Understanding the Human Hand

for Advancing Robotic Manipulation

Workshop at Robotics: Science and Systems 2009
Organizers: Ravi Balasubramanian, Yoky Matsuoka
Website and Proceedings Production Chair: Dillon Eng
Seattle, WA

Overview Speakers Abstracts

Contributors Abstracts Workshop Schedule

 SpeakersAbstracts Back

ProfessorPeterAllen
LowDimensionalDataDrivenGrasping......5-6

Prof.GeorgeBekey
Thehistoryofgraspingwithroboticandprosthetichands.......7

Dr.LynetteJones
HumanHandFunction:TheCouplingofSensoryandMotorSystems.89

Prof.DerekG.Kamper
Transmissionofmusculotendonforcestotheindexfinger.....1011

Prof.HaruhisaKawasaki
ForceSensationoftheHumanFingerwhenUsingaMultiFingeredHapticInterface.....12-13

Prof.SusanLederman
ScientificApproachestotheStudyofHumanHandFunction.....14-15

Prof.GerryLoeb
RobustBiomimeticTactileSensingandGripControl.....16-17

Prof.MarcoSantello
SynergisticControlofHandMusclesThroughCommonNeuralInput.1820

Prof.LuigiVillani
Originalapproachestointerpretationlearning,andmodelling,fromtheobservationofhumanmanipulation.2122

Prof.VeronicaSantos
Developmentofartificialgripreflexesthatutilizetheadduction/abductioncapabilitiesofhumanfingers.........................................2324

Prof.FranciscoValeroCuevas
Theneuromuscularsystemsdoesordinarymanipulationtasksthehardway:Lessonsforroboticmanipulators?...............................3536

 Contributors Abstracts Back

G. Stillfried and P. van der Smagt

Human hand kinematics based on MRI imaging..2728

S. Sueda and D.K. Pai

Dynamic Hand Simulation with Strands..........29

O. Celik, Q. Gu, Z. Deng, and M.K. OMalley

Movement Intermittency and Variability in Human Arm Movements...3031

A. Israr, H. Kapson, V. Patoglu, and M.K. OMalley

Effects of Force and Displacement Cues while Adapting in a Rhythmic Motor Task....3233

M. De Gregorio, K. Bair, and V.J. Santos

Characterizing reflexive grip responses to rotational perturbations of an object in precision grasp...3435

R.S. Dahiya, M. Gori, G. Metta, and G. Sandini

Better Manipulation with Human Inspired Tactile Sensing..3637

L.C. Miller and J.P.A. Dewald

Quantification of wrist/finger flexion forces and EMGs as a function of limb loading in chronic hemiparetic stroke....3839

D.R. Faria, R. Martins, and J. Dias

Human reachtograsp generalization strategies: a Bayesian approach.4041

C.M. Goehler, R.F. ff. Weir, and W.M. Murray

EMGdriven Forward Dynamic Simulations of the Hand performing American Sign Language...4243

J. Iqball, N. Tsagarakisl, and D. Caldwelll

Design Optimization of a Hand Exoskeleton Rehabilitation Device..4445

K. Dermitzakis and A. HernandezArieta

Anthropomimetic approach to the design of a prosthetic robot hand......4647

Q. Fu, W. Zhang, and M. Santello

Learning of the anticipatory mapping between digit positions and forces in twodigit object manipulation...4849

M.E. Maitland and M. Epstein

Analysis of Finger Position During Two and ThreeFingered Grasp: Possible Implications for Terminal Device Design......5051

M. Ittyerah
Does practice affect hand ability? The reliance on frames of reference......5253

L.Y. Chang and N.S. Pollard

Video survey of pregrasp interactions in natural hand activities.....5455

B. Hannaford and J. Dosher

Haptic Exploration of Spheres....5657

T. Feix, R. Pawlik, H.B. Schmiedmayer, J. Romero, and D. Kragic

A comprehensive grasp taxonomy..5859

E. Greenwald, J. Pompe, S. Hsiao, and A. Okamura

Quantification and Reproduction of Human Hand Skin Stretch and its Effects on Proprioception......6061

C. HerreraRincon, B. Jorrin, A. SanchezJimenez, C. Avendao, and F. Panetsos

Cytochrome oxidasebased morphological evaluation of the modifications of the cortical
volume due to electrical stimulation of sectioned peripheral nerves: a shortterm study.....6264

W. Merlo
Threedimensional, Continuousmotion Ball Joint Technologies for Prosthetic Wrist Applications....6567

R. Balasubramanian, B. Dellon, S. Pradhan, S. Gibbs, and Y. Matsuoka

Quantifying the Dimensions of Human Hand Dexterity Through a Survey of Daily Tasks...68-
 Overview Back

Recentadvancesinthehumanscienceshaveenergizedthefieldofroboticstowardpersonalassistantsandbrainmachine
interface.Thereisanincreasedinteresttosolvetheroboticmanipulationquestion:Canwebuildrobotichandsthatcan
accomplishourdailymanipulationtasks?Thehumanhandisadeptatmanydiversetasksinvariedcontexts,including
powerandprecisiongrasping,twisting,andtapping.Butwestilldonotknowwhatfeaturesofthehumanhandenable
suchcapability.Forexample,dobiomechanicalfeatureslikethecomplextendonhood,synergisticmuscleactuation,and
boneshapesmakethedifference?Orisittheneuralcodingofmovement?Importantlyforrobotics,weneedto
understandwhatfeaturesshouldbeincludedinfuturerobotichands.Thisworkshopisaforumforresearcherstodiscuss
manipulationviewedinlightofthehumanhandsfeaturesandhopestopushthestateoftheartoftherobotichand.We
expectthattheworkshopwillbringtogetherresearchersfromdiverseareassuchasrobotics,biomechanics,
neuroscience,andanthropologists.

 Low-Dimensional Data-Driven Grasping
Peter K. Allen Matei T. Ciocarlie Corey Goldfeder Hao Dang
Department of Computer Science, Columbia University, New York, USA

1 Grasp Synthesis the vector a Rb containing the amplitudes along

each subspace axis. In previous work [3], we have dis-
In general, automatic grasp synthesis can be thought cussed the feasibility of finding good grasps for dex-
of the task of finding the combination of hand posture terous hands by searching a subspace defined by two
(intrinsic degrees of freedom, or DOFs) and position eigengrasps. This implies a significant dimensional-
(extrinsic DOFs) that produces a stable grasp, ac- ity reduction of the grasp quality function domain,
cording to a given grasp quality metric. From this which can be expressed as
perspective, it can be approached as an optimiza-
tion problem, seeking to maximize the value of the Q = f (a, w), a R2 (2)
grasp quality Q expressed as a function over a high-
dimensional domain: However, this low-dimensional subspace is only use-
ful as long as it contains the hand postures needed
Q = f (p, w) (1) for stable grasps of a large variety of objects. The
results presented in [3] show that, in general, pos-
If d is the number of intrinsic hand DOFs, p Rd tures where the hand conforms perfectly to the sur-
represents the hand posture and w R6 contains face of the target can not be found in eigengrasp
the position and orientation of the wrist. The tra- space. However, by searching this subspace we can
ditional form for specifying a hand posture is to set usually find a posture that is very close to a desired
a value for each individual DOF of the hand. For grasp. The eigengrasp space can therefore be thought
complex hands, such as the human hand modeled in of as a pre-grasp, or planning space: the best pre-
our study with d = 20, the high dimensionality of the grasps found in this subspace have a good chance of
posture space makes direct searches for good grasps producing stable grasps by simply closing each fin-
intractable. To make this tractable, we perform the ger until motion is stopped by contact with the ob-
grasp planning task in a subspace of highly reduced ject. This suggests a two-stage grasp planning algo-
dimensionality. Our approach is based on the results rithm: the first stage searches the low-dimensional
of Santello et al. [8], who have shown that the range eigengrasp subspace, while the second stage tests the
of postures that humans use in everyday grasping ex- resulting pre-grasps and outputs the best solutions.
hibits significant clustering in the d-dimensional DOF We have applied this idea to online automated grasp-
space. However, while human grasping subspaces can ing where the reduced subspace is searched to find a
be determined through user studies, defining similar stable grasp for objects [1].
subspaces for non-anthropomorphic robotic hands is
an interesting open problem. In previous work [4],
we have applied this concept to five hand models, 2 A Database of Grasps
using the results of Santello et al. for the anthro-
pomorphic models (such as the DLR and Robonaut Our low-dimensional grasp planner has allowed us to
hands) and empirically derived subspaces for the non- create the Columbia Grasp Database, a freely avail-
anthropomorphic ones (such as the Barrett hand). able collection of hundreds of thousands of form clo-
We consider a hand posture subspace defined by a sure grasps for thousands of 3D models [5]. Our pri-
number of d-dimensional basis vectors called eigen- mary interest is in using an objects 3D geometry
grasps; the implication is that these vectors can be as an index into the database. Given a new 3D ob-
linearly combined to closely approximate most com- ject, we can find geometric neighbors in the database,
mon grasping positions. By choosing a basis com- and the accompanying stable grasps for these similar
prising b eigengrasps, a hand posture p placed in the objects. If the number of objects to be grasped in
subspace defined by this basis is uniquely defined by the database is very large and comprehensive then

1
robotic grasping becomes a pre-computed database
lookup. While we have not yet achieved this level of
performance it is our directional goal. 1 1 1 2 3 5
The most direct way to construct a grasp database
is to collect grasping data from human volunteers.
We could gather a large set of example objects, outfit
an army of graduate students with grasp-capture de- 0.135 0.152 0.150 0.049 0.028 0.095

vices such as datagloves, and record the results. Un-

fortunately, this approach is prohibitively time con-
suming for large scale data acquisition. More impor- 1 1 1 1 2 4

tantly, data collection from humans can only produce

grasps with the human hand. Since many popular
robotic hands cannot be easily mapped to the hu-
0.110 0.075 0.098 0.075 0.083 0.121
man hand, a useful database should include grasps
with multiple hands.
The Columbia Grasp Database was created us-
ing GraspIt!, a publicly available grasp planning 2 2 3 3 5 5

and analysis tool developed by our group [7]. The

database is intended to be used in conjunction with
GraspIt! or a similar simulation tool; as we have
0.028 0.016 0.022 0.072 0.029 0.046
shown in previous work [6, 2], planning results ob-
tained in simulation can be successfully applied to
real robotic hands performing grasping tasks. Figure 1: Three example models and their grasps
Even for grasp planning algorithms that do not from neighbors in the grasp database
rely on simulation, an environment such as GraspIt!
is an important tool for evaluation, as grasp qual-
volume 5024, pages 104113. Springer Lecture Notes in Com-
ity measures are generally impossible to compute in puter Science, 2008.
physical experiments. Part of our motivation in pro-
[2] M. Ciocarlie and P. K. Allen. On-line interactive dexterous
ducing a grasp database was to provide a benchmark grasping. In Eurohaptics, 2008.
for robotic grasping tasks. Using a common bench-
[3] M. Ciocarlie, C. Coldfeder, and P. Allen. Dimensionality re-
mark will make it possible to directly compare grasp duction for hand-independent dexterous robotic grasping. In
planning algorithms, which is currently difficult to IEEE-RSJ Intl. Conf. on Intelligent Robots and Systems,
2007.
do.
We briefly discuss a new database-backed grasp [4] M. Ciocarlie, C. Goldfeder, and P. Allen. Dexterous grasp-
ing via eigengrasps: A low-dimensional approach to a high-
planning algorithm based on the data we have col- complexity problem. In Robotics: Science and Systems Ma-
lected (see fig. 1). Using this algorithm, we illustrate nipulation Workshop - Sensing and Adapting to the Real
World, 2007.
the usefulness of a database for grasping, and high-
light some of the lessons learned during its construc- [5] C. Goldefeder, M. Ciocarlie, H. Dang, and P. K. Allen. The
columbia grasp database. In IEEE Intl. Conf. on Robotics
tion. We also provide execution results over the entire and Automation, 2009.
set of objects in the database at their primary scale.
[6] D. Kragic, A. Miller, and P. K. Allen. Real-time tracking meets
We believe this to be one of the most comprehensive online planning. In Intl. Conf. on Robotics and Automation,
tests found in the grasp planning literature, demon- 2001.
strating the use of the database as a benchmarking [7] A. Miller, P. K. Allen, V. Santos, and F. Valero-Cuevas. From
tool. robot hands to human hands: A visualization and simulation
engine for grasping research. Industrial Robot, 32(1), 2005.

[8] M. Santello, M. Flanders, and J. F. Soechting. Postural hand

2.0.1 Acknowledgments. synergies for tool use. Journal of Neuroscience, 18(23):10105
10115, 1998.
This work was funded in part by NIH BRP grant
1RO1 NS 050256-01A2.

References
[1] M. Ciocarlie and P. K. Allen. On-line interactive dexterous
grasping. In Haptics: Perception, Devices and Scenarios,

2
 The history of grasping
with robotic and prosthetic hands
George A. Bekey

Professor of Computer Science Emeritus

University of Southern California

Abstract

The human hand is an amazing and complex instrument, capable of grasping objects of
arbitrary shape (as well as having numerous other functions). Attempts to replicate some
of these functions with both amputees and robots make clear how difficult it is to replace
the hand by an electromechanical (or pneumatic) device. In this talk we will review the
fascinating history of these devices. Prosthetic hands began with a simple hook
(associated with pirates in the Middle Ages), but grasping originated with the invention
of the Dorrance parallel hook in 1912, which is still in common use. Robot hands
began with parallel jaw grippers in industry in the 1960s, followed by the introduction of
systems with 3 or more fingers in the 1980s.

During the past 20 years both robot hands and prosthetic hands have undergone dramatic
improvements. Commercial grippers for robots with 4 and 5 fingers are now available,
and some are incorporated in humanoids. For industrial grasping a number of non-
anthropomorphic grippers have been developed. Dramatic improvements in prosthetic
hands have come about during the past decade, largely in research laboratories. Recently
these developments have been accelerated by the DARPA program in upper extremity
prosthetics. The parallel history of robot and prosthetic hands will be examined with
attention to such issues as control vs. autonomy, cost, reliability, and applications. There
will be few if any equations, so anyone will be able to grasp this talk.
 Human Hand Function: The Coupling of Sensory and Motor Systems
Lynette Jones
Department of Mechanical Engineering,
Massachusetts Institute of Technology

From an evolutionary perspective the hands of primates are often considered the absolute
bedrock of mammalian primitiveness with a skeletal structure that changed little over 65 million
years. The hand evolved relatively early with capabilities that preceded the development of the
cerebral structures required to make use of its potential. The importance of the evolution of
cortical neuronal hardware that processes incoming sensory information from the hand and via
corticospinal neurons directly controls movements of the fingers cannot be underestimated in
analyzing the versatility of primate hands. One major evolutionary change that resulted in a
remarkable increase in the functionality of the hand was the development of the opposable thumb.
Opposition involves flexion, abduction and medial rotation of the thumb so that the top of the
thumb can make contact with the tips of the fingers. This movement is essential for exploring
and manipulating objects and in the absence of the thumb, for example following amputation, it
is estimated that the hand loses 40% of its functional capacity. The extensive area of contact
between the pulps of the thumb and index finger is a uniquely human characteristic that reflects
the more distal position of the thumb and its longer length relative to that of the index finger.
Another feature of the human hand that distinguishes it from other primate hands is the presence
of rough horseshoe-shaped tuberosities on the distal phalanges to which the soft tissues of the
palmar pads are attached. These pads facilitate the distribution of pressure during grasping and
allow the fingertips to conform to uneven surfaces. Further evidence of the specialization of the
hand comes from the relative mass of the musculature devoted to the thumb, which makes up
about 39% of the weight of the intrinsic muscles within the hand.
Human dexterity has been studied experimentally from a number of perspectives ranging
from detailed analyses of the sensorimotor control involved in object manipulation, to kinematic
studies of skilled activities such as typing and piano playing. The manipulative skill of the
human hand is most apparent in the tight coupling that occurs between the motor and sensory
systems when the hand reaches and grasps an object. Within 70 ms of contact, the forces used to
grasp an object are modified by feedback from mechanoreceptors in the skin. Numerous studies
have recorded grip forces as the index finger and thumb grasp and lift an object and shown how
these forces vary as a function of the geometry and weight of the object, the shape and texture of
the contact surface, and the friction between the hand and the object. The crucial role played by
cutaneous mechanoreceptors in the modulation of grip forces has been highlighted in
microneurographic studies in which the activity of isolated tactile afferent fibers is recorded as
people grasp and lift objects. These experiments show that tactile afferents encode the timing,
magnitude, direction and spatial distribution of fingertip forces, and the friction between the skin
and object. When these inputs are eliminated following local anesthesia, the hand is unable to
compensate rapidly as an object slips between the digits, and grip forces are considerably larger
than those used normally. In addition, the predictive control of grip force based on previous
experience is impaired when the expected sensory feedback is not available.
Kinematic analyses of more complex manual tasks such as typing and piano playing have
offered insight into the internal representation of motor activities and have shown that muscles
controlling movements of the fingers are not controlled independently. Although keyboard tasks
are often thought of as serial activities, analyses of the sequencing of finger movements indicate
that there is an anticipatory component to performance particularly when consecutive keystrokes
are executed by different hands. Skilled typists commit themselves to typing a particular letter
approximately three characters in advance of the current keystroke and can visually process up to
eight characters in advance of the character being typed. For expert typists, the interval between
successive keystrokes is typically 100-200 ms, but intervals as short as 60 ms are not infrequent.
Piano playing and sending Morse code are performed at rates comparable to those reported for
typing. Experienced Morse code operators can send code at a rate of 20-30 words per minute,
and at faster tempos the inter-note intervals of pianists are often 80-100 ms, and for brief periods
of time, such as when playing trills, they can produce 20-30 notes per second. The similarity in
peak movement speeds recorded from skilled typists and pianists suggests that they are
performing near the mechanical and neural limits of the human hand.
Skilled manual activities rely on sensory feedback from cutaneous, muscle and joint
mechanoreceptors for successful execution. The ability to perceive finger movements and sense
the position of the fingers is not limited to a single class of receptor, but is derived from a
number of redundant sources. In muscle, spindle receptors respond to changes in muscle length,
with spindle primary afferents being more sensitive to the velocity of muscle contraction and
secondary afferents displaying much greater position sensitivity. The discharge rates of these
receptors are not simply a function of changes in muscle length, but also reflect the activity of
their own motor innervation, called the fusimotor system, which can regulate the sensitivity of
muscle spindles. To decode signals from muscle spindle afferents the central nervous system
must have access to information regarding the level of fusimotor activity to distinguish between
afferent discharges that are proprioceptively significant from those that are the result of
fusimotor activity. Mechanoreceptors (SA II) in hairy skin also discharge in response to finger
movements which typically stretch the loosely connected skin on the dorsum of the hand; the
majority of these receptors respond to movements of more than one joint and so the information
that they provide the CNS is ambiguous with respect to movement direction and amplitude.
However, the ensemble response from a population of SA II receptors can probably provide a
population vector that differentiates individual finger movements.
In contrast to the tactile sensory system in which higher densities of mechanoreceptors are
associated with superior tactile acuity, higher spindle densities do not appear to be associated
with superior sensory acuity. For the muscles of the hand, the number of spindles has been
estimated to vary from 12 to 356. There is no evidence indicating that the detection of
movements or changes in limb position is superior as one goes from proximal to distal joints, and
when expressed in terms of the absolute angular rotation of the joint, the ability to detect
movements is better at more proximal joints such as the elbow (0.08 at 20/s) than the distal
joints of the hand (0.88 at 20/s). The superior performance of more proximal joints is not
surprising, as they move more slowly than distal joints and rotation of these joints results in a
larger displacement of the end-point of the limb than the same angular rotation of a distal joint.
The forces generated by muscles are sensed by Golgi tendon organs normally found at the
junction between muscle tendon and a small group of muscle fibers. These receptors are less
numerous than spindle receptors and some muscles of the hand do not appear to have any tendon
organ receptors. Results from a number of experiments suggest that the perception of force is
derived from internal neural correlates of the descending motor command, and that peripheral
feedback calibrates these signals to indicate whether the motor command is adequate for the task.
Collectively, peripheral receptors together with central feedback systems provide the information
required for the dexterous performance of the hand.
 Transmission of musculotendon forces to the index finger
Sang-Wook Lee, Hua Chen and Derek G. Kamper, Member, IEEE

AbstractThis article reviews work completed by the orientations of the bands in the net model are assumed
authors to examine how musculotendon forces from index fixed at the outset and are not re-configured during finger
finger muscles impact finger kinetics and kinematics. motion. Contributions of specific muscles to the different
Results were obtained from in vivo studies, computer bands are estimated and assumed to remain constant. In
simulations, and cadaver investigations. Passive joint actuality, however, the bands of the extensor hood move
impedance and finger posture had particularly strong with respect to the phalanges [4] and the transmission of
effects on force transmission in the fingers.
force throughout the hood is not known.
Thus, examination of how musculotendon forces are
I. INTRODUCTION
transmitted to index finger segments and joints is
F or the hand, the mapping of musculotendon forces to
joint torques and forces is quite complex. Typically, 21
warranted. The results of a series of such studies
conducted in our laboratory are presented.
degrees-of-freedom (DOF) are ascribed to the fingers and
thumb. If one distinguishes among the different II. METHODS AND PROCEDURES
compartments, controlling different digits, of some of the
In our laboratory, we have used a combination of in
extrinsic muscles, over 30 different muscles control the
vivo studies, computer modeling, and cadaver
hand.
experiments to examine the transmission of tendon forces
The index finger alone is comprised of three joints:
into finger kinetics and kinematics.
metacarpophalangeal (MCP), proximal interphalangeal
The in vivo studies involve either percutaneous
(PIP), and distal interphalangeal (DIP) with a total of four
stimulation or fine wire electroymyography of the
DOF. Motion about these joints is controlled by 7
muscles of the index finger. Two 55-m stainless steel
different muscles: the extrinsic finger flexors {flexor
intramuscular electrodes are inserted into each muscle
digitorum superficialis (FDS) and flexor digitorum
with a 27-gauge hypodermic needle. Kinematics are
profundus (FDP)}, the extrinsic finger extensors
recorded with a video-based infrared motion capture
{extensor digitorum communis (EDC) and extensor
system (OptoTrak 3020, Northern Digital, Inc.,
indicis (EI)}, and the three intrinsic hand muscles {first
Waterloo, Ontario, Canada). Kinetics are recorded with a
dorsal interosseous (FDI), first lumbrical (LUM), and the
6 DOF load cell (20E12A, JR3, Inc., Woodland, CA).
first palmar interosseous (FPI)}. All of these muscles are
Computer models simulating finger dynamics have
multi-articular, with some of them crossing all three
been developed using Working Model (Design
joints.
Simulation Technologies, Canton, MI), visualNastran
On the palmar side of the index finger, both FDS and
(MSC Software, Santa Ana, CA), and Simulink
FDP travel through a series of anatomical pulleys which
(MathWorks, Natick, MA). These models include finger
shape the tendon path. These structures impact force
structures such as anatomical pulleys.
transmission as can be observed in cases in which they are
Cadaver experiments are performed with
damaged. On the dorsal side, the situation is even more
fresh-frozen hand specimens, fixed with the Agee
complicated as the five other muscles, EI, EDC, FPI,
WristJack (Hand Biomechanics Lab, Sacramento, CA).
LUM, and, arguably FDI ([1] but see [2] ), interact with
Known loads can be applied to nylon threads attached to
the extensor hood, a dorsal aponeurosis conveying
the tendons of the index finger muscles (see Fig. 1).
muscle force to the finger bones and joints. The extensor
Strain in structures such as the extensor hood can be
hood, in turn, has multiple insertion sites into the finger
measured using a camera-based system and ultraviolet
phalanges. In the past, researchers have used a force net,
light.
Winslows tendinous rhombus, to describe the
III. RESULTS AND DISCUSSION
distribution of forces in the extensor hood [3-4]. The
Both computer simulations and in vivo testing have
confirmed that activation of the extrinsic flexors, FDS
This work was supported in part by grants from the Coleman and FDP, does lead to concurrent flexion of all 3 finger
Foundation, the Whitaker Foundation, and the National Institutes of joints. This occurs despite the existence of only a single
Health (NINDS R01 NS052369). attachment point for each muscle [5-6].
Sang-Wook Lee is with the Sensory Motor Performance Program Passive joint stiffness and damping are crucial to
of the Rehabilitation Institute of Chicago.
Hua Chen was with the Rehabiltiation Institute of Chicago. He is obtaining this concurrent flexion. When the passive joint
now a graduate student at Harvard University. impedance is removed in the computer models, flexion of
Derek G. Kamper is with the Biomedical Engineering Department, the MCP joint, for example, is greatly reduced or even
of the Illinois Institute of Chicago and the Sensory Motor Performance eradicated [5-6].
Program of the Rehabilitation Institute of Chicago (kamper@iit.edu).
 In contrast, forces in the central and terminal slips of
the extensor hood decreased with PIP and DIP flexion.
For a given level of EDC loading, force levels in these
two slips dropped by more than 50% as the
interphalangeal joints were flexed [11].

ACKNOWLEDGMENT
The authors thank Mr. Erik Cruz, Ms. Heidi Fischer,
and Dr. Joseph Towles for their invaluable help with
these studies.

Fig. 1. Cadaver forearm secured with Agee
WristJack. Fingertip is connected to 6 DOF load cell.
Threads are connected to muscle tendons and run back
through a metal plate [7].
The anatomical pulleys also impact finger
movement. Removal of pulley models from the
simulations altered simulation output, although to a lesser
extent than removal of joint impedance [5-6]. This effect
is in accordance with the description of the impact of
pulley releases [8].
The observed flexion pattern, however, does not
follow from use of the geometrically obtained moment
arms for the muscles at each joint [9]. That technique
leads to overestimation of MCP flexion for a given
amount of force input. Additionally, the effective
moment arms measured directly in static finger postures
show a very strong dependence on joint posture [10]. For
example, the EDC and EI moment arms about the MCP
joint doubled as the PIP and DIP joints became more
flexed (Fig. 2).
REFERENCES
[1] D. G. Kamper, H. C. Fischer, and E. G. Cruz, "Impact of
finger posture on mapping from muscle activation to joint
torque," Clin Biomech, vol. 21, pp. 361-9, 2006.
[2] C. Long, "Intrinsic-extrinsic muscle control of the fingers:
electromyographic studies," Journal Bone Joint Surg, vol.
50-A, pp. 973-984, 1968.
[3] F. J. Valero-Cuevas, F. E. Zajac, and C. G. Burgar, "Large
index-fingertip forces are produced by subject-independent
patterns of muscle excitation," J Biomech, vol. 31, pp.
693-703, 1998.
[4] M. Garcia-Elias, K. N. An, L. Berglund, R. L. Linscheid, W.
P. Cooney, 3rd, and E. Y. Chao, "Extensor mechanism of the
fingers. I. A quantitative geometric study," J Hand Surg
[Am], vol. 16, pp. 1130-6, 1991.
[5] D. G. Kamper, T. G. Hornby, and W. Z. Rymer, "Extrinsic
flexor muscles generate concurrent flexion of all three finger
joints," J Biomech, vol. 35, pp. 1581-1589, 2002.
[6] S. Lee and D.G. Kamper, Modeling of multi-articular
muscles: importance of inclusion of tendon-pulley
interactions in the finger, IEEE Trans BME, [in press).
[7] F. J. Valero-Cuevas, J.D. Towles, and V.R. Hentz,
Quantification of fingertip force reduction in the forefinger
following simulated paralysis of extensor and intrinsic
muscles, J Biomech, vol. 33, pp. 1601-1609.
[8] C. K. Low, B. P. Pereira, R. T. H. Ng, Y. P. Low, and H. P.
Wong, The effect of the extent of A1 pulley release on the
force required to flex the digits, J Hand Surg [Br], vol. 23B,
pp. 46-49, 1998.
[9] K. N. An, Y. Ueba, E. Y. Chao, W. P. Conney, and R. L.
Linsheid, Tendon excursion and moment arm of index finger
muscles, J Biomech, vol. 16, pp. 419-425, 1983.
[10] S. W. Lee, H. Chen, J. D. Towles, and D. G. Kamper,
"Estimation of the effective static moment arms of the
tendons in the index finger extensor mechanism," J Biomech,
vol. 41, pp. 1567-73, 2008.
[11] S. W. Lee, H. Chen, J. D. Towles, and D. G. Kamper, Effect
of finger posture on the tendon force distribution within the
finger extensor mechanism, J Biomech Eng, vol. 130, 2008.

