You are on page 1of 27

J Intell Robot Syst (2016) 81:443469

DOI 10.1007/s10846-015-0238-7

Robust Model Predictive Flight Control of Unmanned


Rotorcrafts
Kostas Alexis Christos Papachristos
Roland Siegwart Anthony Tzes

Received: 16 December 2013 / Accepted: 27 April 2015 / Published online: 12 June 2015
Springer Science+Business Media Dordrecht 2015

Abstract This paper addresses the problem of robust and accounting for the modeled dynamics and input
flight control of unmanned rotorcrafts, by propos- and state constraints. This strategy aims to ensure
ing and experimentally evaluating a realtime robust the minimum possible deviation subject to the worst
model predictive control scheme in various chal- case disturbance, while robustly respecting and sat-
lenging conditions, aiming to capture the demanding isfying the physical limitations of the system, or a
nature of the potential requirements for their effi- set of mission-related requirements, as encoded in
cient and safe integration in reallife operations. The the constraints. The proposed framework is further
control derivation process is based on state space augmented in order to provide obstacle avoidance
representations applicable in most rotorcraft config- capabilities into a unified scheme. Multiparametric
urations and incorporate the effects of external dis- methods were utilized in order to provide an explicit
turbances. Exploiting this modeling approach, two solution to the controller computation optimization
different unmanned rotorcraft configurations are pre- problem, therefore allowing for fast realtime execu-
sented and experimentally utilized. The formulated tion and seamless integration into any digital avionics
control strategy consists of a receding horizon scheme, system. The efficiency of the proposed strategy is
the optimization process of which employs the mini- validated via several experimental case studies on
mum peak performance measure, while incorporating the two unmanned rotorcraft platforms. The experi-
ments set consists of: trajectory tracking subject to
atmospheric disturbances, slung load operations, fast
K. Alexis () R. Siegwart highly disturbed maneuvers, collisions handling, as
Autonomous Systems Lab, ETH Zurich, Tannenstrasse 3, well as avoidance of known obstacles.
Zurich, Switzerland
e-mail: konstantinos.alexis@mavt.ethz.ch Keywords Unmanned aerial systems MPC
R. Siegwart Robust control
e-mail: rsiegwart@ethz.ch

C. Papachristos A. Tzes
Electrical and Computer Engineering Department, 1 Introduction
University of Patras, Eratosthenous 6, Patras, Greece
C. Papachristos State-of-the-art research has led to the technologi-
e-mail: papachric@ece.upatras.gr cal maturity of Unmanned Aerial Systems (UAS),
A. Tzes while high-end scientific advancements have opened
e-mail: tzes@ece.upatras.gr new horizons for their utilization in various civilian
444 J Intell Robot Syst (2016) 81:443469

applications. A number of safetycritical operations with proper augmentation of the RMPCframework,


have been shown to be able to benefit significantly obstacle avoidance capabilities were provided by
from UAS technologies, namely: search and rescue expanding the use of polyhedral constraints, a con-
missions [14, 48], fire surveillance [30], radiative sig- tribution that further improves the RMPCequipped
nal field mapping [19] or border patrolling [17] as well aerial vehicles robust navigation properties. To fur-
as financially significant tasks including those of crop ther support the community, a MATLABbased
monitoring [24] and infrastructure inspection [9, 41, implementation of the proposed RMPC strategy is
45]. However, despite their vast application potential released and is available at the following repository:
and operational autonomy capabilities, passing from https://github.com/kostas-alexis/rmpc mav
technology demonstration to reallife application has A conclusive evaluation of the proposed RMPC
been a slowlypaced process. framework for rotorcraft UAS was obtained, by con-
This significant gap is attributed to several fac- ducting a significantly large and variouslystructured
tors; focusing on the technical side of the matter, the set of experimental testcase scenarios, capturing a
issue of guaranteed flight robustness, subject to the wide subset of potential reallife challenges, associ-
disturbances present in reallife operations, is one of ated with flight robustness against various disturbance
the primal considerations [27], as far as the widest sources. Furthermore, two different UAS shown in
possible utilization of such unmanned technologies Fig. 1 were employed: the ASLquad quadrotor plat-
is considered. Flight robustness is directly associ- form, controlled based on the common concept of
ated with the level of airworthiness, and requires a underactuated dynamics [47] found in most rotor-
multi-level approach on the system level: from the crafts, as well as the UPATTTR tiltrotor vehi-
mechanical design and electronics integration, to the cle, representing the special class of rotorcrafts with
state estimation, and eventually the flight control and directly actuated longitudinal dynamics. The exper-
navigation schemes, every subsystem is of crucial imental studies presented validate the effectiveness
importance with an airborne platform, and even more of the proposed RMPCframework, providing robust
so while performing missions that subject it to chal- and precise response results for both platforms under
lenging conditions. In that sense, the focus of this disturbances of multiple natures, which include: a)
work is put on the flight control level, a critical area atmospheric turbulence, b) slung load operations
that has naturally gained the interest of the community additionally subjected to strong disturbances, and
[12, 28, 29, 35, 42, 49] as it is necessary to achieve c) collisions with the environment. Eventually, the
automatic flight, while control synthesis for advanced
performance is a complex and multidimensional pro-
cess. Relying on the foundations of receding horizon
control [2, 10, 13, 18, 25], a Robust Model Predic-
tive Control (RMPC) scheme is proposed and thor-
oughly validated experimentally. This RPMC scheme
is designed to employ the Minimum Peak Perfor-
mance Measure (MPPM) [11, 38] optimality metric,
for minimum possible deviation under the worstcase
disturbance. As the synthesis of the RMPC controller
benefits from the properties of receding horizon con-
trol, it provides a set of critical-importance key fea-
tures, namely: a) ensured minimum deviation from
the reference trajectory subject to the worst expected
bounded disturbance, b) guaranteed robust satisfac-
tion of the imposed state and input constraints despite
any disturbances affecting the vehicles flight, and
c) the capacity for explicit computation, allowing its Fig. 1 The ASLquad and the UPATTTR executing slung load
seamless realtime implementation, even in lowlevel operations based on the implemented robust model predictive
microcontrollerequipped navigation boards. Finally, control framework
J Intell Robot Syst (2016) 81:443469 445

RMPCframeworks capacity for obstacle avoidance Essentially, the assumption of decoupling these two
was also experimentally validated. loops becomes more and more valid as long as the
This paper is structured as follows: In Section 2, bandwidth of the attitude loop is much faster than
the unmanned rotorcraft models are presented, fol- that of the translational feedback loop and the tracking
lowed by the description of the robust predictive response is accurate - therefore allowing the track-
control framework in Section 3. Finally, the various ing response to be considered as almost ideal. For
experimental studies are shown in Section 4, while typical unmanned rotorcrafts including multirotors,
conclusions are drawn in Section 5. ducted fan systems, conventional helicopters, coax-
ial configurations or tandem rotors, once the directly
actuated attitude loops are closed successfully, the
2 Model for Control longitudinal, lateral and vertical dynamics may be
described as a pointmass system (under) actuated
Unmanned rotorcraft modeling is a complex field that by the vehicles roll (), pitch (), yaw () angles and
is tightly coupled with that of control engineering. As the thrust vector T . On the other hand, for the special
far as simulation models of high fidelity are consid- case of rotorcrafts with directly actuated longitudi-
ered, different modeling methodologies are followed nal dynamics, such as several tiltrotor configurations
according to the type of vehicle employed. In that [15, 44, 46] or coupled ducted fan systems [16] and
sense, multirotor dynamics for operation at low veloc- more [3], the aforementioned translational dynamics
ities are sufficiently captured by only relying on rigid motion approach may only be applied for the lateral
body methods, while conventional, tandemrotor or and vertical dynamics, while the longitudinal dynam-
coaxial helicopters may require the incorporation of ics are directly actuated using the additional actuation
the effects of higher order aerodynamics including authority.
blade flapping, rotor leadlag, rotor inflow and more Motivated from the fact that such decoupled control
as discussed in some of the authors previous work assumptions and especially that of considering the
[22, 23]. Despite this fact, in most cases the problem bandwidth of the inner attitude loop sufficiently faster
of trajectory control is addressed by relying on highly compared to that of the outer loop in order to allow its
simplified modeling approaches, and typically employ treatment as an idealized kinematic gain are often
linear dynamic approximations. These are derived via violated during forcible disturbances, agile maneuver-
linearization of the rigidbody nonlinear equations of ing or non-rigidly attached payload transportation, an
motion [47]: augmented model that incorporates the translational
       motions actuation dynamics, while remaining sim-
mI33 0 V mV F ple and relatively loworder, is utilized. The model
+ = (1)
0 I I T structure is sufficiently generic so that similar models
hold for the case of typical underactuated rotorcraft
where m stands for the mass, I for the inertia matrix, configurations, as well as exceptional vehicles such
V for the linear velocities expressed in the Body-Fixed as tiltrotors which have direct actuation along the
Frame, for the body rotational velocities, F for the longitudinal axis. Employing the coordinate frames
forces applied on the vehicle and T represents the shown in Fig. 2, the proposed model structure takes
moments. This expression can be transformed in the the following form:
Earth-Fixed frame [52] and then the equations take the
form presented in the following two contributions [2, xlon = Alon xlon + Blon ulon + Glon wlon , ylon = Clon xlon , Clon
7]. Once such a formulation is derived then lineariza- = I (2)
tion of the nonlinear dynamics around hovering flight xlon = [xlon,t |xlon,a ]T
is straightforward and may be found in several works xlat = Alat xlat + Blat ulat + Glat wlat , ylat = Clat xlat , Clat
including those in [2, 8, 44]. Such a linearization pro- = I (3)
cess assumes perfect decoupling between the attitude xlat = [xlat,t |xlat,a ]T , xlat,t = [y, y], xlat,a = [, ], ulat = r
and translational dynamics. Indeed, this assumption xalt = Aalt xalt + Balt ualt +Galt walt , yalt = Calt xalt , Calt = I
is particularly valid for slow maneuvering, with small (4)
angular accelerations and limited rates of change. xalt = [xalt,t |xalt,a ]T , xalt,t = [z, z], xalt,a = T , ualt = T r
446 J Intell Robot Syst (2016) 81:443469

