A comprehensive study on path planning algorithms in multi-agent robot
soccer domain
Project Progress Report for EE586 Artificial Intelligence*
Kadir Firat Uyanik
Electrical and Electronics Engineering Department,
Middle East Technical University
kadir@ceng.metu.edu.tr
Abstract— The ability to traverse/navigate around without
colliding to the obstacles is one of most essential requirements
for an autonomous mobile robot that is supposed to realize a
particular task by itself. Assuming that the robot is provided
with its current global position and orientation (initial configu-
ration), and goal configuration as well, the problem of enabling
a robot to reach its goal configuration in a reasonable amount of
time, though seems simple, may present intriguing and difficult
issues. Besides, in multi-robot domains where robots operate
collaboratively, highly complex and unpredictable dynamics can
arise. If an adversarial entity and unactuated-physical-body
interactions are also taken into account, the need for making
navigation calculations within tight time constraints becomes
more serious. In this study, several path planing algorithms,
such as Artificial Potential Fields, Visibility Graphs, Rapidly-
growing Random Trees, and their commonly known variations
are going to be compared in terms of various time and safety
criteria. Finally, a hybrid planning method is going to be
proposed that is going to be tested on a multi-agent robot
soccer platform. This report, on the other hand, represents
the current status of this study and shows preliminary results
with the potential field methods.
I. I NTRODUCTION
One particular feature of an autonomous mobile robot that
makes them superior to the industrial robot manipulators is
that they can navigate and traverse around the environment
if the surface is smooth enough. Yet another special feature
is that they can move from an initial position to a goal
position without requiring explicit intermediate supervised
action commands (viz. no or minimum remote control). This
implies that the robot is capable of avoiding obstacles and
unwanted regions during navigation. One basic family of
methods to accomplish this is called Bug-algorithms [1].
Bug1 and Bug2, are known to be one of the earliest
and simplest sensor-based planners, minimizing the compu-
tational burden on the robot while still guaranteeing global
convergence to the target if reachable. However, these algo-
rithms do not make the best use of the available sensory data
to produce relatively shorter paths by utilizing range data.
Although VisBug [2] algorithm uses the range sensor data,
it can only utilize this data to find shortcuts that connects
points on the trajectory found by Bug2 algorithm containing
no obstacles between. TangentBug algorithm(TBA) [3] uses
range data to produce a local tangent graph so as to choose
*This study is a part of EE586 Artificial Intelligence course offered by
Dr. Afsar Saranli.
locally optimal direction while keeping approaching to the
target position. TBA produces paths, in reasonably simple
environments, that approach the globally optimal path as the
sensor’s maximal detection range increases.
Bug algorithms assumes that the robot is a point on a 2D
planar surface, and the obstacles are stationary. Moreover,
bug algorithms were developed for the robots that do not
have access to the global information of the environment.
Hence, these algorithms are not useful for multi-agent robot-
soccer platform where entities are highly dynamic and global
information is already available (pose of each robot and the
position of the ball).
There are several robot motion control and path planning
algorithms that make use of global information, such as
artificial potential fields(APF), visibility graphs(VG), and
rapidly-growing random trees(RRT).
The rest of the document organized as follows: in the
algorithms section, APFs, VGs, and RRTs are explained
briefly; then experimental setup is explained, and lastly
preliminary implementation results are given, and the article
is concluded.
II. A LGORITHMS
A. Artificial Potential Fields
APF[4] is a reactive approach since the trajectories are
not planned explicitly but obtained while executing actions
by differentiating a function what is called potential function.
A potential function is differentiable real-valued function U :
Rm → R. The output of a potential function can be taken as
the energy and its negative gradient as the net-force acting
on the robot.
Approaching to the goal configuration can be thought as
moving robot from a high-value state to a low-value state by
following a ”downhill” path. Mathematically speaking, this
is the the process of following negative gradient, or gradient-
descent, i.e. c. (t) = −∇U (c(t)). The motion terminates as
soon as the gradient vanishes, that is ∇U (q ∗ ) = 0 where
q ∗ is called the critical value of U . This critical point is
either maximum, minimum or saddle point as it shown in
figure 1.Although the maximum is not that critical as soon
as robot doesn’t start movement from this point, but the
local minimum points are the real concern while designing
a potential function. Here, one assumption is that the saddle
points -being unstable points- do not cause any problem due should look like the behavior of a tightened spring, or
to the robot’s inertia and other dynamics. the magnets getting closer to each other so that obstacle
APF methods can vary regarding how they measure the avoidance can be satisfied. In this study following
distance to the goal, and the type of the functions they formula is utilized,
utilize. In this study, additive attractive/repulsive potential,
and navigation potential based methods are investigated. In (
1 1 1 2
2 η( D(q) − Q? ) , D(q) 6 Q?
addition to these methods, a modified version of additive Urep (q) = (3)
potential fields is also investigated. 0, D(q) > Q?
Derivative of the repulsive potential function(3) is the
repulsive force which is,
(
η( Q1? − D(q)
1
) D21(q) ∇D(q), D(q) 6 Q?
∇Urep (q) =
0, D(q) > Q?
(4)
where Q? is the distance threshold for an obstacle to
create a repulsion effect on the robot, and η is the
repulsion gain.
The gain parameters (i.e. η and ξ) and the threshold
parameters(i.e. Q? and d? ) are set empirically.
The total repulsive potential field can be obtained by
summing up the potentials caused by all of the obsta-
cles,
Fig. 1: From left to right: maximum, saddle, and minimum XN