Fig. 2. Effective static moment arms for EDC and EI

about the MCP extension DOF. Data averaged across 5
cadaveric hand specimens. Errors bars indicate one
standard deviation. IP postures (PIP, DIP): (0, 0), (30,
20), (45,30).
 Workshop of RSS 2009

Title
Force Sensation of the Human Finger when Using a Multi-Fingered Haptic Interface

Speaker
Haruhisa Kawasaki
Professor
Faculty of Engineering, Gifu University
E-mail: h_kawasa@gifu-u.ac.jp, Tel & Fax: +81-58-293-2546

Abstract
Haptic interfaces that present force and tactile feelings have been utilized in the areas of
telemanipulation, interaction with micro-nanoscale phenomena, medical training and
evaluation, and others. A multifingered haptic interface (MHI) has greater potential for
effectiveness in these applications than does a single-point haptic interface. In particular,
telemanipulation is necessary when using a humanoid hand robot. When a human manipulates
an object by using his or her hand, the human receives a tactile sensation through a contact
plane between the finger surface and the object surface as a result of the grasping force,
vibration and heat. However, when a human manipulates an object in a virtual environment or
telemanipulates a humanoid hand robot, he or she uses an MHI and receives a tactile
sensation through a finger holder, which connects the human fingertip and the haptic finger of
the MHI. This means that the tactile sensation is derived from all of the contact surfaces
between the human fingertip and finger holder. Therefore, for the MHI to be further
developed, it is necessary to clarify the human tactile sensation that occurs under these
conditions.

The characteristics of the force sensation of the human finger have been reported in several
papers. In most of these studies, the force sensation was evaluated upon performance of the
flexion/extension motion. Humans feel the force sensation during the adduct/abduct motion
and the moment sensation, which often occurs with the grasping of a corner of a long object.
These characteristics have not been previously evaluated.

Our group has developed a multi-fingered haptic interface called HIRO III[1], as shown in
Figure 1. HIRO III consists of an interface arm with 6 degrees of freedom (dof), a haptic hand
with five haptic fingers, and a controller. Each finger has three joints allowing 3 dof. The first
joint, relative to the hand base, allows abduction/adduction. The second and third joints allow
flexion/extension. A three-axis force sensor is mounted in a tip link of each haptic finger. To
manipulate the haptic interface, the operator has to wear a finger holder on his/her fingertips.
The finger holder contains an iron sphere, which is attached to the permanent magnet at the
force sensor tip to form a passive spherical joint. This passive spherical joint makes
adjustments between the human and haptic finger orientations and ensures safety if the MHI
malfunctions. When the operator moves his/her hand, the haptic interface follows the motion
of the operators fingertips and presents a sensation of force. We have evaluated the
characteristics of the following force sensations experimentally using the HIRO III:
1) Ability to recognize the direction and magnitude of a three-dimensional force;
2) Ability to recognize the direction and magnitude of the friction moment.
The experimental results showed that the ability to recognize the force varied according to the
displayed force direction. Moreover, the ability to recognize the frictional moment was not
affected by the roughness of the object surface or the angle between the fingertip and object
surface.

Finally, a small device that presents the frictional moment, which was developed based on the
experiment results for the force and moment sensations, and a virtual object handling system
which can present the normal force and the friction force, are presented.

Fig 1. Multi-fingered haptic interface HIRO III

Reference
1) T. Endo, H. Kawasaki, T. Mouri, Y. Doi, T. Yoshida, Y. Ishigure, H. Shimomura, M.
Matsumura, and K. KoketsuFive-Fingered Haptic Interface Robot: HIRO III, Proc. of
WorldHaptics 2009, pp.458-463, 2009
 Scientific Approaches to the Study of Human Hand Function

Dr. Susan J. Lederman

Queens University

Psychology
Center for Neuroscience
School of Computing

The human hand is a critical component of a highly complex system that includes, and is
controlled by, a very powerful nervous system. This highly flexible system is used to
perform an impressive range of manual functions.

Human hand performance critically depends upon the structure and function of the
human hand, the properties of the physical environment, the nature of the
hand/environment interactions, and finally, how sensorimotor inputs are processed and
represented, both functionally and neurally.

Within this conceptual framework, I will address selected issues relating to the sensory
side of human hand function that have been investigated by touch scientists. For each, I
will outline some of the more successful methodologies that have been employed and
illustrate the types of conclusions that may be drawn, using examples from the scientific
literature.

1) What is the sensitivity and spatiotemporal resolution capacity of the human hand?
This aspect of psychophysical investigations focuses on the detection and
discrimination of threshold-level events. Human hand function is constrained in part
by whether the somatosensory system can detect the occurrence of mechanical,
thermal and/or electrical events. It is further limited by the precision with which it can
discriminate spatial or temporal events. Selective adaptation and masking paradigms
have been used to behaviorally determine the relative contribution of different
sensory channels.

2) What is the relation between supra-threshold stimulus intensity and perceived

magnitude?
This issue involves an aspect of psychophysics that focuses on the nature of the
mathematical function that best describes the relation between supra-threshold
physical intensity and the perceived magnitude of the corresponding sensation or
percept. Psychophysical methods have been used to study human sensation (e.g.,
warmth, cold, pressure, etc.) and the perception of objects and their properties (e.g.,
roughness, compliance, weight, shape, size, orientation, etc.).

3) What is the nature and role of manual exploration in human haptic processing of
objects and their properties?
Video analysis has shown that manual exploration is highly systematic.
Behavioral experiments have revealed the costs and benefits of performing one (or

1
 more) patterns of manual exploration (exploratory procedures) for haptic object
perception. When manual exploration of unfamiliar objects is unconstrained, their
material properties are more perceptually salient than their geometric properties.
Simultaneous execution of two or more exploratory procedures allows perceivers to
integrate valuable redundant property information about the identity of multi-attribute
objects.

4) What are the contributions of spatial and temporal information to haptic object
processing?
It is possible to assess the contribution of different types of information by
manually constraining haptic exploration in space and/or during its time-course,
thereby eliminating certain sources of information. The resulting decrement in
performance signals the contribution of the missing information.

5) What are the psychological dimensions that underlie complex human tactile/haptic
percepts?
Complementary methods involving controlled psychophysical experiments and
multidimensional-scaling procedures have been used to determine the perceptual
spaces that underlie complex touch-related experiences (e.g., surface texture), and
associated physical dimensions.

6) When both vision and touch are used, how is information from both modalities
integrated?
Human hand performance is frequently affected by the simultaneous availability
of information to more than one sensory modality (e.g., touch, vision, audition). How
are multisensory inputs about a common physical event combined? A number of
scientific methodologies have been used to address this question, usually by
comparing data from unimodal to that of bimodal (or multimodal) conditions.

A number of these topics are also relevant to how cutaneous and kinesthetic cues
contribute to human manual kinesthesis and to dextrous grasping and manipulation, as
will presumably be noted in other presentations at this workshop.

2
Robust Biomimetic Tactile Sensing and Grip Control
Gerald E. Loeb, M.D., Professor of Biomedical Engineering, University of Southern California, Los
Angeles, CA 90089, gloeb@usc.edu, in collaboration with Nicholas Wettels, Jeremy Fishel, Gary Lin,
and Nora Nelson at USC and Roland Johansson, Umea University, Sweden

Rehabilitation therapists refer to the hand as the third eye because its function is highly dependent on
the ability of touch to create a minds eye image of the objects that it encounters even without direct
vision. That internal representation includes not just the objects shape but those mechanical properties
that are needed to develop useful strategies for its manipulation: mass, rotational inertia, texture,
hardness, friction, etc. Much of the data to develop that image comes from the tactile sensors of the
glabrous skin of the fingers. The actual signals result from the deformation of the skin and pulp of the
finger tips as they interact with objects according to forces applied via the tendons of the muscles that
actuate the fingers. The shape and viscoelastic properties of the fingertips are a critical determinant of the
transduction process as well as the mechanical interactions that occur during manipulation.

We have developed a novel biomimetic

technology for tactile sensing from
anthropomorphic finger tips. The TAC array
illustrated at right detects deformation of a
viscoelastic skin and pulp over a rigid core in
which all sensors, connections and circuitry are
protected from the hostile environment in which
hands are often used (Wettels et al., 2008). The
core is covered by a silicone elastomer skin that is
inflated with a weakly conductive fluid. Internal
circuits measure the AC impedance of an array of
electrodes mounted on the surfaces of the core. Deformations of the skin from contact with objects or
sliding over the core as a result of tangential forces produces a distributed pattern of impedance changes
similar to the distributed patterns of neural activity associated with biological skin receptors. Controlled
texturing of the inner surface of the skin confers a wide dynamic range of force sensing and impedance
changes (typically 0.1-30N, 10-1000k). By adding a pressure sensor to the fill tube, we can sense the
acoustic spectra of fluidic microvibrations associated with incipient slip and sliding over textured
surfaces. A thermistor on the surface of the core provides temperature compensation and heat flow
measurements to distinguish the thermal properties of contacted objects. The electrodes and signal
processing, digitizing and multiplexing circuitry are mounted on a flex circuit that is molded into the rigid
core. The only component subject to wear and damage is the inexpensive, molded skin, which can be
easily replaced and reinflated with fluid.

We have incorporated a primitive version of an impedance sensing TAC with 6 electrodes onto the thumb
of the Proto1 hand made by Otto Bock HealthCare (Vienna, Austria) for the DARPA Revolutionizing
Prosthetics 2009 program. Normal and tangential forces were extracted by a Kalman filter and used to
drive a simple algorithm for adjusting grip force in a manner similar to that employed by humans
(Johansson and Flanagan, 2007). The hand was able to grip a fragile Styrofoam coffee cup and to adjust
grip force incrementally as water was poured rapidly into it, thereby avoiding crush or slip failures.
The multimodal sensing provided by the TAC produces synergistic benefits. For example, the thermal
and vibration signals obtained during exploration of an object depend on the location, magnitude and
direction of the applied force from the fingertip, which can be determined from the impedance sensors.
Some exploratory movements (such as texture discrimination) actually seem to be driven by their ability
to excite selectively a particular class of afferents. Rather than imitating the exact procedures and
classifications described for humans, it may be more productive to imitate the process whereby any
intelligent organism comes to know and categorize its environment through its senses. We hypothesize
that machines with a sufficiently rich set of receptors and coordinated movements can use artificial
intelligence methods to develop their own exploratory procedures and object classification schemes.

Robots were originally conceived as machines that would work side-by-side with humans in our
environments, rather than the glorified NC-milling machines now used on industrial assembly lines or the
engaging but clumsy anthropomorphic robots demonstrated at trade-shows. Despite these obvious
limitations, however, manufacturing companies have made and profited from huge investments in
industrial robots, starting with the Unimate of the 1950s and progressing to the almost completely
automated assembly lines for automobile chassis (Kamm, 2005). Yet industrial robots are almost
completely absent from the rest of the labor intensive steps of manufacturing automobiles such as
installation of wiring harnesses, upholstery, carpets, etc. Industrial robots are almost invariably
segregated from human workers lest they unwittingly cause injuries. They generally cannot identify or
handle common human tools. They function best in highly structured environments. The opportunities
for robots in military and police missions are obvious but their inability to deal with even the simplest
common objects such as doors and latches is legendary. While lots of tactile sensing technologies have
been developed, none has provided the combination of robustness, dynamic range and multimodality that
is required to support humanlike dexterity in realistic environments. We believe that the TAC technology
meets those requirements but will now require substantial development of biomimetic artificial
intelligence to make use of the distributed set of nonlinear sensors that it provides.

REFERENCES

Johansson RJ and Flanagan JR. Tactile sensory control of object manipulation in human. Handbook of the
senses. Vol: Somatosensation. 2007

Kamm, L. Industrial Robots; Large, Cartesian, Electro-mechanical. International Journal of Advanced

Robotic Systems. 2005. http://www.ars-journal.com/ars/Free_Articles/Lawrence_Kamm.htm

Wettels, N., Popovic, D., Santos, V.J., Johansson, R.S. and Loeb, G.E. Biomimetic tactile sensor array.
Advanced Robotics. 22(8):829-849, 2008.

ACKNOWLEDGEMENT

Initial development of TAC technology was supported by a grant from the National Academy Keck
Futures Initiative. Authors Loeb, Wettels, Fishel and Lin are principals in SynTouch LLC, a start-up
company developing TAC technology with SBIR grant support from the US NIH and Department of
Agriculture. Various aspects of this technology are covered by pending patent applications.
 Synergistic Control of Hand Muscles Through Common Neural
Input

Marco Santello
Arizona State University

In the past two decades important features of the kinetics and kinematics of

human object grasping and manipulation have been characterized, providing significant

insight into how the Central Nervous System (CNS) controls the hand. In particular,

many studies have focused on characterizing coordination patterns (synergies) that

emerge in hand-object interactions and constrain the behavior of the hands multiple

degrees of freedom. Whereas the interpretation of functional significance of synergies

remains controversial, it is clear that they are the by-product of the interaction between

mechanical and neural mechanisms or constraints (for review see Schieber and

Santello, 2004). Linkages among the tendons of finger muscles as well as the multi-

tendoned and multi-articular organization of these muscles are examples of mechanical

constraints limiting the extent to which joints can be independently controlled. With

regard to neural constraints, the poor somatotopy of the cortical representation of hand

muscles has been identified as evidence for a synergistic, rather than individual, control

of hand muscles.

In the past few years we have studied common neural input (CNI) to motor units

of hand muscles to quantify the neural bases of digit force and movement synergies.

Near synchronous discharge of active motor units is an indirect measure of common

synaptic input across motoneurons. If common synaptic input is delivered to

motoneuron pools of different hand muscles or compartments of multi-tendoned muscles

(i.e., the long flexors of the fingers) it would enhance the coupling of forces exerted by

the digits. In this talk I will review main findings from the literature and my laboratory on

synergistic control of hand muscles assessed through analyses in the time and
frequency domains, as well as EMG amplitude correlations. The main focus of my talk

will be on recent evidence indicating that common neural input is distributed in a

heterogeneous and muscle-pair specific fashion, i.e., the neural coupling between

certain muscle pairs is stronger than that between other muscle pairs (Winges et al.,

2004, 2006; Johnston et al., 2005). Specifically, common neural input across motor

units of two antagonist intrinsic muscles of the index finger (First Dorsal Interosseus,

FDI; First Palmar Interosseus; FPI; Winges et al., 2008) is weaker than across intrinsic-

extrinsic muscle pairs (P < 0.001) as well as synergistic extrinsic flexors of the thumb

and index finger (Flexor Pollicis Longus, index finger compartment of Flexor Digitorum

Profundus; Winges et al., 2004, 2006). In contrast, the strength of common neural input

within each of these intrinsic muscles is ~ 3 times stronger than across them (Winges et

al., 2008). These findings have been confirmed by other studies indicating different

degrees of neural coupling across intrinsic vs. extrinsic muscle pairs (Johnston et al.,

submitted). This evidence suggests that the distribution of common neural input across

hand muscles may reflect the degree of independent vs. dependent neural control of

specific muscles and muscle pairs and possibly their long-term adaptation for their role

in manipulation. The extent to which such distribution is fixed or task-dependent will be

discussed through recent findings on the distribution of common neural input as a

function of digit force magnitude and direction.

 Within muscle Across muscles

CIS (extra sync. spikes s-1)

0.8

0.6

0.4

0.2

0
FDI-FDI FPI-FPI FPI-FDI FPL-FPI 2-digit 5-digit
FPL-FDI
FPL-FDP2
 Original approaches to interpretation learning, and
modelling, from the observation of human
manipulation
Francesco Corato , Pietro Falco, Martin Losh, Emilio Maggio , Jakel Rainer , and Luigi Villani
Dip. di Ingegneria dellInformazione, Seconda Universita di Napoli, Aversa, Italy
Email: pietro.falco@unina2.it
Email: fcorato@unina.it
FZI, Karlsruhe, Germany

Email: loesch@ira.uka.de
Email: jaekel@ira.uka.de
OMG plc, Oxford, UK

Email: Emilio.Maggio@vicon.com
Dip. di Informatica e Sistemistica, Universita di Napoli Federico II, Napoli, Italy

Email: lvillani@unina.it

I. I NTRODUCTION positions and contact force measurements will be developed.

The DEXMART project is focused on artificial systems
reproducing smart sensory-motor human skills, which oper-
ate in unstructured real-world environments. The emphasis
is on manipulation capabilities achieved by dexterous and
autonomous, and also human aware dual-arm/hand robotic
systems.
Manipulation cannot be fully planned without human in-
teraction and intervention. Even then, human demonstration,
advice and correction are important parts of robot manipula-
tion learning and execution. Research will target methods and
mechanisms to observe human manipulations in a way that
the observed actions enhance the robots skills (e.g. by new or
optimized skills) and tasks (e.g. due to the optimized, context
and goal depending combination of skills).
A number of challenges have to be tackled about the obser-
vation and interpretation of manipulation activities performed
by humans, first of all setting up a suitable sensory system
and reconstructing motion with high accuracy. Furthermore,
the question of what is the best kinematic model of the human
hand has to be answered.
In the following, we will limit the presentation to learning
goals which can be learned using only observations of the
human hand. In this case, two main questions remain:
1) What can be learned from the observation and how does
the actual learning take place?
2) What are relevant features for the learning and how can
they be selected?
In the DEXMART project, learning is pursued on two differ-
ent levels of abstraction, the so-called action level and activity
level. While the activity level is more abstract, the action level
covers the interfacing between subsymbolic representations
and atomic symbolic operations.
II. O BSERVATION OF THE H UMAN H AND
A. Motion Capture
A number of problems must be tackled to carry out highly-
accurate observation of the human hand:
marker movements due to skin motion with respect to the
bones
marker occlusions

Subsequently to the fundamental questions of observation positioning of the marker-set

and modelling of the human hand, the challenge is to build
systems which are able to learn from the observation of ma-
nipulations demonstrated by a human. Different robot learning
tasks in this context reach from simple recording and replaying
of observed trajectories to the abstraction and generalization
of the results of observed manipulations. Often the observation
of the hand(s) only is not sufficient, but also the context, e.g.
manipulated objects, have to be considered depending on the
complexity of the learning goal: more advanced sensor fusion
algorithms which integrate joint angular positions, marker
Marker movements due to skin motion are a real problem
for hand motion capture as the error can have the same order
of magnitude than the bone lengths. Existing models treat
the marker positional error as a residual covariance. This
assumption simplifies the mathematical formulation. However,
residual error analysis conducted within DEXMART showed
that it may be possible to improve the predictive power
of the kinematic model by explicitly accounting for marker
movements. To this extent, we developed a novel kinematic
calibration procedure that accounts for soft tissue artifacts

(a) (b)
Fig. 1. MRI measurements for the marker RMM4 on open (left column) and
closed (right column) hand. (a)-(b): marker spatial position (yellow arrow)
by allowing the markers to move according to polynomial
functions of the joints angles. To validate the motion marker
model and to increase the knowledge about marker slipping,
an MRI system has been used to capture the human hand in
two different static poses, shown in Fig. 1. The MR images
have been used to reconstruct a three-dimensional model of
the hand bones according to a co-registration procedure whose
details can be found in [4]. The measurement results show that
the marker slipping over the bones is quite significant ranging
between 2.5 mm and 10 mm along the hand axial direction.
Missing marker measurements due to occlusions can heavily
affect the reliability of the joint angle measurements. This
problem was tackled from two directions: (i) investigation
about the optimal number of markers and their positioning;
(ii) study about the fusion of data-glove angular information
with the marker measurements. Regarding the marker-set, two
possible configurations were tested: a minimal marker set
with approximately one marker per segment and a redundant
marker configuration with three or more markers per segment.
To handle the occlusion problem using the minimal marker
set, in [1] a sensor fusion algorithm is presented. It fuses,
through a Kalman-like algorithm, positions in the space of
a set of reflective markers and measurements from low-cost
angular sensors disposed on the finger joints. The algorithm
allows real-time tracking of complex hand movements.
B. Kinematic model of the human hand
Another key topic is the question of what is the best
kinematic model for hand data. There is no generally accepted
methodology for making such a choice. Ideally an objec-
tive comparison between different kinematic models would
require the actual joint angle values measured with a pro-
cedure that guarantees higher accuracy than motion capture
measurements. Unfortunately, acquiring accurate ground truth
data almost always involves invasive procedures. Many have
used intra-cortical (bone) pin mounted markers or capitalized
on external bone fixation already in place [2]. As direct
measurements are impractical, researchers have evaluate other
desirable model qualities such as repeatability or have used
synthetic or semi synthetic data [3]. Although repeatability is a
well established technique for model validation, it does not tell
us how well the model explains the data. To this extent, OMG
studied statistical model selection for hand kinematics and
tested four different algorithms, namely: Bayesian Information
Criterion (BIC), Akaike Information Criterion (AIC), and Con-
sistent AIC (CAIC) and a procedure based on Bootstrapping
the data. Among these methods the results show that BIC,
CAIC and Bootstrapping can provide a good insight on the
number of DoF to assign to each articulation of the hand.
After the selection of the number of DOFs, a quantitative
analysis of joint interdependencies has been done aimed at
studying kinematic synergies during reaching and hand grasp-
ing activity. Four male subjects have been asked to perform
two tasks: a cylinder-grasp with right hand and a voluntary
flexion and extension of each individual finger. From acquired
data, the computation of the correlation coefficient matrix for
all the DOFs has been carried out, aimed at quantifying the
degree of correlation of each DOF with all the others. The
results obtained in this study will be used both for reducing the
complexity of trajectory planning and for improving the sensor
fusion algorithms of kinetostatic data owing to the increased
a priori knowledge about the hand kinematics.
III. L EARNING FROM OBSERVATION
To find the most relevant features for the learning, an
approach in two steps is investigated. In the first step, the
available sensors are exploited to extract and derive as much
features as possible. These features include, but are not
limited to joint angles, trajectories of all joints and finger
tips, velocities of them, statistical features of these points.
To accommodate for two-handed manipulations, features like
correlations of movements of both hands, temporal and spatial
synchronization points are also investigated. Using training
data from fundamentally different actions, ranging from simple
to complex, all features are analyzed and ranked according to
their utility for separating between different action classes. In
the second stage, from the complete set of features only the
most relevant for a current learning task are selected using
an information content-based method for the actual training
process. This minimizes the noise in the training data. A
complementary method for data reduction is to abstract from
sets of fingers, which serve the same purpose in a grasp or
manipulation action, to a single finger which combines force
and contact information of the whole set.
ACKNOWLEDGMENT
The research leading to these results has received funding
from the European Communitys Seventh Framework Pro-
gramme (FP7/2007-2013) under grant agreement no. 216239
(DEXMART project).
R EFERENCES
[1] P. Falco. Sensor Fusion based on the EKF for human and robotic hand
tracking application. Master Degree Thesis, 2008
[2] J. Fuller, L. Liu, M.C. Murphy, and R.W. Mann. A comparison of lower-
extremity skeletal kinematics measured using skin- and pin-mounted
markers. In 3-D Analysis of Human Movement, volume 16, pages 219
242, 1997.
[3] I.W. Charlton, P. Smyth, and L. Roren. Repeatability of an optimized
lower body model. Gait and Posture, 20:213221, 2004.
[4] The DEXMART consortium, Kinematic Modelling of the Human Hand.
Dexmart Deliverable D1.1, 2008.
 Development of artificial grip reflexes that utilize
the adduction/abduction capabilities of human fingers

Veronica J. Santos, Ph.D.

Dept. of Mechanical and Aerospace Engineering, Arizona State University, Tempe, AZ, USA

Man and machine can work together on levels ranging from the earthbound remote control of a
satellite-repair robot to the intimate connection between an amputee and his prosthetic hand.
One common feature for any man-machine system, independent of machine proximity, is the
existence of delays in information flow between man and machine. In the context of a
neuroprostheses, delays can exist because of afferent and efferent signal limitations.
Techniques such as targeted muscle reinnervation (Kuiken, et al., J Amer Med Assoc, 2009) hold the
promise of bringing a conscious perception of tactile feedback to the user and increasing the
number of channels with which a user can control a complex hand and arm. One open question
is whether users will be able to respond quickly enough through man-machine interfaces in the
face of unexpected perturbations.

The extensive psychophysical experiments of Johansson and colleagues on human grip

responses (Johansson and Flanagan, The Senses: A Comprehensive Reference, Somatosensation, 2007) has
fascinated roboticists, many of whom strive to design fingertip control algorithms modeled after
spinally-mediated reflexes in unimpaired humans. Such built-in, survival behaviors could buy
time for higher-level intervention and decision-making by the human operator in inherently
delayed man-machine systems. The standard mantra of robotic grasp research has been if it
slips, then grip harder. However, as artificial finger kinematics become more anthropomorphic
(Adee, IEEE Spectrum, 2009), it becomes increasingly important to specify in which direction to grip
harder. That is, how much adduction/abduction should be used in conjunction with
flexion/extension?

