712
2, APRIL 2017
Dynamical System Based Robotic Motion Generation
With Obstacle Avoidance
Sotiris Stavridis, Dimitrios Papageorgiou, and Zoe Doulgeri
AbstractThe problem of real-time motion generation to a tar-
get with obstacle avoidance is considered. A second-order dynam-
ical system having the target as a unique globally asymptotically
stable equilibrium is modified in the presence of obstacles by an ad-
ditive signal whose design is based on the prescribed performance
control methodology. It is proved that obstacle avoidance and dy-
namical system stability are guaranteed and that the target remains
asymptotically stable. Simulations are utilized to reveal the cases
where the proposed scheme outperforms the modulated dynami-
cal system and the constrained optimization priority framework.
Experimental results further validate the theoretical findings.
Index TermsCollision avoidance, motion control of manipula-
tors.
I. INTRODUCTION
S ROBOTS are progressively coming out of the structured
A industrial setting and into less predictable environments
operating close to humans, it is imperative to control them ef-
ficiently so that they react in real time to unexpected changes.
To this direction, dynamical systems have offered an alterna-
tive to precomputed trajectories as they generate a trajectory
during motion from a differential equation. Instead of a single
trajectory, a dynamical system offers a flow field which can
guarantee convergence to the target as well as automatically
adapt to perturbations that are captured by external sensors, for
e.g. a change in the desired target pose [1], [2]. Dynamical sys-
tems may be further used to encode rich motion patterns learned
by demonstrations [3], [4].
The robustness of the dynamical systems has also been ex-
ploited against perturbations for reactive obstacle avoidance [5],
[2]. Obstacle avoidance has been investigated since the 80s
within the motion planning techniques which compute a colli-
sion free trajectory to the target, assuming a perfect model of the
Manuscript received September 10, 2016; accepted December 26, 2016. Date
of publication January 10, 2017; date of current version February 2, 2017. This
paper was recommended for publication by Associate Editor J.-M. Lien and
Editor N. Amato upon evaluation of the reviewers comments. This work was
supported by the European Communitys Framework Programme Horizon 2020
under Grant 644938 SARAFun.
S. Stavridis is with the Department of Electrical and Computer Engineer-
ing, Aristotle University of Thessaloniki, Thessaloniki 54641, Greece (e-mail:
sotistav@ece.auth.gr).
D. Papageorgiou and Z. Doulgeri are with the Department of Electrical
and Computer Engineering, Aristotle University of Thessaloniki, Thessaloniki
54641, Greece, and also with the Center for Research and Technology Hellas,
Thermi 570 01 Greece (e-mail: dimpapag@iti.gr; doulgeri@iti.gr).
Color versions of one or more of the figures in this letter are available online
at http://ieeexplore.ieee.org.
Digital Object Identifier 10.1109/LRA.2017.2651172
2377-3766 2017 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications standards/publications/rights/index.html for more information.
IEEE ROBOTICS AND AUTOMATION LETTERS, VOL. 2, NO.
robot and the environment. This assumption is relaxed in reac-
tive obstacle avoidance which introduces obstacles as they are
perceived by sensors during the robot motion and subsequently
adapt to avoid collisions [6]. In the dynamical system-based
approach of [5], obstacle avoidance is addressed by locally
modifying the original first order dynamics via a modulation
matrix which deflects the velocity field along the tangent plane
away from the obstacle without affecting system stability. Sad-
dle points or local minima may occur on the obstacles boundary
and at points multiple obstacles intersect the methods fails to
provide a solution [5]. In [2] an additive term is applied to a
dynamic movement primitive (DMP) to stir the motion away
from the obstacle. The method is however restricted to point
mass obstacles.
In a multiple task priority framework generalized for inequal-
ity tasks [7], obstacle avoidance is expressed as one sided in-
equality constraints on the robots velocities. In this framework
a constrained minimization of a quadratic function under equal-
ity and inequality constraints is formulated and solved by a
general purpose optimization algorithm (HQP). The solution
does not avoid local minima. Moreover, the computational cost
of such generalized solution is high despite improvements of
the solver reported in [8]. The elementary form of constraints
like the joint limits and obstacles calls for their explicit con-
sideration which could lead to computationally more efficient
algorithms, for example and in the case of joint limits, the works
of [9], [10].
In this work, an additive term is utilized for the purpose
of obstacle avoidance with a second order dynamical system
having the target as a unique stable equilibrium as in [2]. In
contrast to [2] however, an obstacle is assumed to be contained
in a convex region defined by a generalized ellipsoid. More-
over, the obstacle avoidance signal is designed following the
prescribed performance control methodology (PPC) introduced
in [11]. PPC has been applied in the design of robot position
controllers [12], [13] and as it allows the designer to impose
bounds on the output signals of nonlinear systems has already
been successfully utilized to impose spatial and joint limit con-
straints, [14], [10]. In this work, it is proved using Lyapunovs
theory that obstacle avoidance and modified system stability is
guaranteed and that the target remains asymptotically conver-
gent. It is illustrated in simulation for non-convex scene cases
in which it outperforms the method of dynamical system modu-
lation (MDS) of [5] and HQP [8]. Experimental results utilizing
a KUKA LWR4+ are also provided to illustrate the methods
performance.
STAVRIDIS et al.: DYNAMICAL SYSTEM BASED ROBOTIC MOTION GENERATION WITH OBSTACLE AVOIDANCE 713