critical points are shown where figures at the top are the Urep = Urepi (q)
functions and the below are the gradient vectors of them. k=1

and the movement is realized by following the negative

1) Additive attractive/repulsive potential: The intuition
behind the additive potential functions is attracting the robot
to the goal position while repelling the robot from the
obstacles by superposing these two effects into one resultant
force applied on the robot. This potential function can be
constructed as U (q) = Uatt (q) + Urep (q), that is the sum of
the attractive and repulsive potentials.
• Attractive potential:
The essential requirement in construction of the at-
tractive potential is that Uatt should be monotonically
increasing with the current distance d(q, qgoal ) of the
robot to the goal, qgoal . Two common Uatt functions
are the ones having conical and/or quadratic forms. In
the experiments, combinations of the two are used since
the conical functions suffer from the discontinuity in
the goal position, that is division-by-zero problem in
computational sense.
(
1
ξd2 , d 6 d?goal
Uatt (q) = 2 ? 2 (1)
ξdgoal d − 12 (d?goal ) , d > d?goal
where d = d(q, qqoal ), ξ is the attraction gain, and d?goal
is the threshold after which quadratic function changes
to the conical form.
The gradient of the function (1) is obtained as;
( As in the case of attractive/repulsive potential functions,
ξ(q − qqoal ), d(q, qgoal ) 6 d?goal
∇Uatt (q) = (q−qqoal )
ξdgoal d(q,qgoal ) , d(q, qgoal ) > d?goal
?
(2)
• Repulsive potential:
The behavior of a proper repulsive potential function
gradient of the sum of attractive/repulsive potentials or
simply the following vectorial sum;
−∇U (q) = −∇Uatt (q) − ∇Urep (q)
In this study, the translation velocity assignment is
simply done by utilizing −∇U (q).
2) Navigation Potential: Due to the local minima prob-
lem in the additive attractive/repulsive potential fields, many
other local-minima free potential field methods are proposed
such as wave-front planner [5][6]. However these methods
have discretization problems and that’s why they become
computationally intractable for high dimensional and large
configuration spaces. To solve the local minima problem,
a special function to be constructed which has the only
minimum at the goal position. These functions are called
navigation functions, formally defined in [7][8].
A function φ:Qf ree → [0, 1] is called a navigation
function if it
k
• is smooth (or at least C for k > 2)
• has a unique minimum at qgoal in the connected com-
ponent of the free space that contains qgoal .
• is uniformly maximal on the boundary of the free space,
and
1
• is Morse
navigation functions are used to move the robot from an
arbitrary configuration to the goal configuration in the neg-
ative gradient direction of the potential at that instant. The
1 Detailed discussion can be found in [9].