Currently, we are characterizing reflexive grip responses to rotational perturbations of objects in

precision grasp because such perturbations elicit adduction/abduction corrective responses in
addition to flexion/extension. Rotation of the test object is corrected as early as 65-80ms from
the perturbation, although it is unclear what portion of this early correction phase may be due to
the passive, compliant structure of the hand and what portion may be due to spinally-mediated
reflex responses. In addition, preliminary results suggest a role reversal of the thumb and index
finger depending on the direction of the rotation perturbation.
In parallel, we are preparing to control
robotic fingertips (Barrett Technology,
BarrettHand) using real-time postural
and tactile sensor data (SynTouch
LLC, TAC sensor). We have
designed an adapter (Figure 1) to
replace the distal phalanges of the
BarrettHand with biomimetic TAC Figure 1. The fingertips of an 11DOF robot (left) will be
outfitted with artificial tactile sensors (right).
sensors designed to detect both slow
and fast tactile stimuli (Wettels, et al., Adv Robotics, 2008). This system will be used to replicate the
rotational perturbation experiments performed with unimpaired humans and to test artificial
reflex control strategies for expected (e.g., impact during tool use) and unexpected
perturbations.
 The neuromuscular systems does ordinary manipulation tasks the hard way:
Lessons for robotic manipulators?
Francisco J Valero-Cuevas1,2,6, Madhushudhan Venkadesan5,6,
Kevin G Keenan2,3,6, Veronica J Santos4,6
1Department of Biomedical Engineering, 2Division of Biokinesiology & Physical Therapy, University of Southern California, CA USA
3Department of Human Movement Science, University of Wisconsin-Milwaukee, WI USA
4Department of Mechanical and Aerospace Engineering, Arizona State University, AZ USA
5School of Engineering and Applied Sciences, Harvard University, MA USA
6Sibley School of Mechanical and Aerospace Engineering, Cornell University, NY USA

I. Introduction
Dexterous function arises from the ability of the nervous system to orchestrate numerous muscles and joints to
meet mechanical demands. We find that the neural controller is severely taxed even for ordinary tasks like
making contact with a surface rubbing a surface. These results have implications to (i) longstanding questions
about neuroanatomy and notions of muscle redundancy, and begin to explain the vulnerability of dexterous
function to development, aging and neuromuscular pathology; and (ii) the design of versatile robotic
manipulators.

II. Transitions from motion to force

The neural control of tasks such as rapid acquisition of precision pinch remains unknown. Therefore, we
investigated the neural control of finger musculature when the index fingertip abruptly transitions from motion
to static force production (Venkadesan & Valero-Cuevas '08). Nine subjects produced a downward tapping motion
followed by vertical fingertip force against a rigid surface. We simultaneously recorded 3D fingertip force, plus
the complete muscle coordination pattern using intramuscular
electromyograms from all seven index finger muscles. We found that a
Motion Force
the muscle coordination pattern clearly switched from that for **
70
motion to that for isometric force ~65ms before contact (p=0.0004).

Contact
Mathematical modeling and analysis revealed that the underlying
Angular deviation from

Transition
referernce vector ()

neural control also switched between mutually incompatible

strategies in a time-critical manner. Importantly, this abrupt switch in
underlying neural control polluted fingertip force vector direction
beyond what is explained by muscle activation-contraction dynamics
and neuromuscular noise (p0.003). We further ruled out an 0
["

["

["

["

[+

[+
20

20
40

21

15

90

impedance control strategy in a separate test showing no systematic

,+

0,
0,

0,

0,

,"

+2
70
"

"

"

40

50
35

16

10

]
change in initial force magnitude for catch trials where the tapping
0]

0]

0]

Time (ms)

b
surface was surreptitiously lowered and raised (p=0.93). We
Contact

12
conclude that the nervous system predictively switches between
Normalized vector magnitude of

10
muscle coordination pattern

mutually incompatible neural control strategies to bridge the abrupt 8

transition in mechanical constraints between motion and static force. 6
Moreover because the nervous system cannot switch between control 4

strategies instantaneously or exactly, there arise physical limits to the 2

accuracy of force production upon contact. The need for such a 0

neurally demanding and time-critical strategy for routine motion-to- -400 -100 0
Time (ms)
200

force transitions with the fingertip may explain the existence of Figure 1. Switch in direction (a) and
specialized neural circuits for the human hand. magnitude (b) of the muscle
coordination pattern vector between
motion and static force production.
III. Combination of motion and force
Numerous studies of limbs and fingers propose that force-velocity
properties of muscle limit maximal voluntary force production during anisometric tasks, i.e. when muscles are
shortening or lengthening. Although this proposition appears logical,
our study on the simultaneous production of fingertip motion and A Position 1 Position 2 Position 3
force disagrees with this commonly held notion (Keenan et al. In Press).
We asked eight consenting adults to use their dominant index
fingertip to maximize voluntary downward force against a horizontal 35
surface at specific postures (static trials), and also during an Maximal
30
anisometric rubbing task of rhythmically moving the fingertip static force

Absolute force (N)

along a 5.80.5 cm target line . The metronome-timed flexion- 25

extension movement speed varied 36-fold from slow (1.00.5 cm/ 20

s) to fast (35.97.8 cm/s). As expected, maximal downward 15
voluntary force diminished (44.815.6%; p=0.001) when any motion Maximal force
10
(slow or fast) was added to the task. Surprisingly, however, a 36-fold during finger
movement
increase in speed did not affect this reduction in force magnitude. 5

These remarkable results for such an ordinary task challenge the 0

dominant role often attributed to force-velocity properties of muscle B Maximal
100 static force
and provide insight into neuromechanical interactions. We propose

Normalized force (% maximum)

an explanation that the simultaneous enforcement of mechanical 80 Maximal force
during finger
constraints for motion and force reduces the set of feasible motor movement
commands sufficiently so that force-velocity properties cease to be 60 Flexion
phase
the force-limiting factor. While additional work is necessary to reveal
the governing mechanisms, the dramatic influence that the 40 Extension
phase

simultaneous enforcement of motion and force constraints has on

20
force output begins to explain the vulnerability of dexterous function
to development, aging and even mild neuromuscular pathology. 0

Static

Moderate
Fast

Static
Slow
Moderate
Fast

Static
Slow
Moderate
Fast
IV. Conclusions
1. The nervous system responded to the evolutionary pressures (i.e.,
Movement Condition
mechanical constraints) of transitioning from motion to well- Figure 2. Maximal downward force
directed force production while maintaining finger posture by diminished when motion was added to the
developing a time-critical strategy of switching control laws. This task; but was not affected by movement
suggests roboticists should consider this option carefully before speed or position along the target line.
Absolute (A) and normalized (B) maximal
favoring passive solutions that rely on peripheral endpoint downward forces are shown for static
impedance to make this transition. (ci rcl e s) a n d d yn a mi c a n i so me tri c
2. While the biomechanical system of the fingers is over-actuated for (squares) trials.
some tasks, we find that it is likely not redundant for ordinary
tasks. Creating over-actuated systems may be a key to creating
versatile manipulators.

V. References
Keenan KG, Santos VJ, Venkadesan M, and Valero-Cuevas FJ. Maximal voluntary fingertip force production is not limited by
movement speed in combined motion and force tasks. J Neuroscience, In Press.
Venkadesan M and Valero-Cuevas FJ. Neural control of motion-to-force transitions with the fingertip. J Neuroscience 28: 1366-1373,
2008.

VI. Acknowledgements
We thank R. V. McNamara III for technical support and experiment execution, F. A. Medina for EMG processing, S. Song for
machining and instrumentation, Drs. M. Price and S. Backus for fine-wire electrodes placement. This material is based upon work
supported by the Whitaker Foundation, NSF Grants 0312271 & 0237258, and NIH Grants HD048566 AR050520 and AR052345 to
FVC. Its contents are solely the responsibility of the authors and do not necessarily represent the official views of the National
Institute of Arthritis and Musculoskeletal and Skin Diseases (NIAMS), the National Institute of Childhood and Human Development
(NICHD), the NIH, the Whitaker Foundation or the NSF.
 Human hand kinematics based on MRI imaging
Georg Stillfried and Patrick van der Smagt 1
Institute of Robotics and Mechatronics
German Aerospace Center (DLR Oberpfaffenhofen)

Anthropomorphic robot hands have come to a technical level where understanding exact human hand kinematics
becomes relevant, e.g. the hand/arm system that is presently being developed at DLR (Fig. 2, Grebenstein and van der
Smagt, 2008). Human hand kinematics have been investigated through cadaver hands (e.g. Hollister et al., 1992 and
1995) and optical motion tracking of surface markers (e.g. Cerveri et al., 2005). A problem with the former is that
tissue properties might be altered due to tissue necrosis. With the latter, the motion of the skin relative to the bones
leads to so called soft tissue artifacts (STA) that negatively influence the quality of the results (Ryu et al., 2006). To
allow in vivo measurements and to avoid STA, we recorded finger postures by magnetic resonance imaging (MRI,
Table 1 and Fig. 2). We used a method similar to the one described by Miyata et al. (2005), but with a much larger
number of hand postures, resulting in a model with multi-degree-of-freedom (multi-DoF) joints.
MRI images of the hand were taken in fifty different static postures (Fig. 3). The postures were chosen so that for
each joint, the extreme poses as well as intermediate ones are included; furthermore, the opposition movement
between thumb and fingers is covered extensively. One hand posture is defined as reference. The pose of each bone in
the other postures is described by a rotation and translation from the reference posture (Fig. 4). The first step for
determining the pose is the segmentation of each MRI image, i.e. identification of the point set of each bone. Next, a
statistical method by Hillenbrand (2008) is used to find the rotation and
translation parameters that are necessary to match the point sets of the
same bone. From this, the relative poses with respect to the neighbor
bone are calculated (Fig. 5).
Seven joint models with varying degrees of freedom and intersecting
and non-intersecting axes of rotation are defined to be valid (Fig. 6): A
joint with one rotation axis (1DoF), a joint with one DoF but two
coupled rotation axes (1DoF_2c), a joint with two rotation axes (2DoF),
a joint with two rotation axes that are orthogonal to each other
(2DoF_o), a joint with two non-intersecting rotation axes (2DoF_ni) and
two joints with three mutually orthogonal axes that are oriented with the
bone geometry (3DoF and 3DoF_ni).
For each joint, the parameters of the joint models are adapted
numerically to fit the measured bone poses. This is done by numerically
minimizing the orientational and the translational discrepancy between
the modeled and the measured bone poses, using the fminsearch function
within Matlab. The orientational discrepancy is defined as the rotation
angle of a rotation that is necessary to match the orientation of the
modeled and the measured bone pose (Fig. 7). The translational
discrepancy is defined as the mean distance of the bone surface point in
the modeled and the measured pose (Fig. 8). Fig. 9 DLR's kinematic hand model with
mean errors smaller than 6 and
By setting a limit for the discrepancy values, each joint was assigned one
3 mm. With 2-DoF joints, the first
of the seven joint models. For example, a limit of 6 degrees mean
joint axis is drawn as a red arrow and
rotational discrepancy and 3 mm mean displacement leads to a the second one as a green arrow.
kinematic hand model with 21 DoF in the fingers plus 3 DoF in the
palm, shown in Fig. 9. In this model, the following joint models are used: A two-DoF joint with non-intersecting axes
(2DoF_ni) for the thumb saddle joint, two-DoF joints with orthogonal, intersecting axes (2DoF_o) for the
metacarpophalangeal joints and one-DoF joints (1DoF) for the interphalangeal and the intermetacarpal joints.
Virtual grasping experiments will be conducted with this hand model to see if it has advantages over the simplified
kinematics that were used so far for the design of the robotic hand/arm system.

1 Author e-mail addresses: {georg.stillfried, smagt}@dlr.de

MRI scanner: Philips Achieva 1.5 T
MRI coil: Philips Sense 8-channel
sequence: balanced FFE
resolution: (0.67 mm) physically
(0.38 mm) interpolated
time per shot: approx. 4 min.
Table 1 Parameters of the MRI Fig. 1 Design prototype of DLR's Fig. 2 An MRI scanner unit and a hand
measurements new robotic hand/arm system. inside an MRI sensor coil.

Fig. 3 Hand postures. Fig. 4 Definition of a bone pose by

a rotation and a translation.

1DoF 1DoF_2ca 2DoF 2DoF_o 2DoF_ni 3DoF 3DoF_ni

Fig. 5 Relative poses of the a Fig. 6 Joint types.
joint in two hand postures.

Fig. 7 Orientational discrepancy Fig. 8 Translational discrepancy is defined as the mean distance between
between two poses. surface points. Here five example points are shown.

Cerveri, P.; Lopomo, N.; Pedotti, A. & Ferrigno, G. (2005), 'Derivation of Centers of Rotation for Wrist and Fingers in a Hand Kinematic Model: Methods
and Reliability Results', Annals of Biomedical Engineering 33, 402-412.
Grebenstein, M. & van der Smagt, P. (2008), 'Antagonism for a Highly Anthropomorphic Hand-Arm System', Advanced Robotics 22(1), 39-55.
Hillenbrand, U. (2008), Pose Clustering From Stereo Data, in 'Proceedings VISAPP International Workshop on Robotic Perception', pp. 23-32.
Hollister, A.; Buford, W. L.; Myers, L. M.; Giurintano, D. J. & Novick, A. (1992), 'The Axes of Rotation of the Thumb Carpometacarpal Joint', Journal of
Orthopaedic Research 10, 454-460.
Hollister, A.; Giurintano, D. J.; Buford, W. L.; Myers, L. M. & Novick, A. (1995), 'The Axes of Rotation of the Thumb Interphalangeal and
Metacarpophalangeal Joints', Clinical Orthopaedics and Related Research 320, 188-193.
Miyata, N.; Kouchi, M.; Mochimaru, M. & Kurihaya, T. (2005), Finger Joint Kinematics from MR Images, in 'IEEE/RSJ International Conference on
Intelligent Robots and Systems'.
Ryu, J. H.; Miyata, N.; Kouchi, M.; Mochimaru, M. & Lee, K. H. (2006), 'Analysis of skin movement with respect to flexional bone motion using MR
images of a hand', Journal of Biomechanics 39, 844-852.
 Dynamic Hand Simulation with Strands
Shinjiro Sueda
Sensorimotor Systems Laboratory, University of British Columbia
1 Introduction
We have developed a framework for simulating muscles and ten-
dons using a dynamic modeling primitive, a strand, based on cu-
bic spline curves with inertia [Sueda et al. 2008]. With this frame-
work, we do not require moment-arm data; the tendons are simu-
lated dynamically as stiff hyper-elastic strands simultaneously with
the bones. We are able to simulate the dynamics of the intricate
mechanisms of the finger, which was not possible with previous ap-
proaches. The lumbrical, for example, with its looped connections
to the flexor tendon and the extensor mechanism, can be simulated
effectively using our framework.
The extensor mechanism of the finger is a vital component for the
proper function of the hand [Zancolli 1979; Valero-Cuevas et al.
2007; Wilkinson et al. 2003]. Because of its complexity, how-
ever, finger pathologies and their treatments are extremely difficult
to simulate properly. With our strand-based framework, complex
phenomena such as linked flexion of the interphalangeal joints, or
the dynamics of the extensor hood, Landsmeers oblique band, and
the lumbrical can be simulated at interactive rates.
Figure 1: Flexion (left), extension (top), Swan-neck (middle) and
Boutonniere (bottom) deformities. Included in the simulation are
the extensor mechanism and Landsmeers oblique band. The three
extrinsic tendons (EDC, FDP, FDS) and the two interossei (DI, PI)
are pulled using an external load, and the lumbrical is modeled as
a contractile strand, originating from the FDP and inserting into
the extensor mechanism.
2 Our Approach
Previous approaches based on lines-of-force (e.g. [Delp and Loan
2000]), while effective and robust for large scale muscles and
skeleton, are not well-suited for complex mechanical systems like
the hand, where multiple contact constraints make it difficult to
route muscles and tendons effectively. Our strand-based simulation
framework combines the robustness and simplicity of lines-of-force
models with the accuracy of solid mechanics models. Our strand-
based simulation framework has the following desirable attributes:
e-mail:{sueda, pai}@cs.ubc.ca
1
Dinesh K. Pai
Conceptually simpler tendon routing mechanism. In previ-
ous approaches, wrapping surfaces based on primitives such
as spheres and cylinders are often used to compute the curved
paths of tendons and muscles. These wrapping surfaces, how-
ever, only affect the kinematics, rather than the dynamics, of
muscles and tendons. Our system is capable of handling the
coupled dynamics of muscles, tendons, and bones through
contact forces on arbitrary bone surfaces, which we believe
are critical for accurately modeling the mechanics of the hand.
Furthermore, branching and merging tendons can be easily
modeled using simple attachment constraints. This is impor-
tant for accurate simulation of the lumbrical, an important
component of the extensor mechanism of the finger.
Full dynamics of muscles and tendons. Unlike some systems,
which simulate muscles, tendons, and bones separately, our
system simulates them simultaneously, resulting in a more ac-
curate exchange of forces between them. For example, the
normal force exerted on a bone by a tendon as it wraps around
the bone can be accurately modeled by our system.
Efficient to simulate. Since we carefully exploit sparsity in our
numerical solutions, our method takes seconds (or minutes,
depending on the scene size), rather than hours to simulate,
unlike solid mechanics models.
3 Results
Using the strand-based framework, we are able to simulate various
clinical tests, such as the retinacular-plus test. We are currently
working on the simulation of various pathological deformities, such
as the swan-neck deformity or the Boutonniere deformity, and their
treatments (See Fig 1).
Acknowledgements. This work was supported in part by Canada
Research Chairs Program, Peter Wall Institute for Advanced Stud-
ies, NSERC, Canada Foundation for Innovation, BC KDF, MI-
TACS, and Maya licences donated by Autodesk.
References
D ELP, S. L., AND L OAN , J. P. 2000. A computational frame-
work for simulating and analyzing human and animal movement.
Computing in Science & Engineering 2, 5, 4655.
S UEDA , S., K AUFMAN , A., AND PAI , D. K. 2008. Musculo-
tendon simulation for hand animation. In ACM Trans. Graph.
(Proceedings of SIGGRAPH), vol. 27, 83:183:8.
VALERO -C UEVAS , F., Y I , J.-W., B ROWN , D., M C NAMARA , R.,
PAUL , C., AND L IPSON , H. 2007. The tendon network of
the fingers performs anatomical computation at a macroscopic
scale. IEEE Transactions on Biomedical Engineering 54, 6,
11611166.
W ILKINSON , D. D., W EGHE , M. V., AND M ATSUOKA , Y. 2003.
An extensor mechanism for an anatomical robotic hand. In Pro-
ceedings of the 2003 IEEE International Conference on Robotics
& Automation.
Z ANCOLLI , E. 1979. Structural and Dynamic Bases of HAND
SURGERY, 2nd ed. J. B. Lippincott Company.
 Movement Intermittency and Variability
in Human Arm Movements
Ozkan Celik, Qin Gu, Zhigang Deng, Marcia K. OMalley
Human reaching and tracking movements exhibit a multi-
peaked speed profile which is commonly interpreted as evi-
dence for submovements. As the movement gets slower, the
individual peaks stand out more distinctly, accompanied by an
increase in the number of the peaks. This phenomenon corre-
sponds to non-smoothness or intermittency in the movement.
In this study, we evaluate two potential sources proposed in
the literature for the origins of movement intermittency and
conclude that intermittency is more likely due to noise in
the neuromuscular system as opposed to a central movement
planner that generates intermittent plans. We also discuss our
results relevance to movement variability and noise models.
Doeringer and Hogan [1] proposed two possibilities as the
source of intermittency: (1) A central movement planner that
utilizes (simple) submovements to generate plans for complex
movements. (2) Noise getting imposed on the plan along the
neuromuscular circuitry. Although they did not arrive at a final
conclusion on the source of intermittency, they stated that
interpreting peaks in tangential speed profiles as incomplete
blending of submovements would lead to a conclusion favoring
the central planner option.
We hypothesize that movement intermittency evaluated at
various locations along the human arm during a constant speed
tracking task will increase in the distal direction. If verified,
this increase in distal direction indicates noise as the major
source of intermittency since the most natural explanation is
a noise mechanism; the intermittency would be amplified due
to additional noise being interjected (through joint rotations
connected in a serial fashion) along the arm in the distal
direction. In contrast, intermittency in the original movement
plan would be expected to produce similar intermittency levels
for all joints along the arm. We have designed an experiment
to test our hypothesis.
In the experiment, five participants completed a circular
tracking task by tracking a pointer on an LCD screen with their
fingertip. All participants provided informed consent approved
by the Institutional Review Boards of both institutions. Motion
capture markers were attached on the bony parts of the arm on
the shoulder, elbow, wrist joints and on the nail of the index
finger as illustrated in Fig. 1. 3D trajectories of the markers
were recorded using a Vicon motion capture system with ten
high resolution MX 40 cameras at a sampling rate of 50 Hz.
O. Celik and M. K. OMalley are with the Department of Mechanical
Engineering and Materials Science, Rice University, Houston, TX 77005 (e-
mails: celiko@rice.edu, omalleym@rice.edu)
Q. Gu and Z. Deng are with the Department of Computer Science,
University of Houston, Houston, TX 77004 (e-mails: ericgu@cs.uh.edu,
zdeng@cs.uh.edu )
Tracking pointer,
r=1 cm
Tracking path
circle, r=13 cm
Shoulder
marker
3. Intercept
2. Cue
Elbow
1. Hold marker
Wrist
marker
Fingertip
marker
4. Track
4. Tr
Fig. 1. An illustration of a participant in the experimental setting with
attached motion capture markers. Insert at bottom left: The trajectory that
the tracking pointer follows as a trial progresses through the four phases: (1)
pointer at the center during the hold phase, (2-4) pointer changing colors and
moving on the tracking path with constant speed during the cue, intercept and
track phases.
The experiment consisted of two sessions of 25 trials each.
Computer screen orientation (vertical/horizontal) was changed
between the sessions. The screen was in a normal position in
the vertical case and it was lying on the table facing up in
the horizontal case. The order of presentation of these cases
to each participant was determined randomly. The speed of
the pointer to be tracked changed randomly among trials, but
it was constant within a trial. Tracking speed had five levels:
2.5 cm/s, 3.75 cm/s, 5 cm/s, 12.5 cm/s and 25 cm/s. Each
speed level was experienced five times in each session. Each
trial consisted of four phases as schematically shown in the
insert on the bottom left on Fig. 1. Our visual interface design
was adapted from Pasalar et al. [2]. A total of 250 trials (12
of which were identified as incomplete trials and excluded
from analyses) constituted the data set corresponding to 5
participants 5 speed levels 5 trials for each speed level
2 orientation levels.
In addition to the original data set, we generated a decoupled
set where the positions of each of the four markers were
replaced by the difference in position of two consecutive
markers, proximal position subtracted from the distal one.
The purpose of the decoupling was to be able to create fair
comparison conditions for intermittency of different markers.
To better illustrate the coupling effect, note that movement
of the shoulder would be a base movement for the elbow,
hence effecting the kinematic variables in elbow data which
is recorded with respect to a fixed world coordinate system.
Both data sets were filtered with a zero phase-shift second
order low-pass Butterworth filter with a cut-off frequency of
5 Hz. Velocity in each axis was calculated using a backward
 70
Coupled Data Set
60 Decoupled Data Set
Movement Intermittency
50
40
30
20
10
0 possibility that central planning generates movement plans
Elbow Wrist Fingertip
Marker Location
Fig. 2. The mean and standard errors of movement intermittency for all
trials corresponding to the markers are plotted to illustrate the increasing
intermittency trend in distal direction. The trend is more pronounced in the
decoupled data set. Based on the results of the paired sample t-test, for all
pairs of markers within the same data set, the distal marker intermittency
is significantly higher than that of the proximal one (p < 0.05). Movement
intermittency is quantified as number of peaks in tangential speed profile. Note
that the reason for high variance presented in the plot is that because all trials
with all five different speed levels are lumped together for this plot, while the
t-test follows a paired sample procedure where each data point for a marker
is compared with the other marker under completely equivalent conditions.
difference method. As the measure of intermittency, we used
the number of peaks in the tangential speed profile, similar to
the measure used by Kahn et al. [3].
To compare the movement intermittency of different mark-
ers, we used a standard two-tailed paired-sample t-test, results
of which are given in Fig. 2. Within the decoupled data set, a
significant difference between means exists for elbow vs. wrist,
elbow vs. fingertip and wrist vs. fingertip [t(237) < 0, p <
0.001 in all cases]. Similar results are obtained using the
original data set. All six pairwise comparisons of movement
intermittency at different points on the arm resulted with
significantly higher intermittency for the more distal point.
Results of a multivariate ANOVA analysis (details not given
due to space constraints) for the fingertip marker using the
decoupled data indicate that intermittency decreases for higher
speed levels, and the main effect of speed on intermittency is
found to be statistically significant. Data for the main effect
of tracking plane orientation and for the interaction effect of
speed by orientation on intermittency failed to show signifi-
cance. These results imply that intermittency characteristics of
movements highly depend on the average movement speed but
are considerably less sensitive to the orientation of the actual
movement plane in the task space.
The fact that significant increases in movement intermit-
tency are observed for points located along the distal direc-
tion on the arm supports a neuromuscular noise alternative
versus a discrete central movement planner hypothesis as an
explanation for the movement intermittency.
One alternative explanation for the observed increase of
intermittency in distal direction could be the biomechanics
of the arm. Proximal parts (such as the upper arm) of the
arm has more inertia than the distal portions (hand or fingers).
Impedance properties are also different for different joints of
the arm. The arm acts as a low-pass filter for the torques
applied on the joints, and the proximal parts are expected to
have a lower cut-off frequency than the distal parts. This can
lead to the varying levels of intermittency along the arm, even
when the torques applied on the joints are identical in terms
of the intermittency content. However, we believe that since
the speed conditions in our experiment are concerned with
considerably slow movements, the range of speeds covered in
our study is below the cut-off frequencies of the arm joints,
hence an explanation based on arm dynamics would not be
valid.
It should be noted that our results do not rule out the
with different intermittency levels for different joints. However
we believe that plans with similar intermittency content for
different joints along the arm are more plausible than plans
that are tailored specifically for each individual joint.
In what follows we briefly discuss our results in relation to
recent findings on movement variability. There is a similar
debate on the origins of movement variability in the neu-
roscience field. Proposed sources are central planning noise
versus execution noise. van Beers et al. [4] demonstrated that
execution noise is the major factor causing movement vari-
ability by using a noise model based on three noise sources:
signal dependent noise (SDN), constant noise and temporal
noise. Our finding that noise increases in the distal direction
is an indication of execution noise domination and supports
this argument. In contrast, Churchland et al. [5] argued that
movement planning noise should also be considered as a
significant factor on overall variability, indicating that such
planning variabilities can involve controller customizations
for the upcoming movement. This idea was motivated by
the observation that changing movement directions in the
experiments in [4] led to change in kinematics and dynamics
of the arm. However, in our experiments, the kinematics
and dynamics do not change since the same circular path
is tracked in all trials, yet the movement intermittency is
significantly different for different speed levels. This further
supports findings reported by van Beers et al. [4].
R EFERENCES
[1] J. A. Doeringer and N. Hogan, Intermittency in preplanned elbow
movements persists in the absence of visual feedback, J Neurophysiol,
vol. 80, no. 4, pp. 17871799, 1998.
[2] S. Pasalar, A. V. Roitman, and T. J. Ebner, Effects of speeds and force
fields on submovements during circular manual tracking in humans,
Experimental Brain Research, vol. 163, no. 2, pp. 214225, 2005.
[3] L. E. Kahn, M. L. Zygman, W. Z. Rymer, and D. J. Reinkensmeyer,
Effect of robot-assisted and unassisted exercise on functional reaching
in chronic hemiparesis, in Proc. IEEE International Conference of the
Engineering in Medicine and Biology Society (EMBC2001), vol. 2, 2001.
[4] R. J. van Beers, P. Haggard, and D. M. Wolpert, The role of execution
noise in movement variability, Journal of Neurophysiology, vol. 91, no. 2,
pp. 10501063, 2004.
[5] M. M. Churchland, A. Afshar, and K. V. Shenoy, A central source of
movement variability, Neuron, vol. 52, no. 6, pp. 10851096, 2006.
 Effects of Force and Displacement Cues while
Adapting in a Rhythmic Motor Task
Ali Israr, Hakan Kapson, Volkan Patoglu and Marcia K. O'Malley
Abstract This paper explores the effects of magnitude and
phase cues on human motor adaptation. Participants were asked
to excite virtual second-order systems at their resonance fre-
quencies via a two-degree of freedom haptic interface, with vis-
ual and visual plus haptic feedback conditions. Their motor
adaptations were studied through catch trials. The results indi-
cate that, i) humans adapt to a nominal virtual system resonant
frequency, ii) humans shift to higher and lower natural frequen-
cies during catch trials regardless of feedback modality and
force cues, iii) humans can detect changes in natural frequency
when gain, magnitude, and phase cues are manipulated inde-
pendently, and iv) humans are able to detect changes in natural
frequency when the feedback (visual or visual plus haptic) is de-
layed such that the phase shift between the nominal system and
catch trial system is zero.
I. INTRODUCTION
Haptic guidance schemes are incorporated in virtual envi-
ronments to improve performance and to reduce training du-
ration and user workload. Virtual fixtures, record-and-play,
shared control, and error-based guidance schemes have
shown potential to improve user performance during task
completion and to accelerate learning rates, by guiding the
user to perform the task in a preferred manner (see [1] and the
references therein). Our prior works have focused on the ef-
fects of various forms of haptic assistance on both perfor-
mance enhancement and training for manual control tasks and
we have proposed shared control as the most general active
haptic guidance scheme [1,2]. Our results indicate that the
performance of manual control tasks is influenced by the us-
ers ability to identify the dynamic parameters of the manipu-
lated system [1]. Hence, it is suggested that a better under-
standing of human response and adaptation to varying system
dynamics can enable improved design of haptic guidance
schemes [2].
In the literature, it has been shown that humans can adapt
their feed-forward control commands over time [3,4]. This
adaptation can be viewed as successful training of a new
skill. Control parameter adaptation during object manipula-
tion was observed by Huang et al. [5] in a recent study of on-
line control during object manipulation. They investigated a
simple rhythmic object manipulation task in a virtual envi-
ronment and determined that participants could identify and
excite distinct virtual system natural frequencies with visual
only, haptic only, or combined visual and haptic feedback.
Mechatronics and Haptics Interface Laboratory, Rice University, 6100 Main
Street MS 321, Houston, TX 77005. Email: israr@rice.edu.
They also observed that participants appeared to tune control
parameters of a general feedback strategy.
We hypothesize that the ability to learn new motor skills
depends on the ability to form a control model, and to tune
control parameters. The objective of this work is to determine
which cues (displacements or forces) are primary for identifi-
cation of a controlled systems dynamic behavior. To this
end, we explore the ability of humans to resonate second-
order virtual mechanical systems with the same or distinct
natural frequencies. We investigate the effects of magnitude
and phase cues, and feedback modality, on the humans abili-
ty to maintain or adapt their control commands for such a
manually excited resonance task.
II. METHODS
Participants sat in front of the monitor screen that displayed
two rectangular masses and held a force feedback joystick
(IE2000, Immersion Corp.) with their dominant hand (Fig. 1).
The motion of one mass, m1, was directly coupled with the
joystick motion. The second mass, m2, was connected to m1
by a virtual spring (k) and damper (b), thus indirectly control-
ling the motion of m2 with the joystick motion. The dis-
placement transfer function (the mapping between the dis-
placement of m2 and the displacement of m1) and the imped-
ance TF (the mapping between the force applied by the mo-
tors and the displacement of m1) were derived and used for
haptic rendering. The visual cues were updated at a rate of 60
Hz on the computer screen while the haptic rate was set at a
typical 1000 Hz. Force and displacement data of m1 and m2
were collected at a rate of 50 Hz.
Fig. 1. A pictorial explanation of the dynamical task.
Six healthy male students of Rice University (18-32 years
old, avg. 22 years) participated in the study. They were asked
to smoothly oscillate m1 at the natural frequency of the virtual
second-order dynamic system by moving the joystick along
one of its axes in a sinusoidal manner throughout 10-second
long trials. They were told that if the excitation was at the
natural frequency of the virtual second-order system, the am-
plitude of oscillations of m2 would grow largest for constant
amplitude excitation of m1. At the end of the trial, a message
indicating if participants hand oscillation was greater than,
lower than or within 5 percent of the system natural frequen-
cy was displayed as performance feedback.
At the beginning of the experiment, participants were
overtrained with a nominal system of 1 Hz natural frequency.
Once they adapted to the dynamics of the nominal system
then in random catch trials the parameters of the dynamical
system were altered such that the resulting target systems had
either the same or different natural frequency with altered
magnitude and/or phase cues. The target systems were pre-
sented in three sets of test sessions, in which one out of ten
trials was a catch trial. In the first set, the natural frequency of
the system was same as that of the adapted nominal system
and the magnitude cue was changed by either changing the
damping ratio or the dc-gain. The phase cue was changed by
incorporating the delay between the states of the two masses.
In the second set, the natural frequency of the system was
changed. The catch trial systems had either both magnitude
and phase cues altered or one of the two cues altered. In the
third set, the impedance magnitude function of the target and
nominal systems were matched while keeping the same or
different natural frequency. All target systems were tested six
times in random order and with two type of sensory feedback,
i.e., visual-alone (V) and visual combined with haptic (V+H)
feedback.
The spectrogram (time-frequency trajectory) profiles of
hand displacement data were obtained and the frequency with
the largest intensity at each sample instant was extracted. A
plot of the frequency as a function of the duration of a trial
was termed as the hand motion frequency profile while con-
trolling the nominal or the target system. Frequency trajecto-
ry error is introduced as a measure of performance, defined
as the absolute difference between the hand frequency profile
and the natural frequency of the target system accumulated
across a 10 second trial, i.e., 500 samples. Statistical tests
such as analysis of variance and Students t-test were used to
determine differences among systems, and sensory feedback
at a 95% significance level (i.e., =0.05).
III. RESULTS
An exemplary time-frequency profile of the hand motion for
the target systems with different natural frequency compared
to the nominal system is shown in Fig. 2.
For target systems with the natural frequency same as that
of the nominal system, participants in general did not diverge
from the natural frequency when magnitude cue was altered,
despite matching the impedance or displacement transfer
function. Effects of feedback on frequency trajectory errors
data failed to show significance. When both the magnitude
and phase cues were changed for target systems of different
natural frequency, the hand trajectory quickly adapted to the
new dynamics and converged to the systems natural fre-
quency. A repeated measure ANOVA showed that the fre-
quency trajectory errors were significantly lower with the
Fig. 2. Average time-frequency trajectory profiles of the target sys-
tems when both magnitude and phase cues were altered. The error
bars show standard error of the mean. Open and filled symbols
represent hand excitation profiles with V and V+H feedback, respec-
tively.
nominal system and larger when exciting the systems of high
natural frequency. Further analysis showed that the trajectory
errors were affected by the feedback at the high natural fre-
quency and were not affected at the low natural frequency.
When either magnitude or phase cue was changed in target
systems with different natural frequency, the errors substan-
tially increased indicating inability of participants to quickly
converge to the natural frequency. The errors were relatively
higher when exciting the high natural frequency than when
exciting the low natural frequency. An interesting observation
was that for target systems with low natural frequency the
effect of feedback failed to show significance on the trajecto-
ry error data, whereas, V feedback yielded significantly larger
errors than V+H feedback when exciting the high natural fre-
quency. This result indicates that when one cue was changed,
a human ability to adapt to the low natural frequency was not
affected by the feedback, and haptic cue help participants
identify and converge to the high natural frequency.
REFERENCES
[1] M. K. OMalley, A. Gupta, M. Gen and Y. Li. Shared Control in
Haptic Systems for Performance Enhancement and Training. Journal of
Dynamic Systems, Measurement & Control, 128:75-85, 2006.
[2] Y. Li, V. Patoglu and M. K. O'Malley. Negative Efficiency of Fixed
Gain Error Reducing Shared Control for Training in Virtual Environ-
ments. ACM Transactions on Applied Perception, 6(1), 2009.
[3] J. L. Patton and F. A. Mussa-Ivaldi. Robot-Assisted Adaptive Train-
ing: Custom Force Fields for Teaching Movement Patterns. IEEE
Transactions on Biomedical Engineering, 51: 636-646, 2004.
[4] J. B. Dingwell, C. D. Mah and F. A. Mussa-Ivaldi. Manipulating
Objects with Internal Degrees of Freedom: Evidence for Model-Based
Control. Journal of Neurophysiology, 88:222-235, 2002.
[5] F. C. Huang, R. B. Gillespie, and A. D. Kuo. Visual and Haptic
Feedback Contribute to Tuning and Online Control during Object Ma-
nipulation. Journal of Motor Behavior, 39:179-193, 2007.
 Characterizing reflexive grip responses to
rotational perturbations of an object in precision grasp
Michael De Gregorio, M.S., Kevin Bair, and Veronica J. Santos, Ph.D.
Dept. of Mechanical and Aerospace Engineering, Arizona State University, Tempe, AZ, USA
INTRODUCTION
Spinal cord reflexes are often credited with mediating corrective fingertip forces for stabilization during
unexpected/expected perturbations to a grasped object (e.g., Johansson and Westling, Exp Brain Res,
1987). Prior work documents afferent activity, muscle activity, and kinetic data. However, there are very
little data available on hand kinematics during these perturbations. Of these data, most focus on tasks that
elicit flexion and extension corrective responses (e.g., lifting objects). This work investigates the human
hand response to rotational perturbations that elicit abduction and adduction of the fingers as well as
flexion and extension (to a lesser degree) in both the kinetic and kinematic arenas.
METHODS
In this study, we used a Vicon motion-capture system with retro-reflective markers to collect kinematic
data on the thumb and index finger of the dominant hand of an unimpaired subject (24 markers) and our
test object (3 markers). In addition, the time response of the first dorsal interosseous (FDI) was measured
for each subject using surface EMG (Biopac EMG 100C). Our rigid test object is instrumented with two
six degree-of-freedom load cells (ATI Nano-25 F/T), each with independent grip plates (one for the
thumb and one for the index finger). Subjects were instructed to grasp the parallel-faced object in a
precision grasp with one digit centered on each grip plate and to hold the object in midair. We then
subjected the test object to an unannounced rotational perturbation (clockwise or counterclockwise) using
a mass and pulley system (100 or 150 grams). The subjects task was to right the object after the onset
of perturbation. Using MATLAB, we extracted joint angles and fingertip forces for both digits.
RESULTS AND DISCUSSION
The kinetic analysis revealed that for clockwise perturbations the thumb acts as a pivot point and the
index finger acts as the primary corrective agent. Figure 1 shows that the grip force (force in the z
direction) is
Fx
slightly greater in Finger Tip Force Reponse Fy
Thumb
Force (lbs)