Fig. 1. Transformation and potential induced by the obstacle avoidance signal.

(a) Transformation k h T ((t)). (b) Potential 12 k h 2 .

II. DYNAMICAL SYSTEM MODIFICATION Fig. 2. Obstacle avoidance parameter selection for two different initial
FOR OBSTACLE AVOIDANCE positions p 0 1 (blue) and p 0 2 (red).

Consider a task involving the position of a robot p =

[px py pz ]T which is desired to reach a target position pT 3 and let d0 = d(p0 ) and dT = d(pT ) be the distance from the
to be provided as input by a perception module. Let the ini- center of the initial position and the target position respectively.
tial position be p0 = p(0) 3 and consider the case of free Let us also define the lower bound for d(p) as the proportion of
space motion. A dynamical system having a globally asymptot- this distance that lies within the obstacles volume, given by:
ically stable equilibrium at pT can generate motion p(t) towards d(p)  f (p)d(p) (6)
the target pT with velocity v(t). We assume the availability of
a dynamical system described by a second order differential with
1
equation with a unique stable equilibrium at the target, e.g a f (p) = h(p) 2 m . (7)
DMP, and focus on the adaptation of such a system for obsta-
cle avoidance. Without loss of generality consider the linear It is clear that f (p) < 1, t 0 means an evolution of p outside
dynamical system model described by the differential equation the obstacle.
p = p (p pT ) where , + are positive constant We further specify an upper bound for d(p), which is de-
gains. A bio-inspired example of such a system is VITE [15]. fined such that the distance between this upper bound and d(p)
It is straightforward to show that this second order dynamical remains constant:
system has a unique stable equilibrium at the target. Similarly to d(p)  d(p) + (8)
[2] we consider an additive control input. Our aim is to design
this signal in a way that guarantees obstacle avoidance while re- where

taining the asymptotic stability at the desired target. The system d0 d(p0 ) + , if d0 d(p0 ) dT d(pT )
in state space is written as: = (9)
dT d(pT ) + , otherwise
p = v (1)
with a positive control parameter specifying the allowed over-
v = v (p pT ) + uh (t) (2) shoot at the initial or target position. Notice that is defined
by either the first or the second line of (9) given an initial state
where uh (t) 3 is the obstacle avoidance signal.
involving a stationary obstacle. Next, let us define the following
Consider now an obstacle which is assumed to be contained in
distance which will serve as an additional scalar attractor for
a 3D convex region centered at ph = [xh yh zh ]T and enclosed
the system:
by the surface:
dm (p) = d(p) + dT d(pT ). (10)
h(p) = 1 (3)
Notice that dm (pT ) = dT and hence the target belongs to this
with
 2m attractor. These scalar quantities are illustrated for the two cases
 p x x h 2m p y y h  p z z h 2m
h(p) = a + b + c (4) identified in (9) in Fig. 2.
Further define the distance from this attractor as a new scalar
where a, b, c  and m N + . The above equation includes variable (t) on which we will impose bounds utilizing the PPC
spheres, ellipsoids and for m 2, shapes that approximate methodology:
cuboids.
(t) = d(p) dm (p). (11)
Let us define the current end-effector distance from the center
of the obstacle: Specifically, we seek to enforce the following inequality:
d(p) = p ph  (5) (dT d(pT )) < (t) < d(pT ) dT . (12)
714 IEEE ROBOTICS AND AUTOMATION LETTERS, VOL. 2, NO. 2, APRIL 2017