Fig. 3: A screenshot from the Webots simulator, while robot
is tested in a three-obstacle between robot and the goal case.
navigation potential functions can be constructed as follows:
(d(q, qgoal ))2
φ= (5)
(λβ(q) + d(q, qgoal )2κ ) κ
1
where β(q) is a repulsive-like function calculated as
βi (q) = (d(q, qi ))2 − ri 2
for each obstacle having a radius of ri . The critical point
in choosing Bi (q) is that QOi = q | βi (q) 6 0; this means
βi (q) is negative inside an obstacle and it is positive outside
in the free space, which requires
2 2
β0 (q) = (d(q, q0 )) − r0
where q0 is the center and r0 is the radius of the sphere
world.
In this report, it is assumed that the configuration space is
bounded by a sphere centered at q0 and the other obstacles
are also spherical centered at qi .
The λ in the equation (5) is necessary for bounding the
navigation function so that it has 0 at the goal and 1 on the
boundary of any obstacle.
The κ on the other hand, has an effect of making the
navigation function has the form of a bowl near to the goal.
Increasing κ causes the other critical points(local minimum
etc.) shift toward the obstacles and resulting in a negligible
repulsive behavior compared to the overwhelming influence
of the attractive field. The effect of increasing κ is shown in
the figure2, for the case of having three obstacle in front of
the robot and a goal position beyond the obstacles as it is
illustrated in the figure 3.
Once a ”proper” κ value is decided, robot’s configuration
or -in this study- velocity is updated with the negative
gradient of the navigation potential(equation (6)) function
similar to the what’s been done in the attractive/repulsive
potential functions.
3) Modified Additive-Potential Function: To deal with
local minimum problem in additive functions one idea would
be applying random forces on the robot once it is detected
that the robot stops without reaching to the goal position
(viz. getting stuck at a local minimum). Similar method is
proposed in [10], called random walk.
Another solution to the local minima problem would be
using repulsive forces in a different form. For instance, if
the robot is imagined as a fluid or gas particle flowing
towards the goal, robot will be able to pass through obstacles
standing between the robot and the goal position. Which can
be realized by applying tangential attractive/repulsive forces
when the robot is at a certain distance from an obstacle.
However, this method will not be successful in the case
of confronting narrow passages, causing collision to the
obstacles. Another problem this method will cause is the
oscillatory movements in locally cluttered environments in
which there might be paths that are longer in the sense of
euclidean distance but shorter in time since robot can reach
higher speeds in obstacle-free paths.
One solution to the oscillatory movement problem would
be grouping the obstacles together and having virtual obsta-
cles which may contain narrow passages, and the obstacles
that are sufficiently close to each other as it is shown in
figure 4 and 6. This method with the visibility testing is
proposed in [11], and I have successfully implemented this
method(given in the algorithm 1), and tested in the simulation
environment in one of my earlier studies. Although tests are
done in highly cluttered environments, generated paths were
close to the optimal paths and no local minima problem is
observed.
Algorithm 1: Obstacle Grouping
while preObstList.size() > 0 do
closestObst ⇐ getClosestObst()
if isVisible(closestObst)=TRUE then
tempObst ⇐ closestObst
repeat
if isLinkable(tempObst,preObstList[i])=TRUE
then
linkObsts(tempObst, preObstList[i])
tempObst ⇐ tempObst→Neighbor
end if
until preObstList.size()= 0
end if
virtualObst ⇐ closestObst
groupObsts(virtualObst)
postObstList.push back(virtualObst)
end while
B. Visibility Graphs
A visibility graph is a graph consisting of intervisible
locations, typically in the euclidean plane. Nodes of the graph
represent a point location/vertex, and edges represent visible
connections between the vertices. Visibility also refers to the
condition when each edge stays in the free configuration
space (viz. no obstacle inter-penetration). In other words,
visible lines are the line segments that a robot supposed
to follow to reach its goal position without colliding any
obstacle on its path.
 2 1
!
2d∇d[(λβ(q) + d2κ ) κ )] − κ1 (λβq + d2κ ) κ −1 (λ∇β(q) + 2κd2κ−1 ∇d)
−∇φ = − 2 (6)
(λβ(q) + d2κ ) κ
PN  QN 
where d = d(q, qgoal ) and ∇β(q) = i=0 ∇βi (q) j=0,j6=i βi (q) . ∇βi (q) = 2(q − qi ) for i > 0, ∇β0 (q) = −2(q − q0 ).

Fig. 2: From left to right and top to bottom, κ is changed between 1 and 6, and navigation potential function is calculated
for the sphere-world by sampling the world for the specific configurations(viz. 2D positions) varying in polar coordinates
r = [0.01, 2.0] and θ = [0.01, 2π] with the discrete sampling steps of 0.01. As it is clearly shown that κ doesn’t effect the
navigation potential considerably after a certain value, which is κ = 4 in this case.