Thumb
the thumb than in Fz
4 Fx
Thumb
the index finger, Fy
Index
2 Index
whereas the lifting Fz
Index
0
force (force in the 1.5 2 2.5 3
y direction) is EMG of First Dorsal Interosseous
Voltage (V)

slightly greater in 0.08

the index finger 0.06
0.04
than in the thumb. 0.02
This is reiterated 0
1.5 2 2.5 3
by the center of Center of Pressure
Position (in)

pressure data 0.4 Thumb

which shows that 0.2 Index
after perturbation, 0
the index finger 0.2
moves upward in a 1.5 2 2.5 3
Time (s)
corrective motion
as the thumb Figure 1: Fingertip force response, surface EMG of the first dorsal interosseus (rectified, filtered,
moves initially toward the smoothed), and y-position of fingertip center of pressure are shown for the 100g, CW case for a single
center of the object 5sec trial of a right-hand dominant subject. The random perturbation occurred at t=1.99s (vertical
line). The black markers indicate the moment of maximum test object rotation. Center of pressure data
(position=0) and then shows that the index finger performs the corrective movement in this case, whereas the thumb tends to
act as a pivot.
slowly corrects upward. This relationship, of the thumb acting as a pivot, holds true across all subjects
(n=3), all masses, and all trials for the clockwise perturbation direction. For perturbations in the
counterclockwise direction, the opposite trend holds true. This suggests that each digit plays a different
role depending on the direction of the rotational perturbation.
As expected, the rotational perturbation elicited adduction/abduction responses in both the thumb and
index finger in addition to flexion/extension (Figure 2). Interestingly, the kinematic analysis shows that
peak index finger joint angles were synchronized with peak object rotation angles (~65-80ms after
perturbation) while thumb joint angles lagged behind. This was true regardless of perturbation direction
and was consistent across all subjects (n=3) and masses. The lag between maximum object rotation and
joint angle response (peak or valley) ranged from 2ms for the MCPFE and MCPAA of the index finger to
23-24ms for the CMCFE and CMCAA of the thumb. The standard deviation ranged from 2.7ms for the
MCPFE and MCPAA of the index finger to as much as 39.1ms for the DIPFE of the index finger.
MCP FE MCP AA PIP FE DIP FE
Index Joint Angles (deg)

2 Lag: 2 (2.7) ms
64 2 52 8
Rotation Angle (deg)

0 63 50 6
0
2 48
62 4
2
Test Object

4 46
61 44 2
6 4
60 42 0
8 Trial 1 6 40
10 Trial 2 59 38 2
Trial 3 58 8 36 4
12 Trial 4
57 10 Lag: 2 (2.7) ms 34 6 Lag: 16 (39.1) ms
14 Trial 5
32
0.1 0 0.1 0.2 0.3 0.1 0 0.1 0.2 0.3 0.1 0 0.1 0.2 0.3 0.1 0 0.1 0.2 0.3 0.1 0 0.1 0.2 0.3
time (s)

CMC FE CMC AA MCP AA MCP FE IP FE

Thumb Joint Angles (deg)

2 21
30 30 32
1 21.5
0 31
32 22 29
1 30
22.5
2 28 29
34 23
3
4 23.5 27 28
36 24
5 26 27
6 24.5
38 26
7 25 25
Lag: 24 (8.2) ms 8 Lag: 23 (5.7) ms 25.5 Lag: 9 (9.6) ms Lag: 9 (8.9) ms 25 Lag: 9 (4.2) ms
40 24
0.1 0 0.1 0.2 0.3 0.1 0 0.1 0.2 0.3 0.1 0 0.1 0.2 0.3 0.1 0 0.1 0.2 0.3 0.1 0 0.1 0.2 0.3
time (s)
Figure 2: Test object rotation angle, index finger joint angles (top) and thumb joint angles (bottom) are shown for the 100g, CW
case for a single right-hand dominant subject. The object rotation angle was used to set t=0 (vertical line). The black markers
indicate the moment of maximum object rotation due to the perturbation. Positive joint angles indicate flexion or adduction
while negative joint angles indicate extension or abduction. The lag between maximum object rotation and joint angle response
(peak or valley) across the five trials is shown as mean(std) wherever possible. Joints: CMC = carpometacarpal, MCP =
metacarpophalangeal, IP = interphalangeal, PIP = proximal interphalangeal, DIP = distal interphalangeal; Motions: FE =
flexion/extension, AA = adduction/abduction.

CONCLUSIONS
This study suggests that each digit plays a different role depending on the direction of the rotational
perturbation. Furthermore, it was shown that these perturbations elicit kinematic responses involving
adduction/abduction in addition to flexion/extension. However, additional analysis is still required to
further understand the delayed thumb kinematic responses. Future studies may look at the effects that
cutaneous anesthesia has on the response of the fingertips to rotational perturbations and the
randomization of our experimental trials through the use of stepper motors. Characterizing reflexive grip
responses for adduction/abduction in conjunction with flexion/extension will be critical to designing
autonomous reflexes for artificial hands, especially as prosthetic and robotic finger kinematics become
more anthropomorphic.
 Better Manipulation with Human Inspired Tactile Sensing
Ravinder S. Dahiya, Monica Gori, Giorgio Metta, Giulio Sandini
Robotics, Brain and Cognitive Sciences Department,
Italian Institute of Technology, Via Morego 30, Genoa, 16163, Italy

Introduction: Understanding what properties of the human hand can to be incorporated in robotic hands has been
an active area of investigation for a long time. Good strides have been made in designing robotic hands and a number
of working dexterous robotic hands have also been built [1, 2]. However, the use of touch sensory (both,
cutaneous/tactile/extrinsic as well as kinesthetic/intrinsic) information for dexterous manipulation still lags the
mechanical capability of such hands. This work presents how extrinsic touch sensing, the cutaneous/tactile analogous
of human sense of touch in robotics, can help improving the manipulation capability of robotics hands. Some features
of human cutaneous sense, namely, the role of skin biomechanics and skin microstructures, the spatio-temporal
response, information coding and transfer are presented as they may help extending the usage of tactile sensing in
robotic manipulation. If introduced, such features can help extending the intrinsic touch sensing and tendon driven
based gross manipulation capability of present day robotic hands to precise manipulation. The discussion is followed
by presenting the POSFET (Piezoelectric Oxide Semiconductor Field Effect Transistor) based tactile sensing arrays,
inspired from cutaneous/tactile sense of touch in humans, for the fingertips of humanoid robot iCub [3].

Tactile sensing for Manipulation by humans and robots: It is difficult to hold or safely manipulate real
world objects without physically touching them. The sense of touch is of essence to any manipulative task. Robots
guidance and force based control has mainly depended on the kinaesthetic information from intrinsic tri-axial or 6D
force sensors located close to wrist and on the actuation of tendon-driven fingers by motors located in the forearm.
However, transmission dynamics such as friction, backlash, compliance, and inertia make it difficult to accurately
sense and control endpoint positions and forces based on intrinsic sensors and actuator signals alone which point
towards the insufficiency of kinaesthetic information for manipulation in robotics [4]. Thus, there is need for
augmenting the kinesthetic information with the tactile information. In humans, the impaired tactile sensibility makes
manipulation difficult as brain lacks the information, about mechanical contact, needed to plan and control
manipulation which is centered on the mechanical events that mark transitions between consecutive action phases
[5]. Signals from tactile afferents play decisive role during such transitions. As an example, various phases of a
grasping action, namely reaching, loading, lifting, holding, replacing and unloading, are characterized as discrete
sensory events by specific tactile afferent responses. The FA (fast adapting) receptors respond to transient stimulation
- FAI responds at end of reaching and unloading phases and FAII responds at beginning of loading and unloading
phases. Similarly, SA-I (slow adapting) and SA-II afferents respond when static forces are applied to the object. The
activity of receptors during various phases of grasp gives an idea of contact timing, contact site, direction of contact
forces and shape of contact zone. Brain uses such tactile afferent information when humans manipulate objects and
similar information is needed by robots for manipulating objects. Measuring material properties such as hardness,
temperature etc., in addition to the measurement of contact forces, with tactile sensors can also be useful for
manipulating real world objects.

Human tactile sensing for better design of Tactile Sensors: Designing of a meaningful robotic tactile
sensing system should be guided by a broad but integrated knowledge of how tactile information is encoded and
transmitted at various stages of interaction via touch sensing. In this context, various studies on human tactile sensing
provide a good starting point. Such studies are also important due to the lack of any rigorous robotic tactile sensing
theory that can help in specifying important system parameters such as sensor density, resolution, location, and
bandwidth etc. - which are also likely to be task or application dependent.
Human skin structure is quite complex with tactile information elaborated by different kind of mechanoreceptors -
embedded in the skin at specific locations and depths and transducing signals with specific spatio-temporal
characteristics [6]. Density of various receptors too varies with body site. As an example, FA-I receptors have higher
density [5] than SA-I receptors on fingertips, thus, reflecting the importance of extracting spatial features of dynamic
mechanical events and supporting the need for dynamic tactile sensors in robotics.
The mechanoreceptors are not just transducers. Both independently and as a group they also involve some local
processing. Different firing rates of mechanoreceptors helps their independent coding of contact events. When
considered as a group, the relative timing of their first spikes provides precise information about the shape of the
contacted surface as well as the direction of the force exerted on the hand, and it does fast enough to account for the
speed with which tactile signals are used in object manipulation tasks [5]. Such processing of data is useful as it helps
optimum usage of the limited throughput of the nervous system. For robotics, these results not only underlie the
importance of having tactile sensing arrays on robotic hand, but, also local processing of the data collected by them.
Minimizing the data by way of local processing not only helps optimum usage of the limited computational resources
of a robot, but also facilitates the speedy transfer of contact information for any control task.
The elasticity of skin varies with depth, which can influence the intensity of the tactile signal that the receptors
receive. Further skin contains some ridge like patterns, comprising of papillary ridges or fingerprints, intermediate
ridges and limiting ridges [7]. Both papillary ridges and intermediate ridges are believed to affect the response of
various receptors in the skin to a different degree [7-9]. At their tips, intermediate ridges house the Merkel cell
complex and hence they are believed to influence the response of SA-I receptors [8]. Similarly in addition of
enabling better grip [10], the fingerprints or papillary ridges are also believed to enhance the tactile sensitivity of
Pacinian Corpuscles and hence help feeling fine textures[9]. Mimicking complex human skin structure - with
receptors embedded at specific depths and locations, performing different functions and responding optimally in
 Fig. 1: Concept of POSFET
Tactile Sensing array and the
SEM picture of implemented
POSFET tactile sensing device.
Like mechanoreceptors in human
skin, each POSFET device is
capable of sensing and
processing the touch information
at same site. The output of
POSFET touch sensing device is
linear (with a slope of 50 mV/N)
for tested range of normal forces
(0.15-5 N)

different frequency ranges, is a challenging technological issue. Nonetheless, adding functional equivalent of a
mechanoreceptor (e.g. Pacinian Corpuscles) to an ordinary tactile sensor, by using soft protective rubber cover
patterned with fingerprint like microstructures can help in broadening the usage of tactile sensing and in bringing the
level of tactile sensitivity and acuity, that human possess, to robotic devices.

Human Inspired POSFET Tactile Sensing Arrays: It is desirable to have tactile arrays or distributed tactile
sensors with density and spatial distribution of taxels (tactile elements) varying with body site. For the sites like
fingertips a large number of fast responding (of the order of few milliseconds) taxels are needed in a small space (~ 1
mm spatial resolution). Further, local processing and use of less number of wires are also desired. Keeping in view
such facts, tactile sensing array using POSFET touch sensing devices have been developed. Designs of, both,
POSFET devices and the array are inspired from cutanous sense in humans. Tactile sensing device is fabricated by
spin coating a piezoelectric polymer (PVDF-TrFE) film on the gate area of MOS (Metal Oxide Semiconductor)
devices. A force applied on piezoelectric polymer generates charge, which in turn modulates the charge in the
induced channel - thereby converting force in to voltage. Contrary to conventional approaches - where transducers
and conditioning electronics are separate entities connected through wires each POSFET touch sensing element
presents an integral unit comprising of transducer and transistor. As shown in Fig. 1, each POSFET element, as an
integral sensotronic unit, is capable of sensing and partially processing the touch signal at same site as is done
by the receptors in human skin. Further, absence of any wire between transducer and the transistor can help solving
the wiring complexity, which is one of the major issues hindering the wide usage of distributed tactile sensing. A
system on chip or system in package with on-chip conditioning electronic circuitry and local processing will further
improve the overall performance of POSFET tactile sensing arrays and utilization of the tactile data in a control task.
To match the spatial resolution and acuity of human fingertips, the size of each touch element is 1 mm x 1 mm and
the center to center distance between adjacent elements is 1 mm. POSFET elements have linear response up to 5 N
and constant gain over tested frequency range of 2.13 KHz. In present format POSFET tactile sensing arrays use a
plain thin rubber cover i.e. one without any microstructure like fingerprints or intermediate ridges. However,
POSFET tactile sensing arrays in future will have the cover patterned with microstructures, as in human skin.

Conclusion: The ways in which biological sensory systems are structured and process information to control
behavior may not always lead to the best engineering solutions for robots, nevertheless they provide useful insights
into how behaving organisms respond to dynamically changing environments, and also provide a comprehensive
multi-level conceptual framework within which to organize the overall task of designing the sensors for robotic
systems. The approach may bring up new ideas that can help improving the level of tactile sensitivity and acuity of
robots to the human range. With this premise, human inspired POSFET tactile sensing arrays have been developed
and presented here. The POSFET tactile sensing arrays are good for dynamic contact events like FA receptors in
the skin. However, the structure can be modified to include other mode of transduction that is sensitive to static or
quasistatic contact events.