It is straightforward to ascertain that (12) is equivalent to d(p) > We are now ready to propose the following signal for obstacle
d(p) > d(p) t 0. Notice that by construction (0) satisfies avoidance:
(12). By defining the open set, T
uh = kh (t) (23)
= {(t)  : (dT d(pT )) < (t) < d(pT ) dT }
(13) with kh > 0 a control parameter. Notice that kh can be used
we can guarantee obstacle avoidance if (t) t 0. To to regulate the flatness of the transformed error away from the
this aim, a transformed variable is utilized: boundaries and thus tune the strength of attraction to dm (p)
in a proportional way [see Fig. 1(a)]. So by tuning this gain it
(t) = T ((t)) (14) is possible to avoid the formation of local minima and induce
where the transformation function T () is a smooth, strictly motions that may not temporarily advance towards the target, as
increasing function defining a bijective mapping: shown in the cases presented in the Section III. Further notice
that determines the acting direction of the obstacle avoidance
T : (, ) (15) signal and is in general the resultant of two vectors (21). In
T (0) = 0. (16) fact, its norm, 1 + f 2 (p) tan2 1, amplifies the obstacle
avoidance signal as we come closer to the obstacle surface to
A suitable transformation based on the natural logarithm is a level dependent on the vicinity of the current position to the
illustrated in Fig. 1(a) and given by: main ellipsoidal axes (governing the value of cos ); its direc-
tion forms angle tan1 (f (p) tan ) with p ph , hence on the
1+ surface (f (p) = 1) is in the direction of nh coming closer to
dT d(pT )
T ((t)) = ln . (17) p ph as we move away.
1 We are now ready to state the main stability result:
d(pT ) dT
Theorem 1: Consider the obstacle ph , a, b, c. The obstacle
According to the PPC methodology it is sufficient to ensure avoidance signal (23) applied to the dynamical system (1), (2)
the boundedness of (t) in order to guarantee the evolution of guarantees i) (t) t 0 and hence obstacle avoidance
(t) t 0. ii) asymptotic convergence to a stationary state iii) asymptotic
Taking into account (11), (10), the time derivative convergence to an invariant set containing the target iv) the
of (17) yields: boundedness of all closed loop signals.
T Proof. Consider the generalized state [(p pT )T p T ]T of
= (d d(p)). (18) the closed loop system p = p (p pT ) kh (t) T ,

(20) and the candidate Lyapunov function:
The time derivative of (5) and (7) respectively yields d(p) =
(pp h ) T 2m +1 T 1
pp h 
p and f (p) = f (p) nh (p)
p, where: V = U + pT p (24)
2
1 where
nTh (p) = h(p) (19)
2m 1 1
U= (p pT )T (p pT ) + kh 2 (25)
is the normal vector of the level set h(p) = c, for some constant 2 2
c , at the current position. Substituting the above into (18), with the latter term representing the potential field induced by
taking into account (7) and that (p ph )T nh (p) = h(p) yields: (23) which can be viewed as a barrier Lyapunov function [see
T T Fig. 1(b)]. It is easy to establish that the derivative of (25) yields:
= p (20)
V = pT p 0, t 0. (26)
where:
The latter implies a bounded transformed variable, i.e. (t)
p ph nh (p) L and hence (t) t 0 [11]; moreover, limt p =
= w1 (p) + w2 (p) (21)
p ph  nh (p) 0 leading to a stationary state. Invoking Lasalles invariance
with principle, it is clear that the system solution will converge to the
following stationary invariant set:
w1 (p) = 1 f (p) T
(p pT ) + kh (t) = 0. (27)
f (p)
w2 (p) = (22)
cos This set is consisted of the desired target pT which clearly sat-
(pp h ) n h (p)
T isfies (27) since at p = pT , (t) = 0 = (t) from (11), (10), as
where cos = pp h n h (p)
. well as isolated positions {p : p = pT (27)} which are unde-
Notice that owing to the definition of nh (p) (19) and h(p) sired local equilibria. 
(4), 0 < cos 1 p, and since f (p) > 0, w2 (p) > 0. More- Remark 1: Notice that undesired local equilibria cannot con-
over, f (p) < 1 outside the obstacle, hence 1 > w1 (p) 0 hav- tain positions belonging to the scalar attractor dm (p), i.e. po-
ing the value of zero on the obstacles surface. For spheres sitions satisfying d(p) = dm (p) (10), since then (t) = 0. A
= nn hh (p)
(p) = pp h  .
pp h
necessary condition for satisfying (27) away from the target,
STAVRIDIS et al.: DYNAMICAL SYSTEM BASED ROBOTIC MOTION GENERATION WITH OBSTACLE AVOIDANCE 715