The standard visibility graph is defined in a 2-D polygonal

configuration space where the nodes includes the start and
goal location in addition to the vertices of the configuration
space obstacles, and the edges that connect two distinct line-
of-sight nodes, i.e.;
eij 6= ∅ ⇐⇒ svi + (1 − s)vj ∈ cl(Qf ree )∀s ∈ [0, 1].

Fig. 4: Closest obstacle and its neighbor obstacle are linked

to each other and nwe virtual obstacle makes obstacle C
invisible to the robot (adapted from[11])

Fig. 6: The thin solid lines delineate the edges of the visibility
graph for the these three obstacles (filled polygons). The blue
line segments represent the shortest path from start to goal
position.
Fig. 5: Obstacle 3, 4 and 5 are linked to each other.They General algorithm for a VG can be summarized as follows:
form a virtual obstacle (adapted from [11])
• Expand the obstacles by utilizing Minkowsky Differ-
 ence[13] if the robot has a polygonal representation
(otherwise approximate it)
• Construct edges between each node pairs considering
the visibility criterion
• Search graph via a shortest-path finding algorithm, e.g.
A*, Dijkstra’s etc.
If the robot is assumed to be a point robot, visibility graph
construction is nothing but picking a node pair and checking
if the edge between these nodes crosses the obstacles edges.
Let V = v1 , v2 , ...vn be the set of vertices of the polygonal
obstacles in the configuration space as well as the start and
goal configurations. Therefore, general graph construction
requires: t ∈/ [0, 1] and u ∈/ [0, 1] where t and u represents
a scalar value, fraction of the line segments between two
arbitrary vertices. More formally [14]:
v k vi × vl vk
t= (7)
vj vi × vl vk
vk v i × vj vi
u= (8)
vj vi × vl v k
∀vi vj , ∀vk vl , vi 6= vj , vk 6= vl (9)
For instance; if vj vi × vl vk = 0 then these line segments
are parallel, and also if vk vi = 0 these lines are collinear.
Otherwise t and u are going to be calculated as scalar values.
If they are not in the interval of [0, 1], then these lines are not
intersecting with each other. Since selecting vi vj has O(n2 )
complexity and there are O(n) intersection tests this VG
construction method has an overall complexity of O(n3 ).
A better approach would be applying the rotation plane
sweep algorithm which has the complexity of O(n2 logn),
which is shown in the figure 7. Sweep line algorithm also
Fig. 7: Rotation sweep line algorithm. Adapted from [9].
discards adding needless edges but only considers supporting
and seperating lines which reduces the number of edges.
Once the visibility graph is constructed, an optimal path
search algorithm such as A* can be utilized to obtain a set
of edges that are on the shortest path.
C. Rapidly Growing Random Trees
An RRT is basically a data structure and algorithm which
efficiently searches nonconvex high-dimensional spaces. The
way RRT is constructed can be given as follows:
• Start with the initial state as the root of a trees
• Pick a random state in anywhere or in the direction of
the target
• Find the closest node in the current trees
• Extend that node toward the target if possible
There are many variations of RRTs. In this study, RRT-
base [15], RRT-connect [16], ERRT [17], and Bidirectional
Multi-Bridge ERRT[18] is going to be studied.
To mention them briefly:
• RRT-connect expands two trees towards each other,
one from the start node, and the other from the goal
node. They interchange the roles of expanding-towards-
the-other and expanding-randomly. The algorithm ter-
minates ones a connection is established between trees.
• ERRT makes use of the idea of way-point caching.
Therefore search operation becomes biased towards the
paths found in the earlier successful executions.
• Multi-Bridge ERRT is similar to the RRT-connect but
it continues to execution although a connection is estab-
lished between the trees until a prespecified number of
connections are established. Then the resultant structure
is not a tree anymore. To find the shortest path an
optimal graph search algorithm is used, such as A*.