References:
[1] G. Bekey, Autonomous Robots: from Biological Inspiration to Implementation and Control. Cambridge: MIT Press, 2005.
[2] Touch Bionics, "The i-LIMB Hand," http://www.touchbionics.com/, 2007.
[3] www.robotcub.org, "iCub," 2009.
[4] R. S. Dahiya, G. Metta, M. Valle, and G. Sandini, "Tactile Sensing: From Humans to Humanoids," IEEE Transactions on
Robotics, 2009 (in press).
[5] R. S. Johannson and J. R. Flanagan, "Coding and use of tactile signals from the fingertips in object manipulation tasks," Nature
Reviews Neuroscience, vol. Advance Online Publication, pp. 1-15, 2009.
[6] K. O. Johnson, T. Yoshioka, and F. Vega-Bermudez, "Tactile functions of mechanoreceptive afferents innervating the hand," J
Clin Neurophysiol, vol. 17, pp. 539-58, Nov 2000.
[7] G. J. Gerling and G. W. Thomas, "Fingerprint Lines may not directly affect SA-I mechanoreceptor response," Somatosensory
and Motor Research, vol. 25, pp. 61-76, 2008.
[8] N. Cauna, "Nature and functions of the papillary ridges of the digital skin," Anatomical Record, vol. 119, pp. 449-468, 1954.
[9] J. Scheibert, S. Leurent, A. Prevost, and G. Debregeas, "The Role of Fingerprints in the Coding of Tactile Information Probed
with a Biomimetic Sensor," Science, vol. 323, pp. 1503-1506, 2009.
[10] T. Maeno, K. Kobayashi, and N. Yamazaki, "Relationship between the Structure of Human Finger Tissue and the Location of
Tactile Receptors," Bulletin of JSME International Journal, vol. 41, pp. 94-100, 1998.
Title: Quantification of wrist/finger flexion forces and EMGs as a function of limb loading in chronic
hemiparetic stroke
Authors: Laura C. Miller1,2, Julius P.A. Dewald1-4
Departments of Biomedical Engineering1, Physical Therapy and Human Movement Sciences,2 Interdepartmental
Neuroscience Program3, and Physical Medicine and Rehabilitation4, Northwestern University, Chicago, IL, USA
Introduction
The expression of the flexion synergy in the paretic upper limb of moderately to severely affected
hemiparetic stroke survivors may cause disabling coupling between shoulder abductors and elbow, wrist
and finger flexors [1]. This would explain the hypertonia in the wrist/finger flexors and weakness in the
wrist/finger extensors frequently observed following stroke. Robotic assistive devices for the hand, such
as extension-aiding gloves and exoskeletons, have the potential to recover hand function, as do robot-
aided rehabilitation protocols, which can involve sophisticated and precise active or passive mechanical
manipulations of the hand/wrist. However, better identification of the physiological and neural
mechanisms underlying stroke-induced movement synergies must be made before these technologies can
reach their full potential. Understanding the underlying mechanisms and exact expression of synergies is
crucial for the creation of novel therapeutic interventions that are physiologically and impairment based,
and possibly more effective than conventional therapies.
Studies that have quantified elements of hand dysfunction following stroke have focused on the hand
in isolation, so the behavior of the hand during movements of more proximal joints in the upper limb has
not been adequately characterized [2-8]. Sukal et al. quantified reductions in active elbow extension for
greater levels of shoulder abduction, according to the flexion synergy [9]. These results agree with the
abnormal coupling described under static conditions between shoulder abduction and elbow flexion in
earlier studies [10, 11]. If abnormal torque coupling exists between the elbow and shoulder then it would
be reasonable to hypothesize that this abnormal coupling would also be present at the wrist and fingers,
consistent with clinically described stereotypical movement patterns following hemiparetic stroke [1, 12].
This study measured isometric forces generated by the paretic, non-paretic, and healthy control fingers
and wrist during lifting and reaching, at 7 levels of shoulder abduction loading, using the ACT3D robotic
device [13].
Methods
Eight individuals with chronic hemiparetic stroke (mean age 62 9 yrs, range 22-270 months post-
stroke) and three healthy controls (mean age 42 14 yrs) participated in this study. All stroke-affected
participants had severe to moderate upper limb impairment according to the Upper Limb Fugl-Meyer
Motor Assessment, with scores ranging 10-37 (mean 20.6) of a possible 66, and had severe to moderate
hand impairment according to the Chedoke-McMaster Hand Assessment, with scores ranging from 2-5
(mean 2.9) of a possible 7.
Each participant placed the arm in the ACT3D robotic device [13]. The rigid forearm orthosis of the
ACT3D had been modified to include the Wrist/Finger Force Sensing module (WFFS) [14], which can
measure isometric forces generated by the fingers and wrist and by the thumb. Participants were
instructed to move the limb into a home position of 85 shoulder abduction, 90 elbow flexion, and 40
shoulder flexion and to hold for 1s. Timing and limb positioning were verified using real-time visual
feedback on the ACT3D monitor. After holding the home position, participants were instructed to reach as
fast as possible to a virtual target displayed on the ACT3D feedback monitor, while keeping their hand
relaxed during the reach. The virtual target was located in the sagittal plane through the shoulder. 11 trials
were completed at each of 7 active shoulder support levels: while supported on the ACT3D haptic table
and while lifting up at 0, 5, 15, 25, 35, and 50% of maximum shoulder abduction force. Surface
electromyographic (EMG) data were also recorded from 10 muscles in the upper limb.
Results
 Forces generated during two time windows were selected for analysis: while holding the home
position (LIFT) and at maximal reach (REACH). Forces were smoothed using a 250 ms moving average
filter and were normalized to maximum volitional flexion force. Normalized wrist/finger flexion forces
generated by the paretic and non-paretic limbs are shown below. Separate two-way repeated measures
(RM) ANOVAs for LIFT and for REACH
revealed significant main effects of support
condition and of limb on flexion forces
generated, as well as significant
interactions between the main effects (p <
0.0001 for all). Additionally, when the
effect of task (LIFT vs. REACH) was
included in the two-way RM ANOVA with
support level, there was a significant main
effect of task for the paretic limb (p =
0.007). A two-way RM ANOVA also
revealed significant main effects of support
level (p = 0.0002) and limb (p = 0.03) on
EMG data recorded over the flexor
digitorum superficialis muscle during LIFT.

Discussion
Wrist/finger flexion forces and EMG generated by the paretic limb drastically increased with required
shoulder abduction force (LIFT) and increased further with concurrent elbow extension and shoulder
flexion (REACH). When participants had to generate 50% of their maximum shoulder abduction force
and then reach out, wrist/finger flexion forces were generated that were near volitional maximum levels
(91.7 40.8%, mean SD). Therefore, studies using robotic devices that focus on the hand and wrist in
isolation provide an incomplete picture of movement dysfunction following stroke, because they do not
consider the abnormal finger/wrist flexion coupling that occurs with increasing levels of shoulder
abduction activity in the paretic upper limb.
1. Brunnstrom S, Movement therapy in hemiplegia: a neurophysiological approach. 1970, New York: Harper and Row.
2. Cruz E G, Waldinger H C, and Kamper D G, Kinetic and kinematic workspaces of the index finger following stroke. Brain,
2005. 128(5): p. 1112-1121.
3. Kamper D G, Fischer H C, Cruz E G, and Rymer W Z, Weakness Is the Primary Contributor to Finger Impairment in
Chronic Stroke. Archives of Physical Medicine and Rehabilitation, 2006. 87(9): p. 1262-1269.
4. Kamper D G, Harvey R, Suresh S, and Rymer W Z, Relative contributions of neural mechanisms versus muscle mechanics
in promoting finger extension deficits following stroke. Muscle & Nerve, 2003. 28(3): p. 309-318.
5. Kamper D G and Rymer W Z, Quantitative features of the stretch response of extrinsic finger muscles in hemiparetic stroke.
Muscle & Nerve, 2000. 23(6): p. 954-961.
6. Kamper D G and Rymer W Z, Impairment of voluntary control of finger motion following stroke: Role of inappropriate
muscle coactivation. Muscle & Nerve, 2001. 24(5): p. 673-681.
7. Kamper D G, Schmit B D, and Rymer W Z, Effect of Muscle Biomechanics on the Quantification of Spasticity. Annals of
Biomedical Engineering, 2001. 29(12): p. 1122-1134.
8. Li S, Kamper D, and Rymer W, Effects of changing wrist positions on finger flexor hypertonia in stroke survivors. Muscle
& Nerve, 2006. 33(2): p. 183-190.
9. Sukal T, Ellis M, and Dewald J, Shoulder abduction-induced reductions in reaching work area following hemiparetic
stroke: neuroscientific implications. Exp Brain Res, 2007. 176: p. 594-602.
10. Beer R, Given J, and Dewald J, Task-dependent weakness at the elbow in patients with hemiparesis. Arch Phys Med
Rehabil, 1999. 80(7): p. 766-772.
11. Dewald J and Beer R, Abnormal joint torque patterns in the paretic upper limb of subjects with hemiparesis. Muscle &
Nerve, 2001. 24: p. 273-283.
12. Twitchell T, The restoration of motor function following hemiplegia in man. Brain, 1951. 74: p. 443-480.
13. Sukal T, Ellis M, and Dewald J, Use of a Novel Robotic System for Quantification of Upper Limb Work Area Following
Stroke., in 27th IEEE EMBS annual international conference. 2005: Shanghai, China.
14. Miller L, Ruiz Torres R, Stienen A, and Dewald J, A Wrist and Finger Force Sensor Module for Use During Movements of
the Upper Limb in Chronic Hemiparetic Stroke. IEEE Transactions on Biomedical Engineering, 2009. In Press.
 Human reach-to-grasp generalization
strategies: a Bayesian approach
Diego R. Faria, Ricardo Martins and Jorge Dias
Institute of Systems and Robotics
Department of Electrical Engineering
University of Coimbra - Polo II
3030-290 Coimbra, Portugal
Email: {diego,rmartins,jorge}@isr.uc.pt

Abstract In this work we present a general structure of an ini- II. BAYESIAN A PPROACH
tial Bayesian framework to describe the mechanisms underlying
the human strategies that define the appropriate characteristics Based on the studies about the grasp model [8] and studies
of the reach-to-grasp movements to specific contexts, objects and of neuroscience of grasping [4], we intend to develop a
how these strategies can be extended and replicated to other Bayesian framework to generalize the human strategies to
contexts and objects. The Bayesian framework uses information
extracted from data about the pose of the hand, fingers and
perform a reach-to-grasp tasks and to transfer and integrate
head acquired by a magnetic tracker device, finger flexure data this knowledge to robotic platforms.
acquired by a data glove, as well as, data about eye gaze and
saccade movements of the subject.

I. I NTRODUCTION

Some of the most performed actions by humans in their

daily activities involve the manipulation of objects from one
place to another and its in-hand manipulation to adjust the
pose of the object with the final goal of the action which will
be performed with it.
Typically, the global hands trajectory during a manipulation
task can be segmented in different stages: reach, lift, transport
and release[6]. We focus our attention in the reach stage
(reach-to-grasp movement). The conclusions of several pre-
vious studies suggest that the human decision making process
of how (type of grasp [5], hand trajectory characteristics and
maximum grip aperture [4][10]) the reach-to-grasp movement
is performed, is influenced by different factors, such as the
presence or absence of visual feedback [9] or the eye-hand
coordination in object manipulation with and without obstacles
[7].
We intend to develop a Bayesian framework which will
describe what are the specific and elementary factors which
are involved in the planning and control of a reach-to-grasp
movement. This can contribute to the description of the human
strategies that are used to extract the physical properties of
the object which will be grasped and how humans relate
those highlighted physical properties with the objects possible
applications (affordance). Our main concern is to provide a
framework that can define a general description of the human
strategies (generalization), instead of a model that is specific
and restricted to the learnt object/context. An initial structure
of the framework is presented in section II. The experiments
carried out to validate this initial framework will be used to
refine and detail it in the future.
Figure 1 - Overview of the expected inputs and outputs of the
Bayesian framework integrated in a robotic platform.
Figure 1 shows a schematic representation of the global
overview of the expected inputs and outputs of the framework.
The robotic platform must have the ability to estimate the
identity (Oi ) (dimensions, shape), position (Op ) and relative
orientation (Oo ) of the object to be manipulated. The robotic
platform must be instructed about the task (T ) which should
be performed with the object. In this initial approach the
considered possible tasks are restricted to the indication of
what type of grasp must be performed, such as top-grasp or
side-grasp. In future developments, it is expected to increase
the set of possible tasks and automatically establish the relation
between the objects identity and its affordances. The type of
task can be indicated through a pre-defined instruction or it
can be extracted from a human demonstration. In the latter
situation, the reach-to-grasp movement can be classified as
top-grasp or side-grasp by the Bayesian classifier developed
in a previous work [5]. This set of information is combined
in the Bayesian framework, which estimates the region of the
object which will be grasped (R) and several characteristics
of the trajectory and behaviour of the robotic hand during
the approach to the object: hand trajectory curvature (Hc ),
hand orientation (Ho ) and the level of fingers flexure (F ).
These expected outputs of the Bayesian framework can be
written using the Bayesian formalism as P (R|op , oo , oi , t),
P (Hc |op , oo , oi , t), P (Ho |op , oo , oi , t), P (F |op , oo , oi , t), re-
spectively. These informations are transmitted to the robot
motor control modules which are responsible by its execution.
A. Learning
The learning phase corresponds to the acquisition of knowl-
edge from human demonstrations of the required tasks per-
formed with different objects (variable identity, size and shape)
in different contexts (positions, orientation).
Six magnetic sensors of the Polhemus Liberty device [2] are
used to provide the pose (6 DoF) of the fingers and head. Five
of them are attached to a data glove [1] on the fingertips region
and the other one is attached to the head-mounted eye-tracker
device. The head-mounted eye-tracker device, such as [3], is
used to extract information about the regions of the object
which are observed during the reach-to-grasp movement. A
magnetic sensor is also attached to the grasped object, in order
to acquires its initial position and orientation.
This acquired knowledge is represented by Learned His-
tograms. For each implemented task, object orientation and
object identity, four types of learning tables are defined:
Trajectory Curvature Learned Table (CLT), Hand Orientation
Learned Table (HLT), Finger Flexure Learned Table (FLT)
and Grasped Region Learning Table (RLT). The definition
of the CLT, HLT are described in more detail in a previous
work [5]. The CLT represents the probability of a specific
curvature in each of the eight segmented parts of a normalized
hand trajectory. Similarly, the HLT and FLT represent the
probability of a specific hand orientation and level of fingers
flexure, respectively, in each of the eight trajectory segments.
The RLT represents the probability of the object being grasped
in each of three defined regions: the upper, middle or bottom
part of the object. The probability is determined based in the
principle studied in [7], which relates the observed regions
of an object during a manipulation task with the region of
the object where the grasping is performed. The mapping of
the observed regions of the object is made by an eye-tracker
device such as [3].
B. Generalization Strategy
As previously referred, for all types of learned fea-
tures we have learned tables (mean histograms calculate
from all observations): P (R|op , oo , oi , t), P (Hc |op , oo , oi , t),
P (Ho |op , oo , oi , t), P (F |op , oo , oi , t).
For a estimated object identity, orientation, position and
pre-defined task, it is possible to determine the appropriated
learned table for each of the wanted features: hand trajectory
curvature, hand orientation, finger flexure and grasped region.
The estimated distance between the robotic platform and the
object is segmented in eight parts, as the data collected during
the construction of the learned tables. For each of these
segments, it is determined the most probable value of each
of the features referred before, based in the learned tables
previously identified.
For instance, to generalize the trajectory curvatures, we get
the higher probability of the features in each hand displace-
ment, in a curvature learned table (CLT) for the top-grasp , so
that we have the curvatures: U-U-U-UR-DR-DR-D-D, and for
each curvature we have associated the (r, , ) information,
that is, the angles of each curvature and the information of
forward or backward direction. The same happens for the hand
orientation, eg. for top-grasp we have the features with higher
probability in each hand displacement: T-T-S-T-T-T-T-T (T:
top orientation, S: side orientation, more details see [5]) and
the angles of the hand plane. The same for the finger flexure
(joint angles).
The learned information: (r, , ) curvatures; angle of the
hand plane ( hand orientation); the fingers joint angles acquired
from P (F |op , oo , oi , t) and the region of the object to grasp
given P (R|op , oo , oi , t), is mapped to the robot referential to
the robot reproducing the same task according with its DoF.
III. C ONCLUSION
In this preliminary work, we present an initial approach
to a framework to represent the human strategies to perform
reach-to-grasp tasks. The analyses that were made will support
and guide our future works. We focus our attention in the
analysis of the visual behavior (conjugation of head and eyes
movements) of the subjects, in the tracking of hand, fingers
and head pose, as well as, the flexure level of the fingers along
the reach-to-grasp movements.
ACKNOWLEDGMENT
This work is partially supported by the European project
HANDLE ICT-23-16-40. Diego Faria is supported by Por-
tuguese Foundation for Science and Technology (FCT).
R EFERENCES
[1] http://www.5dt.com/products/pdataglove5u.html. Accessed in 20 April
2009.
[2] http://www.polhemus.com/. Accessed in 20 April 2009.
[3] http://www.smivision.com/en/eye-gaze-tracking-systems/products/iview-
x-hed.html. Accessed in 20 April 2009.
[4] U. Castiello. The neuroscience of grasping. Nature Reviews Neuro-
science, 6(9):726736, 2005.
[5] D. R. Faria and J. Dias. 3d hand trajectory segmentation by curvatures
and hand orientation for classification through a probabilistic approach.
In (to appear) 2009 IEEE/RSJ International Conference on Intelligent
Robots and Systems. IROS09.
[6] J. R. Flanagan, M. C. Bowman, and R. S. Johansson. Control strate-
gies in object manipulation tasks. Current Opinion in Neurobiology,
16(6):650659, 2006.
[7] R. S. Johansson, G. Westling, A. Bckstrm, and J. R. Flanagan. Eye-
hand coordination in object manipulation. The Journal of Neuroscience,
21(17):69176932, 2001.
[8] E. Oztop, N. S. Bradley, and M. A. Arbib. Infant grasp learning:
a computational model. Experimental Brain Research, 158:480503,
2004.
[9] M. K. Rand, M. Lemay, L. M. Squire, Y. P. Shimansky, and G. E.
Stelmach. Role of vision in aperture closure control during reach-to-
grasp movements. Experimental Brain Research, 181(3):447460, 2007.
[10] M. K. Rand, Y. P. Shimansky, A. B. M. I. Hossain, and G. E. Stelmach.
Quantitative model of transport-aperture coordination during reach-to-
grasp movements. Experimental Brain Research, 188(2):263274, 2008.

EMG-driven Forward Dynamic Simulations of
the Hand performing American Sign Language
Craig M. Goehler, Richard F. ff. Weir, and Wendy M. Murray
Index TermsAmerican Sign Language postures, dexterous
hand control, forward dynamic simulations
I. INTRODUCTION
P ERSONS with recent hand amputations expect modern hand
prostheses to function like intact hands. Current state-of-
the-art electric prosthetic hands are generally single degree-of-
freedom (opening and closing) devices that are controlled
using only two muscle signals. As a result, most state-of-the-
art devices fail to meet users expectations and are under-
utilized or rejected. Overall, performance of sophisticated
hand prostheses is limited by the ability to control them.
Ultimately, the controller for a multi-degree-of-freedom
prosthetic hand must allow the user to seamlessly command
the artificial hand. The long-term goal of this work is to
evaluate the feasibility of utilizing EMG-driven forward
dynamics simulation methods in a controller for a multi
degree-of-freedom prosthetic hand. Providing an individual
with the ability to assume any possible hand posture is an
important, necessary and major intermediate step toward
realizing the goal of enabling dexterous manipulation.
A control source in prosthetics is any body-generated signal
(biomechanical or bioelectric) that can be transduced into an
electric signal to control the prosthesis. The most common
bioelectric signal used as a control source is the
electromyogram (EMG) [1]. EMG signals can be readily
detected and amplified. We propose to use forward dynamic
simulations to predict the motions that would occur in the
intact system (without amputation) given the EMG signals
measured from the residual muscles of an amputees limb.
Based on the simulated movements, the prosthesis controller
would actuate the joint motors to match predicted positions
and velocities.
Dexterous hand function encompasses the fine motor skills
C. M. Goehler is with the Sensory Motor Performance Program at the
Rehabilitation Institute of Chicago, Chicago, IL 60611 USA (corresponding
author; e-mail: c-goehler@northwestern.edu).
R. F. Weir is Director, Biomechatronics Development Laboratory at the
Rehabilitation Institute of Chicago, & with Departments of BME and PM&R
at Northwestern University, and the Research Service at the Jesse Brown VA
Medical Center (URL: http://www.smpp.northwestern.edu/Weir/index.shtml).
W. M. Murray is with the Sensory Motor Performance Program at the
Rehabilitation Institute of Chicago, the Departments of BME and PM&R at
Northwestern University, and the Research Service at the Edward Hines Jr.
VA Hospital (e-mail: w-murray@northwestern.edu).
1
of the hand and fingers. As such, we will evaluate the
suitability of EMG-driven forward dynamics simulation
methods for use in a prosthetic controller by first evaluating
the ability to simulate complex, multi-degree of freedom hand
postures using EMG signals collected from only the extrinsic
muscles of the hand (the long flexor and extensor muscles
whose origins and muscle bellies are located in the forearm).
The manual alphabet of American Sign Language (ASL) was
selected as the metric. These postures represent a continuum
of well-defined static postures and postural transitions that are
representative of a variety of fine motor skills [2].
In this study, EMG-driven forward dynamic simulations
using the extrinsic muscles of the hand were performed for the
W and L ASL hand postures. These simulations illustrate
our first step to understanding control of hand mechanics
utilizing only muscles present in the amputee limb.
II. FORWARD DYNAMIC SIMULATION PROCEDURE
A. The Musculoskeletal Model
In the forward dynamic simulations, differential equations
that describe muscle activation dynamics, muscle-tendon
contraction dynamics and skeletal dynamics are combined
with a representation of the musculoskeletal geometry to
compute the position, velocity, and acceleration of the body
over time. Bone geometry, joint kinematics, muscle paths and
muscle force-generating parameters were obtained from a
musculoskeletal model of the upper extremity [3]. The
simulations utilize the 22 degrees of freedom and 24 muscles
that make up the wrist, thumb and fingers of the upper
extremity model.
B. EMG Acquisition
Intramuscular EMG signals were obtained from a previous
study in which the data were collected from multiple muscles
using fine-wire electrodes while subjects generated ASL
postures [4]. Subjects were visually cued to shape their
dominant hand into each of the 33 ASL static letters and
numbers (J and Z were omitted). The W and L
postures were chosen for the initial simulations.
An assumption was enforced such that the EMG signal
from one slip (muscle path) of a particular finger muscle was
taken to be the signal for all slips of that finger muscle.
Consequently, 16 recorded EMG signals were transformed
into 24 input signals through use of similar assumptions.

C. Simulation Process
The collected EMG data was used to drive the forward
dynamics simulations. The muscle-contraction dynamics
were implemented such that the muscle-tendon force produced
by each muscle was computed from the system states: the
muscle fiber lengths, muscle activations, and the generalized
coordinates and speeds that influence muscle-tendon length
and velocity. Muscle activation dynamics were modeled by a
first-order differential equation. Passive joint torques and
damping were implemented at each joint in order to simulate
ligaments and to limit joint motion at extreme joint angles.
The state variables for the simulation include the 22
generalized coordinates, 22 generalized velocities, 24 muscle
fiber lengths and 24 muscle activations, for a total of 92 states
(Fig. 1).
The 24 input signals from the 16 recorded EMGs of the
extrinsic muscles defined the neural excitation input to each
muscle. The initial conditions were defined by the zero
position of the musculoskeletal model at rest. The equations
Fig. 1. Flowchart of the forward dynamic simulation process.
of motion were numerically integrated using a variable-step
integrator over the time range of the movement, solving for
the motion of each moving segment at each time step. This
process resulted in a time-series output file that contains each
of the state variables.
III. SIMULATION RESULTS
The final poses for the L and W ASL posture mechanics utilizing only muscles present in the amputee limb.
simulations illustrate the potential for the forward dynamic
simulation approach to predicting hand movement (Fig. 2).
These poses reflect the ability to flex or extend individual
fingers along with producing complicated thumb movements.
The decision to use the recorded EMG signal of one slip for
all four slips of certain finger muscles appeared to be well
founded for the W posture. However, the final pose for the
L posture was achieved through modifying this assumption
for one of the flexor muscles of the index finger. Initially, the
final pose resulted in the index finger flexed similar to the
other three fingers. But when the input to this particular
finger flexor was changed, the reported pose was achieved.
This highlights the importance in obtaining EMG signals for
2
Fig. 2. Final poses for L and W ASL posture simulations.
each individual slip of each individual muscle in order to
achieve dexterous motion.
The negative effects of not including the intrinsic muscles
of the hand include inaccurate abduction/adduction angles for
the fingers as well as the inability to fully extend the distal
segments of the fingers.
IV. CONCLUSION
The observed differences between the individual finger and
thumb movements required to achieve the final L and W
poses support the choice to utilize ASL postures as a useful
metric of dexterous hand function. Examining a variety of
differing ASL postures will aid us in our final goal of
understanding the control of the human hand.
However, we learned that achieving an even greater level of
control of the hand will require EMG data for each individual
slip of each extrinsic muscle as well as a method to
incorporate the effects of the intrinsic muscles into
simulations based solely on extrinsic muscles. It was also
noted that the time needed to complete the forward dynamic
simulations was extensive and the development of real-time
methods would be required to implement these results into a
real-time prosthetic controller.
EMG driven forward dynamic simulations using only
extrinsic muscles represent a method of studying hand
movements that will lead to a better understanding of
dexterous hand motion along with the development of a more
sophisticated control paradigm for prostheses along with
robotic or mechanical hands of any nature. In general, this is
a positive first step towards understanding the control of hand
REFERENCES
[1] D. Childress, and R.F. Weir, Control of Limb Prostheses, in Atlas of
Amputations and Limb Deficiencies, 3rd ed. Rosemont, IL: American
Academy of Orthopaedic Surgeons, 2003.
[2] T.E. Jerde, J.F. Soechting, and M. Flanders, Biological constraints
simplify the recognition of hand shapes, IEEE Transactions on
Biomedical Engineering, Feb 2003, 50(2), pp.265-269.
[3] K.R. Holzbaur, W.M. Murray, and S.L. Delp, A model of the upper
extremity for simulating musculoskeletal surgery and analyzing
neuromuscular control, Annals of Biomedical Engineering, June 2005,
33(6), pp. 829-840.
[4] A.B. Ajiboye, Neuromotor Muscle Synergies for EMG Pattern
Recognition of Prehension Grasps for Control of Multifunctional
Myoelectric Prostheses. PhD Dissertation, Evanston, IL: Department of
Biomedical Engineering, Northwestern University, 2007.
 Design Optimization of a Hand Exoskeleton
Rehabilitation Device
Jamshed Iqball,2, Nikos Tsagarakisl, Darwin Caldwelll
lItalian Institute of Technology, Genova, Italy
2University of Genova, Italy
Corresponding author: jamshed.iqbal@iit.it

full extension to full flexion while a 2D trajectory followed

I. INTRODUCTION
The hand is the main body part used for physically
manipulating objects and is vital for the maintenance of
independent living. After hand injuries or strokes,
rehabilitation therapy may be required to allow the hand to
recover its strength and functionality. These therapy
by finger-tip is illustrated in figure 2. Starting from full
extension (Figure 1a), the finger is flexed by actuating every
movable joint with a corresponding value and every
phalanges is moving accordingly until full flexion (Figure
1d) is achieved.
0.3

procedures are usually executed manually by 0.15 0.25

physiotherapists or occasionally using simple passive 0.1 0.2

assistive devices. However, there is an increasing belief that 0.05 0.15

0.1
alternative therapies using robotic hand exoskeletons or 0
Y -ax is

Y -axis
0.05
-0.05
assistive devices could have considerable potential in 0
-0.1

improving the medical outcomes. -0.15

-0.05

Such a hand exoskeleton should be capable of following -0.2

-0.1

the hand motion permitting and ensuring natural movements -0.25

-0.15

of the fingers without any constraints. The device should be -0.15 -0.1 -0.05 0 0.05 0.1
X-axis
0.15 0.2 0.25 0.3 0.35 -0.2 -0.1 0
X-axis
0.1 0.2 0.3

capable of exerting perpendicular forces to fingers 0.3

(a) (b)
phalanxes throughout their motion in an adjustable manner. 0.25
0.25

The exoskeleton system should have enough dexterity to 0.2

0.2

0.15

cover operational workspace of the human hand. This work 0.15

0.1

presents a novel design of an under-actuated exoskeleton 0.1

0.05
Y -axis

Y -axis

0.05

system that can address the drawbacks found in the majority 0

0

of the hand exoskeleton systems such as weight, size and -0.05

-0.05

-0.1

ergonomics. The work will show initially through simulation -0.1

-0.15

how the device design has been optimized to maximize the -0.15
-0.2
-0.1 0 0.1 0.2 0.3 0.4
workspace and minimize the motion constraints. -0.2 -0.1 0
X-axis
0.1 0.2 0.3
X-axis

