evolution, the toolset presented in this document is based on an
application layer protocol (Schwartz 2015) with interesting properties:
It allows peer-to-peer communication of structure data, based on an easily
modiable object dictionary.
It minimizes network load, by using extremely lightweight messages.
It is very easy to use with existing binary serialization utilities provided by
general-purpose programming languages.
It is platform and vendor agnostic.
In the context of this study, the aforementioned protocol is used to normalize the
communication between smart sensors and industrial robot controllers in order to
deploy vocal control and three-dimensional machine vision to such machines via
phones, computers and other Human-Machine Interfaces (HMIs) (Fig. 1).
3 Applications Developed with Smart Sensors

3.1
Intelligent Contextual Multi-environment Sensing
3D sensing appears to be the ideal platform to offer real-time control, feedback, or

assisted operation in any environment. However, many challenges make current 3D
technologies impracticable in most situations. The technical limitations of 3D
sensing, mostly camera-based systems are numerous: from random behaviour under
Fig. 1 Example of connectivity allowed by the proposed sensor toolset
393
Fig. 2 Techisplay 3D-I single tile controller platform linked to an ABB IRB1200 via HAL
light varying conditions; unacceptable resolution/lag compromise; unrealistic

post-processing computation needs; non-deterministic shape recognition success
rate; non-deterministic timing; or simply a slow sensing rate. Additionally, a robot
cell needs much more than 3D sensingideally it needs security sensors, calibration, telemetry, motion tracking, feedback control, and machine vision with shape
recognition or human machine interfaces. This results in unrealistic investments and
complexity to setup multiple sensing devices that should work seamlessly.
Techisplay 3D-I technology (Fig. 2) is an industrial grade and affordable sensing and
control platform that overcomes most of these problems. This new type of optical
sensors benets from exceptional light radiations sturdiness and can work from pitch
black to outdoors lighting (even working under direct varying sunlight exposure) with
steady specications. The sensing is operated at speeds exceeding 1000 frames per
second which permits sensing of objects in motion and even vibrations. The new type of
sensing data generated allows a different algorithmic treatment, resulting in
ultra-low-latency three-dimensional shape recognition with a success rate approaching
the theoretical 100 %. The sub-millimetre resolution can be even ner when varying
sensing timing for critical operations such as calibration or precision assembly.
The ability to use a generic 3D sensing platform allows the fusion of calibration
and telemetry (real spatial coordinates), human machine interfacing (joystick simulation, gestural interface), machine vision (shape recognition and feature tracking),
motion tracking (trajectories tracking and prediction), vibration sensing, impact
sensing, and security (both using shape recognition and context). The parameterization of 3D sensing and 3D control as opposed to point cloud, pixel based
acquisition, statistical analysis or post-processing, is to simplify the use of sensors.
In a professional environment it transforms a flow of raw information into a legible
stream of controls and parameters, usable by both machine and human users.
3.2
Noise-Free Vocal Control Based on Natural Language
Vocal recognition, as supported by advances in articial intelligence, is becoming an

increasingly viable mode of communication between man and machine. In most cases
394
T. Schwartz et al.
the machines in question are our phones, tablets, or computers but the same technology
can easily be extended to robotic manipulatorsthe trick is in the digital interpretation
of our instructions. While audio feature and voice recognition algorithms are extremely promising, they suffer from surrounding noise problems when used in real
conditions. That affects recognition performance and success rate drastically.
Work in this direction has been signicantly helped by successful research and
drastic improvements in noise-reduction technologies, including Soundisplays O.
D.O (Fig. 3). O.D.O sensor is primarily a novel ultra-sensitive acoustic contact
sensor. It potentially transforms any surface or material it is connected to into an
acoustic sensor. The acoustic sensor has the useful property of ignoring airborne
acoustic waves except when directly applied at very close distance, leading it to
behave as a very unique ultra-proximity microphone that is ideal for audio analysis,
vibration analysis and speech recognition as it allows audio recognition algorithms
to approach their ideal performance.
Beyond these technical sensing aspects, the content of the communication
between humans and machines is open to experimentation. The solution proposed
by the authors to the problem of inter-comprehension is a common language, both
simple enough to be parsed in real-time, yet complete enough to retain a high level
of abstraction for the user (Fig. 4). Dear Mister Robot (DMR) (Andraos 2015) is
just such a language. DMR is a restructuring and simplication of the English
needed to control robots in an extensible, intuitive and robust way that is ideal for
both human-machine collaboration and teaching.
At the heart of DMR is a concise list of keywords and a set of new parts of
speech; presented in full in the DMR dictionary along with denitions tied to the
WordNet (Fellbaum 1998) semantic dictionary and usage examples. An exemplary
DMR phrase could be something as simple as Move Left 12 cm. This sentence is
so easily interpreted by a human lector that to describe its intended effect is almost
redundant. Yet the same cannot be said of a numeric parser. To align a human
understanding of the language with that of a machine all DMR keywords are stored
within a limited set of command types: movement, reorientation, tool action, wait,
stop, and conditional. Each command type has predened parameters, some of
which are required to validate a command, while others are optional but increase the
users control of a given action. For example, in a movement command, such as the
Fig. 3 Techisplay O.D.O contact acoustic sensor and ultra-proximity microphone ( Bertrand
Arnold Soundisplay Limited 2014)
395
Fig. 4 Dear Mister Robot (DMR), parsing of a command
one previously shown, we require only a direction (left) and a distance (12). In fact,
in this case, the addition of units (cm) is optional but its omission is only advised
with an a priori knowledge of the machines settings or when the units are implied
by the commands context e.g. Move Left 12 cm then Down 4, whereby the 4
infers the cm from the previous command.
Raw DMR sentences are sent to a proprietary parsing algorithm to extract all the
useable information and execute commands in real time. While this
command-by-command vocal control is useable, collaboration in this manner is
exceedingly tedious and does not particularly improve upon the current scripting for
control paradigm. DMR, therefore, introduces tasks as a means of containing,
combining, and sequencing commands. Tasks are the centre point of DMRs
extensibility allowing users to create and name command sequences on the fly with
strongly-typed variables, I/O based conditionals, and a plethora of other subtle
extensions. On its most basic level a task is just a list of commands, but by having
the ability to recursively nest tasks within one another, and within conditional
commands, we can start to produce increasingly complex robotic behaviours.
3.3
Sensor Fusion for Natural Interaction: Co-speech

Gestures and Object Identication
When we interact with one another an enormous amount information is inferred based
on what we can see, feel, smell, taste, or hear. This contextualisation of our speech is
found in the physical environment, gesticulation, or even body language. Take for
396
T. Schwartz et al.
example the simple request pass me that brick; unless we can see the speaker
pointing, or the brick in question is otherwise clearly dened, the task is difcultly
completed. Similarly, come here is more often than not associated with a co-speech
gesture such as a waving of the hand. Both of these can be considered context for the
speech and are used to remove ambiguity or add extra information to a phrase.
Both gestural and vocal control give a user intuitive, real-time control of a
manipulator but, as they have been presented thus far, require the user to be in continuous communication with their machine and have either a comprehensive knowledge of the environment in which they are working or a means of accessing contextual
data e.g. through vision. The step that we would like to take here is one towards
adaptive automation; a means of giving our machines a knowledge of their own
environments and capabilities to react to what they are sensing. By increasing the
robots comprehension of its environment we can quickly start adding levels of
abstraction to our communication methodologies. The rst tests done to push our
interfacing capabilities with DMR were through a 3D camera mounted on a 6-axis
robots tool and a small piece of code that enabled the robot to identify bricks, their
colours, and their locations in space relative to the end-effector. Exclusively using
DMR and the accompanying DMR Interface we were then able to instruct the arm to
pass me a green brick. The process of robotic control is then incredibly similar to the
way in which our subconscious directs our own bodies. Both entities move towards the
object, orientate themselves appropriately to be able to pick it up, and upon detection of
sufcient proximity, close a hand or gripper to gain possession of the object in question.
In these rst experiments the robot is exclusively using vision, so has a much less ne
motor control than we have with a combination of sight and touch, but the vagueness of
our commands remain similar and that is where our new modes of interaction shine.
The combination of these systems permits two very distinct benets compared to
a traditional programming model. The rst is the ability to communicate in a
contextually enriched language; allowing us to push towards task-based robotics,
higher levels of abstraction, and avoid having to program every little movement or
action (Wahl and Thomas 2006). The second advantage of contextual awareness,
especially that in real-time, is the ability to adapt to changing environments. This
affords us the possibility of breaking robots out of their impeccably calibrated
factories and into the real world. The continued addition of these robotic senses
allow us to leave the robot in a state of autonomous work until a task is achieved
an essential property of any collaborator.
4 Workshop Scenario
The primary aim of the workshop is to put all the tools detailed above at the
disposal of the participants and enable them to experiment with different means of
robotic interaction. As a case study, we will focus on the collaborative assembly of
3D puzzles composed of voxelised parts generated automatically from a design of
greatly reduced complexity, as exemplied by Song et al. (2012). The participants
397
will be invited to collaborate with a team of small-payload industrial arms by either

teaching or directing them to nd the most effective manner to build these interlocking forms. The multi-robot cell available will host three robots and a pool of
sensors, all sharing the same network to enable continuous communication with one
another and increase each elements contextual awareness.
This exercise will implement a pre-created framework including tools for 3D
scanning, custom object recognition, vocal commands and tagging, and gesture
assignment. This support will avoid the need for intensive programming over the
course of the workshop and ideally make it entirely redundant. Through the
gradually-increasing complexity of the designed puzzles we hope to demonstrate to
the participants that collaboration with machines is not only possible but intuitive,
responsive, and advantageous.
5 Conclusion
This paper presents a set of unied real-time software utilities and corresponding
smart sensors aiming to ease the integration of natural human-machine interaction
for on-site collaborative robotics.
With an adequate communication protocol and linguistics framework used in
conjunction with context-aware sensing, we expose how robots can be turned into
intelligent assistive devices that can adapt to human complex operations, changing
environment parameters, and even unforeseen situations. The fluid operation of such a
system aims to enhance human perception speeds and offer a real sense of interaction
where technology disappears as a process and appears as continuous assistance.
Finally, we propose a case study to be trialled during the conference workshop
as an opportunity to pool our expertise and resources to partially validate the
denition and resolution of these technical issues, and to gather preliminary results
that will further this common research effort.
References
Andraos, S 2015, DMR: A Semantic Robot Control Language, Proceedings of the eCAADe
Conference 2015, Vienna.
Fellbaum, C 1998, WordNet: An Electronic Lexical Database. MIT Press, Cambridge, MA.
Schwartz, T, 2012, HAL: Extension of a Visual Programming Language to Support Teaching and
Research on Robotics Applied to Construction, Rob|Arch 2012, Springer, Vienna, pp. 92101.
Schwartz, T. 2015. A Generic Communication Library for Human-Robot Interaction on
Construction Sites, Design Modelling Symposium 2015, Ed. by Thomsen, M.R. et al.,
Springer International Publishing, pp 341-351.
Song, P, Fu, C W, and Cohen-Or, D 2012, Recursive Interlocking Puzzles, ACM Transactions on
Graphics (TOG), vol. 31, no. 6, pp.128.
Wahl, FM and Thomas U 2006, Robot Programming from Simple Moves to Complex Tasks, Institute
for Robotics and Process Control, Technical University of Braunschweig, Braunschweig.
Stigmergic Accretion
Abstract This paper posits a model of generative fabrication in which agent-based

models imbue physical material with digital agency. We demonstrate a process in
which real-time feedback is developed between industrial robots and multi-agent
algorithms to explore the generative potential of the interaction of computational
and material agency. This design research represents an inversion of material
agency, from which two key concepts have emerged: parallelism, and stigmergic
robotics. Rather than encoding material behavior within digital models, physical
material takes on digital behaviors through an inversion of material agency.
Parallelism describes a hybrid of digital and material behaviors through the
closeness of their interaction. Stigmergic robotics collapses design and fabrication
processes into a single operation where the robot operates as an extension of the
digital agent generating form through a series of design behaviors operating directly
on physical material.
Material agency
Multi-agent algorithm Stigmergic robotics
1 Introduction
The relationship between the behavior of physical material and digital algorithms is
increasingly being explored within experimental architectural practice. Within this
context we demonstrate a process in which real-time feedback is created between
industrial robots and multi-agent algorithms that hybridises their behaviors and
R. Snooks (&)
RMIT University, Kokkugia, Studio Roland Snooks, Melbourne, Australia
e-mail: roland@kokkugia.com
G. Jahn (&)
RMIT University, Elseware Collective, Melbourne, Australia
e-mail: gwyllim.jahn@rmit.edu.au
399
400
imbues physical material with a wild (Kwinter 2011) digital agency. This process
inverts the prevailing approach to material computation (Menges 2012) in which
material behaviors are encoded within digital algorithms for the purposes of form
nding, optimization or simulation. Instead we posit a model in which the digital
knows nothing of the physical, but instead operates at such a closeness of feedback
that the digital and material are caught in a mutual process of volatile reaction.
1.1
Volatile Characteristics of Computation
Volatile self-organizational processes of computationmaterial or digitalgenerate expressions and characteristics specic to their emergent processes. The generative capacity of these processes have been exploited by artists as diverse as Perry
Hall, Casey Reas (Fig. 1), Anish Kapoor and Roxy Paine.
Hall (2015) is concerned with the material behavior of paint over time through
thermodynamic, magnetic and chemical interactions. The reactions he provokes are
intended to tease out the generative capacity of the materialnot as a pure
expression of material, but one that is curated. Reas is concerned with the generative capacity of digital computation, programming the underlying behaviors for the
generation of pattern and form. The ambition of our research is to draw out the
emergent characteristics of a hybrid generative capacity between the inherent
shape-generating capabilities of matter (De Landa 2002) and abstract digital
behaviourscharacteristics that are neither a retreat to nature nor derived from the
digital. Robotics becomes the interface between these two realms of behavior.
While the relationship between machine behavior and that of material is illustrated
in Kapoors Between Shit and Architecture and Roxy Paines Scumak pieces, this
relationship is linearthe machine constrains material. Our concern instead is the
Fig. 1 Perry Hall: Tidal Empire (Animist), 2011, Oil, acrylic and custom paints lmed live using
a RED Epic digital cinema camera Perry Hall; Casey Reas: Process 13 (B), 2010, Casey
Reas. (left to right)
401
continuous feedback, facilitated by vision systems, between the behavior of

material and the algorithmic behavior executed by the robot. Within architecture,
this relationship between volatile material behavior and autonomous robotic fabrication has been conceptually established by Francois Roche.
1.2
Designing in Material
The resolution of scanning technology and simulations through General-Purpose