III. E XPERIMENTAL S ETUP
Webots[19] or Gazebo simulation environments are going
to be used by heavily utilizing ROS[20] graph concepts in the
design of the system. If time permits, the winner algorithm
can be run on the actual system with real robots8 for a
previously defined test scenario.
Fig. 8: An actual soccer Fig. 9: Simulated robot in
robot, from the team Robo- Webots simulation environ-
Turk. ment.
IV. E XPERIMENTS
In this section preliminary results for the APF methods are
given for three scenarios, one-obstacle direct-free path, three-
obstacle indirect-free path, two-obstacle direct-free path.
A. Case: One-obstacle direct-free path
In this experiment, we would like to see if the robot
can traverse around a single obstacle to reach the goal
position by following a preferably linear path. Normally,
since the obtacle is not entirely in the path of the robot, we
would expect that the robot almost-directly goes to the target
position. But it is shown that without doing a visibility check
basic additive potential fields may cause perturbances in the
path or even oscillations depending on how the attraction
and repulsive gains are set. Navigation functions, on the
other hand, may give a better path if κ is properly chosen.
Since there is not a theoretical way to select best κ, selection
process may require serious intuition and effort.
Fig. 10: Left: additive att/rep potential function, right: navi-
gation potential function.
B. Case: Three-obstacle indirect-free path
Fig. 11: Left: additive att/rep potential function, right: navi-
gation potential function.
As it is shown in figure 13, basic att/rep potential function
method suffers from the local minima problem and robot gets
stuck at some point. On the other hand, navigation function
finds a path to the goal, though far from being optimal.
Figure 12 shows how the path length varies with different κ
values.
C. Case: Two-obstacle direct-free path
Although this case is similar to first case, it shows the
severity of the additive att/rep potential fields which even
Fig. 12: From top left to the bottom right κ values are 3,4,5
and 7. The time it takes to reach to the goal is tκ3 =no
value, tκ4 = 62s, tκ5 = 53s and tκ7 = 45s in simulation
time. That’s why κ is chosen to be 7 in the experiments.
Please notice that the viewpoints of these figures are different
than each other, careful inspection will reveal the fact that
as κ values decrease, the length of the path increases since
the shape of the bowl=like potential function becomes more
spreaded-out.
TABLE I: Overall results
Additive Navigation
Experiment time(sec.) path(m) time(sec.) path(m)
Case 1 2.5 1.25 2.5 1.26
Case 2 ∞ ∞ 45.52 1.97
Case 3 ∞ ∞ 1.5 1.09
may not result a complete solution, although the only thing
that the robot should do is going just forward.
Navigation function, on the other hand, could find a path
between the obstacles, though the movement was a little
oscillatory. One reason of the oscillation is the delay between
the communication nodes(=processes) since the PC that I’ve
used for the experiments was a standard laptop.
The overall results for the three experiment cases are given
in the table I. Although additive potential functions are lack
of completeness, in most of the cases, they provide shorter
paths than the paths navigation functions return.
V. C ONCLUSIONS AND F UTURE W ORK
In this report, several path planning algorithms are in-
vestigated and preliminary test results are given for various
APF methods. It can be deduced from the test results
that the APF with navigation function is not suitable for
robot soccer domain since it may require a great deal of
time to speed the robot up until finally robot is started to
being gravitated towards the goal position. However, additive
attractive/repulsive potential function based APF with the
obstacle grouping extension gives promising results.
In this study, I’m at the stage of implementing VG
and various RRT algorithms. Once the implementations are
finished, APF w/ obstacle grouping, VG, and RRT-based