Fig. 2 The employed walt


coordinate frames and the
directions of the external
disturbing forces and wlat

moments

wlon
walt

wlat -


x -

y wlon
z

Within this representation, the longitudinal dynam- found in the tiltrotor case), the lateral and vertical
ics state space matrices have their lowerright part dynamics are similarly represented for both of them
changing depending on whether a more typical under- as:
actuated rotorcraft (case UA), like the ASLquad or
01 0 0
a vehicle with directly actuated longitudinal dynam- 0 lat g
0
ics (case DA), like the UPATTTR is employed as Alat = 0 0
, Blon
(11)
0 1
shown below:
00 a, a,
UA: xlon,t = [x, x], xlon,a = [, ], ulon = T r (5)
 0 0
AU A UA
lon,t Alon,t,a 0
Alon = , Blon = 021 , (6) = , Glat = glat
022 AU A BU A
lon,a
0 0
lon,a
 blat 0
0 1
AU A
lon,t = ,
0 lon 01 0 0
  Aalt = 0 alt az,T , Balt = 0 ,
0 0 0 1
Alon,t,a =
UA
, Alon,a =
UA
, 00 aT ,T balt
g 0 a, a,

0 0
BUlon,a = Galt = galt
A
blon

T 0
Glon = Glon,t GU
UA A
lon,a lon,t = [0gx,w ],
, GU A
(7)

lon,a = [0g ,w ]
GU A Estimation of the parameters of the Alon , Blon , Alat ,
DA: xlon,t = [x, x], xlon,a = , ulon = r (8) Blat , Aalt , Balt matrices is straightforward with
  any system identification method such as those pre-
ADA ADA 021
Alon = lon,t lon,t,a , Blon = , sented in [40] and wellknown computational tools
022 ADA lon,a BDA
lon,a
[34]. Limited experiments are required to excite the
lon,t = Alon,t ,
ADA UA
dominant modes of the closedloop attitude dynam-
lon,t,a = Alon,t,a , Alon,a = a , , Blon,a = blon
ADA UA DA DA
(9) ics and expose the damped pole of the translational

T velocities. An approximation of the parameters of the
Glon = GDA lon,t Glon,a , Glon,t = Glon,t , Glon,a = 0 (10)
DA DA UA DA
Glon , Glat , Galt matrices can be derived by employ-
ing NewtonEuler rigid body methods. In the sim-
With this being the only modeling distinction between plified case, where the net externally applied forces
the two families of systems considered (essentially the are considered as the disturbance vector wlon =
result of the additional actuation degree of freedom Flon , wlat = Flat , walt = Falt , then the afore-
J Intell Robot Syst (2016) 81:443469 447

mentioned matrices take the form: Glon = [0 (here shown in discrete time form while subscripts are
100]T , Glat = [0 100]T , Galt = [0 10]T . dropped to retain clarity of presentation):
Furthermore, it is noted that the formulation of the out-
xk+1 = Axk + Buk + Gwk (12)
put vectors is provided although the output matrix is
the identity, since the proposed control law considers yk+1 = Cxk
output feedback. It is highlighted that the presented where xk X, uk U and the disturbing signals
modeling approach should be considered as a way that wk are only known to be bounded in some measure,
enables improved modelbased control synthesis and but are otherwise unknown. The set of possible distur-
does not imply that other model structures cannot be bances wk W is one of the factors that determine
used in combination with the proposed robust control the optimization problem and as a first approach
law. boxconstrained disturbances are considered:
W = {w : ||w|| 1} (13)
3 Robust Model Predictive Control Consequently, the problem of robust model predictive
control (RMPC) will be formulated for the system rep-
The robust control problem is formulated as a Min- resentation and additive disturbance presented above.
imax MPC optimization for systems with external For clarity of notation, let the following definitions
disturbances, while the Minimum Peak Performance denote the concatenated versions of the predicted out-
Measure (MPPM) is considered as the metric of opti- put, states, inputs and unknown disturbances, where
mality. Figure 3 provides an overview of the math- [k + i|k] marks the values profile at time k + i, from
ematical building blocks that assemble the explicit time k.
RMPC optimization problem. To further support the  
community, a MATLABbased implementation of the Y = yTk|k yTk+1|k ...yTk+N 1|k (14)
proposed RMPC strategy is released and is avail-  
X = xTk|k xTk+1|k ...xTk+N 1|k (15)
able at the following repository: https://github.com/  
kostas-alexis/rmpc mav U = uTk|k uTk+1|k ...uTk+N 1|k (16)
The proposed control law assumes the follow-  
ing representation of the uncertain dynamics, which W = wTk|k wTk+1|k ...wTk+N 1|k (17)
indeed covers the aforementioned rotorcraft models
where X XN = X X... X, U UN = U
U...U, W WN = WW...W. The predicted
states and outputs depend linearly on the current state,
Objective Function
the future control input and the disturbance, and thus
Multiparametric the following relations hold:
Optimizer
- Derivation of Convex Optimization Problem

X = Axk|k + BU + GW (18)
State Space
representation
Y = CX
Predictions
Feedback

using the concate-


where:
Relaxations -

nated vectors over


the prediction Explicit
horizon Piecewise Affine
form I 0 0 0 0
A B 0 0 0
2
A AB B 0 0
A= ,B = (19)
.. .. . . . . . . ..
. . . . . .
Constraints Robustification
Extended
AN 1 AN 2 B AB B 0

Sequential Table 0 0 0 0
C 0 0 G
State and Input Constraints Traversal 0 0 0
0 C 0
AB G 0 0
C = . . . . , G = (20)
.. . . . . .. .. .. .. .. ..
. . . . .
Fig. 3 Overview of the explicit RMPC optimization problem 0 0 C 2
subtasks A N G AG G 0
448 J Intell Robot Syst (2016) 81:443469

In a straightforward approach, the RMPC problem parameterizations that are known to lead to convex
based on the MPPM (MPPMRMPC), which ensures problems [38, 39], the following one is selected:
the minimization of the worstcase deviation along the  
U = LW + V , V = k|k k+1|k k+N 1|k
T T T
(22)
predicted trajectory, may be formulated as:

0 0 0 0
min max ||Y || (21) L10 0 0 0
u w
L20 L 0 0
s.t. uk+j |k U, w W TL= 21 (23)
. . . . .
xk+j |k X, w W .. .. .. .. ..