for d(p) = dm (p) is the colinearity of and p pT . This may

occur whenever the current position and the target are aligned
with one of the ellipsoidal axis and scalar parts cancel out. In
fact, assume = (p pT ) for some ; then, the following
condition should be satisfied + kh (t) T = 0 for the local
equilibria to exist.
Remark 2: Notice that all methods for real time obstacle
avoidance suffer from the existence of local equilibria in the
form of saddle points or local minima [6]; when the solution is
trapped in one, algorithmic mechanisms are utilized to enforce
the solution to escape from it.

A. Extension to Multiple Obstacles

It is straightforward to extend the proposed solution to more
than one obstacles by superimposing control signals as the one
defined in (23). In particular, for avoiding obstacles centered
at phi the obstacle avoidance signal is given by:



Ti
uh = khi i (t)i . (28)
i=1
i

The stability proof follows the same line involving now addi-
tional potential terms for each i in the Lyapunov function. The
second term of the equilibrium set (27) is for multiple obstacles
the sum of the obstacle avoidance signals in (28). Notice that
when an obstacle is dynamically introduced, a new term is added Fig. 3. Potential fields induced by the proposed signal. (a) Potential fields
in (28). On the other hand when the desired target is changed induced by each of 3 obstacles ( = 0.3). (b) Sum of potential fields ( = 0.3).
during the motion of the robot, (28) should be recomputed for (c) Sum of potential fields ( = 0.015).
all obstacles.
The potential fields induced by each term of (28) are defined
inside the barriers ( i , i + i ) where i  di (pT ) dTi [see
Fig. 1(b)] that are affected by the choice of i via i (9).
Fig. 3(a) illustrates level sets of potentials for 3 obstacles in
2D from high levels close to the obstacle surface to the lowest
level for each scalar attractor containing pT , getting high again
beyond the attractor and towards the upper barrier. These po-
tential fields define a constrained space depicted in Fig. 3(b).
Notice that the proposed methodology ensures by construc-
tion that the initial position belongs to all these potentials [see
Fig. 3(b)]. Due to the dynamic nature of the problem it is pos-
Fig. 4. Example of obstacle avoidance in the C-space. (a) Task space.
sible for a local minimum to occur if the constrained space is (b) C-space.
not connected, as it is in this example the case of = 0.015
[see Fig. 3(c)]. However, parameter i can be utilized to avoid
such cases [see Fig. 3(b)]. In practice, when a local mini- B. Extension to the Configuration Space
mum is detected (i.e. the velocity becomes zero and the po- The proposed method can be extended to the n-dimensional
sition is not at the target), increasing of khi and i leads to the configuration space (p n ) given that obstacles are mapped
solutions escape. to C-space obstacle regions that should subsequently be fitted
Remark 3: As the proposed signal for obstacle avoidance within n-dimensional ellipsoids. As an example consider a 2
affects the dynamical system by its additional attraction it is dof planar robotic arm [see Fig. 4(a)] moving from the initial
desirable to reduce its effect for obstacles that are left behind or configuration q0 to the target qT in the presence of two obsta-
are at a far distance. Algorithmic mechanisms can be used again cles mapped to two C-space obstacle regions [see Fig. 4(b)].
for this purpose. These can either take the form of direct deacti- Fig. 4(a) and (b) include the path generated by the robots mo-
vation (refer to the simulation example) or smooth elimination tion with and without the proposed obstacle avoidance signal,
e.g. by utilizing the distance from the lower bound d(p) d(p) shown in solid red line and dashed gray respectively. Notice that
in a multiplicative exponential term. the obstacle ellipsoids are avoided in the configuration space,
716 IEEE ROBOTICS AND AUTOMATION LETTERS, VOL. 2, NO. 2, APRIL 2017

Fig. 5. Robots motion (2 obstacles). Black ellipse: obstacle #1, Green ellipse: obstacle #2, red dot: Initial position, blue dot: target position. (a) t = 0.
(b) t = 2 s. (c) t = 4 s. (d) t = 7 s. (e) End of motion.

hence no collision will occur with any of the links of the robot
and the obstacles.