computing on Graphics Processing Units (GPGPU) are increasingly acting to blur
the distinction between physical and digital reality (Sheil 2014), with implications
for robotic fabrication, simulation and design. Alisa Andrasek speculates that the
found physics of complex material processes can now be encapsulated through
computational physics and large data sets, embedded into simulations and, consequently, physical constructions (Andrasek 2012). Similarly, the accessibility of
microcontroller platforms and consumer sensor hardware brings designers closer to
the machines and processes that allow them to act in material. Systems responsible
for materializing or sensing a design are beginning to come into contact with those
responsible for design decisions governing the form, pattern or structure of the
designed object, resulting in new paradigms for robotic control and design.
1.3
Robotic Design Paradigms
A convergence of control and design systems can be witnessed in the recent proliferation of bespoke robotic control plugins and hardware. The vast majority of
these systems extend the functionality of computational design software (such as
Rhinoceros, Grasshopper or Maya) by generating robotic instructions from existing
digital models (that Menges (2015) would describe as an instructional approach).
In contrast, behavioral approaches typically incorporate feedback to respond to
constraints and design requirements during fabrication, demonstrated in the recent
work of Casey Rehm (Rehm 2015) or Adam Fure (Abrons et al. 2014).
Designers face signicant challenges working exclusively within either of these
domains. Instructional approaches tend to be limited by the capacity to model and
anticipate material constraints, behavior and tolerances, and the accessibility of
particular operations and their associated patterns of use can result in hegemonous
design practice (Leach 2015). Behavioral approaches tend to be equally limited by
the challenge of encoding high level design decisions within rudimentary sensory
systems, microcontroller hardware, and robotic communication platforms. Real
time control of KUKA industrial robots, for instance, currently requires proprietary
software and the development of third party protocol to communicate with the
robot. In response to these challenges, we have developed a distributed, real-time,
402
object oriented control platform in which responsibility for allocating and updating
robotic tasks and behaviors can be distributed across multiple design agents
(software packages, sensors, simulations or humans) in parallel.
2 Paralellism
Rather than conditioning simulated models with material behavior or crudely
approximating digital complexity with robotic instructions, we posit the parallel
operation of computation and material that creates a hybrid or infusion of digital
and material behaviors through the closeness of these behaviors enabled by
real-time operation. Robotic parallelism benets from the unconstrained behavior of
both the computational agent and physical material. Encoding material constraints
and behaviors into the algorithmic model either renders material behavior as a
simplistic caricature that connes the design process to quantiable behaviors, or
over constrains the algorithm to the empirical model. Parallelism exploits contingencies and complexities of physical material, enabling the most volatile behavior
of material to participate in the generative process.
The ambition of this is to match the speed and sophistication of material processes through digital computation, where each responds to the other at a ne-grain
scale of operation and material manipulation (deposition and subtraction). This
establishes a stigmergic relationship between material and computation, where the
robotic agent scans and responds to the deposition and removal of material (Fig. 2).
Fig. 2 Stigmergic depositions, detail of robotically fabricated foam model
403
3 Stigmergic Robotics
This approach draws from the logic of multi-agent systems, and in particular
stigmergic behavior that generates complex order through the accretion and
re-organization of matter. Stigmergic systems such as those that underlie the formation of termite mounds or the self-organization of ant trails, operate through the
indirect communication of agents. In these systems the agent (termite/ant) interacts
directly with its environment (pheromone deposition). It is through the feedback of
the agent altering the environment and the environment affecting the agent that
complex order emerges. This self-organizing logic provides the basis for a stigmergic fabrication approach (Snooks 2012). While these systems typically involve
many simple agents, there is no need for direct interaction between agents as the
environment becomes the substrate of communication. Consequently the same level
of complexity can be achieved through the interaction of the environment and either
a single agent or a large population of agentsassuming the number of operations
is equivalent.
The capacity of industrial robots to run autonomously for extended periods of
time allows for either a single, or low population of robots, to form complex
stigmergic assemblages. In this paper we describe two strategies: a robot depositing
and interacting with a volatile material, and secondly two robots interacting with a
volatile materialone depositing material while another removes material.
This approach to stigmergic robotics has arisen from our long-standing design
research involving multi-agent algorithms and behavioral approaches to formation,
and has further developed within what is an emerging discourse on real-time
robotics and sensor feedback systems. This emerging domain encompasses the
parametric updating of models based on sensor feedback to the generation of tool
operations based on scanned matter. The original contribution our research offers is
the real-time interaction of material and computational agency, where the formation
of architecture is directly driven by their behavior, rather than an a priori parametric
model or generative algorithm.
3.1
Control
The technical workflow for stigmergic robotics involves a feedback between the
robot, deposition and subtractive tools, material behavior, vision system and
multi-agent algorithm. Our series of design experiments use two KUKA Agilus
KR10 R1100 SIXX robots controlled by a real-time server through KUKA RSI,
expanding foam and plastic extruded from end-effectors, Microsoft Kinect 2 and
Structure IO sensors and a Java and Processing based algorithmic environment
(Fig. 3).
404
Fig. 3 Control system
3.2
KUKA RSI
KUKAs proprietary Robotic Sensor Interface (RSI) software allows an external

computer to communicate directly with the robot controller via XML strings sent
using a User Datagram Protocol (UDP) server over a local ethernet connection. RSI
provides functions for making corrections to the robot position and digital IO
signals, and sends information to the control PC detailing robot joint angles, global
position, velocity and IO values in 4 ms intervals. Any correction made to the
position of the robot is executed within a single interval, limiting the ability of RSI
to make substantial modications to the existing position or motion of the robot
without exceeding the limits of the joint motors. However, the pilot project below
demonstrates that provided a stable connection can be maintained between the robot
controller and the external PC, most of the functionality that would typically be
found within a simple KRL toolpath program can be reliably executed through the
single RSI_MOVECORR() command by handling path planning and network
communications on an external server application. This allows for non-linear design
workflows whereby adjustment to robotic behaviors can be made by designers on
405
the fly, in addition to facilitating the integration of KUKA industrial robots within
more complex robotic systems that may incorporate dedicated computer vision
processors, large sensor networks, or remote operations.
3.3
Server Application
A C# application was developed that runs two UDP servers concurrently to handle
communication with external machines (External Server) and the robot (RSI
Server). The RSI Server communicates with the KUKA controller over a local
ethernet connection and maintains data transfer using RSI XML formatting. The
External Server handles sporadic communication with other control systems over an
internet connection. Communication between the two server instances occurs
through reference to thread safe dictionaries, allowing updates to the position, speed
or IO signals of the robot to be passed to the KUKA controller by any program
capable of sending XML strings over UDP. The RSI Server calculates acceleration
and interpolated trajectories to a given target position and stores them as a list of
correction commands that can be executed at 4 ms intervals and independently of an
external connection. Any update to the target position provided by the External
Server will trigger a recalculation of the RSI Server path. A consequence of this
server architecture is that responsibility for issuing robot commands is shifted from
the KRL program to an external control system that is wholly independent of the
procedural limitations of KRL. Although we demonstrate an approach that
dynamically assigns robot tasks and behaviors within a multi agent simulation in
Java, the C# application could be integrated with other platforms such as
Grasshopper/Firefly or Python over UDP.
4 Polymorphic Behaviors (AgentRobot)

Our strategy of stigmergic robotics explores computational behaviors that each
describes a simple design procedure. It is the relative weighting and negotiation
between these often competing procedures that creates the behavior of the robot and
the emergent characteristics produced in this process of formation. De Landa (2000)
argues that complex formations emerge when material negotiates differences in
intensity. A difference of intensity exists at both the level of the competing computational behaviors as well as competing robots. Our experiments leverage the
capacity of multiple robots to interact simultaneously on the formation of physical
materialeach embodying different intentions. Examples of which include robots
competing to deposit or remove material, or robots that operate in a volatile fashion
versus those which attempt pragmatic resolution.
The robot is conceptually and syntactically an extension of the agent. The
robots repertoire of behaviors expands on that of the generative design agent
406
extending their capability into the physical realm through deposition and subtractive behaviors in response to real-time sensor feedback. The technical implementation of these behaviors are outlined as follows:
Hashed Scanning: The digital representation (environment) of the physical
workspace of the robot and multi-agent system is represented as a hashed collection
of points that map a colorvalue to a discretized coordinate in space (Fig. 4). These
values can be selectively updated at up to 60 frames per second by querying the
buffer of a Kinect 2 scanner positioned above the workspace, or a Structure IO
scanner xed to the end effector of the robot. The combination of scanners allows
the robot to build up a 3 dimensional representation of its environment and avoid
areas of occlusion. Behaviors allow the hashed values of the environment to be
selectively updated by masking scanned points using a color range, or by cropping
scanned points to a bounded region of space.
Feature Analysis: The robot-agent is made aware of its local environment through
feature analysis and detection behaviors (Fig. 5). These behaviors iterate over
collections of points within a bounded region of space to identify local geometric
features (high or low points, plane approximation, normal approximation, curvature
analysis) or nd suitable conditions (closest point, closest value).
Feature Exaggeration: Features within the environment are manipulated and
exaggerated by associating agent behaviors with the actuation of end effectors that
Fig. 4 Point cloud representations of robot workspace
Fig. 5 Agent feature analysis and reaction behaviors
407
Fig. 6 Emergence of form and pattern in a digital model through feature exaggeration
control material in the physical workspace of the robot (Fig. 6). Because scanning
of changes in the physical environment occur at the same rate as behaviors
responding to changes in the digital environment, feedback between agent and
environment takes place through the medium of physical material. The digital
behaviors of the agent are encoded with an awareness of how the manipulation of
physical material will effect the digital environment. Local features can be exaggerated (such as adding to high points or removing from low points), material
behavior can be exaggerated (for example by combining pouring material with
computational behaviors that diagrammatically model fluid dynamics), and fabrication logic can be encoded (by approximating contoured paths, or by detecting and
lling gaps).
Patient Agency: Multi agent systems are typically modeled with continuous
motion, because it is the differentiation of the trajectory of the agents as they
negotiate multiple constraints and conditions that tend to reveal higher orders of
emergent formal possibility. However, when the agent becomes polymorphic with a
single robot, the necessity arises to condition the behavior of the agent to the speed
of the robotic fabrication process. We term these behaviors as patient agency, in
which the digital behavior of the agent is made both semantically and conceptually
to wait for the completion of a physical task by the robot. This enables the robot
and agent to operate simultaneously with the inclusion of specic tasks (such as
deposition or pick and place), without resorting to a sequential process of digital
simulation prior to robotic execution.
5 Materials
Expanding polyurethane foam facilitates experimentation across a spectrum of
material behaviors. By controlling the expansion rate of the foam, the material can
behave either in a highly predictable or highly volatile fashion. Because it is
408
lightweight it can easily be used in an additive or subtractive fabrication process.

Relatively rapid cure times facilitate the construction of novel assemblages over
time. Surface texture and opacity makes it easy to mask as the workpiece in scans.
The extrusion of thermoplastic (HDPEHigh Density Polyethylene) is comparatively stable. While the plastic shrinks considerably as it cools, it is sufciently
stable to build several consecutive layers. The limited tensile strength of the plastic
during its extrusion enables bridges or catenary arches to be strung between xed
points. This capacity enables an importantly different array of formal possibilities to
the expanded polyurethane.
6 End-Effectors
The actuation of end effectors takes place through UDP over a wi network and is
operated through the same control architecture that communicates with the robot.
An Arduino running a client program listens for string commands that can then
operate motors connected to a motor driver. Four basic tools will be used in the
workshop that facilitate the primary operations of this research:
a. Foam deposition gun: This off the shelf tool has been modied to be controlled
by a stepper motor through a calibrated gear rack (Arduino controlled). A low
feed rate allows foam to be extruded with a relatively high degree of control.
Controlled bursts allow material to be placed, while a high feed rate results in
highly volatility deposition.
b. Air turbine motor: This pneumatic spindle is used for subtractive operations
including milling, brushing and swarf cutting.
c. Plastic welder extruder: This thermoplastic deposition tool enables pelletised
plastic to be extruded with consistent velocity based on a screw drive mechanism. The velocity of the extruder is controlled directly through the digital IO
ports of the robot.
d. Heat gun: A heat gun is used to melt previously deposited plastic. The
behaviors developed for this equate to feature smoothing operations.
7 Conclusions and Speculations

The intention of this research is to explore a hybrid space of formal characteristics
and expressions that emerge from the closeness and parallelism of material and
robotic-agent behavior. This is primarily posited as a design approach that is
potentially independent of fabrication. The robot is considered an extension of the
agent, one that allows the computational agent to directly negotiate with the full,
and perhaps unknowable, generative capacity of material. While the space of
algorithmic and material expression has been heavily explored in the previous few
409
decades, our concern is to imbue physical material with a digital agency. We

speculate that as this digital agency has the capacity to negotiate competing conditions (such as design goals, material behavior, local conditions or fabrication
constraints) it facilitates the exploration of broader design spaces and thus leads to
greater formal diversity and novelty. Future advancements of this work will extend
competitive behavioral models and explore self-organization of additive and subtractive behaviors driven by limitations on the total amount of material within the
autonomous fabrication environment.
References
Abrons, E, Fure, A, Dubor, A, Diaz, G, Camprodon, G, and Wolking, A 2014, Sense-It, in
Mcgee, W and Ponce De Leon, M (eds), Robotic Fabrication in Architecture, Art and Design,
Springer International Publishing, Switzerland.
Andrasek, A 2012, Open Synthesis, Log, vol.25, pp. 4554.
De Landa, M 2000, Deleuze, Diagrams, and the Genesis of Form,Amerikastudien / American
Studies, vol.45, pp. 3341.
De Landa, M 2002, Philosophies of Design: the Case of Modelling Software, VERB:
Architecture Bookazine, Actar, Barcelona.
Hall, P, 2015, About - Perry Hall Studio, http://www.perryhallstudio.com/about. Accessed 26 Jul
2015.
Kwinter, S 2011, The Computational Fallacy, inMenges, A and Ahlquist, S (eds), Computational
Design Thinking: Computation Design Thinking, Wiley, London.
Leach, N 2015, There is No Such Thing as Digital Design, in Gerber, DJ and Ibanez, M(eds),
Paradigms in Computing: Making, Machines, and Models for Design Agency in Architecture,
eVolo Press, Los Angeles, pp. 148158.
Menges, A 2015, Coalescence Of Machine And Material Computation, in Yuan, P, Menges, A
and Leach, N (eds), Robotic Futures, Tongji University Press, Shanghai, pp 1722.
Menges, A 2012, Material Computation: Higher Integration in Morphogenetic Design,
Architectural Design, vol. 82, pp 1421.
Rehm, C 2015, Automaton: Designing Intelligence, in Yuan, P, Menges, A and Leach, N (eds),
Robotic Futures, Tongji University Press, Shanghai, pp. 170-175.
Sheil, B, 2014. High Denition: Negotiating Zero Tolerance, Architectural Design, vol. 84,
pp. 819.
Snooks, R 2012, Roland Snooks / Disguincio & Co, City Vision, no.7 Autumn/Winter, pp. 48.
Sensors and Workflow Evolutions:

Developing a Framework for Instant
Robotic Toolpath Revision
Alexandre Dubor, Guillem Camprodom, Gabriel Bello Diaz,
Dagmar Reinhardt, Rob Saunders, Kate Dunn, Marjo Niemel,
Samantha Horlyck, Susana Alarcon-Licona,
Dylan Wozniak-OConnor and Rodney Watt
Abstract This paper examines the potential for creative practitioners to adopt
robotic fabrication processes augmented with the introduction of sensors. Typically,
the outcomes of a fabrication process are predetermined, however, with the introduction of sensors, design and fabrication process may be interrupted by real-time
feedback. In such a system, design roles and authorship become secondary to the
process of manipulating data, such that new rules of design can be introduced and
developed in response to materials. Hardware and software such as Arduino,
Grasshopper3D, Rhinoceros3D and Processing have opened up new strategies of
hacking, coding and robotic manipulation that can be embedded in robotic fabrication processes. The addition of sensors provides feedback about material location
and characteristics, work environment and co-workers, so as to support architectural
dialogue. This paper proposes a framework for designing new protocols for human
interaction and machine response in robotic fabrication systems.
Keywords Human-machine interaction (HMI)

Collaborative processes
Robotically-assisted design creativity
Generative fabrication
Material
feedback Robotic design workflow
A. Dubor (&) G. Camprodom

Institute for Advanced Architecture of Catalonia (IAAC), Barcelona, Spain
e-mail: alex@iaac.net
G. Camprodom
e-mail: guillem@iaac.net
G.B. Diaz
Make It Locally, Seattle, USA
e-mail: gabriel@makeitlocally.com
D. Reinhardt R. Saunders M. Niemel S. Horlyck S. Alarcon-Licona
D. Wozniak-OConnor R. Watt
Sydney, Australia
411
412
A. Dubor et al.
1 Introduction and Motivation

Progress in robotic fabrication and manufacturing has accelerated in recent years
through research in industry, practice, construction and manufacturing (Gramazio
and Kohler 2014). Robotic fabrication labs are now embedded in professional
practices, educational institutions and research centers across architecture, art and
design. While robotic fabrication has extended previous automation processes of
the automotive industry towards complex and singular fabrication solutions, the
challenge is now to expand the negotiation of robotic processesto influence
toolpath options and dene new material processesin short to introduce a form of
design thinking (Moggridge 2007) for robotics with the goal of enhancing creativity
and the evolution of design processes, models, and techniques. In this paper we ask:
How do robots and humans work together to explore material agency? How does
the application of robotics expand design affordances or intuition?
Robotic fabrication processes enable designers and architects to explore the
boundaries between digital and material worlds. Beyond optimization criteria or
parametric design, new design strategies such as generative design and collaborative design are enabling new ways of approaching material exploration through
robotics. Open source software and hardware enable new forms of design, yet these
new tools also demand design frameworks dealing with robots, data, sensor technologies and material contingencies. Like computational composites (Vallgarda and
Redstroem 2007), robotic composites posit a challenge: How do we think about
hybrid processes that bridge different hardware (robot, human, end-effector,
material) and software (data, programs, toolpaths, workflows)?
This paper proposes a framework for robotic fabrication, which links data,
workflow, interaction, feedback, material behavior, protocols and time as major
R. Saunders
e-mail: rob.saunders@sydney.edu.au
M. Niemel
S. Horlyck
e-mail: samantha.horlyck@sydney.edu.au
S. Alarcon-Licona
e-mail: susana.alarconlicona@sydney.edu.au
D. Wozniak-OConnor
R. Watt
K. Dunn
University of New South Wales, Sydney, Australia
e-mail: k.dunn@unsw.edu.au
Sensors and Workflow Evolutions: Developing a Framework
413
project constraints. This paper provides an overview of different creative practices

using robotic fabrication augmented by sensor feedback. It examines the feedback
loops involved in these practices and concludes with a proposal for a framework for
designing new protocols for human interaction and machine response.
2 Evolution of Digital Fabrication Workflows

2.1
File to Factory
File to Factory has become more common as the availability of digital tools and
digital fabrication has increased. Designers and artists have used these workflows as
a way to materialize digital objects, allowing them to bridge the gap between digital
and material worlds with an expectation that the machine will materialize their
designed object as it appears on the screen. CAD/CAM (computer-aided design and
computer-aided manufacturing) software has also become increasingly accessible,
making the process of materialization easier.
The gap between digital and material worlds is not a barrier to be overcome but
can also be seen as a place for exploration and experimentation. While materialization is the focus of many practitioners, the classic File to Factory approach
lacks flexibility and the opportunity for feedback as part of an exploratory process.
2.2
Parametric Process
Public interest in digital fabrication and the rising availability of 3D printers has
allowed an increasing number of non-specialists to understand and adapt the logic
and mechanisms behind the materialization process. It is now becoming common
for users to change parameters of a digital fabrication process, e.g., feed-rates and
the nozzle temperatures, to adapt it to their specic requirements. In addition, the
spread of open source hardware and software has empowered hobbyists, artists and
designers to build their own machines, permitting the rise to new types of machines
and fabrication processes. In architecture, industrial robots have proven to be a
robust and flexible research platform, allowing the precise placement of many types
of tool within a large working envelope, allowing the designer to focus on the
design of novel end-effectors and processes.
In parallel, parametric design has gained momentum in many design disciplines
as a way to explore a space of possible designs when the nal outcome is not
precisely dened. This has resulted in a shift from shape design to process design
by empowering designers to manipulate the fabrication parameters as well design
414
A. Dubor et al.
Fig. 1 Common CAD + CAM workflow
parameters. Recent initiatives have brought computer aided manufacturing

(CAM) into parametric software, giving designers access to tools and languages to
manipulate both design and fabrication, considerably decreasing the learning curve
as well as the speed of exploration (Fig. 1).
Introductory digital and robotic fabrication workshops at Institute for Advanced
Architecture of Catalonia (IAAC) and The University of Sydney combined parametric tools, i.e., Rhinoceros3D + Grasshopper3D, with CAM plugins, e.g.,
KUKA|prc, to allow students to explore the potential and limitations of robotic
fabrication processes, e.g., 3D printing. By varying parameters exposed within a
predened process, students are able to learn from materialized results and move
quickly through iterations. Within this parametric workflow, teachers and students
analyze the results of iterations and provide the feedback for material exploration.
Consequently, students are able to achieve signicant results within a day.
2.3
Limitation and Challenges
The division between design and fabrication process is slowly disappearing in favor
of a continuous form of design, which includes fabrication as an essential element.
While providing a great framework for fast iteration and exploration, linear
approaches reach their limit when fabrication becomes more complicated, requiring
lengthy iterations. In addition, complex fabrication processes that use non-static
materials, e.g., clay or polymer, require more precise and sensor feedback to enable
tracking, ne-tuning and synchronization between material, machine and design
(Fig. 2). Sensors thus enable real-time feedback loops that have the potential to
radically change the design process.
415
Fig. 2 Parametric workflow: using parametric design for material exploration
3 (Im)Material Response
When material is understood as relative to time and protocol, material transformations can be considered as a series of actions influenced by a range of variables
or agencies. These include immaterial factors such as velocity, density, mix ratios,
temperature and evaporation. As such, an indeterminate, unpredictable material
self-formation can be considered a material computation. In some cases time may
affect material conditions, e.g., velocity may be a factor affecting toolpaths when
working with semi-liquid materials that exhibit sedimentation.
3.1
Material as Process
In deposition processes, e.g., Free Form Fabrication (FFF) or extrusion-based 3D

printing, materials are processed by the extrusion of a liquid, or viscous, material,
e.g., clay, wax, concrete, polymer. The success of the extrusion is highly dependent
on the material properties being adequately linked to the fabrication protocol, e.g.,
feed rates and toolpaths. As 3D printing, parametric design and CAD/CAM technologies advance so does the need for control, manipulation and development of
suitable materials (Gardiner and Janssen 2014; Friedman et al. 2014).
A coupling of material protocol to sensors can enable new design approaches.
For example, predicting the nal deposition location of a clay extrusion implies
calculating the shear viscosity of clay at the extrusion point. This depends on
environmental conditions (e.g. air temperature, relative humidity) and the time the
material undergoes shear. By obtaining the fluid speed, based on the pressure
exercised on the material, it is possible to calculate the vector that the extrusion will
follow. Finally, by obtaining the deposition plane position we can calculate the nal
extrusion location and adjust the fluid speed to match a deposition target.
416
A. Dubor et al.
Fig. 3 3D printing process informed by sensors (Pylos project, IAAC, 2014)
A model provided with real-time data from direct (extrusion cylinder pressure)
and environmental (air temperature, relative humidity) sensors can deal with
complex material behaviors (Fig. 3). These behaviors are difcult to predict within
digital simulation and make the use of predetermined tool-paths obsolete, as they
have too little tolerance to guarantee a successful outcome. The use of sensors
allows for the bridging of the gap between the expected outcome and reality.
The use of sensors is critical for understanding complex material behavior.
Digital sensors are devices capable of turning physical properties into data.
Traditionally industrial sensing equipment has been tied to specic industry sectors,
making them expensive and difcult to operate. The rise of consumer electronics
such as digital cameras and smartphones has made available low-cost digital sensors for a wide range of physical properties, e.g., temperature, proximity, pressure.
3.2
Material Feedback Sensor Toolkit
The availability of digital sensors makes it possible to assemble an inexpensive

toolbox of sensors useful for digital fabrication. Multiple approaches to sensing can
be quickly tested in order to understand how a material behaves before moving to
more specic, industrial grade solutions. Furthermore, the development of open
source hardware and microcontroller platforms, such as Arduino, has democratized
access to electronics by providing tools and documentation. At the same time
digital fabrication tools allow for the customization of sensors.
The Material Feedback Sensor Toolkit is a rst attempt at establishing a collection of sensors and tools for sensing material behavior. None of the sensors listed
are industrial-grade, instead they were developed for consumer electronics. The use
of consumer-grade sensors can require more work than industrial sensors but this is
compensated by the low cost and extensive range of the sensors available. The use
of these sensors has been made possible due to the work done by the open hardware
community in documenting and exploring the use of these devices.
Reference sensor
Category
Supplier
Cost (USD)
Image based analysis (pixels image)

Webcam
Vision
Various
10+
Fast material and environment structure 3D scanning (point cloud)
Microsoft Kinect or Asus Xtion
Vision
Various
150
Material point tracking or edges recognition without external software
Pixy CMUcam5
Vision
Adafruit
75
Very long to short range detection for collision avoidance
Maxbotix HRLV-EZ4
Distance
Adafruit
35
Distance precision sensors for material to extruder distance
VL6180
Distance
Sparkfun 15
Precisely measure distances up to 40 m
LIDAR-Lite v2
Distance
Sparkfun 115
Material flexion or tool joint movement
Spectra Symbol FS-L-0055-253-ST
Flexion
Sparkfun 10
Water flow meter
Adafruit ID828
Fluid
Adafruit
10
Optical, non-contact, odometer for fluid speed
ADNS3080
Fluid
Avago
20
Extrusion chamber pressure
MS5803-14BA
Fluid
Sparkfun 60
Sense tool force against material from 0.01 to 50 N
Sparkfun/Vernier SEN-12873
Force
Sparkfun 100
Read contactless tags for tools and materials identication
Adafruit PN532 NFC/RFID
Identity
Adafruit
55
Water and dry powder level sensor
PN-12110215TC-8
Level
Sparkfun 40
Sensing material color
TCS34725
Light
Adafruit
10
IR and visible light detection for light sensitive materials
SI1145
Light
Adafruit
10
End effector/tool gravity/acceleration feedback
Adafruit 10-DOF IMU
Position
Adafruit
29
Sensing high temperatures with direct surface contact
Thermocouple + MAX31850K
Temperature Adafruit
20
Low resolution IR temperature sensors for non-contact material temperature MLX90620 or AMG8832 or TMP006 Temperature Various
39
Environmental temperature and humidity
SHT-21
Temperature Sparkfun 15
Weight materials up to 200 kg
Load cell + HX711
Weight
Sparkfun 20
Notes Breakout boards for integrated circuits are available from the suppliers. All sensors are digital. Most come calibrated from factory and report measurements using
I2C or SPI protocol. Open source Arduino compatible driver libraries and documentation are provided by the supplier or on-line
Application of sensor
Table 1 The material feedback sensor toolkit