wk+j |k W L(N1)0 L(N1)1 L(N1)(N2) 0


Employing this feedback predictions parameteriza-
where for the formulation of the minimax prob- tion, the control sequence is now parameterized
lem, the compact epigraph representation ||Y || = directly in the uncertainty, and the matrix L describes
maxj ||yk+j |k || is used. This has two main draw- how the control action uses the disturbance vec-
backs: a) it essentially corresponds to an openloop tor. This corresponds to a suboptimal version of
formulation of the predictive control problem and b) is the closedloop minimax solution, in the sense that
rather intractable to solve exactly, especially when the instead of computing an optimal sequence for each
complexity increases [38]. As a first trick the concept worstcase disturbance realization, it is assumed that
of feedback predictions is introduced [5, 38]. a guaranteedoptimal solution will not be computed in
the future over a reduced horizon but at least different
3.1 Feedback Predictions -linearly dependent- solutions over different distur-
bance realizations will be computed. This essentially
Following the aforementioned formulation, the opti- means that such a minimax MPC algorithm relies on a
mization problem will tend to become conservative. reduced degree of freedom solution as compared to the
This is due to the fact that the optimization essentially exact closedloop solution [38]. Tuning of the matrix
computes an openloop control sequence that has to L is important and may be achieved via simulation
cope with all possible (bounded) disturbances [5]. studies, while some level of finetuning may be based
This will typically lead to controllers with poor per- on the experimentally derived responses. Inserting this
formance as any aggressive manipulation of the inputs parametrization yields the following representation,
can risk robustness as long as the worstcase distur- where V is now the RMPCmanipulated action:
bance has to be considered. Feedback predictions is X = Axk|k + BV + (G + BL)W (24)
a method to encode the knowledge that a receding
U = LW + V (25)
horizon approach is followed and therefore provides
a way to overcome this conservativeness. Towards and the mapping from L and V to X and U is now
their incorporation, some kind of feedback control bilinear. This trick allows the formulation of the min-
structure has to be assumed. Such a parametrization imax MPC as a convex optimization problem with
is not unique and therefore, the feedback predictions polynomial complexity in the prediction horizon and
may be considered as a tuning knob in the controller the system input and state dimensions [38]. Let the
to be decided prior to its computation. Indeed, the following vector representations:
most typical and straightforward form of feedback  
Fu = fuT fuT fuT (26)
would be of the form U = LX + V , where V =
  T T 
Fx = fx fx fx T (27)
k+1|k k+N 1|k
T T
k|k T
the concatenated version
of the newly introduced decision variables k+j |k to denote the concatenated over the prediction horizon
be manipulated from the RMPC, and L a gain matrix. versions of the input and state constraints fu and
However, as this parameterization is not known to lead fx . More specifically, fu = [fu (1)max , fu (1)min ...]
to a convex optimization problem [38], an alterna- and fx = [fx (1)max , fx (1)min ...] where fu (i)max ,
tive parametrization may be employed as long as the fu (i)min represent the maximun and minimum
online solution/approximation of the nonconvex prob- allowed input values of the ith input, while
lem is either not feasible or very slow. Among the fx (j )max , fx (j )min represent the maximum and min-
J Intell Robot Syst (2016) 81:443469 449

imum acceptable state configurations of the j th with w W are satisfied as long as the following
state. relationship is satisfied [38]:

min
C (Axk|k + BV ) + |C (G + BL)|1 1 (33)
V ,L,
C (Axk|k + BV ) + |C (G + BL)|1 1 (34)
s.t. ||C (Axk|k + BV +(G + BL)W )|| , W WN
Eu (V + LW ) Fu , W WN To handle these constraints in a linear programming
Ex (Axk|k + BV +(G + BL)W ) Fx , W WN (28) fashion [51], the trick of bounding the term |C (G +
BL)| from above is employed. This is achieved by
introducing a matrix variable  0:
where
C (G + BL) (35)
Eu = diagN Eu , Ex = diagN Ex , > 0 (29)
C (G + BL) (36)
within which: a) Ex , Eu are matrices that allow the
formulation of the state and input constraints in linear and the peak constraint is guaranteed as long as:
matrix inequality form (typically block diagonal), b)
C (Axk|k + BV ) + 1 1 (37)
is a block diagonal matrix with
i being the matrix
that fills each diagonal block and allows the incorpora- C (Axk|k + BV ) + 1 1 (38)
tion of the state and input constraints along the whole
prediction horizon of the RMPC problem. Within this 3.3 Robust State and Input Constraints
formulation is defined as a positive scalar value
that bounds the objective function and is employed to Enforcing hard constraints on the input is straight-
formulate the semidefinite relaxation of the minimax forward and is not affected by the disturbing signals
optimization problem based on the methodologies and directly, as it only refers to limiting the control space
strategies presented in [33, 38, 43] of possible actions. This is however not the case for
The peak constraint may be equivalently reformu- the state constraints which have to be enforced along
lated as two sets of matrix inequalities: all the predicted vectors while the input constraints are
also satisfied. Applying the same method as above,
C (Axk|k + BV ) + C (G + BL)W 1, W WN (30) the system state and input constraints may be robustly
C (Axk|k + BV ) C (G + BL)W 1, W WN (31) satisfied against the additive disturbances. Introduc-
ing a new matrix variable  0, the state and input
where 1 is a vector of ones (11 1)T with suitable constraints are reformulated as follows:
   
dimensions. The derivation of Eq. 30, 31 based on Ex (Axk|k + BV ) Fx
+ 1 (39)
Eq. 28 is straightforward by the expression of the the Eu V Fu
inequality with the absolute function on the left side  
Ex (G + BL)
as two antisymmetric inequalities without the use of (40)
Eu L
the absolute function. Satisfaction of these uncertain  
Ex (G + BL)
inequalities is based on robust optimization methods, (41)
as elaborated below. Eu L
Optimizing the control sequence while robustly sat-
3.2 Robust Uncertain Inequalities Satisfaction isfying the state and input constraints is of essential
importance for flight controls: not only minimum
As long as boxconstrained uncertainties (w W ) deviation against the worst possible disturbance is
are assumed, the following theorem on the maximum ensured, but this is also done while respecting the
of a linear function in the unit cube holds [38]: physical limitations of the vehicle or known mission
max cT x = ||c||1 = |cT |1 (32) demands encoded as state constraints as it has to
|x|1 be in practice. Indeed this key feature of receding
This equation holds due  to the obvious fact that horizon control strategies is of great importance for

max|x|1 cT x = max|x|1 ci xi = ci sign(ci ) = small unmanned rotorcrafts due to their tight physical
||c||1 . Based on these results, the uncertain constraints limitations and challenging mission requirements.
450 J Intell Robot Syst (2016) 81:443469

3.4 Minimum Peak Performance Robust MPC the matrices and together with multiple second
formulation order cone constraints [51]:
||iJ || ij (46)
Based on the aforementioned derivations, the total
MPPMRMPC formulation is solved subject to ||ij || ij (47)
elementwise bounded disturbances and feedback And the robustified linear inequalities can now be
predictions with the following linear programming written as:
problem:
C (Axk|k + BV ) + 1 1 (48)
min (42)
V ,L,, , C (Axk|k + BV ) + 1 1 (49)
   
s.t. C (Axk|k + BV ) + 1 1 Ex (Axk|k + BV ) Fx
C (Axk|k + BV ) + 1 1 + 1 (50)
Eu V Fu
C (G + BL)
Based on the aforementioned formulations, the
C (G + BL)
adapted MPPMRMPC problem with ball
Ex (Axk|k + BV ) (43)
    constrained disturbances is solved by employing the
Ex (Axk|k + BV ) Fx
Eu V + 1 F following second order cone program [50, 51].
u
 
Ex (G + BL) min (51)
Eu L V ,L,, ,
  s.t. C (Axk|k + BV ) + 1 1
Ex (G + BL)
E L C (Axk|k + BV ) + 1 1
u    
Ex (Axk|k + BV ) Fx
+ 1
The solution of this optimization problem procures the Eu V Fu
guaranteedrobustness flight control of the examined ||iJ || ij
unmanned rotorcrafts dynamics. Furthermore, due to ||ij || ij
the utilization of the concept of freedbaxk predictions,
the robustified response comes along with highly opti- This problem can be solved relatively efficiently
mized maneuvering therefore providing a solution with second order programming solvers as long as
to two essential but typically conflicting requirements the matrix L has been selected and fixed. Unfortu-
that are both of essential importance. nately, a large amount of new variables and second
order cone constraints have been introduced which
3.5 Modifications for BallConstrained Uncertainties indeed increases its computational complexity. Two
experiments were conducted using the aforementioned
The case of ballconstrained uncertainties (w W2 ) formulation that assumes ballconstrained uncertain-
can also be dealt within an analytical formulation, ties.
while a relatively more complex optimization prob-
lem will be generated. Let the following partitioning 3.6 Alternative Feedback Prediction Parametrization
[38]:
The main problem with the aforementioned mini-

1T max formulations is the excessive amount of deci-
 
T T sion variables and constraints. The reason is mostly
C (G + BL) = 2 , i = i1 i2 ...iN (44)
T T T
.. related with the high dimensional parametrization of
.
T the matrix L as given in Eq. 22. In addition to this,
  1   the matrices and introduce an additional large
Ex (G + BL) T T
Eu L = 2 , i = i1 i2 ...iN (45)
T T T
set of variables. Due to the fact that the number of
..
. free variables is very high, a proper alternative strat-
egy has to be employed in order to obtain formulations
As the vectors ij and ij depend on L, the relaxed of lowercomplexity and therefore more applicable in
optimization may only be formulated by introducing largescale problems.
J Intell Robot Syst (2016) 81:443469 451