Fig. 13: Top: additive att/rep potential function, bottom:
navigation potential function.
algorithms are going to be tested for the scenarios pretty
much the same with the ones used to test APF methods as
it is given in this report.
During the tests, there will be several comparison metrics,
such as:
• tplan : The time it takes for the algorithm to generate
sound and preferably-optimal path.
• texec : The overall and individual times it takes for
each robot to follow the path being generated, or to
reach a particular goal configuration from a given initial
configuration.
• saf ety: How safe the robots are while traversing
through this path. This can be calculated by summing up
the distances between the robots while they are getting
closer to each other during their journey. Summation
procedure can be done if the robots are close to each
less than a predefined safety region threshold.
As a feature work, dynamics of the robot and the ball-if
it is being dribbled- can be taken into account to compare
these algorithms again by using the similar metrics as they
are given above. This kind of a planning approach is applied
to the RRT based planner, called behavioral kinodynamic
RRT[21]. Therefore, test scenarios can be enriched by adding
controllable and uncontrollable dynamic robots to the en-
vironment, and a ball dribbling scenario where the ball
possession can also be considered as yet-another comparison
metric.
VI. ACKNOWLEDGEMENTS
Some part of this study, particularly APF case-studies are
done for the homework fulfillment of CS548 Robot Motion
Control and Planning course offered by Dr. Uluc Saranli.
R EFERENCES
[1] V. J. Lumelsky and A. A. Stepanov. Path-planning strategies for a
point mobile automaton moving amidst obstacles of arbitrary shape.
Algoritmica, 2:403-430, 1987
[2] H. Choset and J. W. Burdick. Sensor based planning, part ii: Incremen-
tal construction of the generalized voronoi graph. In IEEE International
Conference on Robotics and Automation, Nagoya, Japan, May 1995.
[3] I. Kamon, E. Rivlin, E.Rimon. A new range-sensor based globally
convergent navigation algorithm for mobile robots. Robotics and
Automation, 1996. Proceedings., 1996 IEEE International Conference
on In Robotics and Automation, 1996. Proceedings., 1996 IEEE
International Conference on, Vol. 1 (1996), pp. 429-435 vol.1.
[4] Khatib, O. 1986. Real-time obstacle avoidance for manipulators and
mobile robots. Int. J. Rob. Res. 5(1):9098.
[5] R. Jarvis. Coliision free trajectory planning using distance trans-
forms.Mech. Eng. Trans. of the IE Aus, 1985.
[6] J. Barraquand, B. Langlois, and J.C.Latombe. Numerical potential
field techniques for robot planning.IEEE Transactions on Man and
Cybernetics, 22(2):224-241, Mar/Apr 1992.
[7] Koditschek, Daniel E., and Rimon, Elon: Robot navigation functions
on manifolds with boundary, Advances in Applied Mathematics 11(4),
volume 11, 412442, 1990.
[8] Rimon, Elon and Koditschek, Daniele; Exact robot navigation using
artificial potential functions IEEE Transactions on Robotics and Au-
tomation. Vol. 8, no. 5, pp. 501-518. Oct. 1992
[9] Principles of Robot Motion by Howie Choset etal., MIT Press, 2005,
ISBN: 0-262-03327-5
[10] J. Barraquand and J.-C. Latombe. Robot motion planning: A
distributed representation approach. Internat. J. Robot. Res.,
10(6):628649, 1991
[11] B. Zhang,W.Chen, M. Fei, An optimized method for path planning
based on artificial potential field, Sixth International Conference on
Intelligent Systems Design and Applications (ISDA’06) Volume 3
(2006)
[12] Lozano-Perez T. and Michael A. Wesley. 1979. An algorithm for
planning collision-free paths among polyhedral obstacles. Commun.
ACM 22, 10 (October 1979), 560-570
[13] M. de Berg, M. van Kreveld, and M. Overmars. Computational
Geometry: Algorithms and Applications. Springer, Berlin, 1997.
[14] Goldman R., Intersection of Two Lines in Three-Space, in Andrew S.
Glassner, ed., Graphics Gems, Academic Press, p. 304, 1990.
[15] S. M. LaValle. Rapidly-exploring random trees: A new tool for path
planning. TR 98-11, Computer Science Dept., Iowa State Univ., Oct.
1998
[16] Kuffner, J.J., Jr.; LaValle, S.M.; , ”RRT-connect: An efficient approach
to single-query path planning ,” Robotics and Automation, 2000.
Proceedings. ICRA ’00. IEEE International Conference on , vol.2,
no., pp.995-1001 vol.2, 2000
[17] Bruce, J.; Veloso, M.; , ”Real-time randomized path planning for
robot navigation,” Intelligent Robots and Systems, 2002. IEEE/RSJ
International Conference on , vol.3, no., pp. 2383- 2388 vol.3, 2002
[18] Based on: Bruce J. R., Real-Time Motion Planning and Safe Naviga-
tion in Dynamic Multi-Robot Environments, PhD. Thesis, 2006
[19] Michel, O. Webots: Professional Mobile Robot Simulation, Interna-
tional Journal of Advanced Robotic Systems, Vol. 1, Num. 1, pages
39-42, 2004
[20] Quigley M., Conley K., Gerkey B., Faust J., Foote T. B., Leibs J.,
Wheeler R., and Ng A. Y. (2009). Ros: an open-source robot operating
system. n International Conference on Robotics and Automation, ser.
Open-Source Software workshop.
[21] Zickler S., Physics-Based Robot Motion Planning in Dynamic Multi-
Body Environments PhD Thesis, 2010, Computer Science Department,
Carnegie Mellon University Thesis Number: CMU-CS-10-115