417
418
A. Dubor et al.
The sensor selection (Table 1) prioritizes low cost, open source drivers and the
existence of good documentation. Most sensors are compatible with the Arduino
electronics platform. The total cost of the toolkit, including wiring and the Arduino
development board, is less than 1000 USD.
4 Coding Intuition: Embedding Sensors

and Logic in Design
Gathering the right data is only half of the process in a feedback workflow. The data
must be turned into decisions and nally actions.
Embedding sensors in fabrication processes is not new. Closed loop control
systems such as the Watt (or centrifugal) governor date back to the origins of the
industrial revolution and have been extensively used in industry since. Closed loop
control systems are based on the idea that an error in a system can be corrected by
continuously measuring the output with a sensor in order to adjust the input based
on a threshold. Originally, control systems were designed in the form of analogue
devices tied to their own mechanics. Digital sensors and microcontrollers allowed
industrial control systems to become cheaper, smaller and more easily programmable. Despite the importance in industry, however, traditional control theory
is focused on process efciency, optimization and safety. Approaching feedback
from an experimental point of view requires a different approach.
When control systems in fabrication are seen from a material instead of a
machine perspective, the design of the controller becomes part of the design process
itself. Consequently, the focus becomes exploring the material by connecting its
behavior to the machine control system using relatively simple logic. This is critical
when we look at how the complexity of modeling certain fabrication processes
using tools like Grasshopper3D can result in signicantly less experimentation.
4.1
Integrating Material Feedback into Design Software
All the sensors in the Material Feedback Sensor Toolkit can be integrated into
parametric design software such as Grasshopper3D to allow designers to integrate
material feedback into their digital design process. For example, in the Magnetic
Architecture, data from a camera informed the decision-making process for each step
according to the materialization of the previous tool path. The experiment used the
Firefly plugin to feed the data from the camera into a Rhinoceros3D + Grasshopper3D
script, which produced code using KUKA|prc. In this setup an iterative logic is
encoded to compute each successive toolpath one step at a time.
Fig. 4 Iterative workflow: embedding sensor feedback in the parametric design
Fig. 5 Behavioral workflow: embedding response to sensor feedback in fabrication logic
Fig. 6 Sense-it 6 axis workshop
419
420
A. Dubor et al.
In this experiment, sensors permitted the integration of a self-organizing process

as part of the fabrication and design (Fig. 4). The feedback loop in this fabrication
process took approximately a minute, limiting opportunities for experimentation.
4.2
Encoding the Logic into the Machine
When working with continuous fabrication processes, e.g., material extrusion in

additive manufacturing, real-time feedback is required. Industrial robots can be
connected in near real-time with parametric design software, e.g., Hal Robotics
streaming supports communication speeds up to 5 Hz, allowing the feedback loop
to be signicantly shortened (Fig. 5). Continuous path adjustments, however, need
even faster reaction times requiring the logic to reside within the robot controller.
The Sense-it 6 axis workshop at ROB|ARCH 2014 explored the combination of
KUKA Robot Language (KRL) scripts with digital sensors and Arduino (Fig. 6).
Participants explored generative fabrication processes where the outcome was not
predetermined but was a result of a dialogue between the robot, tool and material.
In these experiments the threshold values for sensors were dened in Arduino
code and used to trigger digital or analog inputs on a KUKA Robot Controller
(KRC). KUKA IO1 was developed for this experiment to facilitate communication
of the Arduino with the KRC. Using this framework, a simple KRL script (approx.
20 lines) produces a rapid feedback loop (<20 ms) encoding the desired logic. The
material process used, plastic extrusion, is difcult to predict but could be tracked in
real-time using distance and temperature sensors.
The participants demonstrated impressive creativity when inventing fabrication
processes using KUKA IO. Nevertheless, this framework and the specic workflow
used in the workshop have some signicant limitations. The single byte that was
exchanged through the input/output port of the robot controller restricted the control
that the Arduino could have on a running process.
In addition, microcontrollers, such as those used on an Arduino Uno, have
limited processing ability, restricting the types of sensors that could be used.
Finally, traditional robot control languages, such as KRL, are restrictive when
compared to modern scripting languages, which limited the possibilities available to
experienced coders and made it difcult for inexperienced users to code logic to
produce desired behaviors. During the workshop the importance of manual
experiments to understand a materials behavior became apparent. Manual tests
were conducted to simulate the sensor-robot logic and understand what needed to
be scripted. The need for manual experiments may have been avoided with better
support for rapid development of control software for the purposes of material
experimentation.
https://github.com/pral2a/KUKAIO.
421
The most common solution to these limitations is to externalize the controller on

a remote computer giving users the possibility to code the robot motion and
behavior in another language, e.g., Java or C/C+, and communicate through a
faster, machine-specic protocol. An example of such framework is OpenKC,
which is an open source, real-time control software specically designed for the
KUKA Light Weight Robot (LWR), coded in C/C+, and based on the KUKA RSIXML interface. Robot manufacturers are also starting to make their controllers more
accessible to researchers and designers. Universal Robots, and more recently
KUKA, offer APIs to control the motion and get information on a robots state.
5 Designing Protocols for Human-Machine Interaction

The discrepancy between material contingency, digital control, technological limitations and designers creativity reveals the difculties in dening a suitable
interface to interact with in this context. We envision an ideal framework to
facilitate interaction without the technological issues mentioned previously while
providing space for creativity through craftsmanship, e.g., manual experiments, and
generative fabrication, e.g., fabrication responsive to material behavior. The following section discusses design frameworks and workflows that may be experimented with in the workshop at ROB|ARCH 2016.
5.1
Craftsmanship and Digital Fabrication
Investigating ways of depositing materials that have been traditionally formed

either by hand, such as clay, opens up the possibility of investigating the place of
the handmade and the concept of analogue authorship in digital fabrication. For
example, can the author, designer and creative practitioner alter a program that has
been set in motion by interacting with sensors?
lArtisan lectronique by UNFOLD addresses the manipulation of a printing
process through human intervention using sensors (Fig. 7). A separation of the
human hand and the material process, however, caused a delay and a disconnection
in the creative process between a users input and the material feedback. Objects of
Rotation was a project undertaken at the Harvard Graduate Design School allowed
the use of mark-making processes on rotating clay. The clay is unresponsive,
however, and there is no place for the human hand.
Both of these projects go some way to addressing the place of the handmade in
digital fabrication and how creative practitioners may utilize robotics. But there
remains a need to investigate the smooth exchange of design intention between
analogue and digital processesan exchange that opens a space for spontaneous,
reactive authorship in digital fabrication.
422
A. Dubor et al.
Fig. 7 lArtisan lectronique ( UNFOLD, Belgium, 2010)
We propose the exploration of a new framework where a craftsmans intuition

and sensibility can be combined with the power of digital analysis and the precision
of robotic fabrication. We envision this framework being particularly useful for
fabrication involving complex material behaviors such that it remains open-ended
for creative exploration. To test the proposed framework, we are exploring
clay-modeling processes using additive manufacturing.
5.2
Experiment 1: Alternating Manual

and Digital Manipulation
Our rst experiment will introduce 3D scanning and data from other sensors, e.g.,
room temperature or humidity, within a manual fabrication process allowing
information about the manual process to be captured. Computational analysis of the
process may allow improvements in the iteration of a design by providing the
designer with specic information, e.g., geometrical, topological or structural
analysis. In addition, the data may also be used to elicit feedback from remotely
located co-designers or clients.
Having digitized a manual fabrication process the reverse of the process would
be to materialize the data captured. An additive manufacturing process will be
used to reproduce the previously scanned object. This materialization will allow a
network of collaborators to get physical copies of the object and the possibility of
manipulating the object, e.g., by modifying the shape. Using integrated CAM
software, such as KUKA|prc, we can close the loop of digital iteration using a
common platform (Fig. 8) with a feedback loop of minutes or hours.
This experimental setup will allow the reproduction of a manual design task and
generate an augmented fabrication but a clear difference will still exist between
the human produced and the 3D printed copies. These differences will be evident at
the multiple scales, e.g., material continuity, physical behavior, and texture.
423
Fig. 8 Digital craftsman workflow: combining digital and manual fabrication
5.3
Experiment 2: Human Feedback Within Real

Time Process
While additive manufacturing with clay has been used since ancient times, however, 3D printing layer-by-layer is quite different from traditional crafts, requiring a
level of precision that is almost impossible for craftsmen, especially when trying to
have homogeneous material deposition. The introduction of cooperative robots able
to safely share a workspace with humans opens up the possibility of a robot and
human working simultaneously on an object, possibly with the same tool.
We propose using the force feedback sensors of a KUKA LWR iiwa robot to feel
the indication of a user manipulating a tool attached to the robot (Fig. 9). The tool
Fig. 9 Collaborative workflow: coupling human-machine interface with robotic fabrication,

sensor feedback and digital computation
424
A. Dubor et al.
and the robot would be free to move until it reaches one of the constraints dictated
by a model. In the case of 3D printing with clay, a robot might maintain a constant
speed in the XY plane in response to human input. In a similar fashion, movement
can be constrained to a specic height from existing object using data from a
distance sensor attached to an end effector. This would allow a user to move freely
along an extrusion path while maintaining the specic constraints of the fabrication
method, e.g., extrusion speed and layer height. Such real-time feedback needs to be
programed with a fast response rate and therefore requires coding in the robot
language to achieve a feedback cycle of less than 50 ms.
An external link will be used to connect the robot controller to a separate
computer where each robot position is recorded. This data will serve to make a
session reproducible without additional human input but also provide feedback
from digital analysis of the object being produced. This analysis can then be projected back on the workspace or object to provide a non-invasive feedback with a
slower response rate (>1s) to complement the real-time force feedback.
In such a setup, the user is not only exploring the toolpath by moving the tool in
space but also the different parameters of the fabrication process, e.g., by changing
the rules that the robot follows. These parameters and logic become core information in the design researchinformation that can be shared with a community
and continuously adapted.
6 Conclusion
This paper has provided an overview of an evolution of creative processes supported by computational design and fabrication and the potential for future changes
supported by data feedback. The paper has discussed this via a series of case studies
examining different feedback loops and a proposal for a framework for designing
new protocols for human interaction and machine response.
The act of giving a machine freedom to assist the creative process leads to
unexpected and useful information both from the machine and material perspective.
By coupling Human-Machine Interface with robotic fabrication, sensor feedback
and digital computation, new possibilities for creative collaboration are appearing.
Collaboration between robots and human can enhance creativity and innovation by
supporting designer and researcher while exploring complex material system. Such
material exploration through robotic fabrication can gain precision and in depth
information from sensor analysis of the material, the context and the users
movements. The advantages associated with an open-source framework and low
cost sensors may permit widespread adoption of this approach and enhance new
collaboration between researchers and designers.
The creation of a flexible framework for Instant Robotic Toolpath Revision
intends to make such practice more accessible to a wider range of designer and
425
researcher and aims to extend its applications to other elds and industries. The idea
and
technology discussed in this paper will be explored further through a

series of workshop to be held in Sydney and will be the occasion to
apply this framework to a growing number of fabrication processes.
References
Gramazio, F and Kohler, M 2014, The Robotic Touch: How Robots Change Architecture Park
Books.
Moggridge, B 2007, Designing Interactions, MIT Massachussetts.
Vallgarda, S and Redstroem, J 2007, CHI 2007, April 28-May 3, 2007, San Jose, California, USA,
pp. 513522.
Friedman, J, Kim, H and Mesa, O 2014, Experiments in Additive Clay DepositionsWoven
Clay, in McGee, W and Ponce de Leon, M (eds), Robotic Fabrication in Architecture, Art and
Design 2014, Springer International Publishing Switzerland, pp. 261272.
Gardiner, J and Janssen, S 2014, FreeFabDevelopment of a Construction-Scale Robotic
Formwork 3D Printer. McGee, W and Ponce de Leon, M (eds.), Robotic Fabrication in
Architecture, Art and Design 2014, Springer International Publishing, Switzerland, pp. 131146.
Towards Real-Time Adaptive

Fabrication-Aware Form Finding
in Architecture
Dave Pigram, Iain Maxwell and Wes McGee
Abstract This paper identies the disciplinary potential latent in the combination
of algorithmic design and sensor-enabled robotic fabrication to achieve multiple
channels of feedback between design, fabrication and assembly. Three key
methodological shifts are identied. The rst is a shift to fabrication-aware-form
nding. In comparing analogue form nding to digital form nding practices, it is
clear that a greater number and diversity of constraints can be negotiated within an
information-based digital environment. The second methodological shift is to bidirectional negotiation between design and production limits. Robotic fabrication is
highly customizableinitial constraints do not need to be seen as xed. The nal
shift is the introduction of sensor feedback and near real-time control. This permits
the continual redenition of parts during fabrication in response to material-,
dimensional- and assembly-volatility. Taken together, these shifts challenge the
typically linear and compartmentalized nature of the processes linking design with
construction and therefore open up new ecologies of design practice and opportunities for innovation.

Keywords Robotic fabrication Algorithmic design Fabrication-aware form

nding Digital form nding Analogue form nding Sensory robotic control
Adaptive part variation
D. Pigram (&) I. Maxwell

University of Technology Sydney, Sydney, Australia
e-mail: David.Pigram@uts.edu.au
I. Maxwell
e-mail: Iain.Maxwell@uts.edu.au
W. McGee
University of Michigan, Ann Arbor, MI, USA
e-mail: Wesmcgee@umich.edu
427
428
D. Pigram et al.
1 General
This paper posits the possibility of an expanded denition of architectural form
nding by embedding algorithmic design methodologies and robotic fabrication
strategies in the form nding process. Three emerging research trajectories are presented that when seen collectively point towards a highly flexible and integrated
approach to design: Fabrication Aware Form Finding; Robotic Sensor Control; and
Adaptive Part Variation. These interrelated elds of research have led to the realization
of highly productive feedback loops between the once separated domains of design,
part production and assembly. This framework not only consolidates and embeds the
logics, constraints and viability of part assemblies (and production) within the generative cycles of design but, critically, affords increased opportunities for innovation
throughout the entire production chain. Each of the concepts introduced will be situated through the description of a case study. The overarching themes and consequences observed for the discipline of architecture, serve as the basis for a workshop to
be held as part of the Rob|Arch 2016 International Conference.
2 Analogue Form Finding (AFF)

Form nding seeks the discovery of optimal shapes against a known set of design
constraintstypically structural. Exemplary methods include: Antoni Gauds
nested hanging-chain models to resolve highly complex load paths within
compression-only masonry structures and Frei Ottos study of minimal surfaces
such as soap-lm bubbles to realise complex membrane structures. Typically,
Analogue Form Finding (AFF) methods attain a single design solution, via equilibrium, for a given starting state: material substrate, boundary, support conditions
and assumed loading scenario.
3 Digital Form Finding (DFF)

The analogue methods above can and have been widely reproduced digitally
through the implementation of dynamic relaxation (Day 1965). The basis of the
method is to trace step-by-step for small time increments, Dt, the motion of each
node of a structure until, due to articial damping, the structure comes to rest in
static equilibrium (Barnes 1999). Kilian and Ochsendorf (2005) describe a method
for the re-conception of architectural geometry (surfaces, volumes etc.) as a network of weighted particles (nodes) tethered to one another by variable springs
Towards Real-Time Adaptive Fabrication-Aware
429
(members). Within such a computational approach, each constituent particle

negotiates its immediate neighborhood of connecting springs (topology) towards a
state of equilibrium or equal residual force (towards a given spring length).
Dynamic relaxation (DR) is therefore the iterative application of Hookes law of
elasticity: for elastic deformations of an object, the magnitude of its deformation
(extension or compression) is directly proportional to the deforming force or load.
The advantage of implementing form nding methods digitally is that the initial
design constraints, topological conguration and constituent variables of the system
(i.e. node masses, spring lengths and elasticity) can be managed, modied or discontinued during the form nding process. Additionally, embedding variegated
materials (weights, strengths, thickness and densities) is a simple numerical operation applied to either the nodes or members allowing for an accelerated exploration
of numerous design solutions. Clifford and McGees La Vote de LeFevre project
varies element thickness to reach a target vault underside geometry, serving as an
example of a form nding inversion virtually impossible with analogue methods
(Clifford and McGee 2014) (Fig. 1). The digital allows for both an engagement with
and a convenient departure from the pragmatic constraints of the physical world.
Fig. 1 Matter Design La Vote de LeFevre: The size of a components opening is negotiated to
reach a target underside geometry ( Wes McGee and Brandon Clifford)
430
D. Pigram et al.
4 Fabrication-Aware Form Finding (FAFF)