Towards this direction, a cheaper parametrization a last resort safety mechanism that ensures at the
can be derived by exploiting the Toeplitz Structure in control level that the vehicle will not crash into
LX (I BLX ) 1G , where LX = diag(L, ...L). As obstacles or enter nofly zones. Let Pw be a polyhe-
presented in [38], a cheaper formulation may take the dron that describes the total workspace of the robot,
p
following form: Po1 , Po2 , ..., Po guarding polyhedra that enclose the
known obstacles and also provide additional thresh-
0 0 0 0
olds to account for the vehicle dimensions and some
L1 0 0 0
further safety bounds that allow limited penetration.
L2 L 0 0
L= 1 (52) The safe robot workspace is defined as the Pontrya-
. .. . . . . ..
.. . . . . p
gin difference Ps = Pw \ {Po1 , Po2 , ..., Po }. In
LN1 LN2 L1 0 order to enforce operation within Ps , the following
adaptation in the optimization problem is required:
As shown from its matrix structure, such a
parametrization is structurally less expensive. As a min (54)
V ,L,, ,
result it corresponds to a significant improvement
compared to the initial parametrization [38] espe- s.t. [x, y, z] Ps
cially as the prediction horizon increases. To illustrate Inequalities (43)
the benefit of allowing such increased prediction
where all the matrices have to be reformulated for the
horizons, a specific experimenet was conducted using
case of the coupled translational dynamics. The poly-
this alternative parametetrization and is presented in
hedric constraint leads to maximum repelling control
Section 4.
actions as the vehicle enters the obstaclesguarding
polyhedra and guides the vehicle back to the collision
3.7 Obstacle Avoidance
free area. To achieve a smooth repelling action, rela-
tively high prediction horizons are required. Despite
The RMPC framework may also be extended to pro-
the increased complexity of this problem, once an
vide 3D obstacle avoidance capabilities by combining
explicit solution is computed, its realtime imple-
the translational dynamics models in Eqs. 2, 3, and 4
mentation remains very efficient. Finally, it is noted
into a unique statespace representation and introduc-
that such a obstacle avoidance safety mechanism is
ing an additional polyhedric constraint. For this case,
not proposed as an alternative to path planning for
a single optimization problem is solved and the sys-
collisionfree trajectories: it is rather a last resort
tem dynamics take the following general form for both
backup safety action in case the robot is - due to some
rotorcraft configurations considered within this work
error or disturbance approaching a known obstacle
(indices indicating the kind of platform are dropped
or a forbidden nofly zone. To illustrate this approach,
for simplicity of notation):
an example where a collisionfree path is computed
x3D = A3D x3D + B3D u3D
P (53) based on an RRT-star [26]t planner is computed but
x3D = [xlon xlat xalt ], u3D = [ulon ulat ualt ] a forcible disturbance eventually pushes the vehicle
towards an obstacle is presented.
Alon 0 0 Blon
A3D = 0 Alat 0 , B3D = Blat
3.8 Multiparametric Explicit Solution
0 0 Aalt Balt
Note that in order to achieve fast realtime implemen- The presented RMPC strategy requires the solution of
tation an explicit controller is going to be computed a linear or second order cone program at each time
and therefore only obstacles of known dimensions step. Although efficient linear and cone programming
and location are supported within this work. Although solvers exist, the solution time remains considerable
this may be considered conservative, quite often some especially as the state dimensions and the prediction
firstlevel knowledge of the environment is avail- horizon increases. Due to the fact that it is uncommon
able in many reallife applications such as infras- for miniature aerial vehicles to have very powerful
tructure inspection. For the part of the environment onboard processing resources available (or they are
that is known, the proposed framework can provide also encumbered by other demanding tasks such as
452 J Intell Robot Syst (2016) 81:443469

5 10000

4 9000

3 8000

2 7000

1 6000

0 5000

1 4000

2 3000

3 2000

4 1000

5 0

10 10
8 8
6 6
4 4
200 200
2 2
0 100 0 100
2 2
0 0
4 4
6 100 6 100
8 8
200 200
10 10

Fig. 4 Example illustration of the piecewise affine optimizer and the values of the objective function

localization and mapping tasks), an explicitoffline region r = {xk |Hr xk Kr } computed and
equivalent representation of this controller is impor- described in hrepresentation during the explicit
tant, as it enables seamless realtime implementation. controller derivation [31]. The controller is equiv-
Towards this goal, multiparametric approaches may alently translated to a piecewise affine mapping
be employed. Multiparametric programming is related between feedback gains and affine terms Fr , Zr
with optimization problems where the elements to and corresponding polyhedric regions r . Figure 4
be optimized depend on one or more parameters and presents the values of this piecewise affine optimizer
the parameters space can be systematically subdivided as well as the objective function values (value func-
into characteristic regions in which the optimal value tion) for the case of the ASLquad dynamics. For this
as well as the optimizer can be expressed as explicit visualization, the actuating dynamics of the transla-
functions of the parameters. Such a multiparametric tional motions have been neglected in order to achieve
explicit solution becomes possible for the MPC prob- a 2D representation and therefore allow the illustra-
lem due to the fact that the control action of such tion of the piecewise affine optimizer and the value
receding horizon optimization problems takes a piece- function in a 3D plot. This explicit controller is equiv-
wise affine form [32]: alent to its online counterpart, in the sense that the
same state trajectories will produce identical control
uk = Fr xk + Zr , if xk r (55) actions, and therefore share the same stabilizing and
optimality properties [6].
where i , r = 1, ..., N r are the regions of the reced- Such a fact enables fast realtime execution even in
ing horizon control strategy applied on each of to the microcontrollers with very limited computing power.
longitudinal, lateral, or vertical dynamics or their cou- In this framework, the realtime code is described in
pled 3D version. The rth control law is valid if the Algorithm 1 and corresponds to an extension of the
state vector xk is contained in a convex polyhedral table traversal algorithm [31].
J Intell Robot Syst (2016) 81:443469 453

3.9 Yaw and Attitude Compensation 4 Experimental Studies

The aforementioned control strategy is based on 4.1 Initial Simulationsbased Evaluation


the initial modeling representation and the overall
assumption that longitudinal and lateral motions can For initial evaluation purposes of the designed
be treated independently while navigation of the vehi- RMPC controllers, a simulation-based approach was
cle along the horizontal plane can take place by only followed: The nonlinear dynamics models of the
commanding these twoaxes of the system and with ASLquad and the UPATTTR (as elaborated in [8, 44]
no need to coordinate the yaw motion. Therefore, dur- respectively) were employed and augmented to
ing the modeling and control phase, the yaw angle include a pointmass slung load, considered to be
was considered to be zero. It is this reason why now an suspended from the vehicles Center of Mass with
additional computation has to take place, in order to a non-deformable linkage, introducing phasedelayed
compensate for the possibly rotated yaw angle () and disturbances to the UAV as it freely moves like a
the effect of roll and pitch to the projected thrust. More pendulum with respect to the vehicles BodyFixed
specifically, the previously derived urlon , urlat , uralt ref- Frame. Figure 6 indicates the simulation results of the
erence signals should now be rotated according to the implemented RMPC for the UPATTTR, and for a step
following transformation: {x r , y r } = {1, 1}m reference maneuver while carrying
the modeled slung load.

ulon = ulat cos() ulon sin(), ulon =
r, A comparison to non-predictive robust compensat-

ulat = ulat sin() + ulon cos(), ulat = r, ing schemes is also given. The employed controller

, , is the one presented in [36], wherein each i


ualt = ualt /(cos cos ), ualt = T r
{x, y, z, } subsystem input ui = uN i + ui consists
R
(56) N
of a nominal part ui (for nominalsystem track-
, ing) and a robust compensating part uR
It is these final ulon , ulat , ualt , which correspond to i (to restrain
the pitch, roll and thrust references r, , r, , T r , any i disturbances). As indicated in [36], the esti-
that are applied to the system as indicated in mated disturbances i are low-pass filtered to form
Fig. 5. the final robust compensating inputs uR i = Fi i ,
454 J Intell Robot Syst (2016) 81:443469

wlon,wlat,walt
xlonr,xlatr,xaltr r, act1
RMPCs r, attitude act2 Unmanned
x ,x ,x
r=0 control Rotorcraft lon lat alt
r actn Dynamics

Fig. 5 The RMPC scheme for the problem of flight control of unmanned rotorcrafts. The terms act1 , act2 , ..., actn correspond to an
abstract notation that refers to the actuators of each platform

0.5 1.05
1
0.95
0
2.5 3 3.5 4 4.5 5 5.5 6 6.5 7 7.5
0 1 2 3 4 5 6 7

0.5 1.05
1
0.95
0
2.5 3 3.5 4 4.5 5 5.5 6 6.5 7 7.5
0 1 2 3 4 5 6 7

0.2

0.2

0.4

0.6

0 1 2 3 4 5 6 7

1.2

1.05

1
1

0.8
0.95

0.95 1 1.05
0.6 0.2

0
0.4
0.2

1
0.2 0.4

0 0.5
0

0.5
0.2 0
1 y [m]
0.2 0 0.2 0.4 0.6 0.8 1 1.2
x [m]

Fig. 6 Simulated RMPC performance for suspended slung load and a {x r , y r } = {1, 1}m step maneuver
J Intell Robot Syst (2016) 81:443469 455

yielding results as indicated in Fig. 7, wherein the per- and the UPATTTR. These two systems correspond to
formance deterioration with respect to the proposed indicative examples of the class of rotorcraft systems
RMPC scheme is discernable. with underactuated dynamics, as well as the special
rotorcraft configurations with directly actuated longi-
4.2 Experimental Platforms tudinal dynamics. Both of these systems share a sim-
ilar software architecture which is depicted in Fig. 8.
The experiments presented below were conducted Rapid deployment of such controllers is feasible by
using two different vehicles, namely the ASLquad implementing a new block of MATLAB/Simulink