III. SIMULATION RESULTS

Simulations are performed, utilizing the kinematic model
of KUKA LWR4+ 7dof manipulator [16] with the proposed
method, the MDS [5] as well as the HQP method [8] for com-
parison purposes. At all cases and in order to be robust to
algorithmic singularities, a closed loop inverse kinematics
scheme was utilized to map the generated end-effector mo-
tion to the joint space. The end-effector motion is generated
by numerically integrating (1), (2) in real time, initialized with
the measured state of the robot at t = 0 s, utilizing = 8.5 and Fig. 6. End-effector path: Proposed method (blue line), changing target at
t = 4 s (magenta line), HQP (red line), MDS (green line).
= 1, in order to produce the next desired state. In MDS, (1), (2)
with uh = 0 serves as the original dynamical system from which
the velocity v(t) is generated and subsequently modulated. In
HQP two different priority levels are constructed; on the sec-
ond priority level resides the task equality J(q)q = vr , with vr
calculated by integrating (2) with uh = 0. On the first priority
level, the obstacle avoidance inequality expressed by JA q > b is
utilized for each of the obstacles with JA = (pp h)
T

pp h  J(q) where

J(q) is the robot Jacobian and b = d(p)d(p)
Tc with Tc being the
control cycle that is set to 1ms, ensuring d(p) > d(p).
In all simulation runs the initial robot configuration is
q0 = [30 70 0 80 0 30 0] deg corresponding to an initial
end-effector position p0 = [0.494 0.285 0.021] m. First,
the case of two obstacles being present in the scene is con-
sidered for all three methods with the end-effector target
set to position pT 1 = [0.495 0.3 0.18] m. The two obsta-
cles for which m = 1 are centered at ph1 = [0.47 0 0.185]
and ph2 = [0.5 0 0.045] m with their axes length being
[0.06 0.075 0.135] and [0.32 0.15 0.15] m respectively. The
case of the second obstacle introduced at t = 2 s is additionally
tested. Furthermore, the case of dynamically changing the target
from pT 1 to pT 2 = [0.2950 0.3 0.18]T m at t = 4s is included.
For the obstacle avoidance signals = 0.1 while kh1 = 0.01,
kh2 = 0.02 is used. For the MDS method the safety factor is set
to 1.1 while the reactivity factor at = 2. The generated path
with the proposed method is shown in Fig. 6 for both cases of
constant (blue line) and changed target (magenta line) with dif-
ferent time instances from the motion of the arm in the former
case depicted in Fig. 5. Fig. 6 also shows the path generated by
MDS (green line) and HQP (red line), which are close to each
other. The path generated by the proposed method when the
obstacle is introduced at t = 2 s is not included as it is similar to
Fig. 7. End-effector velocity with the proposed method with the two obstacles
scene (blue line) and when the second obstacle is inserted at t = 2 s (red line).
the one with all obstacles present from the start of the motion.
However, end-effector velocity for these two cases depicted in
Fig. 7 shows the velocity change at t = 2 s (subplots).
In order to reveal the different operation and advantages of
our proposed method, we next constructed a non-convex scene
utilizing four spherical obstacles of radius r = 10 cm. The ob-
stacles are centered at ph1 = [0.137 0.148 0.011], ph2 =
[0.137 0.268 0.011], ph3 = [0.137 0.268 0.131] and
ph4 = [0.137 0.148 0.131] m. The target position is now
set to pT = [0.220 0.132 0.121] m and kh = 0.02 in the pro-
posed method for all obstacle avoidance signals. Results de-
picted in Fig. 8 show the trapped HQP and MDS algorithms
while the proposed method adapts the dynamical system so
that it finds a feasible path. Notice how the path generated by
the proposed solution is even turned slightly backwards by the
STAVRIDIS et al.: DYNAMICAL SYSTEM BASED ROBOTIC MOTION GENERATION WITH OBSTACLE AVOIDANCE 717

Fig. 8. End-effector path: Proposed method (blue line), HQP (red line) MDS
(green line). (a) Side view. (b) Top view.