(c) (d)
II. METHODOLOGY Figure 1. Finger model: (a) Full extension (b) Intermediate
An anatomical hand model has been developed and used (c) Partial flexion (d) Full flexion
to obtain the finger angle ranges during full grasping and 0.3

release. Each joint in the model is characterized by a specific

x-position
geometry and parameters such as the minimum and 0.25
y-position
z-position
maximum angle which bound the motion of the system. The 0.2 T4

model also incorporates coupling between Proximal

Position--->

0.15

InterPhalangeal (PIP) and Distal InterPhalangeal (DIP) 0.1

joints. The angle of the distal joint (DIP) is calculated from

the middle joint angle (PIP) [1] using the following formula 0.05

0
T1 T2 T3
DIP = 0.46PIP + 0.083PIP2 -0.05
0 100 200 300 400 500 600
Time --->

Figure 2. Finger-tip position (one cycle) T1 = Flexion phase, T2 = Grasping

Figure 1 shows the graphical representation of finger from phase, T3 = Extension phase, T4 = Full release.
 Figure 3a shows a drawing of the planned exoskeleton Start

(EXO) system attached to a finger model. The link lengths Finger angles

are denoted by L1E, L2E and L3E. The performance of such Finger forward
Finger Model
kinematics
device can be represented by its capability to exert as
perpendicular forces as possible to fingers phalanges [2]. Mapping in EXO frame

Target/attachment point

EXO link lengths

assumption

EXO IK
Theta [1-3]
Worst case (collision)
distance calculation

NO
d >0 &
d > T (EXO
Thickness)

YES
Figure 3. (a) EXO and finger model (b) Kinematic iteration
Dexterity analysis Force Impact Factor
III. RESULTS AND DISCUSSION
This work focuses on finding those link lengths which can
Overall Impact Factor
maximize the performance criterion mentioned in the last
section. The angle values from the finger model are used in
Update EXCEL file
kinematic analysis and optimization algorithm. The EXO
link lengths have been iterated and kinematic analysis
selected those combinations which maximize the NO Trajectory
traversed
performance criterion. Figure 3b shows one such valid
iteration for L1E, L2E, L3E equal to 8, 7, 1.5 cm YES

respectively. The green lines show Meta CarpoPhalangeal Stop

(MCP) and PIP links of human finger whereas EXO links are
represented by pink lines. This valid combination of lengths Figure 5. Flow chart
is then subjected to optimization. Parameters of optimization
included the Perpendicular Impact Force (PIF) factor and the Table 1. Optimization results
Global Isotropy Index (GII). The PIF factor gives how A 1.5 cm
much exerted force is perpendicular to the finger phalanges
INPUT B 5.5 cm
whereas GII gives worst case workspace dexterity and
PARAMETERS 0.5 cm
isotropy throughout the workspace [3]. Based on these T

parameters, an Overall Impact Factor (OIF) has been HAND SIZE MEDIUM
calculated by assigning equal weights to PIF and GII. Figure L1E 6.5 cm
4 shows the OIF results as a function of L1E and L2E while OPTIMIZED
L2E 7.5 cm
Figure 5 shows the flow chart of the optimization procedure LENGTHS
used. The algorithm also incorporates the worst case L3E 1.5 cm
collision avoidance in entire workspace. Table 1 shows
optimization results based on selected input parameters A, B
and T (see Figure 3a). REFERENCES
[1] Kessler G.D, Hodges L.F., Walker N., Evaluation of the CyberGlove
as a whole hand input device, ACM Transactions on Computer-
Human Interaction 2(4), 1995, pp. 263-283.
[2] Y,Fu, P,Wang, S,Wang, Development of a multi-DOF exoskeleton
based machine for injured fingers, IEEE/RSJ International
Conference on Intelligent Robots and Systems (IROS) Sept. 2008, pp.
1946-1951.
[3] L. Stocco, S.E. Salcudean, F. Sassani, Fast constrained global
minimax optimization of robot parameters, Robotica, vol. 16, issue
6, Nov. 1998, pp. 595-605.

Figure 4. OIF v/s link lengths

 Anthropomimetic approach to the design of a prosthetic robot hand
Konstantinos Dermitzakis and Alejandro HernandezArieta
Artificial Intelligence Laboratory, University of Zurich, Switzerland
dermitza@ifi.uzh.ch
A loss of a limb is a common phenomenon
among both humans and animals. Where animals
have to cope with such a loss, humans have
pushed attempts towards limb replacements for
quite a long time. It has only been relatively
recently that such devices, in addition to
providing a cosmetic replacement can also be
sophisticated enough to account for some of the
functions of the lost limb. Upper-limb
replacements in particular are inherently difficult
to realize as the human hand is the most complex
and important tool man possesses.
The prosthetics and robotics research
fields have marked increasing progress towards
functional hand replacements. By satisfying
requirements such as (1) numerous sensory
modalities, (2) high processing capabilities, and
(3) high output torque, current robotic/prosthetic
prototypes are close to attaining functionalities
similar to those of the human hand
(Balasubramanian et al., 2008). In the prosthetics
industry however, such advancements are
usually not realized. Upper limb prostheses used
in the real world pose additional, inherent
requirements which are usually not taken into
consideration by robot designers. Such
requirements, vital to the adoption of the
prosthetic device by its user (Carrozza et al.,
2006), are: (1) reduced power consumption, (2)
low device weight, (3) high dexterity, and (4)
ease of use. In contrast, prosthetics research is
carried out using robotic hand prototypes that
only partially fulfill the above requirements.
Even so, most prosthetic and arguably robotic
hand prototypes cannot be considered optimal.
Size and weight considerations are rarely taken
into account, and power consumption issues are
usually not dealt with; when they are, usual
approaches involve either smaller or fewer
motors, disregarding any dexterity requirements.
Clearly, for upper-limb amputees to
directly benefit from prosthetics research, the
devices being used to conduct this research
should themselves adhere to these real world
requirements. Had the biomechanics of the
human hand been taken into consideration, could
such requirements be met in a better fashion than
traditional engineering approaches? We argue
that by studying the morphology of the human
hand, we can identify and exploit principles that
underlie its functions, creating a device that can
be of use to patients in real world circumstances.
Towards identifying the underlying
functional principles of the human hand, we
focus our attention on two important properties:
(1) the phalangeal curvature and (2) the friction
between tendons and their sheathing under high
loads. Even though both properties have been
well documented in the anthropological and
medical fields, to our knowledge, they have not
yet been exploited in robotics.
The human finger consists of three
segments or phalanges, the distal, middle and
proximal and a palm segment, the metacarpal.
All of the above segments are so similar in
morphology that further discussion will include
the metacarpal segment. Despite the many
evolutionary changes the human finger has
undergone, a characteristic that is common to
numerous arboreal creatures is the curvature of
the finger segments. Such phalangeal curvature
is discussed in paleontology and primate
morphology, where the association between this
curvature and arboreality has been extensively
documented (Richmond et al., 2003; Stern et al.,
1995). Arboreality itself however is not a direct
indicator of the reasons behind such a curved
finger structure despite the fact that most
arboreal creatures possess it. A recent study has
identified, using finite element methods, the
benefits of this morphology (Richmond, 2007).
The results from this study indicate that a curved
bone segment like the one present in siamang
mammals but also humans can provide
significant strain benefits over a similar but non-
curved segment. The strain present on the curved
segment is roughly half that of the straight one,
despite equivalence in lengths, areas, mechanical
properties and loading force conditions in the
two models. This difference in strains can be
attributed to the higher compression forces
present in the curved model. Such a result is
quite important in prosthetics research as it
indicates that a curved finger segment with less
volume than a straight segment will in fact
withstand the same strain with the same loading
forces present. A decrease in bone volume will
be beneficial in at least two ways. (1) Given a
constant bone density, the weight of the device
scales with its volume, satisfying the according
requirement. (2) We can identify a volume
metric demonstrating that a minimal bone
volume is preferred. The empty volume gained
can be efficiently employed by other structures
such as a cosmetic skin or a large volume of
sensory modalities presently absent from most
prosthetic hands in use whilst maintaining the
same weight as a device with a larger bone
volume.
A simple curved bone segment was
created in Solidworks. The curvature of the
bone segment is identified by a radius R of a
circle tangent to the segments joints, denoted by
two smaller circles of radii R1 and R2. To design
a complete robotic finger, the original segment
was replicated, only changing its length and the
diameters of the joints accordingly. Our initial
model shows that when fully realized, a
complete hand will provide sufficient structural
strength to be used in every day activities while
at the same time will help satisfy the weight
constraints of the device.
The presence of friction in most
mechanical devices is unwanted and usually
associated with energetic losses. Such is also the
case with all robotic and prosthetic hands to date,
where hands employing a tendon-driven
transmission mechanism aim towards frictionless
transmission, e.g. by utilizing Teflon cables and
sheaths. In the human tendon-sheath mechanisms
however, frictional forces might be beneficial. It
has been experimentally shown that during high-
load flexion of the interphalangeal joints,
eccentric and concentric forces differ by 9%, a
difference that can be directly accounted to
tendon-sheath friction (Schweizer et al., 2003).
In addition, the property of tendons to compress
when tensioned, but also the orientation of
tendon and sheath fibers further simplify the
appearance of frictional forces (Walbeehm et al.,
1995). Such utilization of friction is present in a
more specialized form in chiropterans (bats),
which employ frictional forces to dangle on their
fingers without the application of muscular force.
Preliminary results on a robotic finger indicate a
higher mechanical advantage of a frictional
(rubber sheaths) tendon transmission system
over a frictionless (Teflon sheaths) one during a
pre-caging task. Utilizing a friction mechanism
in robotic hands can have two alternative
advantages: (1) finger force output under high
loads could be increased at no motor cost; or (2)
it could be achieved using smaller motors, thus
saving both weight and energy.
Designing a human hand replacement
cannot be treated as a purely engineering
problem as is usually the case with prosthetic
research prototypes. Embodiment is well known
to lead to surprising insights (Pfeifer et al., 2007).
And there is much to be gained by studying and
exploiting the biomechanics of the human hand.
We can extract principles and concepts that can
be used towards advancing prosthetic research.
In turn, we can realize devices that will directly
benefit prosthesis users by better satisfying real
use requirements. Our research aims at paving
the way to exposing principles of the human
hand morphology, and using them directly to
implement prosthetic hands. Showcasing the
benefits of this approach will highlight the
fundamental importance of morphology not only
of the hand but of the human body as a whole.
REFERENCES
Balasubramanian R. and Matsuoka Y. (2008).
Biological stiffness control strategies for the
Anatomically Correct Testbed (ACT) Hand. IEEE
International Conference on Robotics and Automation,
737742.
Carrozza M.C., Cappiello G., Micera S., Edin B.B.,
Beccai L., Cipriani C. (2006). Design of a cybernetic
hand for perception and action. Biological
Cybernetics, 95(6):629644.
Pfeifer R., Lungarella M., Iida F. (2007). Self-
organization, embodiment, and biologically inspired
robotics. Science, 318(5853): 10881093.
Richmond, B.G. (2003). Early hominin locomotion
and the ontogeny of phalangeal curvature in primates.
American Journal of Physical Anthropology, 36
(Suppl.):178179.
Richmond B. G. (2007). Biomechanics of phalangeal
curvature. Journal of Human Evolution, 53(6):678
690.
Schweizer A., Frank O., Ochsner P. E., Jacob H. A. C.
(2003). Friction between human finger flexor tendons
and pulleys at high loads. Journal of Biomechanics,
36(1):6371.
Stern Jr., J.T., Jungers, W.L., Susman, R.L., (1995).
Quantifying phalangeal curvature: an empirical
comparison of alternative methods. American Journal
of Physical Anthropology, 97:110.
Walbeehm ET, McGrouther DA., (1995). An
anatomical study of the mechanical interactions of
flexor digitorum superficialis and profundus and the
flexor tendon sheath in zone 2. Journal of Hand
Surgery, 20(3):26980.
Learning of the anticipatory mapping between digit positions and
forces in two-digit object manipulation

Qiushi Fu1, Wei Zhang1, Marco Santello2

1
Department of Kinesiology and 2The Harrington Department of Bioengineering
Arizona State University, Tempe, Arizona

Dexterous object manipulation requires accurate distribution of digit position and

forces. However, none of the previous studies (Johansson and Flanagan, 2009; Cohen
and Rosenbaum, 2004; Friedman and Flash, 2007; Lukos et al., 2007, 2008) have looked
at how central neural system (CNS) learns to control both of these two variables. To
address this question, we asked 12 subjects to reach, grasp, and lift a symmetrical
inverted T shaped object (Fig. 1) using only thumb and index finger. An external
moment was applied by adding a 400 g mass to the left or right side at the base of the
object. The location of the added mass was changed across blocks of 10 trials. Although
subjects were given no cues about the location of the mass at the beginning of each block,
they were aware that the object center of mass would remain the same for 10 consecutive
trials. The only task requirement was to minimize object roll during lift. The vertical
position of each digit center of pressure as well as normal and tangential digit forces were
measured at lift onset (i.e., before object was lifted) to quantify anticipatory control of
these variables. Peak object roll during lift was measured to quantify trial-to-trial learning.

Figure 1: The grip device (left), sensor placement (center), and experimental setup (right)

Subjects showed significant performance improvement only between first and second
trials by significant adjustment to both digit force and digit placement (post hoc
comparison with bonferroni correction, p<0.005, Fig.2). This was achieved mainly by
estimation of external torque from sensory input and generation of a compensatory
moment at object lift onset that approximated the magnitude of external moment (~80%)
with reversed direction. The compensatory moment was analyzed across trial 4-10 for
each subject. We found that it is produced from concurrent modulation of two torque
components which were regulated by digit position and load forces. Importantly,
although significant trial-to-trial variability was found on digit placement and forces,
covariation of load torque and digit position-dependent normal torque resulted in a
relatively stable compensatory torque at object lift onset (average ratio between standard
deviation of distribution of its two components and stand deviation of total compensatory
moment and is 1.96).
These results indicate that the CNS is able to quickly generate sensorimotor memories
that map digit positions and forces to the dynamics of the object. Nevertheless, trial-to-
trial variability associated with digit placement is accompanied by concurrent force
modulation and therefore does not affect the compensatory torque necessary for accurate
object manipulation. We propose that this is accomplished by using sensory information
about digit position shortly after contact to update digit forces such that the required
compensatory torque can be generated.

Figure 2: Object peak roll (A), relative placement of center of pressure (B), load force sharing (C),
and grip force (D).

References:
Cohen RG, Rosenbaum DA (2004) Where grasps are made reveals how grasps are
planned: generation and recall of motor plans. Exp Brain Res 157:486495.
Friedman J, Flash T (2007) Task-dependent selection of grasp kinematics and stiffness in
human object manipulation. Cortex 43:444-460.
Johansson RS, Flanagan JR (2009) Coding and use of tactile signals from the fingertips
in object manipulation tasks. Nature Rev Neurosci 10:345:359.
Lukos J, Ansuini C, Santello M (2007) Choice of contact points during multidigit
grasping: effect of predictability of object center of mass location. J Neurosci
27:3894-3903.
Lukos J, Ansuini C, Santello M (2008) Anticipatory control of grasping: independence of
sensorimotor memories for kinematics and kinetics. J Neurosci 28:12765-12774.
 ANALYSIS OF FINGER POSITION DURING TWO AND THREE-FINGERED
GRASP: POSSIBLE IMPLICATIONS FOR TERMINAL DEVICE DESIGN

Murray E. Maitland and Molly Epstein, University of Washington, Department of

Rehabilitation Medicine, Seattle, Washington, USA

INTRODUCTION: The human hand is a complex mechanical structure that functions

during communication, sensory experiences, manipulation and grasping. To develop an
effective and accessible grasping mechanism there should be a clear understanding of
normal hand movements as they relate directly to grasping. Analysis of human hand
mechanics during grasp has been used to justify the design characteristics of terminal
devices. Conceptual frameworks developed from these studies can inform the development
of artificial mechanisms through design specifications.
Recent studies indicate that only a small part of the hands potential motion is used
during grasp. Mason et al. 4 concluded that much of reach-to-grasp uses a base posture with
small refinements in finger and thumb positions. In another study, Baud-Bovy &
Soechting 1 found that when two fingers oppose the thumb, there is a predictable result of
finger position that produced a balanced lever. Kamper et al 2 found that the thumb used
less than 5% of the available range during a variety of grasping tasks. The findings of these
studies should be expanded to get a broader perspective finger position during grasp.
The current study examined finger positions during two- and three-fingered grasp of
a wide range of geometrical objects. The results from people with normal hands were
qualitatively compared to two terminal devices.

METHODS: Seven subjects participated in the study (2 male, 5 female). A 6-camera video
motion analysis system (Qualisys, Gothenburg, Sweden) was used to collect marker
locations in a calibrated volume approximately 1 m3. The hand biomechanical model was
modified from previous studies.3 Fourteen reflective markers were taped to the
interphalangeal joints, metacarpophalangeal joints and the carpals of the subjects. Seated
subjects were asked to grasp and lift an object, using their preferred two fingers or three
fingers. The 9 objects were chosen based on standard geometric shapes and sizes. The
objects included discs, rods, cubes, rectangles and spheres. The smallest object was a 1.9
cm diameter sphere and the largest object was a 9.5 X 12.5 X 1.7 cm rectangular prism.
Data were collected at 120 frames/s and the markers 3-dimensional coordinates were
exported for further analysis using mathematical software. The three dimensional joint
angles of each of the finger articulations were calculated.
Pearson product moment correlation coefficients (r) were used to determine the
degree of association between distal to proximal joints as well as parallel joints between the
fingers.

RESULTS: For two and three-fingered grasp using this specific set of objects, the thumb
interphalangeal joint and metacarpal phalangeal joint summed ranges of motion was from
20o to 43o for 80% of grasps. In other words, the range of motion requirement for 80% of
grasps in this sample was 23o to position the contact surface of the thumb. The largest
variability of any of the finger joints was the index MCP joint (a range of -44o to 49o).
There appeared to be a predisposition for the distal segments of the thumb and index finger
to be angled rather than being parallel to each other at the point of contact with the object.
The average angle between the index and thumb distal segments was about 30o.
Distal to proximal articulations in separate fingers (thumb, index and long fingers)
had relatively low correlations (r= -0.20 to 0.40). Thumb total range of motion had low
correlation to index and long finger motion. Higher correlations were found with index
finger and long finger parallel joints (r= 0.41 to 0.65) and total ranges of motion (r= 0.74).

DISCUSSION: The available sagittal plane range of motion of index and long finger distal
phalanges relative to the metacarpals is approximately 320 degrees and the distal phalanx
of the thumb can move 180 degrees relative to the 1st metacarpal, but grasping patterns use
only a small proportion of the total range. To develop an effective grasping terminal
device, it might be possible to reproduce the necessary range of motion more easily than
the entire potential range of motion of the hand. For example, data from the current study
suggests that for these objects the range of motion of a terminal device contact surface
should be about 90 degrees.
The current study identified greater correlations between parallel joints compared to
proximal to distal joints correlations. This finding seems to suggest that degrees of freedom
in the hand could be reduced in parallel fingers rather than by controlling proximal to distal
joints.
While it is possible that only the biomimetic approach can accomplish all possible
grips, other approaches with simpler controls or mechanisms have not been explored
thoroughly. In creating an artificial hand for grasping, the optimal configuration may differ
from the anatomical configuration. Alternative articular designs may lead to contact
geometry that is similar to the normal hand but have desirable features in some other way.

CONCLUSIONS: Thumb, index and long finger positions during two and three finger
grasp of standard geometric objects was relatively consistent for this sample of subjects.
They utilized less than half of the available ranges of motion. Index and long finger
metacarpophalangeal joints used the largest ranges of motion. Design criteria for grasping
devices could include these findings depending on the goals for the device. Improved
performance, robustness, cost, and ease of manufacturing might be goals of non-
biomimetic terminal device designs

Reference List

1. Baud-Bovy G, Soechting JF: Two virtual fingers in the control of the tripod grasp . Journal of
Neurophysiology 2001;86:604-615.

2. Kamper DG, Cruz EG, Siegel MP: Stereotypical fingertip trajectories during grasp. Journal of
Neurophysiology 2003;90:3702-3710.

3. Lee S-W, Zhang X: Development and evaluation of an optimization-based model for power-grip posture
prediction. Journal of Biomechanics 2005;38:1591-1597.

4. Mason CR, Gomez JE, Ebner TJ: Hand synergies during reach-to-grasp. Journal of Neurophysiology
2001;86:2896-2910.
Does practice affect hand ability? The reliance on frames of reference

Miriam Ittyerah
Department of Psychology
Christ University
Bangalore 560029, India

Early attempts to explore the effects of practice in a certain hand action have emerged
from relating hand preference to a variety of tasks (Fleishman 1958; Fleishman &
Ellison 1962). Since the degree and direction of hand preference does not differ in the
blind and sighted children, it may be inferred that vision does not determine hand
preference (Ittyerah 1993, 2000; McManus et al 1988). The question of interest in this
investigation is to know the role of vision in the effect of practice with the preferred
and non preferred hands over a period of development, both in blind and sighted
blindfolded conditions. Further, whether practice will show gains only from the
preferred hand remains to be established. In tasks such as type writing, piano playing
or braille reading (Millar 1987) the hands perform as well as each other. These are
however non prehensile movements where the object is manipulated by the hand or the
fingers and not grasped in the hand. In tasks using prehensile movements when the
object is partly or wholly held, the evidence is sparse.

In the present study the hypothesis that blind children will improve in ability with
practice in spatial tasks was tested in a group of 90 congenitally blind and blindfolded
sighted children between the ages of 5 and 15 years. All the children were tested for
hand preference. The hand preference test comprised of 10 items selected from
unimanual and bimanual tasks. The selected tasks assessed the common repertoire of
hand actions in daily life. Four tasks were selected to assess hand ability (sorting,
stacking, finger dexterity, Minnesota rate of manipulation). The tests required an
object to be held partly or wholly within the hand. The instructions required the child
to perform each task separately with the left and right hands and the performance times
were recorded in seconds. This was followed by a practice period of four months for
each group during which time each child was required to practice each task three times
with each hand in front of the experimenter. The children were post tested and the
performance times of each task and hand were recorded in seconds. The blindfolded
sighted children practiced the tasks with their blindfolds. The order of the two groups
of children tested was the same in pretest as in the posttest.

The results of the handedness task showed that most children had a right hand
preference; blind (mean=97.8) and blindfolded sighted groups (mean = 97.8) (F (1,
144) = .0001; p> 0.05), and lateralization increased with age in development (mean
age groups 5-7 = 93.72; 9-11 = 99.6; 13-15 = 100), (F (8, 144) = 4.338; p< .001)
(Ittyerah, 1993, McManus et al 1988). Analysis of variance computed for each task
separately indicated a percentage gain with practice during development for the left
and the right hands of the blind children for the sort task (F (1, 162)= 307.9; p<0.001),
the stack task (F (1, 162) = 48.8; p< 0.001), the finger dexterity task (F (1, 162)=
1400.2; p< .001) and the Minnesota rate of manipulation task (F (1, 162) = 44.7; p<
.001). The left and right hands of both groups of children did not differ in percentage
gain, indicating little or no relationship between hand preference and hand ability.
Thus general laterality does not affect ability (Ittyerah1993, 2000). Percentage gains
were more for the blind than the blindfolded sighted children, indicating that practice

1
improves skill in the absence of vision (Millar & Ittyerah, 1991). The blind folded
sighted had more losses than gains in the post- test, probably because the blindfolds
hampered performance, indicating their dependence on vision in interacting with the
environment.
The finding that the congenitally blind children gained from practice indicates
that performance is not solely confined to benefits from vision. The use of self referent
coding is evident in the blind and to a lesser extent in the sighted blindfolded children.
Self referent coding refers to the use of the body midline as an anchor from which
estimates of distances to the left or the right side of the body can be inferred. Body
centred reference cues are more reliable than information from external cues in blind
conditions. In principle geometric rules can be applied to body centred reference axes
just as they can be applied to external reference cues in sighted conditions. It also
indicates that the losses of the blindfolded sighted children may have been caused by
the difficulty of having to rely on body centred self referent codes that are usually
ignored in their sighted existence (Millar 1981). This indicates that long term haptic
experience tends to elicit more self referent coding than prior visual experience.
Evidence (Millar and Al-Attar, 2005) indicates that vision improves performance in a
haptic spatial task only in so far as it adds cues that are potentially relevant to spatial
discrimination and reference. Millar and Al-Attar (2003) have for example
demonstrated that the left and right hands do not differ in a spatial processing task that
assessed distance. Thus comparisons between the hands in the processing of sensory
inputs (Millar and Al-Attar 2003) must be considered separately from comparisons
between the hands for spatial versus sequential processes (Langdon and Warrington
2000). Therefore processing of sensory inputs from touch and movement is not
expected to differ between the hands. Thus hand actions require a frame of reference to
perform effectively and are independent of vision.

Key words: practice, spatial, proficiency, self reference, blind

References
Fleischman, E. A. (1958). Dimensional analysis of movement reactions. Journal of
Experimental Psychology, 55,438-453.
Fleischman, E. A. & Ellison, G.D. (1962). A factor analysis of fine manipulative tests
Journal of Applied Psychology, 46, 96-105.
Ittyerah, M. (2000). Hand skill and hand preference in blind and sighted children.
Laterality, 5, (3), 221-235.
Ittyerah, M. (1993) Hand preferences and hand ability in congenitally blind
children. Quarterly Journal of Experimental Psychology, 46B, 35-50.
Langdon, D. & Warrinton, E.K. (2000). The role of the left hemisphere in verbal and
spatial reasoning tasks. Cortex, 36, 691-702.
McManus, I. C., Suk, G., Cole, D. R. Mellon, A. F., Wong, J.& Kloss, J.(1988). The
development of handedness in children. British Journal of Developmental
Psychology, 6, 257-272.
Millar, S. & Al-Attar (2005). What aspects of vision facilitate haptic processing?
Brain and Cognition, 59, 258-268.
Millar, S. & Al-Attar (2003). Spatial reference and scanning with the left and right
hand. Perception, 32, 1499-1511.
Millar, S. & Ittyerah, M. (1991) Movement imagery in young and congenitally blind
children. Mental practice without visuospatial information. International
Journal of Behavioural Development, 15,135-146.