0.5 1.05
1
0.95
0
3 4 5 6 7 8 9 10
0 1 2 3 4 5 6 7 8 9 10

0.5 1.05
1
0.95
0
3 4 5 6 7 8 9 10
0 1 2 3 4 5 6 7 8 9 10

0.2

0.2

0.4

0.6

0 1 2 3 4 5 6 7 8 9 10

1.2

1.05

1
1

0.8
0.95

0.95 1 1.05
0.6 0.2

0
0.4

0.2
1
0.2
0.4

0.5
0
0

0.5
0.2 0
1
0.2 0 0.2 0.4 0.6 0.8 1 1.2

Fig. 7 Simulated Robust Compensating Input Scheme performance for suspended slung load and a {x r , y r } = {1, 1}m step maneuver
456 J Intell Robot Syst (2016) 81:443469

Fig. 8 The software


framework deployed on the Joystick
ASLquad Simulink Simulink
RMPC Auto-Code
Attitude
Interface
Attitude
Control Bridge Control
MATLAB

RC
Node
UAS
RC
Perception
Joystick
Perception
Node
Node

functions, which are then autocode generated to ATOM based higherlevel electronics which allow the
C++ functions and embedded to the autopilot soft- fast deployment of multiple control architectures. It
ware framework. To further support the community, has a mass of mq = 0.65 kg and a protective shroud
a MATLABbased implementation of the proposed mechanism for increased safety of the system and the
framework that can be also used for auto-code genera- surroundings. The arm length of the ASLquad is 0.3m.
tion purposes is released and is available at the follow- The UPATTTR setup is depicted in Fig. 9. Its
ing repository: https://github.com/kostas-alexis/rmpc main control unit is based on an Intel ATOM Z530
mav 1.6 GHz CPU, while the handling of the lowlevel
The ASLquad is a quadrotor that employs an in communication and control signals is achieved via
house designed airframe and the electronic compo- interfacing to an ARM Cortex M3based microcon-
nents and actuation subsystems of an AscTec hum- troller. The onboard estimation is based on measure-
mingbird [4]. These electronics include an attitude ment data acquired from the following sensors: a)
autopilot which was then interfaced with the Intel an MTiG Attitude and Heading Reference System

UPAT-TTR
Rotor Servo ARM m/c Kontron pITx WiFi SSD

MTi-G AHRS

Motor ESC

Optical Flow
USB WebCam
RPM Encoder Sonar

Fig. 9 The UPATTTR hardware setup


J Intell Robot Syst (2016) 81:443469 457

(AHRS) for providing KalmanFiltered data for the estimation scheme can be found in [44]. Finally, it is
systems rotational states {, , } and their deriva- noted that the overall mass of UPATTTR is mt =
tives, as well as for the translational accelerations 2.2 kg, its length is 0.75 m, its width 0.6m and its
{x, y, z}, b) a Sonar sensor for absolute altitude z height is 0.2 m.
measurement, c) RPMencoder units which are used In both systems, an attitude controller was already
to estimate the rotors thrust forces via their rota- available either from the autopilot manufacturer
tional speed (based on prior benchtest mapping), and (ASLquad) or based on previous work of a subset
finally d) a groundpointed PS-Eye camera which of the authors (UPATTTR) [44]. For the case of the
is used to acquire sequential frames of the ground ASLquad, position feedback was provided using a
plane as the vehicle performs aerial maneuvers, and Vicon motion capture system while the attitude states
consequently apply an Optical Flow [21] calculation were estimated based on the onboard Inertial Mea-
scheme to estimate its translational velocities (based surement Unit (IMU). On the other hand, for the case
on its pinhole camera model [20] obtained via cam- of the UPATTTR, complete onboard state estima-
era calibration). The aforementioned measurements tion was employed by solely relying on its sensor
are further fused onboard via Exteded Kalman Filter- suite and appropriate sensor fusion algorithms [44], as
ing, to acquire estimates of the vehicles 6-Degrees previously mentioned.
of Freedom absolute pose {x, y, z, , , }, which are As mentioned, identification of the actuation
used in trajectorytracking. Further elaboration of the dynamics of translational motions as well as esti-
specific methodology (hardware and softwarewise) mation of the linear velocities damping parameters
followed to implement this onboard autonomous state are required in order to derive a model of increased

20 0.2

0.15
40
0.1
60
0.05

80 0

0.05
100
0.1
120
0.15

140 0.2
1 0 1 0 10 20 30 40 50 60
10 10 10

0 0.5

20

40
0
60

80

100 0.5
1 0 1 0 10 20 30 40 50 60
10 10 10

30 0.6

20 0.4

10 0.2

0 0

10 0.2

20 0.4

30 0.6

40 0.8
1 0 1 0 10 20 30 40 50 60
10 10 10

Fig. 10 Identification results for the two platforms. The first two rows represent the ASLquad attitude closedloop and longitudinal
velocity dynamics, while the last row represents the UPATTTR longitudinal velocity
458 J Intell Robot Syst (2016) 81:443469

accuracy that will then aid the control synthesis. constraints summarize for both platforms what was
Utilizing wellestablished frequency domain identifi- incorporated in the RMPC computation:
cation methods [37], sufficientquality identification
results were derived as shown in Fig. 10 where a x 1.5m/s
y 1.5m/s
comparison of indicative experimentallyderived fre-
z 1.5m/s
quency and time responses against their modelbased
/4rad
estimates is presented. Similar results hold for the
/4rad
other degrees of freedom and the thrust dynamics
r /4rad
response which is dominated by the motors character- 
 r

istic time constant. Note that the required identifica- I77 077 /4rad
,
tion step is fast and can be efficiently executed with 077 I77 x 1.5m/s

three computercommanded frequency sweeps on the y 1.5m/s

roll, pitch and throttle commands or corresponding z 1/5m/s

piloted commands. The duration of the experiment /4rad

/4rad
should be approximately 8 times the period of the r
slowest expected oscillation and a frequency sweep /4rad
that starts from the slowest frequency up to the high- r /4rad

est expected for the attitude loop should be employed x 1.5m/s
[40]. y 1.5m/s

z 1.5m/s

/4rad
4.3 Experimental Results
r /6rad

 r

In order to evaluate the capabilities of the proposed I66 066 /4rad
(57)
RMPCframework, a set of challenging testcases 066 I66 x 1.5m/s

were considered, namely: a) trajectory tracking sub- y 1.5m/s

ject to wind disturbances, b) slung load operations, z 1/5m/s

a number of which additionally subjected to strong /4rad
r
forcible disturbances, c) unpredicted collisions with /6rad
the environment, as well as d) trajectories that demon- r /4rad
strate the obstacle avoidance capabilities of the control
framework. Most experiments are validated utiliz- Regarding the first experimental testcase, a 80W
ing both unmanned rotorcrafts under consideration, electric fan was pointed to the ASLquad while the
providing results indicative of the flight robustness RMPC was acting in order to track a helical path
achieved by the RMPCframework. The testcases despite the turbulent wind disturbance. As can be
requiring a protective shroud or global positioning observed from the upper part of Fig. 11, the tracking
data were conducted using the ASLquad due to its response remains precise and only a minor influence
mechanical structure and the availability of a Vicon from the external disturbance is observed. In this
motion capture system. Fully autonomous solutions to experiment, the reference trajectory is equally sam-
such missions can be made feasible using advanced pled and a constant velocity is commanded with a zero
SLAM algorithms [53] and/or GPS data fusion. When acceleration profile. A similar experiment was con-
not mentioned otherwise, boxconstrained uncertain- ducted with the UPATTTR, where the turbulent wind
ties and the feedback predictions parametrization disturbance was produced by a 180W electric fan,
introduced in Section 3.1 are employed. For all the this time for a slowlyadvancing timestampedhelical
experimental testcases, the sampling time was set to trajectory. Due to the slowly advancing reference tra-
Ts = 0.08s for the ASLquad and Ts = 0.1s for the jectory and the constant presence of turbulent wind the
UPATTTR, the prediction horizon was set to N = 6 result is less smooth; however the achieved response
for both of them, while the following state and input by the RMPC remains robustly accurate, as shown
J Intell Robot Syst (2016) 81:443469 459

Fig. 11 Trajectory tracking


subject to wind disturbances
using the RMPC framework
applied to both ASLquad
and UPATTTR. The upper 0.5
plot corresponds to the 0.4
ASLquadbased 0.3
experiment, while the lower 0.2

plot refers to the 0.1

UPATTTR data. Video 0

results are available in the 0.1

supplementary video file 0.2

0.3

0.4

0.5

0.6

0.4

0.2

0.2
0.6
0.4
0.4
0.6 0.2

0.8 0
0.2
1
0.4
1.2 0.6

0.5

0.5

1.5

0.5

0.8
0 0.6
0.4
0.2
0.5 0
0.2
0.4
0.6

in the lower part of Fig. 11. For these experimental is forcibly disturbed, highly precise position hold
studies, as well as for all the other cases presented and disturbance rejection results were achieved for
within this paper, a video that contains recordings of both rotorcraft configurations. This is despite the fact
the flights is attached as a supplementary file. that the slung load introduces disturbances that are
Consequently, the capabilities of handling slung phasedelayed compared to the vehicle states and the
loads were evaluated. As shown in Fig. 12 where controller is not augmented to incorporate the loads
the external disturbance is applied onto the slung motion. In the case of the ASLquad a 0.16 kg load was
load, as well as in Fig. 13 where the rotorcraft body utilized with a string length of 0.65 m, while for the
460 J Intell Robot Syst (2016) 81:443469