4.1
Denition
Fabric-Aware Form Finding (FAFF) seeks to embed the constraints of manufacture

material handling and processingwithin the iterative cycles of a generative
computational design framework and should be viewed as a consequential extension of the form nding traditions of architecture. Such approaches demand the
development of specic parametric or algorithmic protocols, implemented via
bespoke software or scripts that extend the functionality of existing software.
Typically, such models mix opportunities of user interaction with underlying orders
of geometric control and/or optimization routines to ensure only valid outputs are
derived. Examples of criteria for validity may include topological limits (allowable
connections), panel shapes (triangles, quads), curvature (planarity, single, double,
syn- or anti-clastic) and maximum and minimum part or angle sizes. As such, FAFF
positions itself in opposition to the typical events of rationalization that traditionally
occur after designing has nished. Within a FAFF paradigm, the transformations
that a design undergoes in order to be built happen within the primary design
process and thus the outcomes of these steps (positive or negative) can feedback
upstream for negotiation with other factors. In this way, the viability of a project
can be embraced by the designer, avoiding later degradation of design intent during
building.
4.2
FAFF Case Study: Utzon/40 Pavilion
The Utzon/40 pavilion serves as a purposeful demonstration of the possibility of a

FAFF design approach. The pavilion completed for the 40th anniversary of the
Sydney Opera House draws inspiration from the material palette and innovations of
Jorn Utzons design while positing the capacity of algorithmic design methodologies coupled to CNC production chains to overcome the challenges of standardization that so problematically impacted the original (Fig. 2).
To do so, an expanded implementation of dynamic relaxation (DR) was
developed that allowed for the simultaneous negotiation of global shape and the
constraints of part manufacture. All critical aspects related to the 5-axis CNC
machining of parts were re-conceived as dimensional and geometric limits (minima
and maxima) influencing the spring-member network that dened the DR model
(Fig. 2). Essential fabrication constraints and their translation included:
Minimum (450 mm) and maximum part length (1200 mm). These dimensions
were governed by the vacuum pods (150 mm 150 mm) used to x the parts
during machining and acknowledged the very practical concerns of minimizing
431
Fig. 2 Completed Utzon 40/Pavilion ( Michael Ford, left), Fabrication Constraints ( Oliver
Bennett, right)
the repositioning of parts, maintaining ve-axis clearances and restricting the

cantilevering of parts within an acceptable range. The slack length is decreased
during the dynamic relaxation process to counter an exceeding of maximum
length and vice versa.
Maximum angular deviation between node normals (15). This parameter
results from the design detail where each members top and bottom edges are
cut as twisted ruled surfaces to eliminate stepping at the joint. Greater node
deviation resulted in very sharp and vulnerable edges. When this value is
exceeded, the slack length of the member between the nodes in question is
contracted to pull the nodes closer together, thus reducing normal variance.
5 Robotics and Fabrication Aware Form Finding

Fabrication with industrial robots differs from the use of dedicated or
single-purpose machines (NC or other) in the obvious and here extremely signicant aspect that they have been designed specically to support many styles of
customization. The use of custom end-effectors and control interfaces for example
can signicantly broaden the range of possible outcomes. It is this aspect that is an
amplier of the potential of more typical implementations of fabricationaware-form-nding where it is often assumed that the constraints are xed and
so feedback takes the form of iterative design changes assessed for viability against
those xed constraints. With robotic fabrication, a system that permits evolution,
the feedback can occur in both directions. Negotiated changes could be demanded
432
D. Pigram et al.
of the fabrication system and not only of the design. The assessment of a proposed
design (or component, process etc.) would not only include an indication of viability but a measure of what changes to the constraints would lead to that design
achieving viability. In this way design loops that understand limitations, can be
extended to also analyze the payoffs from reducing those limitations. There are
many ways that changes to the fabrication system could be made to increase
capability. They may be physical or information based. Examples include:
End-effector redesign
Implementation of sensors to respond to various forms of volatility
Additional empirical tests to better test specic edge cases and verify associated
limits
Algorithm optimisation for conditions prevalent in a given design
Interface changes for example to change the calculation of orientations to relieve
pose issues
Fixture redesign, or allowance for the production of multiple xtures
In the case of architecture, where the cost of many building packagese.g. facade
panelsfar surpasses the cost of most industrial robotic equipment, and certainly
the cost of modications to existing equipment, it is entirely likely that a
cost-benet analysis would reveal that enhancements to a fabrication system would
pay for themselves by absolving the need for changes to a project. Clearly many
other non-nancial barriers remainthe frequent demand for competitive tenders
for examplethough some of these may also be negotiable via feedback. One
might imagine a tender process where companies with increased production
capabilities can bid based on the enhanced project that they could deliver, clearly
understood against the alternate project deliverable by a company with truncated
production capabilities.
6 Sensory Robotic Control (SRC)

6.1
Denition
Sensory Robotic Control (SRC) introduces real-time sensor data as a dynamic

feedback during machining operations. The concept massively challenges the
explicit nature of machine instruction code, namely pre-ordained blind motion, and
instead path planning includes near real-time adjustment or occurs on the fly via
mechanisms of visual, force or other sensing activities.
6.2
433
Case StudySensor Enabled Tooling
Over the past 2 years the robotics lab at the Taubman College of Architecture has
been gradually implementing an external Programmable Logic Controller
(PLC) system to expand the capabilities of its 3 dual Kuka robot workcells.
Numerous manufacturers produce compatible PLC systems, in this case the external
PLC is implemented using Beckhoffs TwinCAT XAE 3.1 software and Beckhoff
hardware, and communicates via the EtherCAT protocol with Kukas KRC4 controller. One example of sensory tooling is a stepper controlled automated ber
placement head (Seyedahmadian et al. 2015). In this tool a NEMA 17 size stepper
is used to feed composite tow through the head, both as restart after cutting at the
end of path, and during placement in velocity synchronous mode. Start and stop
commands are issued via subroutine commands within the KRL (Kuka Robot
Language) program, and the actual motor control is implemented via PLCOpen
function blocks under TwinCAT 3.1.
The second example is the ongoing development of extrusion heads for additive
fabrication processes. In this case a 2 kW servo is controlled via the external PLC to
synchronize with the robot motion. As the exact deposition rate is somewhat
non-linear with the velocity of the motor, a cam table is implemented (also using
PLCOpen function blocks) to adapt the motor velocity to the robot velocity. This
cam table is empirically generated, and corresponds to specic polymer viscosities
and temperatures. While not impossible, such a control framework would be relatively difcult to implement natively in the robot controller. A smaller stepper
motor driven extruder has also been implemented using the same framework.
7 Adaptive Part Variation (APV)

All assembly processes demand attention be paid to dimensional variation and the
provision of tolerances. This is particularly true within architecture where: parts are
either large, come in high populations, or both; cumulative errors are signicant;
and site irregularities are common. Adaptive Part Variation is a strategy where the
real-time redenition and fabrication of parts occurs during the serial process of
assembly thereby allowing detected errors to trigger a conditional design response.
Necessary preconditions to establishing APV workflows are: the ability to progressively measure an assembly in process; to recognise key features within that
assembly; and to update the digital model to register variance (Vasey et al. 2014).
Adaptive Part Variation demands a radical repositioning of design intent away
from static geometric descriptions towards relational frameworks and the denition
of contingent responses to error. One immediate consequence of this shift is a
radical reappraisal of prevailing approaches to the provision of construction tolerances. Typically, architectural details must consistently cater to a predicted
434
D. Pigram et al.
worst-case, APV however, affords the denition of an enriched set of design

responses that can directly address each actually occurring situation.
7.1
Case StudyRobotic Bending, Assembly and Welding
The 2012 Rob|Arch workshop superForm and the the accompanying prior and
consecutive research revolved around the desire to test an Adaptive Part Variation
workflow. The workshop made use of the dual KUKA KR120HA on dual 1500
Linear Axes workcell at the University of Michigans Taubman College. The
synchronous robotic cells bent, cut, positioned, scanned and welded in. diameter
steel rods into a larger assembly. A custom-made, free-standing, robot-tended, CNC
bender was controlled as an additional axis of the rst robot. Once each rod was
bent, this rst robot then used its collet gripper to locate the rod within the growing
assembly. A second robot, equipped with a welding-end-effector, additionally
employed a 1D laser scanner together with a specic search routine to verify part
geometry (Fig. 3). This information was then used to modify both the welding
motion paths and the design of subsequent parts (Vasey et al. 2014).
Long (up to 1800 mm) and thin (1/4 diameter) rods used to amplify inaccuracies caused by bending, guaranteed the need for scanning. Part position measured
at the weld locations varied by up to 20 mm due to the long parts bending being
amplied further from the robots grip. This translated to necessary part modications of up to 35 mm due to the further amplications of error from one long part
to the next.
Fig. 3 1D Laser Scanner 2014 ( Adam Smith) Image Front Cover: Adam Smith
435
8 Future WorkRob|Arch 2016 Workshop

While the work to date on the establishment of sensor-enabled real-time robot
control and its application towards Adaptive Part Variation has proved its viability,
key adaptations were still completed as part of an offline feedback loop. The natural
next step is to combine each and all steps into real-time adaptive frameworks.
Further, APVs implications for design have not been signicantly tested or
demonstrated. Key to this is an exploration of the various styles of response to
cumulative errors. Possibilities include concealing or amplifying the visibility of
errors; producing hierarchies of design intentions that can be shed consecutively if
errors preclude their possibility; contingent detailing for various forms of error; and
adaptive construction sequencing.
The 2016 Rob|Arch workshop will focus on lightweight stereotomy as a means to
address these research extensions. The workshop will employ a material system where
parts can be rapidly produced on demand: swarf-cut Aerated Autoclave Conrete
[AAC]. This will be combined with a construction system where structural form
nding can be coupled to fabrication-aware form nding: dry-stack thin-shell vaulting
and single xture robotic machining. Parts will be scanned during production and
assembly. Finally, this will all be included in an algorithmic design system that can
recompute the parts in order to respond to the inevitable accumulated errors in the low
tolerance construction system: adaptive algorithmic stereotomy (Fig. 4).
Fig. 4 On-demand part production enabling the serial correction or exaggeration of accumulative
part and construction errors
436
D. Pigram et al.
9 Volatility
The conventional procurement of architecture operates on an assumed resilience of
the design against the encroaching scales of economic production and the reductive
and expedient cultures of building delivery (i.e. Key Performance Indicators (KPI)s
of project management consortiums). Scenarios generally characterised by linear
and highly compartmentalised workflows. Each of the thematics introduced here, in
one form or another, seeks to embrace volatility within the ecology of architectural
production. Fabrication-Aware Form Finding (FAFF) seeks the realisation of an
increasingly flexible design model capable of negotiating the often-competing
imperatives for design within its earliest phases. Critically, it seeks the establishment of multiple pathways through which such transactions can be conducted,
allowing the designer to elect the trade-offs of most benet to the greater project.
Sensory Robotic Control (SRC) challenges the explicit command of machines
through the introduction of either on-board or design model feedback mechanisms
of adaptation and response. Here the pre-programmed point-to-point and linear
moves of G-Code, Kuka Robot Language (KRL) or similar are toppled by
increasing environmental, material or task awareness. Finally, Adaptive Part
Variation (APV) challenges the static and predetermined nature of conventional
construction detailing that must consistently cater to the worst-case via generic
construction tolerances. Instead, APV opens the possibility of deploying distinct
conditional responses to local moments of dimensional variance. Observed collectively, the acceptance of volatility suggests that the pursuit of automation could
be re-framed as the search for a new form of design intelligence, where the logics of
design and making intermingle within a reflective design-to-construction model.
10
Conclusion
The traditions of form nding have provided fertile territory for architects to produce sophisticated forms with rational underpinnings that contribute to their viability as well as their highly specic character. The digital implementation of form
nding has allowed an increase in the number and nature of negotiated inputs and
the addition of fabrication constraints has provided both pragmatic and inventive
influence as well as allowing for exploration of design and material territories of
high volatility.
Robotic fabrication in combination with algorithmic design methodologies and
le-to-factory workflows affords new pathways of reciprocity between design and
fabrication. A scenario where discovery and invention may initially occur on either
side of the design and making ledger but must ultimately permeate the entire design
process. The addition of computer vision sensors extends this paradigm, adding a
further feedback loop to include responses to real-time events and dimensional
errors prevalent during construction.
437
The series of shifts outlined in this paper collectively challenge the dominance of
linear and compartmentalized models of the production chain linking design with
construction. Taken together they set the context for new ecologies of design
practice premised on increased reciprocity and provide many opportunities for
innovation.
References
Barnes, M 1999, Form Finding and Analysis of Tension Structures by Dynamic Relaxation, in
Space Structures vol:14, no.2 pp, 89104.
Clifford, B and McGee, W 2014, La Vote de LeFevre: a Variable-Volume Compression-Only
Vault, in (eds) Gramazio, F, Kohler, M, and Langenberg, S Fabricate Negotiating Design and
Making, Verlag.
Day, A 1965, An Introduction to Dynamic Relaxation, in The Engineer, vol. 29, pp. 218221.
Kilian, A and Ochsendorf, JA 2005, Particle Spring Systems for Structural Form Finding, in
International Journal of Shell and Spatial Structures, vol. 46, no.2, pp. 7784.
Seyedahmadian, A, Oliyan, OT and McGee, W Forthcoming 2015, Developing a Computational
Approach Towards a Performance Based Design and Robotic Fabrication of Fibrous Skin
Structures, in International Journal of Shell and Spatial Structures, Future Visions
Symposium, Amsterdam.
Vasey, L, Maxwell, IJ and Pigram, D 2014, Adaptive Part Variation: A Near Real-Time Approach
to Construction Tolerances, in McGee, W and Ponce de Leon, M (eds), Robotic Fabrication in
Architecture, Art and Design 2014, Springer, Cham Heidelberg New York Dordrecht London,
pp 291-304.
Part V
Industry Papers
Direct Robot Control with mxAutomation:

A New Approach to Simple Software
Integration of Robots in Production
Machinery, Automation Systems, and New
Parametric Environments
Heinrich Munz, Johannes Braumann and Sigrid Brell-Cokcan
Abstract The industry-crises of the past have made it clear how existentially
important it is to have flexible, living production facilities. Automation by means
of industrial robotics has proven to be a key technology in this regard. However,
truly dynamic processes can only be achieved when the robots and the environment
to be automatedmachines, handling equipment, etc.are perfectly integrated,
both operationally as well as from the operators perspective. KUKAs
mxAutomation interface now allows a granular remote operation of the robot in
interaction with modern industrial real-time communicationand beyond that also
entirely new, flexible workflows from design to production towards fabricating
highly customizable products in the creative industry.
Keywords KUKA Direct robot control Remote operation Mass-customization

Robotic arms
1 Introduction
Robots are not an end in themselves but fulll a wide variety of tasks that contribute
to an overall automation solution. In industry they are used e.g. for the preparation
or post-processing of tools or work pieces of a production machine, or for handling
tasks, such as loading and unloading.
H. Munz (&)
KUKA Roboter GmbH, Augsburg, Germany
e-mail: Heinrich.Munz@kuka.com
J. Braumann
Association for Robots in Architecture, University for Arts and Design Linz, Linz, Austria
e-mail: johannes@robotsinarchitecture.org
S. Brell-Cokcan
Association for Robots in Architecture, RWTH Aachen University, Aachen, Germany
e-mail: sigrid@robotsinarchitecture.org
441
442
H. Munz et al.
Fig. 1 KUKA KR16 use-cases in the creative industry: Tripod-mounted as part of an

art-installation by Conrad Shawcross (left), milling wooden elements with an entry-level spindle
(right)
While robots have been used for such purposes for decades, it has only been a
few years that accessible visual-programming environments coupled with dynamic
components for robot control have also opened up the use of robotic arms to the
creative industry where they are now being utilized for a huge variety of tasks, from
fabrication and assembly to interactive installations and performance art (Fig. 1).
Today, plugins such as KUKA|prc (Braumann and Brell-Cokcan 2014) allow
architects, artists, and designers to work from design to fabrication within a single
visual programming environment, where they can not only dene an objects shape
and form, but also visually assemble robotic processes and immediately simulate
the robots movements, automatically checking for singularities and unreachable
positions. Rather than dealing with the intricacies of a robot control data le,
movements and other commands are assembled in a visual system that allows
ne-grained control over most properties such as speed and interpolation. The
automatically generated KRL (KUKA Robot Language) code can then be immediately copied to the robot, enabling a rapid turnaround time as well as an efcient
feedback loop. Today, also users of other industries use KUKA|prc to quickly
prototype processes, e.g. in the aeronautical industry.
However, an issue with such a le-based workflow is the delay between writing
the le and executing it. While KRL les representing hundreds of design variations can be generated within a fraction of a second, they have to be manually
copied onto the robots command execution memory before they can be executed.
Ideally, to fully utilize the potential of accessible visual programming environments, we would need a direct link between PC and robot, through which we can
then dynamically stream commands to the robot. A similar challenge can be
observed within the context of Industry 4.0, where one of the main ideas is to
enable machines to communicate with each other in order to form smaller, functional units, rather than just hierarchical structures with a central control unit.
Direct Robot Control with mxAutomation
443
New industrial solutions and interfaces by KUKA thus enable communication

between automation machines or external controllers of any kind and robots. We
expect that building upon these interfaces will also lead to entirely new and flexible
workflows for direct robot control in the creative industry.
2 Control Systems and Communication in Industry

In order to coordinate the operation of external controllers and robots, communication is necessary between their control systems. Originally such tasks were fullled by signals wired in parallel, until conventional eld buses took over these
functions.
With the emergence of modern Ethernet-based real-time eld buses and Time
Sensitive Ethernet Networks (IEEE 802.1 TSN), the technological basis for efcient
communication between external controllers and robots was established. An
important characteristic of such new communication means is their ability to run
several protocols and thus services via the same Ethernet cable or to operate
multiple buses at the same time. In addition to the standard eld buses EtherCAT,
Pronet and Ethernet/IP, the new KUKA KR C4 robot controller also optionally
supports standard UDP communication over Ethernet.
Based on these improved communication mechanisms, the machine builder or
system integrator can solve automation tasks as usual by creating separate machine
programs, robot programs and the corresponding operator screens for the machine
and the robot separately and putting them into operation by using self-dened bits
and bytes on the communication means. For this the builder or integrator, and any
other participants, such as service personnel, require detailed knowledge of robot
programming.
In order to keep this learning curve lower, and to be able to program the
movements of the robot and the machine at just one location, KUKA has created
mxAutomation (KUKA Robotics 2014) so that machine builders or system integrators can integrate the KUKA robot controller into their remote machine controller. All programming and operator tasks for the robot can then be done
exclusively from the machine controller and its operator panel. Special knowledge
of robot programming is no longer necessary.
mxAutomation consists conceptionally of two main parts (Fig. 2). A server
program created by KUKA runs on the robot controller, which waits for commands
to arrive via one of the eld buses or UDP-Ethernet mentioned above. The actual
robot control program runs on the machine controller; this program is created in the
programming language and with the programming methods of the machine controller. In order to make this possible, KUKA provides the mxAutomation client
library for several controllers, which is integrated by the machine rmware programmer into the rmware. The programmed robot commands and their parameters
are put into the dened data format and nally streamed via the communication
means to the mxAutomation server running on the robot controller. The server
444
H. Munz et al.
Fig. 2 mxAutomation
control concept. Software
components provided by
KUKA are marked orange
interprets the data, executes the desired robot commands and sends return parameters, status messages, etc. back to the mxAutomation library on the machine
controller. In order for the data transfer of the robot commands and their return
values to take place as quickly as possible and deterministically, cyclical process
data of the particular eld bus are used exclusively for this. Several robots can be
simultaneously remotely controlled by a single machine controller.
The mxAutomation library has been created in portable Structured Text and is
available in adapted form for Siemens Simatic, and based on that for Sinumerik
840d, as well as for Rockwell PLCs and CodeSys or ProConOS systems, with more
to be added in the near future. Through an adaptation kit other controller manufacturers can integrate the mxAutomation library into their controllers themselves
after concluding a cooperation agreement with KUKA.
The Application Programming Interface (API) of the mxAutomation library is
based on the programming paradigm of PLCOpens Motion Control Function
Blocks (MCFBs) Part 4 and has been specically expanded for KUKA robots.
Virtually all functions which can be directly programmed on the robot are also
available via the mxAutomation library. This includes general functions, such as
reading of the current robot position, speed and acceleration, and also functions
such as reading and writing local robot I/Os and variables. An application example
for the advantageous use of system variables is sensitive gripping of work pieces or
packaged goods, where the use of axis-specic torque limitation is a proven means.
Within mxAutomation, the system variable TORQMON_DEF like several other
robot system variables can be assigned to an mxAutomation variable, which can
then be written via the KRC_SetSysVar function.
Corresponding blocks are of course also available for the axis-specic or
Cartesian motion of the robot with LIN, PTP or CIRC. Approximate positioning is
also possible, as are exact time-distance functions or interrupts.
445
3 Building upon mxAutomation

In fall 2015, the client library in the adaptation kit (see Sect. 2) has been enhanced
by additional programming languages such as C#, and C++. As such,
mxAutomation is no longer depending on a particular brand of controller, but could
be run on basically any computer, from an embedded controller, a compact SoC, a
smart device to a regular laptop or PC, even from the cloud.
Building upon KUKAs mxAutomation C# client sample library as well as
Robots in Architectures KUKA|prc framework we created a custom client software
that communicates with one or more connected KUKA robots running the
mxAutomation server software. First, the native KUKA|prc commands are processed, andrather than being written into a *.src lemapped to their corresponding C# classes in the mxAutomation library, so that e.g. a KUKA|prc LINear
Movement becomes a KRC_MoveLinear-Absolute object.
Once the server is up and running on the robot controller, data are continuously
exchanged as UDP datagrams between the PC running KUKA|prc and the robot
controller at the robots sub cycle time of 4 ms. Command blocks such as the
aforementioned KRC_MoveLinearAbsolute are executed in sequence and either put
into the robots command buffer, or immediately return status values from the robot,
such as current robot positions, the read of variables or I/Os.
By design, the presence of a buffer makes mxAutomation usable even for non
hard real-time applications running i.e. on Microsoft Windows, while at the same
time making it highly resistant against communication timeouts as long as the
buffer is sufciently lled. This is especially important for regular PCs, where
Windows is not real-time capable and processes such as garbage collection may
cause timeouts of more than several 10 s of milliseconds.
Fig. 3 Dynamic stone surface structuring utilizing the mxAutomation interface to allow dynamic
adjustments of angle and force during the structuring process
446
H. Munz et al.
Using task-based programming via async/await we created a resilient,

high-performance mxAutomation controller and implemented it into Grasshopper.
Thus, the user can rst simulate a robot program via KUKA|prc and then simply
reconnect the commands to the mxAutomation component in order to stream them
to the robot. The interface is already utilized within the scope of the AROSU
research project (Fig. 3) (Brell-Cokcan and Braumann 2015).
4 New Interaction Strategies

A communication process between an automation machine and an industrial robot
is usually quite straightforward, with the machine instructing the robotic arm to
perform a certain task and waiting upon its completion. However, in the context of
the creative industry one of the main appeals of direct robot control is to utilize the
feedback of the robot to inform the process. As such, we have identied and
implemented four different dataflow-modes for mxAutomation (Fig. 4):
Default Mode uses a rather large buffer to store and process commands. When
the input data of the mxAutomation component changes, the robots buffer is
wiped and the new commands are streamed for immediate execution.
Adaptive Mode relies on a much smaller buffer with the goal of keeping data
parametric as long as possible. New data does not cause the process to restart,
but rather adapts the existing commands, which only become static once they
are placed in the robots buffer.
Iterative Mode waits for the completion of a set of commands before accepting
new inputs, so that the returned data can be used to inform the next fabrication
step.
Fig. 4 mxAutomation data flow strategies (Braumann and Brell-Cokcan 2015)
447
Real-time Mode does not achieve hard real-time, but rather uses a shallow
buffer. Commands are streamed stepwise towards the given (and possibly
moving) target position.
Together, these four modes enable a wide range of applications, from straightforward fabrication, to iterative assembly processes that are continuously evaluated
and fully interactive installations.
5 Analysis and Outlook

mxAutomations approach provides several advantages over other solutions, where
the machine controller completely takes over control of the robots drives or
motors. Most importantly, mxAutomation works straight away with all KR
C4-based industrial KUKA robot typesfrom KR AGILUS to the KR 1000 Titan.
There is no risk that the machine controller, due to lack of knowledge of the
mechanical limits of robot motors, gear units and mechanical components, will
exceed these load limits, resulting in mechanical damage to the robot. Additionally,
the machine controller retains all advantages and features of the KUKA robot
controller, such as energy-efcient motion algorithms adapted for the specic robot
arms, loads and moments of inertia, sophisticated exception handling routines or the
entire safety functions.
mxAutomation has proven itself to be a simple and efcient tool and an
end-user-friendly robot integration system within a wide variety of different environments. Its open approach and the availability of the mxAutomation library in
many programming languages enable completely new applications, as demonstrated by its quick integration into KUKA|prc, where it now provides a direct and
immediate interface between design and fabrication.
mxAutomation enables a new way of system thinking as it is necessary for
Industry 4.0 or the Internet of Things: not only the robot arm (a component) should
be in the focus of the user but the whole and integrated system (the solution).
References
Braumann, J and Brell-Cokcan, S 2014, Visual Robot ProgrammingLinking Design,
Simulation, and Fabrication Proceedings of the 5th Annual Symposium on Simulation for
Architecture and Urban Design (SimAUD).
Braumann, J and Brell-Cokcan, S 2015, Adaptive Robot Control, Proceedings of the 33rd
eCAADe Conference, Vienna, Austria.
Brell-Cokcan, S and Braumann, J 2015, Towards Adaptive Robot Control Strategies,
Proceedings of the 35th ACADIA Conference, Cincinnati, USA.
KUKA Robotics, 2014, CODESYS Library for KUKA.PLC mxAutomation 2.0.
Collaborative Robotics
Frank-Peter Kirgis, Peter Katsos and Martin Kohlmaier
Abstract Robotic arms are modelled after the human arm and offer speed, accuracy and strength that in its sum by far exceed the capabilities of humans.
Conversely, an area where the technology of industrial robots had not caught up yet
was the sensitivity to external forces. Today, integrated technologies such as ABB
Force Control allow robots to react to the forces that are applied to their
end-effectors, while the ABB YuMi represents an entirely new kind of sensitive
robot that can assemble small parts and collaborate in a safe way with humans.
Keywords ABB
Collaborative robotics Force Control YuMi
1 Introduction
In 1974 the worlds rst microcomputer controlled, electric industrial robot IRB6
was installed in Sweden. Since then, robotic arms have become indispensable for a
great range of industries, from large-scale automotive installations to single robots
that automate tasks at small enterprises. With the development of new interfaces
and programming methods, robots are not exclusive to the shop floor anymore, but
are utilized for diverse and creative applications, from the research that is being
done in the creative industry (Brell-Cokcan and Braumann 2012), to robots working
in hotels and entertaining guests on cruise ships (Fig. 1).
Among the main reasons for todays widespread popularity of robotic arms are
the robots multifunctionality as well as their superhuman strength, speed, and
accuracyenabling a robot to perform tasks that would be dangerous, harmful, or
even physically impossible for a human worker.
F.-P. Kirgis (&) P. Katsos M. Kohlmaier

ABB Switzerland Ltd, 5400 Baden, Switzerland
e-mail: frank-peter.kirgis@ch.abb.com
449
450
F.-P. Kirgis et al.
Fig. 1 ABB RoboScreen

installation on the quantum
of the seas cruise ship:
Six ABB IRB 6620 robots,
each holding a 100 diagonal
LED screen
However, these attributes also make robots inherently risky, as any machine
capable of lifting a 500 kg car chassis can also cause harm to a human. For this
reason, combinations of safety-certied software and hardware, from mechanical
stops and massive fences to highly reliable 3D cameras, are used to keep the robot
within and persons outside its designated workspace. Therefore, robotic arms
commonly work best in custom-tailored environments such as the highly efcient
manufacturing lines of the automotive industry.
Today, as robotic automation enters new elds, new requirements for robotic
arms arise, when companies do not want to set up entirely new spaces for robots,
but rather want them working alongside the human workforce where they are
applied selectively (Naumann and Fechter 2015) for tasks that they can perform
better, safer, or more efciently. Robot manufacturers such as ABB therefore had to
nd ways in which to make robots safe to interact with, without requiring complex
external safety equipment (Przemyslaw et al. 2014).
2 Integrated Force Control