2
 Video survey of pre-grasp interactions in natural hand activities
Lillian Y. Chang and Nancy S. Pollard
Carnegie Mellon University, Pittsburgh PA 15213

The first step in many manipulation actions is object acquisition, where the hand is formed
around the object to achieve a task-specific grasp. A large body of the previous research has
investigated object acquisition in the context of reach-to-grasp actions. The primary focus has
been on the motor coordination of the upper limb and hand shape in the process of reaching
toward the object which is grasped. In many of the reach-to-grasp actions studied previously, the
target of the coordinated reach and preshape motion is a presented object whose placement is
considered fixed in the environment.
In several natural task settings, however, the object is movable in the environment, and the
task does not require the object to be grasped exactly from its presented placement. The object
interaction during acquisition may be more complex, such that the object is moved prior to the
complete formation of the desired grasp. This pre-grasp interaction may serve to adjust the object
configuration to improve the goal grasp. For example, a person may use non-prehensile contact to
slide and re-orient a mug on a table before grasping it by the handle. When grasping a pen off of
a table, the fingers may quickly pivot the pen to orient the tip for the subsequent writing task.
Our previous work [Chang et al., 2008, 2009] has investigated preparatory object rotation as
a specific example of pre-grasp interaction. In preparatory object rotation, the object is pivoted
in the plane of the support surface in order to re-orient the object handle prior to grasping, as in
the mug example above. Instead of completely re-planning a new direct reach-to-grasp action for
novel object orientations, a preparatory rotation strategy first adjusts the object orientation with
pre-grasp interaction and then completes the hand formation of the final desired object grasp.
The present work surveys the broader class of pre-grasp interaction strategies beyond the specific
example of preparatory rotation. Our goal was to develop a taxonomy for classifying the variety
of pre-grasp action primitives which are integrated into complex reach-to-grasp tasks. We were
specifically interested in surveying human hand activity in natural settings in contrast to instructed
tasks within a laboratory environment. In this way, we could capture the richness of pre-grasp
interactions beyond the direct reach-to-grasp actions studied previously in the literature.
In the video survey of human hand activity, we filmed people performing manipulation tasks in
natural settings such as the home or place of occupation. All participants provided informed con-
sent. In all observations, the participants performed manipulation skills which had been practiced
previously as part of their regular occupation. There were a total of 10 sessions of both individual
and group manipulation activities, such that overall 38 people were filmed. The sessions covered
activities for housekeeping, food preparation, office work, and mechanical repair. Specific tasks
include sorting office supplies, washing dishes, and moving furniture.
We found that there is indeed a broad class of pre-grasp interactions where the object is not
grasped directly from its presented placement in the environment. Our framework describes the
survey examples according to two main aspects of the pre-grasp interaction. The first aspect is the
type of object re-configuration resulting from the interaction. The second aspect is the underlying
intent of the interaction to improve the posture quality or grasp quality of the manipulation action.
First, the object reconfiguration is classified by a taxonomy based on the degrees of freedom
which were adjusted by the pre-grasp interaction (Fig. 1). The object motion may be completely
comprised by planar displacement. This is common in examples of non-prehensile pre-grasp inter-
action where the object is primarily supported on a horizontal surface. Alternatively, for a bulky

RSS 2009 Workshop Understanding the Human Hand for Advancing Robotic Manipulation
 Object adjustment

Rigid transform Non-rigid reconfiguration

1-DoF 3-DoF planar 6-DoF Continuous Set

Articulation
rotation displacement tumbling deformation arrangement

Figure 1: Taxonomy for the object reconfiguration aspect of pre-grasp interaction. Examples of rigid planar
transformations included rotation of a cup by its handle and sliding books off the top of a stack. General
rigid tumbling was used to achieve a whole body grasp of a bulky piece of furniture. Pre-grasp interaction
was also observed for non-rigid objects. A hinged bucket handle was rotated to achieve a cylinder grasp, and
a piece of paper was curled to achieve a pinch grasp. Multiple objects were also rearranged as a set, such as
in the scooping interaction with a pile of peelings.

piece of furniture, the pre-grasp tumbling interaction may result in general 6-degree-of-freedom
rigid displacement. In more complex cases, the pre-grasp interaction may cause a morphological
reconfiguration of a deformable or articulated object, such as a bucket with a hinged handle.
Second, the pre-grasp interaction is described by the intent of the object adjustment in relation
to the final grasp. The presented configuration of the object in the environment could be suboptimal
for direct reach-to-grasp object acquisition due to preferences for a particular body posture and/or
grasp. When the handle on a cooking pan is oriented away from the person, a direct grasp of the
handle may be feasible but could require lifting the heavy pan from an uncomfortable body posture
with limited lifting capability. In other scenarios, the intent of the pre-grasp interaction may be to
improve the grasp quality rather the posture quality. This is especially relevant to situations where
environmental clutter occludes the desired grasp contact surfaces, as in the case of a shelved book
where only the spine is exposed in the initial task condition. The observed examples suggest that
the intent of pre-grasp interaction was often a combination of preferences for both posture quality
and grasp quality, and potentially other optimization metrics.
The presented pre-grasp interaction framework suggests several approaches for improving the
dexterity of robotic manipulators. Taking advantage of object movability may extend the effective
workspace by changing the environmental constraints when direct reach-to-grasp actions are of
insufficient posture and/or grasp quality. Non-prehensile pre-grasp interaction could reduce the
load on the manipulator by using shared support with the work surface during the initial interaction
with the object. Moreover, the expense of tuning control parameters for complex manipulators can
be reduced if pre-grasp object reconfiguration enables the reuse of a single well-tuned grasp action
for multiple initial placements. Finally, because pre-grasp strategies are part of natural human
manipulation, incorporating them in the repertoire of assistive or teleoperated manipulators could
facilitate more intuitive control for human operators.

Chang, L. Y., Klatzky, R. L., and Pollard, N. S., 2009. Selection criteria for preparatory object rotation in manual
lifting actions. Journal of Motor Behavior. In review.

Chang, L. Y., Zeglin, G. J., and Pollard, N. S., 2008. Preparatory object rotation as a human-inspired grasping
strategy. In IEEE-RAS International Conference on Humanoid Robots (Humanoids 2008). pages 527534.

RSS 2009 Workshop Understanding the Human Hand for Advancing Robotic Manipulation

Haptic Exploration of Spheres
Blake Hannaford and Jesse Dosher
Department of Electrical Engineering
The University of Washington
I. I NTRODUCTION
Haptic exploration is a complex sensory/motor ac-
tivity involving a high number of degrees of freedom,
multiple sensing modalities, and integration of sensory
information over space and time. Lederman and Klatzky
created a taxonomy of Exploratory Procedures which
are stereotyped hand and finger movements which people
preferentially use to make specific sensory discrim-
inations and object identifications. Their experiments
were mostly aimed at either 1) identification of objects
through haptic interaction and 2) assessment of a prop-
erty such as temperature or roughness.
Psychophysical literature relevant to the present study
has been reviewed in detail in [1]. A review from the
point of view of haptic interface design is available in
[2].
The work reported here studies haptic perception of
sphericity. If an object is placed in the fingers, how well
can humans determine whether or not it is a sphere?
Starting from nearly perfect spheres, such as for exam-
ple, precision industrial balls, a variety of increasingly
subtle distortions can be made.
If a subject is to detect deviations from perfect spheric-
ity, they may use a variety of queues, depending on
how the reference spheres are distorted. These cues may
include
Eccentricity (non-constant diameter)
Texture (local variations in radius)
Changes in curvature (the curvature is constant at
all points on a perfect sphere)
To define these experiments more precisely, we wish
to distort some test spheres in precisely defined ways,
ideally to determine a sensory threshold for a specific
attribute. These cues are not variable independently.
Since a sphere is defined as a surface of constant local
curvature, any amount of eccentricity would also change
local curvature. Texture also changes curvature. We will
define texture to mean radius and curvature changes on
a small scale compared to the spheres nominal radius
which average out to zero within each small region.
II. M ETHODS
A set of seven distorted spheres (test objects) were
produced from highly polished industrial ball bearings
0.4996 inches (12.69mm) in diameter. These balls appear
The authors would like to acknowledge National Science Foundation
grant IIS-0303750
1
quite perfect with a highly polished mirror finish. Seven
additional balls were set aside as controls.
The seven test spheres were ground down on two
opposing sides to achieve a range of diameters along
a single axis of the sphere. Initially, a flat face was
ground on each side. Then the ball was ground by hand
to minimize the curvature over the area between the
center of the flat and a circle 45 below the center of
the flat. A description of the grinding procedure is given
in [3]. After grinding, the balls were polished to achieve
uniform texture around the balls[4].
An equation describing the desired ball contour is
derived as follows. We will parameterize the smoothing
zone with an angle from the pole, . The boundary of
the smoothing zone will then be all the points North or
South of 90 or the intersection of the sphere with
a cone of apex angle 2.
Considering a section through the sphere, a function
which describes this is a polynomial which intersects the
sphere at a point from the pole, and which is tangent
to the sphere at these points, and further passes through
the point x = 0, y = r d where d is the amount by
which the sphere is ground down. This must be an even
function and a suitable candidate is
y = ax4 + bx2 + c (1)
The parameters a, b, c were derived as functions of , r,
and d. However, our hand fabrication methods did not
guarantee that this actual profile could be achieved for a
given diameter.
Brand new 0.4996 balls weighed an average of 8.17g.
After processing, seven control balls weighed an average
of 8.10 grams. All of the test balls weighed either 8.0
or 8.1 grams (0.1g scale resolution) (see Table I).
III. R ESULTS
A. Experimental Procedures
The protocol below was approved by the University
of Washington Human Subject Committee.
6 subjects between the ages of 22 and 53 years were
recruited with 67% male. Each subject read a printed
set of instructions. A plastic box with 7 numbered
bins contained pairs of balls, one test and one control.
Ball diameters are listed in Table I. The subjects were
instructed to take the pair from bin 7 first (containing
the most eccentric ball), and work through the bins
in descending order. They repeated the sequence of
seven bins 10 times for a total of 70 discriminations.
The subjects were instructed to randomize each pair
 2

Ball # Diameter (inches) Weight (g) Serial #

1 0.4980 8.1 155 exists for 0.4770 D .4880. Diameters above that
2 0.4925 8.1 498 range were reported at chance levels by most of the sub-
3 0.4880 8.0 532 jects and below that range were almost always detected
4 0.4835 8.0 784
5 0.4770 8.0 733 by all subjects. In terms of just noticeable difference
6 0.4725 8.0 552 (JND) this is equivalent to 2.3% JND 4.7% where
7 0.4695 8.0 814 JND is the ratio of the distorted diameter to the nominal
TABLE I diameter. These figures are rather low compared to val-
C HARACTERISTICS OF MODIFIED BALLS . C ONTROL BALLS ues of 6-9% obtained in force discrimination experiments
WEIGHTED 8.10 GRAMS . by others[5], [6].
The definition of sensory threshold is somewhat arbi-
trary. It must be greater than 50% to be above chance,
and it must be less than 100% to account for human
Average Correct Vs. Ball Number fallibility. A commonly accepted method, the adaptive
1-up 2-down method converges to a threshold estimate
1.0

at which the subject responds correctly about 71% of

the time[7]. Using this value, our measurement suggests
0.9

a threshold for perception of non-sphericity between

0.4835 D .4880
This preliminary protocol exposed some limitations of
0.8

our experimental method. In the preliminary experiment,

% Correct

two subjects were able to score 100% or near 100%

0.7

on every ball, suggesting that even though the least

modified ball is within 0.4% of the control diameter, the
0.6

population threshold might be near that value. Another

explanation was possible sources of experimenter or
0.5

subject bias. For example, the data sheet used by the

subjects identified the diameter of the test balls and also
0.4

which balls were test/control.

1 2 3 4 5 6
Ball Number
Fig. 1. Results of preliminary experiment. Distribution of correct
discrimination rate vs. test ball.
of balls by shaking in their cupped hands, and then,
without looking at the balls, select the distorted ball.
After making the selection, subjects read the engraved
number and placed a checkmark on a data form by that
serial number in the current trial column. In this manner,
subjects received knowledge of the correctness of their
results. Learning effects were reduced by ensuring that
the subjects completed all balls first before moving to
the next trial.
a) Preliminary Results: A boxplot of the prelim-
inary experiment data (Figure 1) showed a clear trend
towards lower recognition rates as the distorted ball
diameter approached 0.499 inches (the control size).
The normality of the data was evaluated with the R
statistical software package using the qnorm() com-
mand and found to be normal except for a saturation
near 100% recognition. One-sided Analysis of Variance
was computed and the effect of ball number was highly
significant at the 0.05 level.
IV. D ISCUSSION
This preliminary experiment showed that when at-
tempting to judge sphericity of eccentric spheres of
0.4996 inches nominal diameter, a threshold probably
7
Another potential limitation of the experiment is the
limited degree of geometric control of the test spheres
which could be achieve by our hand methods. Although
we attempted to reduce the peak curvature of the modi-
fied balls as much as possible, this might not have been
achieved. Local curvature is likely to be a significant cue
since it can produce local skin deformation which would
strongly stimulate receptors.
Future work includes better characterization and con-
trol of the test objects, a more double blinded and ran-
domized protocol for the experiment, and video taping
and analysis of the types of finger movements used by
the subjects to make the discrimination.
R EFERENCES
[1] S.C. Venema. Experiments in surface perception using a haptic
display. Ph.D. Thesis, April 1999.
[2] KB Shimoga. A survey of perceptual feedback issues in dexterous
telemanipulation. i. finger force feedback. Proc. Virtual Reality
Annual International Symposium, pages 263270, Sept 1993.
[3] Blake Hannaford. Making distorted spheres for psychophysical
experimentation. University of Washington Electrical Engineering
Technical Report, 2009-0001, Rev April- 2009.
[4] Blake Hannaford and Jesse Dosher. Polishing and measuring
distorted spheres for psychophysical experimentation. University
of Washington Electrical Engineering Technical Report, 2009-
0004, April 24 2009.
[5] L. A. Jones. Perception and control of finger forces. In Proc.
ASME Dynamic Systems and Control Division, pages 133137,
1998.
[6] S. Allin, Y. Matsuoka, and R. Klatzky. Measuring just noticeable
differences for haptic force feedback: Implications for rehabilita-
tion. In Proc. 10th Symposium on Haptic Interfaces for Virtual
Environments and Teleoperator Systems, pages 299302, 2002.
[7] J. C. Stevens, E. Foulke, and M. Q. Patterson. Tactile acuity,
aging, and braille readings in long-term blindness. Journal of
Experimental Psychology: Applied, 2(2):91106, 1996.
 A comprehensive grasp taxonomy
Thomas Feix and Roland Pawlik Heinz-Bodo Schmiedmayer
Otto Bock Healtcare Gmbh Vienna University of Technology
1070 Vienna, Austria Inst. for Mechanics and Mechatronics
Email: thomas.feix@ottobock.com 1040 Vienna, Austria
roland.pawlik@ottobock.com Email: hschmied@mail.tuwien.ac.at
Abstract The goal of this work is to overview and summarize
the grasping taxonomies reported in the literature. Our long term
goal is to understand how to reduce mechanical complexity of
anthropomorphic hands and still preserve their dexterity. On the
basis of a literature survey, 33 different grasp types are taken
into account. They were then arranged in a hierarchical manner,
resulting in 17 grasp types.
I. I NTRODUCTION
The design of an anthropomorphic hand is always a compro-
mise between hand complexity and the tasks it is supposed to
accomplish. In general, sophisticated hands with many degrees
of freedom are dexterous but pose significant requirements in
terms of control. Many of the reported taxonomies have been
made with the goal of understanding what types of grasps
humans commonly use in everyday tasks and use this as an
inspiration for designing robotic and prosthetic hands. The
goal of our research is in the same direction: understanding
how to minimize the complexity and maximize the dexterity
of a mechanical hand.
Since there is little consensus in the existing literature on
the grasp types humans use, the first step was to review the
existing literature. The goal was to find the maximal number
of grasp types, which will act as basis for further research.
II. M ETHOD
A. Definition of a grasp
Since grasping in humans is a very broad area, it was
necessary to find a definition of a grasp relevant to our work.
We thus propose the following:
A grasp is every static hand posture with which an
object can be held securely with one hand.
The definition also implies that the grasp stability has to be
guaranteed irrespective of the relative force direction between
hand and object.
Therefore, intrinsic movements are excluded because the
object is not in a constant relationship to the hand. Bimanual
tasks are not relevant because they use both hands. Gravity
dependent grasps are ruled out, because the hand orientation is
vital to the grasp stability. If one turns the hand, the object may
fall down, which shows that it is not independent of the force
direction. Thus grasps being excluded are the Hook Grasp and
the Flat Hand Grasp.
Javier Romero and Danica Kragic
Comp. Vision and Active Perception Lab
Centre for Autonomous Systems
School of Comp. Science and Communication
KTH, SE-100 44 Stockholm, Sweden
Email: jrgn,dani@kth.se
B. Comparison of Taxonomies
To develop the comprehensive taxonomy, several literature
sources were compared. They range from the field of robotics,
developmental medicine, occupational therapy to biomechan-
ics [1],[2],[3],[4],[5],[6],[7],[8],[9],[10],[11],[12],[13],[14].
An excerpt of the comparison table is shown in Fig. 1.
Columns store equal grasps, whereas rows store all grasps
defined by an author. Grasps that are defined by the author
as power, precision or intermediate, are marked with a color
code. Yellow is denoting a power grasp, green a precision
grasp and yellow/green an intermediate grasp as defined in
[15],[16],[17]. Red is marking grasps that are not conforming
to our definition of a grasp.
Fig. 1. The sheet used for comparison of different grasp taxonomies. This is
just a small excerpt of the whole sheet, the complete table can be downloaded
via the Human Grasping Database, [18].
III. R ESULTS
A. The Taxonomy
In total, we have found 147 grasp examples in the consid-
ered literature sources. Out of those 147 examples, we have
detected only 45 different grasp types. A further classification
based on our grasp definition has revealed only 33 valid grasp
types.
The grasps were then arranged in a taxonomy depicted
in Fig. 2. The classification in the columns is done by the
power/precision requirements. The next finer differentiation is
done, depending on whether the opposition type is Palm, Pad
or Side Opposition. The opposition type is also defining the

Fig. 2. Comprehensive Grasp Taxonomy which includes 33 grasp types.
VF 1: In the case of Palm Opposition the palm is mapped into
VF 1, in Pad and Side Opposition the thumb is VF 1. The
only exception to this rule is the Adduction Grasp, where
the thumb is not in contact with the object.
To differentiate between the two rows, the position of the
thumb is used: the thumb CMC joint can be in a either
adducted or abducted position. This is a new feature introduced
in our taxonomy.
B. Merging of grasps within one cell
Since many grasps have similar properties (opposition type,
thumb position etc.), some cells are populated with more
than one grasp. In the literature examples, the sole difference
between such grasps is the shape of the object the hand is
manipulating. This offers the possibility to reduce the set of
all 33 grasps down to 17 grasps by a merge of the grasps within
one cell to a corresponding standard grasp. Depending on
the task, this offers the possibility to choose two different
levels of accuracy of the grasp classification.
As a comparison, the classification of Cutkosky [1] has 15
different grasp types that fit into our definition of a grasp. This
is very close to the amount of grasps the reduced taxonomy
has. However, our comparison shows that even though the
number of grasps is nearly the same, the classification is very
different. When one classifies the grasps from [1] according
to our scheme, the grasps only populate 7 cells, which is a
reduction by more than half. This is only natural, since [1]
mainly differs grasps by the object properties which in our
case is done within one cell.
IV. C ONCLUSION AND F UTURE W ORK
A comprehensive human grasp taxonomy, on the basis
of a comparative literature research, was developed. A total
of 33 different grasp types were identified and arranged in
a taxonomy. The position of the thumb was introduced as
additional attribute, which can be either abducted or adducted.
Depending on the need for precision, the taxonomy offers a
second level of classification which includes only 17 gross
grasp types.
Compared to the existing taxonomies the present one is
more sophisticated, since it incorporates a larger number of
grasp types than the reviewed literature sources. This should
allow a better description of the human grasping capabilities
and therefore will be an excellent starting point for further
research. As next step the grasp types will be modeled in a 20
DoF hand model and the corresponding joint angles will be
determined. This will then act as a basis for an analysis on how
the complex hand model can be simplified, but still preserve a
lot of dexterity. This will be defined via the grasp types it can
to accomplish. Finally this will lead to an anthropomorphic
hand setup which is dexterous and simple in design.
ACKNOWLEDGMENT
This research is supported by the EC project GRASP,
IST-FP7-IP-215821.
R EFERENCES
[1] M. R. Cutkosky, On grasp choice, grasp models, and the design
of hands for manufacturing tasks, Robotics and Automation, IEEE
Transactions on, vol. 5, no. 3, pp. 269279, 1989.
[2] S. B. Kang and K. Ikeuchi, Grasp recognition using the contact web,
in Intelligent Robots and Systems, 1992., Proceedings of the 1992
lEEE/RSJ International Conference on, vol. 1, 1992, pp. 194201.
[3] D. Lyons, A simple set of grasps for a dextrous hand, in Robotics
and Automation. Proceedings. 1985 IEEE International Conference on,
vol. 2, 1985, pp. 588593.
[4] J. M. Elliott and K. J. Connolly, A classification of manipulative hand
movements. Developmental medicine and child neurology, vol. 26,
no. 3, pp. 283296, June 1984.
[5] S. J. Edwards, D. J. Buckland, and J. D. McCoy-Powlen, Developmental
and Functional Hand Grasps, 1st ed. Slack Incorporated, October 2002.
[6] C. E. Exner, Development of hand skills, in Occupational therapy for
children, C. J. Smith, Ed., 2001, pp. 289328.
[7] N. Kamakura, M. Matsuo, H. Ishii, F. Mitsuboshi, and Y. Miura,
Patterns of static prehension in normal hands. The American journal
of occupational therapy. : official publication of the American Occupa-
tional Therapy Association, vol. 34, no. 7, pp. 437445, July 1980.
[8] I. A. Kapandji, Funktionelle Anatomie der Gelenke. Schematisierte und
kommentierte Zeichnungen zur menschlichen Biomechanik. Thieme
Georg Verlag, January 2006.
[9] G. Lister, The hand: Diagnosis and surgical indications. Churchill
Livingstone, 1984.
[10] C. L. Taylor and R. J. Schwarz, The anatomy and mechanics of the
human hand. Artificial limbs, vol. 2, no. 2, pp. 2235, May 1955.
[11] W. P. Cooney and E. Y. Chao, Biomechanical analysis of static forces
in the thumb during hand function, J Bone Joint Surg Am, vol. 59,
no. 1, pp. 2736, January 1977.
[12] D. B. Slocum and D. R. Pratt, Disability evaluation of the hand, J.
Bone Joint Surg, vol. 28, pp. 491495, 1946.
[13] K. H. Kroemer, Coupling the hand with the handle: an improved
notation of touch, grip, and grasp. Human factors, vol. 28, no. 3, pp.
337339, June 1986.
[14] C. M. Light, P. H. Chappell, and P. J. Kyberd, Establishing a standard-
ized clinical assessment tool of pathologic and prosthetic hand function:
normative data, reliability, and validity. Archives of physical medicine
and rehabilitation, vol. 83, no. 6, pp. 776783, June 2002.
[15] J. R. Napier, The prehensile movements of the human hand. The
Journal of bone and joint surgery. British volume, vol. 38-B, no. 4,
pp. 902913, November 1956.
[16] T. Iberall, G. Bingham, and M. A. Arbib, Opposition space as a structur-
ing concept for the analysis of skilled hand movements, Experimental
Brain Research Series, vol. 15, pp. 158173, 1986.
[17] T. Iberall, Human prehension and dexterous robot hands, The Interna-
tional Journal of Robotics Research, vol. 16, no. 3, pp. 285299, June
1997.
[18] Human grasping database, http://web.student.tuwien.ac.at/
e0227312/, April 2009.
 Quantification and Reproduction of Human Hand Skin Stretch
and its Effects on Proprioception
1* 1* 2 3
Elliot Greenwald , Jeffrey Pompe , Steve Hsiao , and Allison Okamura
1
Dept. of Biomedical Engineering, 2Dept. of Neuroscience, 3Dept. of Mechanical Engineering
The Johns Hopkins University, Baltimore, MD USA
*Both authors contributed equally to this work