0.2
, r
xy
0.1 x
y

0.1

0.2
0 1 2 3 4 5 6 7 8 9 10

r
z
1.1 z

0.9

0.8
0 1 2 3 4 5 6 7 8 9 10

0.5
r

0.5
0 1 2 3 4 5 6 7 8 9 10

0.1
r

0.05

0.05

0.1
0 1 2 3 4 5 6 7 8 9 10

0.2
xr,y
r
0.1 x
y
0

0.1

0.2
0 2 4 6 8 10

1.1
zr
1.05 z

0.95

0.9
0 2 4 6 8 10

0.1
r

0.05
r

0.05

0.1
0 2 4 6 8 10

0.1
r

0.05

0.05

0.1
0 2 4 6 8 10

Fig. 12 RMPC performance against disturbance of the slung experiment, while the four lower plots refer to the respec-
load which induces unpredicted disturbances on the vehi- tive UPATTTR experiment. Video results are available at the
cle. The four upper plots correspond to the ASLquadbased supplementary video file
J Intell Robot Syst (2016) 81:443469 461

0.5
x,yr
x
y

0.5
0 1 2 3 4 5 6 7 8

r
z
1.1 z

0.9

0.8
0 1 2 3 4 5 6 7 8

0.3
r

0.2

0.1

0.1

0.2
0 1 2 3 4 5 6 7 8

0.05
r

0.05
0 1 2 3 4 5 6 7 8

r
x ,yr
0.6 x
0.4 y

0.2
0
0.2
0 1 2 3 4 5 6 7 8 9 10

1.1
r
z
z

0.9
0 1 2 3 4 5 6 7 8 9 10

0.2
r

r


0

0.2
0 1 2 3 4 5 6 7 8 9 10

0.1
r

0.1
0 1 2 3 4 5 6 7 8 9 10

Fig. 13 RMPC performance against disturbance of the vehi- while the four lower plots refer to the respective UPATTTR
cles body during slung load operations for both platforms. The experiment. Video results are available at the supplementary
four upper plots correspond to the ASLquadbased experiment, video file
462 J Intell Robot Syst (2016) 81:443469

UPATTTR a 0.33kg load, suspended at 0.45m, was achieved when continuously tracking a square refer-
employed. ence trajectory for 10 times while the same slung load
Once the capabilities of the proposed RMPC to is attached onto the quadrotor platform. As shown,
handle slung load operations even while subjected efficient and robust results were derived, indicative
to strong disturbances were verified, the problem of of the capabilities of the proposed control scheme
trajectory tracking during slung load operations (i.e. to effectively execute such operations. Similarly to
transportation and delivery of goods in a Search the helical trajectory tracking testcase, the reference
andRescue scenario or as part of a product delivery trajectory is equally sampled and a constant veloc-
mission), was examined. Figure 14 presents the results ity is commanded with a zero acceleration profile.

Fig. 14 The upper part of


the figure presents
trajectory tracking during
slung load operations using
the ASLquad: the controller 1.5

tracks the same square 1.4

trajectory for 10 times with 1.3

minimal deviations from the 1.2

1.1
reference. The lower part of
1
the figure presents
0.9
trajectory tracking during 0.8
slung load operations using 0.7

the UPATTTR: the 0.6

controller tracks the same 0.5

square trajectory for 5 times


with minimal deviations
1
from the reference. Video
results are available at the
0.5
supplementary video file
0.6
0 0.4
0.2
0

0.5 0.2
0.4
0.6

1.5
1.4
1.3
1.2
1.1
1
0.9
0.8
0.7
0.6
0.5

0.5

0.6
0 0.4
0.2
0
0.5 0.2
0.4
0.6
J Intell Robot Syst (2016) 81:443469 463

The UPATTTR was also employed to conduct an more agile maneuvers, increased prediction horizon
analogous experiment, and the results are shown in and proper commanding of the velocity and attitude
the bottom subset of Fig. 14. For both these exper- profiles was considered. Towards employing a longer
iments, ballconstrained uncertainties are assumed, prediction horizon (N = 16) while keeping the com-
and the multiparametric solution produces control putational complexity of the multiparametric problem
actions identical to that of the second order cone at a sufficiently low level, the alternative parametriza-
programming problem formulated in Subsection 3.5. tion of the feedback predictions was employed. More-
In the last presented trajectory tracking experi- over, in order to compute the velocity and attitude
ments, a slow response was commanded, as is typi- profiles, a simulation of the closedloop response was
cally expected during slung load operations. However, derived by employing a controller slightly more agile
more agile maneuvers could also be considered even than the one used in practice (higher penalization of
during slung load operations; such a sequence proves the Q matrices). The derived simulation response of
indicative of the RMPC frameworks performance in the velocity and attitude profiles were then fed to the
handling disturbances during a more aggressive oper- realvehicle along with the position step responses.
ation. To this purpose, two yaxis step responses Figure 16 presents the new result for a 3m step
were conducted. In the first case, a step reference of response and a peak velocity of more than 2m/s. As
1.25m was commanded with the velocity and attitude shown, an improved response with no overshoot was
references being conservatively set to 0. The results achieved.
are shown in Fig. 15 and as depicted, a satisfac- In a similar direction, the UPATTTR was
tory response was derived although some overshoot employed to conduct an xstep response along the
is observed. To alleviate this problem and even allow directly actuated longitudinal dynamics during slung

0.2 r
x
0.1 x

0
0.1
0.2
0 1 2 3 4 5 6 7 8 9 10

2
r
y
1.5
y
1

0.5

0
0 1 2 3 4 5 6 7 8 9 10

zr
1.1 z

0.9

0.8
0 1 2 3 4 5 6 7 8 9 10

0.4
r

0.2

0.2

0.4
0 1 2 3 4 5 6 7 8 9 10

0.1
r
0.05

0.05

0.1
0 1 2 3 4 5 6 7 8 9 10

Fig. 15 Lateral step response during slung load operations: minimal overshoot is observed. Video results are available at the
supplementary video file
464 J Intell Robot Syst (2016) 81:443469

xr
0.1
x
0.2

0.3

0.4

0.5
0 1 2 3 4 5 6 7 8

4
yr
3 y

0
0 1 2 3 4 5 6 7 8

1.5
zr
z

0.5
0 1 2 3 4 5 6 7 8

0.4

0.2

0.2

0.4
0 1 2 3 4 5 6 7 8

0 1 2 3 4 5 6 7 8

Fig. 16 Lateral fast step response during slung load operations with appropriate velocity and attitude references: no overshoot is
observed. Video results are available at the supplementary video file

load operations, and the recorded results are depicted contact with the environmental surface without tip-
in Fig. 17. The step response presents practically no ping over. Once such a behavior is ensured, a safe
overshoot while the employed velocities are relatively maneuver to fly away from the obstacle may be exe-
low (due to the limitations of the onboard opticflow cuted. As shown in Fig. 18 a safe physical interaction
based velocity estimation). Essentially this wide set response is achieved. It is noted that the reference is
of experiments validates that such complex tasks may set close to the obstacle, and therefore the approach-
be safely executed by relying on such robust flight ing velocity is small. If this is not the case, then
controls, even without the need for additional state the main risk is related with harsh collisions, the
feedback and specialized controllers tailored to the effect of which may be compensated with the design
specific task. This is a major advantage that allows the of a compliant protective shroud and/or specialized
versatile operations of such aerial robots in a variety controllers [1].
of applications not fully defined during their design Furthermore, obstacle avoidance capabilities of the
phase. proposed RMPCframework were evaluated, utilizing
As a third testcase, the scenario of colliding with a testcase where prior knowledge of the environment
the environment is considered. In this scenario the exists; the respective results are shown in Fig. 19.
ASLquad is driven towards a reference point, with- For the purposes of this experiment the ASLquad
out knowledge of an obstacle that blocks its motion. was employed, since ground truth data were available
Consequently, the vehicle undergoes an undesired and from the Vicon external tracking system. As long as
unpredicted collision with the environment. In such the workspace map is known, a guarding polyhedron
a case, the robust control has to be able to provide Po , is defined and the augmented RMPC strategy is
control actions that smoothly react and as long as applied. As demonstrated, a last resort obstacle avoid-
a protective shroud is available keep the vehicle in ance mechanism is achieved, ensuring the safety of
J Intell Robot Syst (2016) 81:443469 465