A rst step towards making a robot safe to work with is to make it sensitive to
outside forces. By default, an industrial robot will follow its programmed path and
only stop if any axis reports a torque value exceeding nominal values. However, for
many applications the torque values reported by current industrial robots are not
ne-grained enough to ensure a safe and reliable process.
ABBs Integrated Force Control is thus designed as an add-on for existing
industrial robots, consisting of a force sensor that reports all six components of
force and torque. On the software side, the sensor feedback can be incorporated into
robot programs through a specic set of RAPID instructions, providing real-time
tactile feedback.
Main applications of force control are grinding, polishing, and deburring
(Fig. 2)all applications where the actual physical object can deviate from the
given digital data.
451
Fig. 2 ABB robot equipped

with a Force Control sensor
for sensitive grinding
By adjusting in real-time for the reported force that the tool applies to the
workpiece it is possible to reduce tool-wear and improve the surface nish. The user
can choose between two strategies: Force Control Pressure keeps the tool at a
constant speed and adjusts the toolpath, while Force Control Speed Change stays
on the programmed path while adjusting the tools speed.
While Force-Control opens up entirely new applications for industrial robots, it
does not automatically make a robot safe to work with, as the sensor provides only
data relating to the tool itself.
Creating a collaborative robot that is safe in its own required ABB to move away
from existing robotic design notions towards new design concepts and kinematics.
3 ABB YuMi
The ABB IRB 14000 YuMi (Fig. 3) was created as a robot for the fourth industrial
revolution, towards enabling the sensitive assembly of affordable, miniaturized,
mobile electronics (refer also to Stolt et al. 2012) in collaboration with human
workers (Fig. 4). It is equipped with a large range of sensors that allow the YuMi to
see and recognize parts, grasp them carefully, and then to snap them together
without breaking them.
Conceptually, the YuMi is modelled after a human worker, consisting of two
seven-axis robotic arms that are controlled by a single controller, which is contained
within the robots base, weighing a total of 38 kg. The arms themselves are also
constructed similar to a human arm, with a strong magnesium skeleton covered by a
floating plastic casing, which in turn is wrapped in soft padding. In the case of a
collision, the sensors in each axis will detect the force and stop the robot within
milliseconds, while the soft construction ensures that neither human nor robot are
hurt.
452
F.-P. Kirgis et al.
Fig. 3 ABB IRB 14000

YuMi collaborative robot
Fig. 4 YuMi robot using its

vision system to collaborate
with a human worker
Despite its low weight and focus on safety, the YuMi can reach a speed of
1.5 m/s and 0.02 mm repeatability. Each arm can manipulate a maximum payload
of 0.5 kg within its reach of 559 mm.
Its software builds upon existing ABB controllers, but allows a new kind of
interactive teachingso called Lead Through Programmingin addition to
IRC5 RAPID as in other ABB robots.
4 Outlook
The ABB YuMi provides an outlook into the future of industrial robotics as it
contains a great number of new technologies that will become increasingly
important for users worldwide. Safe robots will enable companies to simply set up
new machines within their existing workspace, without requiring additional
453
hardware. These advances in technology will also have a profound impact on the
creative industry, from ensuring a safe environment when teaching robotics to
students, to deploying large-scale, sensitive robots on construction sites.
References
Brell-Cokcan, S and Braumann, J (eds), 2012, Rob|Arch Robotic Fabrication in Architecture, Art,
and Design, Springer-Verlag, Vienna.
Naumann, M and Fechter, M 2015, Robots as enablers for changeability in assembly
applications, Proceedings of the 15 Internationales Stuttgarter Symposium, Springer-Verlag,
Berlin.
Stolt, A, Linderoth, M, Robertsson, A and Johansson, R, 2012, Force controlled robotic assembly
without a force sensor, Proceedings of the 2012 IEEE International Conference on Robotics
and Automation (ICRA), Saint Paul, USA.
Przemyslaw, A, Rossano G and Shah, J 2014, Toward Safe Close-Proximity Human-Robot
Interaction with Standard Industrial Robots, Proceedings of the 2014 IEEE International
Conference on Automation Science and Engineering (CASE), Taipei, Taiwan.
Flexible Gripper Design Through Additive

Manufacturing
Marcel Nagel, Felix Giese and Ralf Becker
Abstract Custom-made gripper jaws for particular object geometry involve a

complex and expensive tooling process. Employing Rapid Tooling Technology via
3D printing solutions and additive manufacturing can signicantly improve the work
flow. The design and development process however still leaves room for improvement. By streamlining this process within a browser-based web tool for customized
gripping ngers SCHUNK GmbH & Co. KG shows an improved (semi-)automatic
approach for efcient creation of form tting grippers in record time.
Keywords SCHUNK Gripper technology Rapid prototyping Rapid tooling

Additive Manufacturing Customization 3D printing
1 Introduction
A common gripping principle employed within industrial production consists of an
actuator and two gripper ngers. Therefore a number of two-nger gripper modules
employing electrical or pneumatic actuation with different forces and torques are
available on the market. These systems however still require tting gripper nger
and jaw design for each specic product. Simple jaw designs work best with simple
work piece geometries that are easy to grip e.g. with parallel edges. For the pur-
M. Nagel (&) F. Giese R. Becker

SCHUNK, Lauffen am Neckar, Germany
e-mail: marcel.nagel@de.schunk.com
F. Giese
e-mail: felix.giese@de.schunk.com
R. Becker
e-mail: ralf.becker@de.schunk.com
455
456
M. Nagel et al.
poses of individualized production in architecture, art and design a higher measure

in flexibility is needed. Jaws for form t gripping have many advantages with
regards to force distribution and adjusted gripping. Nevertheless these form tting
gripper ngers require a more sophisticated design process.
2 3D Printing
The signicant advantage of additively manufactured components lies in the
tremendous degree of freedom with regards to geometry. Designers neither have to
pay attention about demolding chamfers nor to undercuts and are thus able to open
up new possibilities for construction. A variety of channels to supply power, signals
or air pressure can for instance be integrated. Complex components that can only be
manufactured with extensive efforts through conventional means can be accomplished relatively easily with the help of 3D printing. Furthermore, movable parts
such as hinges or joints, can be printed in one piece, while the creation of complex
movable parts as well as mixed material manufacturing are still areas of current
research. Production time is signicantly shortened since the components can be
directly generated from the CAD system. Particularly in case of individualized
solutions with low quantities, the additive manufacturing process saves time and
reduces costs. As a side effect, based on the low weight of the printed polyamide or
metal components, new areas of application, such as in the eld of tooling for
smaller robots and handling systems, are made possible.
3 Flexible Gripper Design

For the last several years; a number of modules have already been produced from
the light weight and robust polyamide PA12 using a state of the art selective laser
sintering machine at SCHUNK (www.schunk.com) without the need for chipping
or machining tools. The product range spans from one-piece grip arms with closed
impulse membrane for hygienic use to individually formed and optionally soft and
flexible gripper ngers up to tool changing systems. If required even supply
channels for different media can be integrated. Additionally, complex contours or
individual inscriptions are possible without added cost. Since polyamide is chemically continual and food compatible, it can also reliably be used in environments
with aggressive media or in the food industry.
Flexible Gripper Design Through Additive Manufacturing
457
Fig. 1 Form tting gripper jaws from eGRIP
4 Automated Individuality and Potential Cost Reduction

Additively manufactured grip arm components require special solutions. Through
the development of a browser-based design tool named eGRIP (www.eGRIP.
schunk.com) the process has signicantly been optimized. Using 3D printing in
combination with eGRIP the development of form tting gripper jaws has been
automated from design to ordering and production (Fig. 1).
Comparable with online photo development, the user congures the gripper jaws
by uploading a CAD le (STEP or STL) of the part to be gripped and its parameters. The user is only required to specify some basic parameters such as the gripper
type, jaw-blanks, weight of the object, position and orientation of the object and
nger length. To be sure that the custom designed ngers with the gripper module
does not interfere with anything else in the application, a CAD le (STL) can be
downloaded consisting of the gripper module, the gripping ngers and the object to
be gripped. This can then be loaded into a CAD system for simulation. Figure 2
illustrates the development process employing the eGRIP web interface.
Usually, a lot of construction and production time is necessary for complicated
gripper nger design. Experienced design engineers require between 2 to 8 h for the
development. The web tool eGRIP is able to reduce this time requirement to about
15 min. The required know-how integrated into eGRIP was developed exclusively by
SCHUNK with a patent pending. The production of the additively manufactured
458
M. Nagel et al.
Fig. 2 The eGRIP web interface
components within eGRIP is performed by Materialise Ltd. The web tool is able to
reduce the design phase by 97 %. Manufacturing and delivery can be reduced by up to
88 %. Additionally the cost for gripper jaws is reduced up to 50 %. These results are
based on extensive testing over a period of 6 months with different customers including
experienced design engineers. Based on feedback the web tool interface is available in
german and english and has been designed to be user friendly and easy to use.
5 Outlook
Currently only selected gripper modules can be used in combination with eGRIP
designed ngers. Future work will expand the number of supported gripper types.
The current web tool also requires translation into more different languages to make
the service available to a larger customer base. Form-tting gripper ngers preserve a
dened orientation and position of the gripped object while grasping and are almost
universally applicable. However, they must be designed for a specic product.
Flexible Gripper Design Through Additive Manufacturing
459
For this reason, increasing the speed of additive manufacturing results in new
possibilities for individualized manufacturing. By automating the design process,
an individualized use of form tting gripper ngers is made possible. Future work
will focus on a deeper analysis of graspable materials and form specic wear of the
gripper jaws.
Individual Serialism Through the Use

of Robotics in the Production
of Large-Scale Building Components
Martin Krammer
Abstract In collaboration with ETH Zurich and industry partners, Swiss rm

ERNE AG Holzbau has developed one of the largest robots for building component
manufacturing in Europe. This multifunctional 7-axes machine can manufacture
large building components from many different materials on an industrial scale,
making it possible to mass produce complex shapes in an economically feasible
way. Due to comprehensive planning and production within a digital chain, size
restrictions only become an issue when a length of 48 m is reached.
Keywords Gantry robot
Sequential roof
ERNE AG Holzbau
Large-scale fabrication
1 Introduction
The Swiss timber construction rm, ERNE AG Holzbau, provides building solutions using system and modular construction methods, windows and facade systems
as well as interior solutions. Digital planning involving Building Information
M. Krammer (&)
ERNE AG Holzbau, Laufenburg, Switzerland
e-mail: martin.krammer@erne.net
461
462
M. Krammer
Fig. 1 Interior gallery, soft, robot assembly of 48,624 elements, The Sequential Roof,
Gramazio Kohler Research, Institute of Technology in Architecture, ETH Zurich, 2015
Modeling (BIM), engineering, an extensive knowledge of wood, steel, metal, glass

and hybrid construction, traditional craftsmanship, and the knowledge of a General
Contractor are combined with industrial production to enable the realization of
complex construction projects at a high level of quality and with cost certainty. At
the beginning of 2015, the gantry robot was put into operation for the rst time,
opening up a new chapter in the prefabrication of large-scale building components
at Erne. The aim is to provide large-scale, complex construction parts made in
workshop conditions but with the precision and efciency that comes with using
robotics in production (Fig. 1).
2 Building Construction Challenges Robotics

Thinking in prefabricated (construction) elements has got a long history within the
classical timber construction industry, as pre-industrial carpentry was already the
trade with the highest level of element prefabrication.
Individual Serialism Through the Use of Robotics
463
It is therefore only logical for a timber construction rm to develop an industrial

production method for building parts that reflects the level of prefabrication needed.
Konrad Wachsmannn, who designed Einsteins house in Caputh in 1929, and Jean
Prouv, who designed the BCC house in 1942, are pioneers of prefabricated
industrial construction. But their systems never spread as far as they could have.
Furthermore, industrial prefabrication suffers from negative connotations by many
architects. Prefabricated houses and standardized catalogue building components
are perceived as too uniform and backwards in their design.
So how could a uniform, industrial construction method be used to create results
that are economically sound and aesthetically appealing? The answer lies in a
digital, three-dimensional planning chain, in further intelligent development
through the optimization of design by engineering experts and, last but not least, in
the implementation of robotics that makes it possible to create individual shapes for
building elements, even in the case of series production, and to achieve the sizes
and quantities needed in the building industry.
The design does not need to have a parametric regularity but can consist of many
different shapes, materials and elements combined. An additive assembly process is
used as an elementary addition to the usual subtractive methods, such as milling or
grinding.
3 The Story Behind the Processor How Does a Robot

Become a Construction Robot?
Ernes original aim was to automate the production of timber frame walls and
facade elements. However, the dimensions of such wall and facade construction
components meant that using an articulated robot would have been out of the
question, for both technical and nancial reasons. The minimum requirements were
set to be a cantilever of over 4.00 m and a load capacity of 250 kg.
Today, the established gantry robot can be used for supporting structures, walls,
roofs, facades and concrete formwork construction up to 48.00 m in length, 1.40 m
in height and a width of 5.60 m. The interchangeable head allows a large selection
of tools to be used, ranging from tools for lifting, grasping, holding and nailing to
those for milling, cutting, gluing and many more (Figs. 2 and 3). Seven axes,
including two parallel axes, allow for the greatest possible level of flexibility.
Regarding both material and choice of tools, there is hardly anything that this robot
cannot do.
464
M. Krammer
Fig. 2 Guedel gantry robot, equipped with gripping device and nailing aggregate ERNE AG
Holzbau
Fig. 3 Guedel gantry robot, equipped with gripping device and nailing aggregate ERNE AG
Holzbau
465
Fig. 4 Guedel gantry robot, production of truss girders ERNE AG Holzbau
4 The Sequential Roof