Fig. 10. Scene captured by vision. Green dots indicate the center of the
obstacles.

Fig. 9. End-effector velocity p with the proposed method.

action of the additional to the target scalar attractor to circumvent

the obstacles. The reason why HQP and MDS cannot in this case
escape the local minimum is that these methods generate mo-
tions that always advance towards the target. In the case of MDS
increasing the reactivity factor is proved ineffective since this
factor amplifies the modulated field up to a bound and not pro-
portionally. The end-effector velocity of the proposed scheme
is shown in Fig. 9. Furthermore, Fig. 8 shows the path gener-
ated (orange line) with obstacle avoidance signal deactivation
with the first deactivation occurrence denoted by the asterisk.
The mechanism utilized for ensuring that the obstacle is left
behind is the simultaneous satisfaction of inequalities nT p > 0
and p pT  < f (p)(p ph ) (pT ph ); the first ensures
a motion away from the obstacle and the second ensures that
the distance to the target is less than the targets distance from a
point on the obstacle surface related to the current position.
Fig. 11. End-effector path, without obstacles (red line) and with obstacles
IV. EXPERIMENTAL RESULTS (blue line). (a) Side view. (b) Top view.
Experimental verification is conducted utilizing a KUKA
LWR4+ 7dof manipulator with the proposed methodology im- was also fitted with an ellipsoid of rank m = 2 with a = 3 m,
plemented in C++ utilizing ROS and the FRI library with a 2 ms b = 3 m, c = 0.1 m with its center below the surface by
control cycle. Ellipsoids are fitted around obstacles captured 0.1 m at the target x, y coordinates and the respective signal
by an RGB-D camera. The control parameters of the original gain was set to kh = 0.01. The initial robot configuration is
dynamical system are: = 10.5, = 4, while for the obsta- q0 = [135 60 0 60 0 60 0]T deg corresponding to the ini-
cle avoidance signal we have utilized = 0.1, kh = 0.1 for tial tool position p0 = [0.483 0.483 0.13]T m, and the target is
every obstacle. Five objects were placed on the supporting sur- selected to be pT = [0.216 0.383 0.14]T m. Notice that the
face as shown in the camera view Fig. 10 with the fitted ellip- initial configuration of the robot is selected (with elbow up) to
soids shown in Fig. 11(a) and (b) in side and top view where minimize the probability that a collision with any link will oc-
three ellipsoids are used for the lamp. The supporting surface cur in the scene, since the method was implemented solely for
718
Without (red line) and with obstacle presence
Fig. 12. End-effector velocity p.
(blue line).
Fig. 13. Obstacle avoidance signal u h .
the motion of the end-effector in 3D space. Moreover, similarly
to the simulations the numerical integration of the dynamical
system is done on-line, initialized with the measured state of the
robot at t = 0 s; the closed loop inverse kinematics solution is
utilized to map the end-effector state to joint state. The path of
the end-effector is shown in Fig. 11(a) and (b) for both cases of
motion generation with and without the obstacles presence.
Translational velocity of the end-effector is shown in Fig. 12
while the obstacle avoidance signal is shown in Fig. 13. Notice
that translational velocities are calculated by simple numerical
differentiation of the position provided by KUKA software. No-
tice the effectiveness and the smoothness of the motion achieved
by the proposed method.
V. CONCLUSION
In this paper we propose an obstacle avoidance signal de-
signed according to the prescribed performance control method-
ology, which affects the motion generated by a second order
dynamical system having the target as a unique stable equilib-
rium. The proposed scheme, utilizing generalized ellipsoids for