2
1.5 x
r

1 x
0.5
0
0 2 4 6 8 10 12 14 16

0.2
0.1 y
r

0 y
0.1
0.2
0 2 4 6 8 10 12 14 16

1.2
1.1 z
r

1 z
0.9
0.8
0 2 4 6 8 10 12 14 16

0.2 r
0.1
0
r

0.1
0.2
0 2 4 6 8 10 12 14 16

0.1
r

0
0.1
0 2 4 6 8 10 12 14 16

Fig. 17 Lateral step response of the UPATTTR during slung load operations: a slow but with minimum overshoot response is derived.
Video results are available at the supplementary video file

the system. Note that complete onboard autonomy for The final testcase demonstration consists of a
the execution of such an experiment is feasible via the more comprehensive operational scenario for an
utilization of SLAM systems [53] or possibly GPS. unmanned rotorcraft platform: the ASLquad executes

Distance from the reference

Collision Obstacle
Position

Fig. 18 RMPCbased handling of unpredicted collisions with the environment. Video results are available at the supplementary video
file
466 J Intell Robot Syst (2016) 81:443469

Fig. 19 Evaluation of the


obstacle avoidance
capabilities that are
incorporated on the Robust
MPC framework. Video
results are available at the
supplementary video file

slung load navigation, while having to avoid an obsta- environment, with the proposed framework providing
cle, and additionally operating within a windy envi- the advantage of having the presence of known obsta-
ronment. In this context, a collisionfree trajectory cles encoded and incorporated into the control compu-
is derived first using an implementation of the RRT tation structure, further augmenting the UASs flight
planner as presented in [26]. The computed reference safety properties. While the RRT planner provides
path is shown in blue color in Fig. 20. From a control an offline generated reference trajectory that does not
perspective, this example illustrates the RMPC capa- intrude into the obstacles polyhedric envelope, the
bility to handle the slung load operation in a windy evolution of the online experiment provides no such

Fig. 20 Demonstration of a
complex mission: the
ASLquad tracks a collision
free trajectory computed
using RRT methods while
carrying a slung load in a
turbulent environment. The
external disturbances lead
to a deviation towards the
obstacles and the
augmented RMPC acts in
order to ensure that the
control actions drive the
vehicle back to the
collisionfree area. Video
results are available at the
supplementary video file
J Intell Robot Syst (2016) 81:443469 467

guarantees: at a specific time during navigation, the 2. Alexis, K., Nikolakopoulos, G., Tzes, A.: Model predictive
vehicle enters the polyhedron that bounds the obstacle quadrotor control: attitude, altitude and position experimen-
tal studies. IET Control Theory Appl. 6(12), 18121827
as a consequence of the external distbances. Relying
(2012)
on the obstacle avoidance capabilities of the RMPC, 3. Alexis, K., Tzes, A.: Revisited Dos Samara Unmanned
the vehicle reacts accordingly and avoids the obsta- Aerial Vehicle: Design and control. In: 2012 IEEE Inter-
cle in time. Furthermore, the helicopter is forcefully national Conference on Robotics and Automation (ICRA),
pp. 36453650 (2012)
disturbed during the path navigation using a stick.
4. Ascending Technologies GmbH, http://www.asctec.de/
5. Bemporad, A.: Reducing conservativeness in predictive
control of constrained systems with disturbances. In: Pro-
5 Conclusions ceedings of the 37th IEEE Conference on Decision and
Control, vol. 2 (1998)
6. Bemporad, A.: Modeling, control, and reachability analy-
The challenge of increasedreliability flight control sis of discretetime hybrid systems. University of Sienna
for rotorcraft UAS was addressed by the design, devel- (2003)
opment and experimental evaluation of a robust model 7. Bouabdallah, S., Siegwart, R.: Full control of a quadrotor.
predictive control framework. The control scheme In: 2007 IEEE/RSJ International Conference on Intelligent
Robots and Systems, pp. 153158 (2007)
ensures the minimum deviation for the worstcase dis- 8. Bouabdallah, S.: Design and control of quadrotors with
turbance subject to the physical limitations of the sys- application to autonomous flying, Ph.D. dissertation, STI
tem was elaborated in theory, relaxations for efficient School of Engineering, EPFL, Lausanee (2007)
convex optimizationbased solutions were discussed 9. Burri, M., Nikolic, J., Hurzeler, C., Caprari, G., Siegwart,
R.: Aerial service robots for visual inspection of thermal
and the principles for its realtime implementation in power plant boiler systems. In: 2012 2nd International
explicit form were separately studied. The experimen- Conference on Applied Robotics for the Power Industry
tal implementation of the derived RMPC was evalu- (CARPI), pp. 7075 (2012)
ated utilizing two unmanned rotorcraft configurations, 10. Camacho, E.F., Bordons, C.: Model Predictive Control.
Springer (2003)
namely one employing the typical underactuated actu- 11. Campo, P.J., Morari, M.: Robust model predictive control.
ation principle, and one characterized by the excep- Am. Control Conf. 1987, 10211026 (1987)
tional feature of direct actuation over its longitudinal 12. Cheviron, T., Plestan, F., Chriette, A.: A robust guidance
dynamics. The set of disturbance sources examined and control scheme of an autonomous scale helicopter
in presence of wind gusts. Int. J. Control 82(12), 2206
included: wind gusts, external forceful disturbances (2009)
while additionally carrying slung loads, and physical 13. Dalamagkidis, K., Valavanis, K., Piegl, L.: English
collisions. The derived results conclusively indicate Autonomous autorotation of unmanned rotorcraft using
the disturbanceattenuation and responseprecision nonlinear model predictive control. J. Int. Robot. Syst.
57(1-4), 351369 (2010)
effectiveness of the developed framework. The addi- 14. Doherty, P., Rudol, P.: A UAV search and rescue sce-
tional feature of obstacle avoidance was incorporated nario with human body detection and geolocalization. In:
into the control structure and experimentally studied, Advances in Artificial Intelligence, vol. 4830/2007, pp. 1-
resulting in sufficiently safe maneuvers while navigat- 13 (2007)
15. Flores, G., Escareno, J., Lozano, R., Salazar, S.: Quad-
ing into areas with obstacles. The overall evaluation tilting rotor convertible mav: Modeling and real-time hover
confirms the capacity of such a robust control strategy flight control. J. Int. Robot. Syst. 65, 457471 (2012)
to effectively handle forcible disturbances and ensure 16. Forte, F., Naldi, R., Serrani, A., Marconi, L.: Control of
the minimum possible deviation from a reference tra- modular aerial robots: Combining under- and fully-actuated
behaviors. In: IEEE Conference on Decision and Control,
jectory, while respecting the imposed input and state pp. 11601165 (2012)
constraints. 17. Girard, A., Howell, A., Hedrick, J.: Border patrol and
surveillance missions using multiple unmanned air vehi-
cles. In: 43rd IEEE Conference on Decision and Control,
2004. CDC. (2004)
References
18. Guerreiro, B., Silvestre, C., Cunha, R.: Terrain avoidance
nonlinear model predictive control for autonomous rotor-
1. Alexis, K., Huerzeler, C., Siegwart, R.: Hybrid modeling craft. J. Int. Robot. Syst. 68(1), 6985 (2012)
and control of a coaxial unmanned rotorcraft interacting 19. Han, J., Chen, Y.: Multiple uav formations for cooper-
with its environment through contact. In: 2013 Interna- ative source seeking and contour mapping of a radia-
tional Conference on Robotics and Automation (ICRA), tive signal field. J. Intell. Robot. Syst., 110 (2013).
pp. 53975404, Karlsruhe, Germany (2013) doi:10.1007/s10846-013-9897-4
468 J Intell Robot Syst (2016) 81:443469