In 2013, the ETH Zurich asked Erne to realize a parametric roof design for the
Institute of Technology in Architecture. A competent development partner was
found in Guedel AG in Langenthal, a company that had already realized a woodwork robotic system, albeit on a smaller scale. Guedel specializes in manufacturing
robots and robotic systems in the automobile and logistics industries. A partner for
the control and operation software was found in ROB Technologies in Zurich.
These two companies provided just the expertise that Erne needed and their
cooperation resulted in one of the largest robotic systems for building component
manufacturing in Europe.
Small timber elements, a locally available building material, and a total of
800,000 nails were used to create the 2308 m2 free-form timber roof. It was made
out of 120 long trusses of 15 m length, and 48 short truss girders of 3 m length, and
was constructed out of a total of 48,624 elements (Fig. 4).
The roof was developed at the Institute of Technology in Architecture (ITA) at
the ETH Zurich. It was designed at the Chair of Architecture and Digital
Fabrication and planned by the Arch-Tec-Lab AG in Zurich. The design of the roof
structure, the structural dimensioning of the nodes and members as well as the
feeding in of the production data were all connected by an integrated digital
planning process. The robot was controlled via the relevant interfaces, which were
directly connected to the 3D model.
466
M. Krammer
Fig. 5 Truss girders, lifted on site with a construction period of two weeks for a total of 2308 m2
roof structure. The Sequential Roof, Gramazio Kohler Research, Institute of Technology in
Architecture, ETH Zurich, 2015
The architects Rhino model was sent to the structural engineering consultants,
who used a nail-positioning generator to create the nail image required for the
structure. Erne checked the data upon arrival using a Solibri Model Checker and
then prepared it for use in robot machining via a BTL interface. This all means that
the distances between planning and execution are amazingly short and direct.
Finally, the robot is also used for quality assurance in production. The system
uses a photogrammetric system to compare every node and the constructed nail
image with the target plan, to ensure that everything has been built as planned. After
fabrication, the trusses are then delivered to the site (Fig. 5).
4.1
Gantry Robot Data
7 axes (2 parallel)
Processing pieces up to 48 m long, 5.60 m wide and 1.40 m high
3 zoneseach zone is 14 m long
Zones can be combined with each other
Maximum speed in all axes 120 m/min (7 km/h)
Maximum load-bearing capacity at the gripper/suction head 250 kg
Tools: sawing, milling, screwing, nailing, lifting, holding plus many more
Machines both timber and steel and other materials
The robot calculates the sequence of construction independently
467
Acknowledgments The Sequential Roof is developed by: Gramazio Kohler Research, Institute of
Technology in Architecture, ETH Zurich (design and development), Institute of Technology in
Architecture and Arch-Tec-Lab AG (planning) together with the following experts: Dr.
Lchinger + Meyer Bauingenieure AG, SJB. Kempter Fitze AG, Estia SA, Chair of Structural
Design (Prof. Dr. Josef Schwartz), ETH Zurich, Institute of Structural Engineering (Prof. Dr.
Andrea Frangi), ETH Zurich, ROB Technologies AG and the team members: Aleksandra Anna
Apolinarska (project lead construction), Michael Knauss (project lead building project), Jaime de
Miguel (project lead preliminary project), Selen Ercan, Olga Linardou. The project is made
possible and assigned by: ETH Zurich Infrastrukturbereich Bauten.
PowerMILL Robot: Simplifying

the Complex Issue of Robotic
Programming
Robert Simunic
Abstract Exciting new applications for robots are being developed each day.
Some of these are in the traditional manufacturing environment, with the more
innovative opportunities for robots being found in the arts, architecture, multimedia,
and digital design industries. As these new opportunities are explored, then so too
does the supporting software need to evolve to support what are often ground
breaking new applications. PowerMILL Robot, from Delcam, is one such software
system that provides an easy to use computer interface allowing the programmer to
design, analyze, and simulate in a single virtual environment. The benet this offers
to designers is the ability to overcome traditional manufacturing barriers.
Keywords Delcam
PowerMILL Robot
Programming
Simulation
Verication Additive manufacturing Subtractive manufacturing
1 Introduction
PowerMILL Robot, developed by Delcam Ltd, was created to eliminate the complex and laborious task traditionally associated with programing robots and generating robot code. This is achieved by offering a complete virtual workflow, where
a 3D CAD model is opened within the software, toolpaths generated and overlaid
onto the CAD model, followed by complete simulation and verication of the entire
robotic cell in native robot language. The simplied workflow makes it easy to
program, simulate, review and rene toolpaths, whilst also enabling robots to
achieve levels of capability similar to many CNC machines.
R. Simunic (&)
Delcam, Melbourne, Australia
e-mail: robert@delcam.com.au
469
470
R. Simunic
2 Workflow
The modern programming workflow (Fig. 1) developed by PowerMILL Robot
offers a consistent approach to being able to manufacture a wide variety of parts.
With the workflow remaining constant it allows the programmer to be easily
trained, become more skillful and adaptable in a very short amount of time.
Generate and Calculate Toolpaths. The 3D model can vary from an engineered CAD model as in composites manufacture, to laser scanned and reverse
engineered model in STL format, such as those typically used in sculpting applications. A wide variety of toolpaths can be generated to suit a variety of end
effectors, common subtractive machining, additive and deposition technology, and
nishing techniques.
Select the Robot and Cell Conguration. A wide variety of robot brands are
supported, with up to 18 axes in a single cell.
Set the Part Position. One of the most critical elements of successfully using
the robot is aligning the virtual datum with the physical datum.
Fig. 1 PowerMILL Robot programming workflow
PowerMILL Robot: Simplifying the Complex Issue
471
Simulate and Analyse the Toolpaths Relative to the Robot. Toolpaths are
typically created in a generic format, meaning that they could be used on a wide
variety of machine tools, or robots. The simulation and analysis process ensures that
the toolpaths will successfully run on the specic robot cell. This process checks for
any collisions, singularities, or toolpaths running out of the robots range.
Convert Toolpath to Native Robot Language Code. Robot offers the ability to
congure the postprocessor to convert the virtual toolpaths into native robot code,
including any unique custom functions specic to each cell.
Start Machining. The code prepared for the robot can be immediately run with
no further operator input. As many of the programs run into millions of lines of
code it is common for the robots to run unattended in a lights out mode, enabling
time efciencies to be achieved by running the robot overnight.
3 Applications
There are a wide variety of applications and industries supported by PowerMILL
Robot, with the most typical being subtractive machining using a rotating cutting
tool (Fig. 2, left). However, a wide variety of end effectors are supported allowing
for a range of different tools to be used. This can include the use of nishing and
polishing discs, hot wire, plasma cutters, as well as the newer additive technologies
of laser cladding (Fig. 2, right) and plastics extrusion.
Fig. 2 Digital model and nished GRC sculpture at Monument Park Melbourne, as manufactured
by GRC Environments (left). The Welding Institute in the UK generate toolpaths to repair aircraft
engineer turbine blades and impellers for automotive turbo chargers by laser cladding inconel and
stainless steel (right)
472
R. Simunic
4 Outlook
One of the most exciting opportunities in the digital age of manufacturing is where
designers, creators or artists are not limited by traditional rules of manufacture;
transfer of shapes to fabrication; or production time. Materials can be blended to
offer an amazing array of mechanical properties, and the physical size of parts is
supported through the ability of robots to offer large format manufacturing for a
fraction of traditional machinery prices. This breadth of capability is also being
increasingly supported through digital design tools, which in turn is supported
through PowerMILL Robotsimplifying the task of robotic programming.
Erratum to: The SPIDERobot:

A Cable-Robot System for On-site
Construction in Architecture
Erratum to:
The SPIDERobot: A Cable-Robot System for On-site
Construction in Architecture in: D. Reinhardt et al. (eds.),
Robotic Fabrication in Architecture, Art and Design 2016,
DOI 10.1007/978-3-319-26378-6_17
Due to an unfortunate misunderstanding some authors were not listed in Chapter 17.
The updated original online version for this chapter can be found at
DOI 10.1007/978-3-319-26378-6_17
J.P. Sousa (&) C.G. Palop
Faculty of Architecture, University of Porto + CEAU/DFL, Porto, Portugal
e-mail: jsousa@arq.up.pt
C.G. Palop
e-mail: cpalop@arq.up.pt
E. Moreira A.M. Pinto G. Veiga
INESC TEC, Porto, Portugal
e-mail: eduardo.j.moreira@inesctec.pt
A.M. Pinto
e-mail: andry.pinto@fe.up.pt
G. Veiga
e-mail: germano.veiga@inescporto.pt
E1
E2
J. Lima
INESC TEC + Polytechnic Institute of Bragana, Bragana, Portugal
e-mail: jlima@ipb.pt
P. Costa P. Costa A. Paulo Moreira
INESC TEC + Faculty of Engineering, University of Porto, Porto, Portugal
e-mail: paco@fe.up.pt
P. Costa
e-mail: pedrogc@fe.up.pt
A. Paulo Moreira
e-mail: amoreira@fe.up.pt
J.P. Sousa et al.
Scientic Committee
Sean Ahlquist, Taubmann College of Architecture and Urban Planning,

University of Michigan, USA
Kristy Balliet, Knowlton School of Architecture, The Ohio State University,
USA
Sascha Bohnenberger, Bollinger + Grohmann, Melbourne, Australia
Johannes Braumann, Association for Robots in Architecture/University for Arts
and Design Linz, Austria
Sigrid Brell-Cokcan, Robots in Architecture/Chair of Individualized Building
Production, RWTH Aachen University, Germany
Jan Brninghaus, Chair of Individualized Building Production, RWTH Aachen
University, Germany
Michael Budig, Architecture and Sustainable Design, Singapore University of
Technology and Design, Singapore
Daniel Davis, WeWork, USA
Ryan Luke Johns, GREYSHED, USA
Chris Knapp, Abedian School of Architecture, Bond University, Australia
Branko Kolarevic, Faculty of Environmental Design, University of Calgary,
Canada
Andrew Kudless, Architecture, California College of the Arts, United States
Jason Kelly Johnson, Future Cities Lab, San Francisco, USA
Wes McGee, Taubmann College of Architecture and Urban Planning,
University of Michigan, USA
Andrew Payne, LIFT Architects, Cambridge, USA
Steffen Reichert, ICD Institute for Computational Design, University of
Stuttgart, Germany
Tobias Schwinn, ICD Institute for Computational Design, University of
Stuttgart, Germany
Marshall Prado, ICD Institute for Computational Design, University of Stuttgart,
Germany
Tim Schork, MADA, Monash University, Australia
Marc Aurel Schnabel, School of Architecture, Victoria University of
Wellington, New Zealand
Art and Design 2016, DOI 10.1007/978-3-319-26378-6
473
474
Scientic Committee
Gregor Steinhagen, IPS Institut fuer Produktionssysteme, Technical University

of Dortmund, Germany
Oliver Tessmann, DDU (Digital Design Unit), Faculty of Architecture,
Technische Universitt Darmstadt, Germany
Aaron M Willette; Manufacturing Design; Apple Inc.; USA.

Rob Fab in Arc Art and Des 201

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Rob Fab in Arc Art and Des 201

Uploaded by

Copyright:

Available Formats

Dagmar Reinhardt

Robotic Fabrication in Architecture,

Dagmar Reinhardt Rob Saunders

Funded by KUKA Robotics and the Association for Robots in Architecture

Library of Congress Control Number: 2015955855

Foreword by the Association for Robots

Foreword by the Association for Robots in Architecture

With the increasing proliferation of previously arcane robotic knowledge, new

Foreword by the Association for Robots in Architecture

From Robotic Fabrication to Creative Robotics

Research and Discussions

Keynotes, Research and Project Papers

computational modeling to robotic manufacturing of joints, deposition of material,

Robots at the Sagrada Famlia Basilica: A Brief History

Building a Bridge with Flying Robots . . . . . . . . . . . . . . . . . . . . . . . . .

Autonomous Robotic Assembly with Variable Material Properties. . . . .

An Integrated Modelling and Toolpathing Approach

Fabric Forms: The Robotic Positioning of Fabric Formwork . . . . . . . . 107

Path Planning for Robotic Artistic Stone Surface Production . . . . . . . . 123

Fabrication-Aware Design of Timber Folded Plate Shells

BotBar: A Platform for Multi-disciplinary Design Education. . . . . . . . . 251

Stigmergic Accretion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 399

Direct Robot Control with mxAutomation: A New Approach

Scientic Committee . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 473

Robots at the Sagrada Famlia Basilica:

Abstract Cutting stone by hand to the architects precise measurements is an

1 Enter the Robots: The Nave Columns (1989 Onwards)

Fig. 1 Nave columns as built and a geometrical description of their derivation

Robots at the Sagrada Famlia Basilica

Robots at the Sagrada Famlia Basilica

2 Lessons from the Stonemasons Ingenuity: Passion

This period coincided neatly with the introduction of rapid prototyping to

Robots at the Sagrada Famlia Basilica

3 7-Axis Robotic Stone Cutting: The Passion Faade

Robots at the Sagrada Famlia Basilica

Robots at the Sagrada Famlia Basilica

Fig. 16 Jordi Barbany and

Robots at the Sagrada Famlia Basilica

and beyond. These pataphysical machines articulate symmetricallythrough weird

Fig. 1 Stargate. New-Territories/Francois Roche-Stephan Henrich/Broomwitch/2009/ http://

Antipersonal Nymphomanic Wanderer

Fig. 2 Wanderer. New-Territories/Francois Roche-Stephan Henrich/He shoot me down/2009/

areas such as post-military zones with unreachable infrastructure interstices.

Introverted Eczemental Recycler

Instructions: Transforming informal heaps into deformed ones, this machine

Fig. 3 Recycler. New-Territories Academia/Francois Roche-Stephan Henrich/Adam Orlinski/

Difference and Repetition/Intricate Randomizer

Fig. 4 Randomizer. New-Territories Academia/Francois Roche-Stephan Henrich/Guo and

Body. Builder. Shitter

Instructions: Rent an agile hyper-proteined device, shitting liquid concrete in a

Fig. 5 Shitter. New-Territories/Francois Roche-Stephan Henrich/An architecture des humeurs/

Fig. 6 OCD Packer. New-Territories/Francois Roche-Stephan Henrich/Welostit/2006/ http://

Precautions for use: In order to achieve high efciencies in ordering, numbering,

Instructions: Rent this under-seawater device that behaves as an extractor removing

Fig. 7 Cyclothymia. New-Territories Academia/Francois Roche-Stephan Henrich LAB

Bulimic Enclosure Weaver

Fig. 8 Enclosure Weaver. New-Territories Academia/Francois Roche-Stephan Henrich/Qi Su and

Instructions: Rent this devicethe Astrolabe Stuttererto ascertain the level of

Fig. 9 Astrolabe Stutterer. New-Territories/Francois Roche-Stephan Henrich/TBWND/2010

Building a Bridge with Flying Robots

A. Mirjan (&) F. Gramazio M. Kohler

Building a Bridge with Flying Robots

Fig. 1 A suspension footbridge structure obtained through a simulation tool