the obstacles, is proved to guarantee collision avoidance and
IEEE ROBOTICS AND AUTOMATION LETTERS, VOL. 2, NO. 2, APRIL 2017
system stability retaining the asymptotic stability of the target.
It is shown in simulations to avoid traps in non convex scenes
as opposed to dynamical system modulation and priority level
constraint optimization techniques. Experimental results with
KUKA LWR4+ demonstrate its smooth effective performance
in an environment involving a variety of obstacles.
REFERENCES
[1] M. Hersch and A. Billard, Reaching with multi-referential dynam-
ical systems, Auton. Robots, vol. 25, no. 12, pp. 7183, 2008,
doi: 10.1007/s10514-007-9070-7.
[2] H. Hoffmann, P. Pastor, D. H. Park, and S. Schaal, Biologically-inspired
dynamical systems for movement generation: Automatic real-time goal
adaptation and obstacle avoidance, in Proc. IEEE Int. Conf. Robot.
Autom., 2009, pp. 25872592, doi: 10.1109/ROBOT.2009.5152423.
[3] S. M. Khansari-zadeh and A. Billard, Learning stable nonlinear dynam-
ical systems with gaussian mixture models, IEEE Trans. Robot., vol. 27,
no. 5, pp. 943957, Oct. 2011.
[4] A. J. Ijspeert, J. Nakanishi, and S. Schaal, Learning attractor landscapes
for learning motor primitives, in Proc. Adv. Neural Inf. Process. Syst.,
2002, vol. 15, pp. 15471554.
[5] S. M. Khansari-Zadeh and A. Billard, A dynamical system approach to
realtime obstacle avoidance, Auton. Robots, vol. 32, no. 4, pp. 433454,
2012, doi: 10.1007/s10514-012-9287-y.
[6] B. Siciliano and O. Khatib, Springer Handbook of Robotics. New York,
NY, USA: Springer-Verlag, 2007.
[7] O. Kanoun, F. Lamiraux, and P.-B. Wieber, Kinematic control of redun-
dant manipulators: Generalizing the task-priority framework to inequal-
ity task, IEEE Trans. Robot., vol. 27, no. 4, pp. 785792, Aug. 2011,
doi: 10.1109/TRO.2011.2142450.
[8] A. Escande, N. Mansard, and P.-B. Wieber, Hierarchical quadratic
programming: Fast online humanoid-robot motion generation, Int.
J. Robot. Res., vol. 33, no. 7, pp. 10061028, 2014, doi:
10.1177/0278364914521306.
[9] F. Flacco, A. D. Luca, and O. Khatib, Motion control of redundant robots
under joint constraints: Saturation in the null space, in Proc. IEEE Int.
Conf. Robot. Autom., 2012, pp. 285292.
[10] A. Atawnih, D. Papageorgiou, and Z. Doulgeri, Kinematic control of
redundant robots with guaranteed joint limit avoidance, Robot. Auton.
Syst., vol. 79, pp. 122131, 2016, doi: 10.1016/j.robot.2016.01.006.
[11] C. Bechlioulis and G. Rovithakis, Robust adaptive control of feed-
back linearizable MIMO nonlinear systems with prescribed performance,
IEEE Trans. Automat. Control, vol. 53, no. 9, pp. 20902099, Oct. 2008.
[12] Y. Karayiannidis and Z. Doulgeri, Model-free robot joint position regula-
tion and tracking with prescribed performance guarantees, Robot. Auton.
Syst., vol. 60, pp. 214226, 2012.
[13] Y. Karayiannidis, D. Papageorgiou, and Z. Doulgeri, A model-free con-
troller for guaranteed prescribed performance tracking of both robot
joint positions and velocities, IEEE Robot. Autom. Lett., vol. 1, no. 1,
pp. 267273, Jan. 2016, doi: 10.1109/LRA.2016.2516245.
[14] A. Theodorakopoulos, G. A. Rovithakis, and Z. Doulgeri, An impedance
control modification guaranteeing compliance strictly within preselected
spatial limits, in Proc. 2015 IEEE/RSJ Int. Conf. Intell. Robots Syst.,
2015, pp. 22102215, doi: 10.1109/IROS.2015.7353673.
[15] D. Bullock and S. Grossberg, Neural dynamics of planned arm
movements: Emergent invariants and speed-accuracy properties during
trajectory formation, Psychol. Rev., vol. 95, no. 1, pp. 4990, 1988.
[16] V. Bargsten, P. Zometa, and R. Findeisen, Modeling, parameter iden-
tification and model-based control of a lightweight robotic manipula-
tor, in Proc. 2013 IEEE Int. Conf. Control Appl., 2013, pp. 134139,
doi: 10.1109/CCA.2013.6662756.