ABSTRACT

Local stretch of the skin is thought to contribute to the sense of

body segment position and movement, also known as
proprioception. A Vicon motion capture system and markers
placed on the skin were used to quantify skin stretch during
different grasping movements. The skin between the index finger
and the thumb stretches by about 25% during a whole hand grasp,
and the skin behind the knuckle of the index finger stretches by
about 30% for a pinching movement. A linear actuator was used
to artificially stretch the skin, reproducing the natural skin stretch
and altering the sense of finger position. This preliminary study
motivates the development of artificial proprioception displays for
prosthetics and haptic interface.
1. INTRODUCTION
The mechanism by which the brain determines the relative
positions of ones body parts is still unknown [1-2]. Cutaneous
receptors that sense skin stretch may provide the brain with a code
that gives rise to proprioception [3-7]. Experiments have been
performed in which the experimenter produces illusions of
movement in the subject by manipulating skin stretch around
joints [3]. However, to our knowledge, no one has measured the
degree to which the skin is stretched during movement, using a
motion capture system.
2. METHODS
To determine the extent that skin stretch plays a role in grasp
proprioception we designed a series of experiments to quantify the
amount of stretch that naturally occurs during different grasps and
pinch movements. We then built a device to recreate the relative
skin stretch that was observed. The device artificially stretched the
skin during a passive proprioceptive task. The aim was to
determine whether the artificial stretch affected the perception of
hand conformation.
2.1 Quantification of Skin Stretch
Skin stretch was quantified by using a Vicon motion capture
camera system (http://www.vicon.com/). The system consisted of
nine cameras outfitted with IR optical filters and an array of IR
LEDs, and a set of reflective dots. The dots, arranged on the hand
of the subject, reflect the IR radiation emitted by the LEDs. All
other light is filtered so that the system only recognizes the dots.
In software, the images taken from the nine cameras are used to
construct a three-dimensional representation of the markers. From
the reconstruction, the distance between dot pairs and local strain
of the skin is calculated.
Qualitatively, we determined two areas where skin is stretched
significantly during grasping movements: (1) behind the knuckle
of the forefinger, and (2) on the hand between the forefinger and
the thumb. We then used the motion capture system to quantify
skin stretch in these regions, by measuring changes in distance
between two adjacent dots during a grasping movement. We also
measured skin stretch for pinch grasps. The subject pinched
calipers opened to distances ranging from 20 to 90mm.
Figure 1. Device to generate skin stretch. The red arrow
points to the linear actuator, and the blue arrow to the end-
effector attached to the skin.
2.2 Artificial Skin Stretch
To artificially generate skin stretch, we mounted a Firgelli
(http://www.firgelli.com) linear actuator onto the subjects arm
using a hook-and-loop strap, and adhered an end-effector to the
subjects hand using double-sided tape and Smith+Nephew
electrode prep spray (Fig. 1). A string connects the actuator and
end-effector. The actuator was attached in the fully extended
position with little slack when the subjects hand was in the
position of maximum stretch. The actuator could thus be retracted
to provide artificial skin stretch and extended to maintain purely
natural skin stretch during trials.
2.3 Passive Proprioception
An experiment to test passive position proprioception of the
index finger was developed. A flat rod was taped to the volar side
of the subjects right index finger, beginning in the middle of the
proximal phalange and extending past the length of the finger.
The linear actuator and end-effector were also attached. The
subjects then rested his arm on a table with the right index finger
hanging off the edge. A protractor was positioned between the
index and middle finger, allowing measurement of the proximal
joint angle. The horizontal surface of the table was classified as 0
rotation and the direction straight down as 90.
With no visual cues, the experimenter moved the subjects index
finger randomly in the range of motion back and forth to prevent
any relative position cues. During this time, the linear actuator
was extended and retracted to disrupt cues of the status of the
artificial skin stretch. After a few seconds of these distracters, the
subjects finger was stopped at a specific angle with the motor
either fully extended or retracted (ON/OFF). The subject then
guessed the angle at which the finger was resting, purely from
proprioceptive cues.
3. RESULTS AND DISCUSSION
We identified regions of interest that showed maximum skin
stretch during natural grasp motions and for different pinch
lengths. These stretches were then artificially recreated using the
linear actuator attached to one of the two regions of most stretch.
We then confirmed the ability of the linear actuator to generate a
 Figure 2. Skin stretch at different pinch sizes.
relative skin stretch approximately with the same magnitude
found during natural motion. Finally, the index finger passive
proprioceptive task was performed and the amount of angle error
estimation was compared to trials with and without the presence
of artificial skin stretch.
3.1 Natural and Artificial Skin Stretch
Various grasps and pinches were repeated for different
configurations of the IR reflective dots on the hand. We found two
distinct regions of interest. The first was the area behind the
knuckle. Maximum skin stretch in this area occurred when the
index finger was fully flexed. The second area, the skin between
the thumb and index finger, showed maximum stretch for the
largest pinch sizes. Both areas had a non-linear stretch pattern,
with maximum stretch increase at ~70% of the range of motion.
For a whole-hand grasp (grasping a water bottle), the maximum
skin stretch between the forefinger and the thumb was 3.5mm for
dots initially spaced 13.6mm apart, corresponding to a 25%
change in distance. For the various static pinch sizes an increase
in stretch was seen as the pinch size was increased (Fig. 2). The
total stretch recorded was almost 4mm for dots initially spaced
13mm apart, which corresponded to a stretch of 30%.
Artificially induced skin stretch of the same regions of
maximum stretch was performed while measuring the amount of
stretch with the Vicon system. We found a maximum stretch of
1.5mm for dots initially spaced 14.5mm apart, corresponding to a
10% change in distance. Intermediate values could also be
obtained by only partially retracting the linear actuator.
The actuator was also used to change the skin stretch measured
initially during different pinches of calipers. These results are seen
by the 40mm** bar in Fig. 2. The artifical stretch was performed
while the subject held the calipers at 40mm. The data shows that,
with the artifical skin stretch device, the relative skin stretch
increased from 1.03mm to 2.08mm (50%). This validates the
capability of the device to induce a signicant amount of skin
stretch in physiologically relevant configurations and ranges.
3.2 Psychophysical Results
Preliminary psychophysical experiments were performed as
described in Section 2.3 on only three individuals. The small
sample size (n=3) limits many significant statistical findings, but
trends were observed that warrant discussion and future
investigation.
We found a trend in the increase of angle estimate error
correlating to the presence of artificial skin stretch from 4.4 with
the artificial stretch to 2.3 (p=0.2). In addition, subjects'
performance may be affected by the time they took to decide on
the answer. The distribution in angle estimate error was normal
about 0 (+/- 10), when the linear actuator was off. However,
artificial skin stretch seemed to bias the error in the positive
direction for angles less than 70, and bias error in the negative
direction for angles greater than 70 (Fig. 3). Artificial skin stretch
tended to cause subjects to overestimate small angles and
Figure 3. Angle error and standard deviation of passive
proprioception with the linear actuator OFF and ON.
underestimate large angles. The large drop in standard deviation
and error direction at this point (70), corresponds to the
maximum rate of skin stretch increase as measured with the
motion capture device, at ~70% of the total movement range. This
may be the angle at which natural skin stretch is most rapidly
changing and thus most susceptible to artificial perturbation.
Because the hand moves in a limited range of motion, it is natural
to assume the skin stretch fits this range of motion to achieve the
maximum amount of stretch across physiologically relevant
ranges.
4. CONCLUSIONS
Our experiment allowed quantification of a relatively subjective
proprioceptive mechanism candidate through the measure of
relative skin stretch for different hand grasps. Similar stretches
were artificially imposed on subjects during a passive index finger
proprioceptive discrimination task. The findings indicate that
artificial modification of skin stretch may alter subjects
perception of joint angle. The alteration, however, does not appear
to follow any signification pattern in terms of pushing the
direction of error in proprioception except at values around where
the skin naturally begins to stretch most rapidly during grasping
motions. This would support the notion that skin stretch plays an
important role in proprioception, especially at intermediate joint
angles where the muscle and joint afferents are least effective.
Further tests must be completed to determine the level of impact
these skin receptors have on proprioception as well as developing
methods to artificially stretch the skin in a more complex manner.
This will allow alteration of proprioception purely through
artificial skin stretching, and provide a better understanding of the
neural code underlying the perception of joint angle and hand
conformation.
REFERENCES
[1] K. O. Johnson, et al. Tactile functions of mechanoreceptive afferents
innervating the hand. J Clin Neurophysiol 17(6):539-58, 2000.
[2] D. F. Collins, et al. Cutaneous receptors contribute to kinesthesia at
the index finger, elbow, and knee. J Neurophysiol 94(3):1699-706,
2005.
[3] B. B. Edin. Quantitative analyses of dynamic strain sensitivity in
human skin mechanoreceptors. J Neurophysiol 92(6):3233-43, 2004.
[4] B. B. Edin and N. Johansson. Skin strain patterns provide
kinaesthetic information to the human central nervous system. J
Physiol 487 (Pt. 1): 243-51, 1995.
[5] K. O. Johnson. The roles and functions of cutaneous
mechanoreceptors. Curr Opin Neurobiol 11(4): 455-61, 2001.
[6] E. Moberg. The role of cutaneous afferents in position sense,
kinaesthesia, and motor function of the hand. Brain 106 (Pt. 1): 1-19,
1983.
[7] A. B. Vallbo, et al. Somatosensory, proprioceptive, and sympathetic
activity in human peripheral nerves. Physiol Rev 59(4): 919-57,
1979.
 Cytochrome oxidase-based morphological evaluation of the modifications
of the cortical volume due to electrical stimulation of sectioned peripheral
nerves: a short-term study

Celia Herrera-Rincon1, Beatriz Jorrin1, Abel Sanchez-Jimenez1, Carlos Avendao2 and Fivos
Panetsos1

1
Neurocomputing and Neurorobotics Research Group, Complutense University of Madrid.

Avda Arcos de Jalon 118, 28037, Madrid - Spain

2
Department of Anatomy, Histology and Neuroscience, Faculty of Medicine, Autonomous University of
Madrid. C. Arzobispo Morcillo, 2, 28029, Madrid Spain

corresponding author (fivos.panetsos@opt.ucm.es)

Amputation of a member implies the loss of sensory perception and motor skills of the
patient. Neural prostheses and artificial members, devices aiming at restoring these lost
capabilities, will be directly connected to the Central Nervous System (CNS) and, in
order to be effective, they have to code sensory information as the human body does and
to communicate it by means of electrical pulses through some interface with a
peripheral nerve.

In the present communication we address the effect electrical signals generated by

artificial members should exercise to the CNS and in particular to the somatosensory
cortex.

Any peripheral modification of the sensory input (from a temporary local anesthesia to
the amputation of a limb) provokes modifications to both, the anatomic substrate and
the physiology of brain structures that were receiving this sensory input1. Our first
hypothesis was that electrical stimulation from artificial members will be an additional
to the peripheral nerve sectioning manipulation of the peripheral input (now an artificial
one) and, consequently, will provoke additional modifications to the CNS. Our second
hypothesis was that CNS alterations after the nerve sectioning would be less

1
pronounced when artificial electrical stimulation was applied than when no stimulation
was delivered.

To test our hypotheses we manipulated the tactile sensory input from the whiskers of
young adult rats: we sectioned the left trigeminal sensory nerve of three animals and we
applied 12 hours/day electrical stimulation of 100s-long pulses of 3.0V, at 20Hz
through chronically implanted microelectrodes. After 4 weeks, animals were sacrificed
and brains were removed and sectioned. In rodents, the whiskers on the animals nose
are topographically represented by clusters of neurons called barrels in the contralateral
cortex in a one-to-one relationship (barrels are functional modules that process input
signals arriving from the periphery)2. To evaluate the alterations, in each animal, we
compared the cortex receiving input from the manipulated nerve to the cortex receiving
input from the intact nerve. We also compared the two cortexes of control animals
(intact, n=3) and the two cortexes of animals with sectioned peripheral nerve
(equivalent to common amputees, n=3).

As a measure we used the cortical volume of the barrel cortex (whiskers A1-A5 and E1-
E5) expressing Cytochrome Oxidase, a key mitochondrial enzyme that catalyzes the
final step of oxidative metabolism, generating the ATP necessary for neuronal
functioning3. Because of the tight coupling between neuronal activity and energy
metabolism, activity alterations within the CNS are reflected by regional changes of the
level of energy metabolism4. As such, it serves as a marker for direct visualization of
the oxidative capacity of specific cells and their processes. In control animals, we
obtained 4.36 0.20 and 4.28 0.03 for the right and left cortex respectively, with a
difference between cortexes of 1.8 3.81. Animals with sectioned peripheral nerve
show a very significant loss of the cortical volume (3.31 0.00 versus 3.98 0.00),
with a difference of -16.9 3.35. Finally, animals with sectioned peripheral nerve
submitted to electrical stimulation show a clear reduction of the loss of cortical volume
(-6.6 3.46, 4.03 0.21 for the right versus 4.31 0.01 for the left cortex). Measures
have been obtained using stereological methods5, values correspond to mean SD in
mm3. Variance analysis show significant differences between the theree groups:
Control-Sectioned p< 0.001; Control-Stimulated p=0.009; Sectioned-Stimulated
p=0.003.

2
Our results show that, on a short-term, electrical stimulation has an effect on
maintenance of organization in cortical structures.. In animals with transected nerve, the
comparisons of the contralateral and ipsilateral hemispheres, with respect to the
transected nerve, show a very significant loss of the cortical volume. Electrical
stimulation of the transected peripheral nerve reduce this loss in a percentage >50%.
Consequently, advanced neuroprostheses could help in the maintenance of the cortical
activity near its normal values preventing the direct effects of the deafferentation. We
are currently studying the effect of long-term stimulations as well as the effect of neural
stimulation with a delay between amputation and electrode implants.

(1.) JH Kaas, MM Merzenich, HP Killackey: The Reorganization of Somatosensory Cortex Following

Peripheral Nerve Damage in Adult and Developing Mammals. Annual Review of Neuroscience, 1983,
Vol. 6, pp. 325-356.

(2.) ME Waite and DJ Tracey: Trigeminal sensory system. In: The Rat Nervous System, edited by G.
Paxinos, San Diego:Academic Press, 1995, p. 705-724.

(3.) MT Wong-Riley: Cytochrome oxidase: an endogenous metabolic marker for neuronal activity.
Trends in Neurosc., 1989, Vol. 12, pp. 94-101.

(4.) R Machn, B Blasco, R Bjugn, and C Avendao. The size of the whisker barrel field in adult rats:
minimal nondirectional asymmetry and limited modifiability by chronic changes of the sensory input.
Brain Res. , 2004, 1025:130-138.

(5.) HJG Gundersen, P Bagger, TF Bendtsen, SM Evans, L Korbo, N Marcussen, A Mbller, K Nielsen,
JR Nyengaard, B Pakkenberg, FB Sbrensen, A Vesterby, MJ West: The new stereological tools: dissector,
fractionator, nucleator and point sampled intercepts and their use in pathological research and diagnosis,
APMIS 96, 1988, 857 881.

3
 Three-dimensional, Continuous-motion Ball Joint
Technologies for Prosthetic Wrist Applications
Werner Merlo
Medical Bionics Inc., 51203 Range Road 265
Spruce Grove, Alberta, Canada T7Y1E7
Telephone: (780) 987 3245, Fax: (780) 987 0075
E-Mail: werner@medicalbionics.com
http://medicalbionics.com

In this abstract, two revolutionary joint technologies that have been developed at Medical
Bionics Inc. are introduced. The CIVIC robotic hybrid and the Passive Locking (PL) joint
technology are based on the Robo-Flex core and can be applied to revolutionize any robotic
or manually operated joint system.

C I V I C - CONTINUOUS INTERNAL VARIABLE INTERLOCKING CONCEPT

The hybrid version CIVIC (Continuous Internally Variable Interlocking Concept) shown in
(Figure1) and best described as a three dimensional rotational mechanical gear drive
mechanism, is a variant of the prior Robo-Flex PL joint technology described below, shown in
(Figure 2). Unlike its PL predecessor which cannot move into a new locking position without
first being disengaged, CIVIC features a permanent linkage between the two engaging
surfaces. This linkage enables a continuous flow between locking positions and significantly
increases the potential of the CIVIC concept for other innovative applications, including
robotics.

CIVIC Know-How: A ball joint, connected to a Terminal Device (TD) and held in place by a
shell-like enclosure is able to move completely around its own axis and/or tilt in all directions
within its confines. A portion of the ball is comprised of ball race protuberances that are
permanently interlinked with a crown of pressure sensitive, spaced actuators. Bound only to
the race protuberances, the actuators control the movement of the ball joint. When the actuator
assembly is moved within a small horizontal plane, it directs the angular and rotational
deviations of the ball-race and consequently the TD attached thereto. Every degree of tilt
and/or axial rotation by the TD is directly proportional to the directional changes made by the
actuators. In essence, the two components move in tow.

This system incorporates complete and reliable three-dimensional function in a single unit, and
can be used for a multitude of robotic applications. CIVIC can be powered through a
complete range of locking positions to accurately and continuously control the positions of
pitch, yaw and roll while maintaining its 3D interlock. The current prototype spans 26 degree
rotations in the pitch and yaw directions on a 360 degree perimeter; however, it is possible to
extend the range of lateral rotations to 180 degrees by increasing the surface of the ball race
and the area covered by the actuators.

A finite-element analysis showed that a two-way actuator penetration of actuators and

protuberances as depicted in depicted in Figure 1 is more efficient than a one-way penetration
scenario, in which the same actuators protrude into a labyrinth of grooves in the ball.
According to the analysis, overall joint strength doubles by increasing the diameter of
actuators (and presumably conical protuberances) by 26%. We are currently working on
optimizing the actuators, the ball race and its protuberances for maximizing the range of
motion.

Figure 1: The CIVIC system involving continuous interlocking 3D rotational motion

Current robotic joint mechanisms have limited dexterity. The robotic wrists of such
sophisticated projects as the Mars Spirit Rover, the Dexterous Manipulator Arm attached to the
Space Station or unmanned Reconnaissance Vehicles are still comprised of three individual,
albeit highly complex, unidirectional joint systems to control pitch, yaw and roll. Besides
extremely high manufacturing and maintenance costs, the controls of these systems are difficult
to synchronize and operate efficiently. Also, the extreme weight and bulk of these mechanisms
are often a hindrance. By contrast, CIVIC could revolutionize the robotic joint industry as a
technology capable of emulating real life-like 3D movements while controlling pitch, yaw and
roll from one single, compact joint module.

ROBO-FLEX PL - 3D LOCKING BALL JOINT

Medical Bionics Inc. developed the production ready Robo-Flex prosthetic wrist. The ball joint
based joint mechanism enables the smooth motion through a wide angular range, and a step-
less 3D locking mechanism that can be activated at the desired working angle. That wrist has
been tested extensively by amputee soldiers at the Walter Reed Medical Center, Washington
DC, during the past two years.
The Robo-Flex passive-locking ball joint technology (Figure 2) works as follows. A rounded
object such as a ball or part thereof has a surface covered with polygonal patterns of spaced-
apart protuberances. The spaces between protuberances are cavities. When an assembly of
closely-spaced, spring-loaded actuators is imprinted against the protuberances, the actuators
emulate a mirror image of the opposing surface. Actuators contacting protuberances are
pushed back, while unobstructed actuators penetrate into the cavities between protuberances.
Once trapped in a cavity, the actuators are unable to move in any direction, thereby freezing
the spatial orientation between actuators and ball surface against forces of pitch, yaw and roll
and enabling a large load carrying capacity. As soon as the actuator assembly is retracted
from the protuberances, the unhindered ball can be freely moved to another angle. The
actuators, compressed to various depths while engaged with the protuberances, regain their
fully extended position ready to imprint the protuberances at a different locking position.

Figure 2: Robo-Flex passive-locking 3D ball joint


Quantifying the Dimensions of Human Hand
Dexterity Through a Survey of Daily Tasks
Ravi Balasubramanian, Brian Dellon, Sujata Pradhan, Spencer Gibbs, and Yoky Matsuoka
AbstractThe concept of dexterity and its dimensions have
been difficult to define through the years, and various research
communities have different definitions of dexterity. In this paper,
we present a novel approach to understand peoples perception of
dexterity and its different dimensions in daily tasks. The
approach entails a video-based survey where people score
different tasks along pre-specified dimensions of dexterity. The
results show how the dimensions contribute to an overall
dexterity score as well as the importance of each dimension.
Index TermsDexterity, Hand function, Daily tasks.
I. INTRODUCTION
Human hand dexterity has been a challenging construct to
define due to the inherent structural complexity of the human
hand and the variety of tasks it performs. Some of the existing
definitions of dexterity include versatile capacity, the ability to
carry out harmonious movements, finding an efficient motor
solution, and the ability to correctly solve a motor problem in
any situation [Latash et al., 1991]. Although these attempts
have resulted in conceptual definitions of dexterity, it is still
unclear what the key components or dimensions of dexterity
are. While previous work has considered task features such as
force magnitude, distance traveled, and speed as dimensions of
dexterity [Jones, 1998], dexterity includes even more
dimensions such as the element of internal models and range
of joint motion. Another barrier that has resulted in ambiguity
is that dexterity has been defined differently in different
domains such as robotics, physiotherapy, hand surgery, and
physiology. The overall goal of this project is to identify
various dimensions that contribute to dexterity, their relative
importance, and a ranking of the dexterity of various daily
tasks. We approach this problem by measuring peoples
perception of the dexterity in various daily tasks through a
survey.
II. METHODS
A. Survey structure
The survey, deployed on the internet, consists of a series of
questions asking the participant to "score" several daily human
Ravi Balasubramanian (bravi@cs.washington.edu), Brian Dellon, and
Yoky Matsuoka are with the Department of Computer Science and
Engineering at the Univesity of Washington (UW), Seattle, WA. Spencer
Gibbs is with the Department of Electrical Engineering at UW, Sujata
Pradhan with the Department of Rehabilitation Medicine at UW. Contact
author: Ravi Balasubramanian.
1
hand-based tasks according to pre-defined dimensions related
to the overall concept of dexterity.
A pool of twenty common tasks that require fine motor
control and are typically associated with manual dexterity was
created. Some example tasks were hammering, card shuffling,
orange peeling, and keyboarding. Tasks that involved a
variety of grasps (palmar, hook, pincer) as well as multiple
movement patterns (discrete movements, continuous
movements, repeated movements, sequentially linked
movements) were included, based on work by Kimmerle et.
al., 2003. The tasks were precisely defined using short videos
created specifically for this purpose. To help the subject
understand the roles played by the hand in the task, we also
identified how many fingers were being used. Subjects were
asked to provide scores on exactly the task shown in the video
and not on how he/she might do the task. The subject was
asked to give scores even if they were not familiar with the
task.
The ten pre-defined dimensions of dexterity were familiarity
with the task, force magnitude, force modulation, stiffness
magnitude, stiffness modulation, speed of movements,
coordination between joints/timing with an external clock,
range of joint motion, the importance of sensory feedback
(vision, tactile, proprioceptive), and the amount of practice
required.
Each survey was fifteen pages long, received responses on
ten tasks, and took about thirty minutes to complete. The
survey was structured such that each page required the subject
to score, on a scale of 05, all the tasks for each dimension.
The final page asked for the overall dexterity score for each
task. We also collected some demographic information at the
beginning of the survey. Three equivalent surveys were
created using 10 tasks from the task pool.
B. DATA ANALYSIS
Two statistical techniques were used to analyze the data
collected, namely linear regression and principal component
analysis. Linear regression was used to find a set of
coefficients wi that can predict dexterity D given the ten
dimensions, xi, i={1,2,..,10}, of the task:
D = wi xi .
Regression analysis will show how each dimension xi
influences dexterity---if the regression coeffecient wi for
dimension xi is positive, it implies that dexterity increases as
the value of that particular dimension increases and vice versa.
Principal component analysis measures the inherit
relationship between the dimensions chosen and which

combination of dimensions are important.
III. RESULTS
The number of subjects was 26 (43% female, 77% in the
1835 age range, 62% from the field of engineering, 11%
from the health sciences). The dexterity scores for some tasks
were as follows: keyboarding 3.9, playing the violin 4.8,
hammering, 1.8, screw driver 2.8, rolling a ball 4.0. Overall, it
was noticed that playing the violin was ranked the highest in
the overall dexterity score, while hammering was ranked the
lowest.
The regression coefficients (and their standard deviation in
brackets) for the ten dimensions of dexterity were computed to
be {0.01 (0.1), -0.06 (0.1), -0.02 (0.1), -0.11 (0.2), 0.1 (0.2), -
0.05 (0.1), 0.39 (0.1), 0.21 (0.1), 0.23 (0.2), 0.30 (0.1)} (r2
value 0.98). It was noticed that the contribution of the first six
dimensions to dexterity were significantly smaller than the last
four dimensions. Fig. 1 shows the correlation between the
different dimensions, and it was noticed, for example, that
force magnitude and practice were negatively correlated.
Figure 1: The correlation between the various dimensions of
dexterity (white represents high positive correlation, black
represents high negative correlation).
Principal component analysis of the data showed that the
first principal component explained 55% of the variance in the
data, the first two principal components 75%, the first three
principal components 85%, and the first four principal
components 91%. Fig. 2 shows a biplot of the different
dimensions expressed in terms of the first two principal
components.
Fig. 2: A biplot of the dimensions of dexterity represented in
2
the first two principal components.
IV. DISCUSSION
A. Relationship between the dimensions of dexterity
The correlation between dimensions (see Fig. 1) was
intuitive. It was noticed, for example, that tasks that require
more sensory feedback require more practice. Also more
practice is required for tasks that require large forces. In
contrast, it was noticed that movement speed was negatively
correlated with stiffness magnitude.
Principal component analysis provided interesting insights
into the key components of dexterity. The first principal
component was dominated by dimensions such as practice,
coordination/timing, sensory feedback, and speed. All these
dimensions represent involuntary aspects of task performance;
for example, a person does not have direct control over the
sensory feedback inherent to a task. In contrast, the second
principal component was dominated by dimensions such as
force and stiffness magnitude. These dimensions represent
voluntary aspects of task performance.
B. Impact on robotics and rehabilitation research
Dexterity in robotic manipulation is currently defined using
robot state spaces and generating control strategies to travel
from one state to another (Bicchi, 2000). Our work provides a
more intuitive definition of dexterity in terms of daily tasks
and using a human-specified metric. Second, it has been
difficult to quantify the complexity of daily tasks and design
cheap robots or prosthetics that perform tasks of specific
complexity levels. This survey makes headway by identifying
peoples perception of task complexity.
These results will also provide rehabilitation professionals
with a framework for progression from objectively quantified
simple activities to more dexterous activities for rehabilitation
during recovery. It may also be incorporated into current
assessment tools of hand function to quantify severity of
functional deficits.
C. Limitations
These results are based only on the peoples perception of
the tasks and the dimensions. It would be better to use
quantitative measurements for each dimension to derive a
dexterity score. Still our survey sheds light on the importance
of different dimensions when combining the dimensions to
give an overall dexterity score. A larger task pool and a
sample space will also provide strong statistical results.
REFERENCES
M.L. Latash, M.T. Turvey, N. Bernstein, 1991, Dexterity and Its
Development, Lawrence Erlbaum
Lynette Jones, 1998. Manual Dexterity in Psychobiology of the hand,
Ed. K. J. Connolly, Cambridge University Press.
M. Kimmerle, L. Mainwaring, and M. Borenstein (2003). The functional
repertoire of the hand and its application to assessment. American
Journal of Occupational Therapy, 57, 489498.
 3

A. Bicchi, 2000. Hands for Dexterous Manipulation and Robust

Grasping: A Difficult Road Toward Simplicity, Transaction of Robotics
and Automation, 2000.
 UnderstandingtheHumanHandforAdvancingRoboticManipulation
WorkshopatRobotics:ScienceandSystemsConference,2009 Back
TentativeSchedule

RaviBalasubramanianandYokyMatsuoka

Time Event Speaker PresentationTitle

8:20AM Welcome RaviBalasubramanian
Topic:PhysiologyofHand
Presentation FranciscoValeroCuevas Theneuromuscularsystemsdoesordinarymanipulation
8:30AM
tasksthe``hard''way:Lessonsforroboticmanipulators?
Presentation MarcoSantello SynergisticControlofHandMusclesThroughCommon
9:00AM
NeuralInput
9:30AM Presentation DerekKamper Transmissionofmusculotendonforcestotheindexfinger
10:00AM Presentation RaviBalasubramanian HandResearchatUniversityofWashington
10:30AM CoffeeBreak
Topic:StudyingHumanHandBehavior
Presentation LynetteJones HumanHandFunction:TheCouplingofSensoryand
11:00AM
MotorSystems
Presentation LuigiVillani Originalapproachestointerpretationlearning,and
11:30AM
modelling,fromtheobservationofhumanmanipulation
PosterOne Posterpresenters
12:00PM Minute
Madness
12:30PM Lunch
Poster Posterpresenters
1:15PM
Presentation
Topic:HumanlikeRoboticSensing
2:00PM Presentation GeraldLoeb RobustBiomimeticTactileSensingandGripControl
Presentation VeronicaSantos Developmentofartificialgripreflexesthatutilizethe
2:30PM
adduction/abductioncapabilitiesofhumanfingers
Presentation HaruhisaKawasaki ForceSensationoftheHumanFingerwhenUsinga
3:00PM
MultiFingeredHapticInterface
3:30PM CoffeeBreak
Topic:HumanlikeRoboticHands
4:00PM Presentation GeorgeBekey Thehistoryofgraspingwithroboticandprosthetichands
4:30PM Presentation PeterAllen LowDimensionalDataDrivenGrasping
Rock Allparticipants
5:00PM
Turning!deas
RockTurning Allparticipants
5:35PM
Presentations

Note:Unfortunately,Prof.SusanLedermanisunabletoattendtheworkshop.