20. Hartley, R.I., Zisserman, A. Multiple View Geometry in 38. Loefberg, J.: Minimax approaches to robust model pre-
Computer Vision, 2nd. Cambridge University Press (2004) dictive control, Ph.D. dissertation, Linkoping University,
21. Horn, B., Schunck, B.: Determining optical flow. Artif. Int. Linkoping, Sweden, 2003. [Online]. Available: http://www.
17, 185204 (1981) control.isy.liu.se/research/reports/Ph.D.Thesis/PhD812.pdf
22. Huerzeler, C., Alexis, K., Siegwart, R.: Configurable real 39. Cannon, M., Li, S., Cheng, Q., Kouvaritakis, B.: Effi-
time simulation suite for coaxial rotor uavs. In: 2013 Inter- cient robust output feedback mpc. In: Proceedings of
national Conference on Robotics and Automation (ICRA), the 18th IFAC World Congress, vol. 18, pp. 79577962
pp. 309316. Karlsruhe, Germany (2013) (2011)
23. Huerzeler, C.: Modeling and design of unmanned rotorcraft 40. Tischler, M.B., Remple, R.K.: Aircraft and Rotorcraft Sys-
systems for contact based inspection, Ph.D. dissertation, tem Identification: Engineering methods with Flight-Test
Swiss Federal Institute of Technology Zurich (2013) examples. American Institute of Aeronautics and Astronau-
24. Hunt, E.R., Hively, W.D., Fujikawa, S.J., Linden, D.S., tics (AIAA)
Daughtry, C.S.T., McCarty, G.W.: Acquisition of nir-green- 41. Merz, T., Kendoul, F.: Beyond visual range obstacle avoid-
blue digital photographs from unmanned aircraft for crop ance and infrastructure inspection by an autonomous heli-
monitoring. Remote Sens. 2(1), 290305 (2010) copter. In: IEEE/RSJ International Conference on Intelli-
25. Alexis, K., Nikolakopoulos, G., Tzes, A.: Switching model gent Robots and Systems (IROS), 2011, pp. 49534960
predictive attitude control for a quadrotor helicopter subject (2011)
to atmospheric disturbances. Control Engineering Practice 42. Nicol, C., Macnab, C., Ramirez-Serrano, A.: Robust adap-
(2011). doi:10.1016/j.conengprac.2011.06.010 tive control of a quadrotor helicopter. Mechatronics 21(6),
26. Karaman, S., Frazzoli, E.: Incremental sampling-based 927938 (2011)
algorithms for optimal motion planning, arXiv preprint 43. Orukpe, P., Jaimoukha, I.: A semidefinite relaxation for
arXiv:1005.0416 (2010) the quadratic minimax problem with application to hinf
27. Kendoul, F.: Survey of advances in guidance, navigation, model predictive control. In: 2007 46th IEEE Conference
and control of unmanned rotorcraft systems. J. Field Robot. on Decision and Control, pp. 177181 (2007)
29, 2 (2012) 44. Papachristos, C., Alexis, K., Tzes, A.: Model predic-
28. Kerma, M., Mokhtari, A., Abdelaziz, B., Orlov, Y.: Non- tive hoveringtranslation control of an unmanned tri
linear h inf control of a quadrotor (uav), using high order tiltrotor. In: 2013 International Conference on Robotics
sliding mode disturbance estimator. Int. J. Control 85(12), and Automation, pp. 54055412, Karlsruhe, Germany
18761885 (2012). doi:10.1080/00207179.2012.709656 (2013)
29. Kim, J., Kang, M., Park, S.: Accurate modeling and robust 45. Rathinam, S., Kim, Z., Sengupta, R.: Vision-based moni-
hovering control for a quadrotor VTOL aircraft. J. Int. toring of locally linear structures using an unmanned aerial
Robot. Syst. 57(1), 926 (2010) vehicle1. J. Infrastruct. Syst. 14(1), 5263 (2008)
30. Kingston, D., Beard, R., Holt, R.: Decentralized perime- 46. Mehra, S.G.R.K., Prasanth, R.K.: xv-15 tiltrotor flight con-
ter surveillance using a team of uavs. IEEE Trans. Robot. trol system design using model predictive control. In: IEEE
24(6), 13941404 (2008) Aerospace Conference. Snowmass at Aspen, CO, USA
31. Kvasnica, M., Rauova, I., Fikar, M.: Automatic code gener- (1998)
ation for real-time implementation of model predictive con- 47. Seifried, R.: Dynamics of Underactuated Multibody Sys-
trol. In: 2010 IEEE International Symposium on Computer- tems: Modeling, Control and Optimal Design. Springer
Aided Control System Design (CACSD), pp. 993998 (2014)
(2010) 48. Rudol, P., Doherty, P.: Human body detection and geolocal-
32. Kvasnica, M.: RealTime Model Predictive Control via ization for uav search and rescue missions using color and
MultiParametric Programming: Theory and Tools. VDM thermal imagery. In: Aerospace Conference, 2008 IEEE,
Verlag (2009) pp. 18 (2008)
33. Laraki, R., Lasserre, J.: Semidefinite programming for 49. Ryan, T., Kim, H.: Lmi-based gain synthesis for simple
minmax problems and games. Math. Program. 131(12), robust quadrotor control. IEEE Trans. Autom. Sci. Eng.
305332 (2012). doi:10.1007/s10107-010-0353-y 10(4), 11731178 (2013)
34. Lennart, L.: System Identification Toolbox T M Users 50. Self Dual Minimization: SeDuMi, http://sedumi.ie.lehigh.
Guide, Mathworks, vol. 3. Apple Hill Drive (2012) edu/
35. Liu, H., Bai, Y., Lu, G., Shi, Z., Zhong, Y.: Robust tracking 51. Stephen Boyd, L.V.: Convex Optimization. Cambridge Uni-
control of a quadrotor helicopter. J. Intell. Robot. Syst., 1 versity Press (2004)
14 (2013). doi:10.1007/s10846-013-9838-2 52. The MathWorks Inc., Aerospace Blockset 3 Users Guide
36. Liu, H., Li, D., Kim, J., Zhong, Y.: Real-time imple- (2009)
mentation of decoupled controllers for multirotor air- 53. Weiss, S., Achtelik, M., Lynen, S., Chli, M., Siegwart, R.:
crafts. J. Intell. Robot. Syst. 73(1-4), 197207 (2014). Real-time onboard visual-inertial state estimation and self-
doi:10.1007/s10846-013-9916-5 calibration of mavs in unknown environments. In: 2012
37. Ljung, L.: System Identification: Theory for the User, 2nd. IEEE International Conference on Robotics and Automa-
Prentice Hall, Inc., NJ, USA (1999) tion (ICRA), pp. 957964 (2012)
J Intell Robot Syst (2016) 81:443469 469

Dr. Kostas Alexis is a Senior Postdoctoral Researcher at the Anthony Tzes is Professor of the Electrical & Computer Engi-
Autonomous Systems Lab at ETH Zurich working in the direc- neering Department of the University of Patras (UPAT) in
tion of flight control and autonomous planning for aerial robots. Greece. He is a graduate of UPAT (85) and received his doctor-
Kostas obtained his PhD in the field of aerial robotics, con- ate from the Ohio State University (90). From 1990 till 1999 he
trol and collaboration from the University of Patras, Greece in was with Polytechnic Institute of NYU. His research interests
2011 and from the beginning of his research efforts he has been include cooperative control of networked mobile robots, surgi-
the author of more than 40 scientific publications, has received cal robots, and control engineering applications. Prof. Tzes has
several best-paper awards in the field of control and has par- received research funding from various organizations including
ticipated in multiple large-scale, multi-million Euros research NASA, the National (US) Science Foundation, and the Euro-
projects. He is passionate about robotic research and appli- pean Union (FP6, H2020). He has been the a member of the
cations and has the vision of facilitating aerial robots as an Greek committee of the European Control Association (EUCA),
integral part of a new kind of automation that operates within member at several committees of the International Federa-
the dynamic physical world. tion of Automatic Control (IFAC), and until 2008 Greeces
national representative to EUs FP7s thematic area Regions of
Knowledge, Research Potential and Coherent Development of
Christos Papachristos is a PhD Candidate at the Electri- Policies. He has served in various positions (Program Chairman
cal & Computer Engineering Department of the University (MIM00), Organizing Committee Chairman (ECC07), Chair-
of Patras (UPAT) in Greece. He works in the field of aerial man (MED2011)), and as IPC-member at several international
robotics, focusing on the design, implementation, flight con- conferences. He is the leader and principal investigator of
trol, autonomous state estimation, and perception of unmanned the Applied Networked Mechatronics Systems group. He has
aerial systems. He received his Diploma from the University authored more than 70 (200) papers published in international
of Patras, Greece in 2009 and began his research on umnanned technical journals (conferences) and served in the editorial
tiltrotor systems in 2011 with 20 scientific publications and a board of several journals.
participation in large-scale research projects at the European
and National level.

Roland Siegwart is full professor for autonomous systems at


ETH Zurich since July 2006. He has a Diploma in Mechani-
cal Engineering (1983) and PhD in Mechatronics (1989) from
ETH Zurich. In 1989/90 he spent one year as postdoctoral fel-
low at Stanford University. After that he worked part time as
R&D director at MECOS Traxler AG and as lecturer and deputy
head at the Institute of Robotics, ETH Zrich. In 1996 he was
appointed as associate and later full professor for autonomous
microsystems and robots at the Ecole Polytechnique Fdrale de
Lausanne (EPFL). During his period at EPFL he was Deputy
Head of the National Competence Center for Research (NCCR)
on Multimodal Information Management (IM2), co-initiator
and founding Chairman of Space Center EPFL and Vice Dean
of the School of Engineering. In 2005 he hold a visiting posi-
tion at NASA Ames and Stanford University. Roland Siegwart
is member of the Swiss Academy of Engineering Sciences and
board member of the European Network of Robotics (EURON).
He served as Vice President for Technical Activities (2004/05)
and is currently Distinguished Lecturer (2006/07) and AdCom
Member (20072009) of the IEEE Robotics and Automation
Society. He is member of the Bewilligungsausschuss Exzellen-
zinitiative of the Deutsche Forschungsgemeinschaft (DFG).
He is coordinator of two European projects and co-founder of
several spin-off companies.

You might also like