2014 IEEE international Symposium on ROBOTICS AND MANUFACTURING AUTOMATION (IEEE ROMA2014)

December 15 16, 2014, Kuala Lumpur, Malaysia

Runtime Reduction in Optimal Multi-Query Sampling-Based Motion Planning

Weria Khaksar, Khairul Salleh bin Mohamed Sahari, Marwan A. Ali, Firas B. Ismail and Moslem Yousefi
Centre for Advanced Mechatronics and Robotics (CAMARO), College of Engineering
University Tenaga Nasional (UNITEN), Kajang, Malaysia.
{Weria, Khairul, Firas, Moslem, Marwan}@uniten.edu.my
AbstractSampling-based motion planning algorithms have
been successfully applied to various types of high-dimensional
planning tasks. Recently an extension of PRM algorithm called
PRM* planner has been proposed which guarantees asymptotic
optimal solutions in terms of path length. However, the high
runtime of sampling-based algorithms is still a serious
disadvantage. In this paper, a new extension of PRM planner is
proposed which incorporates the variable neighborhood radius
feature of PRM* and the sampling radius of low-dispersion
sampling in order to improve the cost of the generated solutions
in terms of path length and runtime. The performance of the
proposed algorithm is tested in different planning environments.
Furthermore, the proposed planner is compared to the original
PRM and the PRM* approaches and shows significant
improvement.
KeywordsSampling-Based, Motion Planning, PRM, lowdispersion

I.

INTRODUCTION

One of the fundamental problems in robotics is motion

planning. This problem can be stated as moving a robot
between a pair of start and goal configurations without
collision with present obstacles. A simplified version of the
motion planning problem is planning a collision-free path for a
robot made of an arbitrary number of polyhedral bodies among
an arbitrary number of polyhedral obstacles, between two
collision free positions of the robot. This instance of the
problem has been shown to be PSPACE-Complete [1]. In cases
where the problem is more complex (e.g. taking into account
the physical properties, and actuator limitations in a real robot)
it is not known if the problem is even decidable except for
some particular cases [2].
In order to solve practical motion planning problems,
sampling-based motion planning algorithms have been
developed such as the Probabilistic Roadmap (PRM) [3] and
Rapidly Exploring Random Trees (RRT) [4]. These planners
are unique in the fact that planning occurs by sampling the
configuration space of the robot [5]. Generally, sampling-based
algorithms can be divided into two main groups namely multiquery and single-query. In the multi-query class, the
configuration space is sampled in a learning phase and then the
resulted graph can be used to solve any given query. In the
single-query class, the sampling will be used for a specific
query and normally a tree structure will be grown rooted at the
start configuration.
Since the development of sampling-based algorithms,
reducing the cost of the generated solutions has been one of the

IEEERAS 2014

challenging requirements. The cost of a solution can be the

length of the generated trajectory or the runtime of the planner
for computing such solution. Unfortunately, these two types of
costs are contradictory and improving one of them normally
results in worsening the other. Several attempts have been
made to reduce the cost of the solutions in sampling-based
motion planning. The problem of computing optimal motion
plans has been proven to be very challenging even in basic
cases [6]. In the context of sampling-based motion planning,
the optimality properties have not been systematically
investigated, and most of the relevant work relies on heuristics
[7-12]. However, optimality guarantees of these algorithms are
only ensured up to the grid resolution and since the number of
grid points grows exponentially with the dimensionality of the
state space, so does the (worst-case) running time of these
algorithms. Furthermore, the classical sampling-based motion
planning have been proven to be enable to provide optimal
solutions [6].
Recently, an improved version of PRM planner has been
proposed which provides asymptotic optimal solutions [6].
This algorithm (PRM*) is able to provide optimum trajectories
with a sufficient set of inputs (nodes). It considers a variable
neighborhood radius around each sample and connections are
made only for nodes within that region. This planner has been
specifically designed for reducing the length of the generated
paths without considering the runtime of the planner as a part
of the cost function. This is the main disadvantage of the
asymptotic optimal planners.
In this paper, an improved version of PRM algorithm is
introduced that combines the asymptotic optimal feature of the
PRM* algorithm and the concept of dispersion. The main
objective is to get similar results with a much smaller graph.
The proposed planner incorporates the neighborhood radius of
the PRM* and the sampling radius of the dispersion-based
PRM aiming to provide similar level of optimality while the
size of the graph is reduced. In the following sections, first
different versions of the PRM planner are discussed. Next, the
proposed planner is introduced. Later, the proposed algorithm
is simulated to evaluate and compare its performance in
different environments. Finally, the paper is concluded with
some suggestions for future research on this topic.
II.
A.

BACKGROUND

Algorithm
The PRM algorithm has originally been designed for
motion planning in high-dimensional configuration spaces. A

103

typical PRM consists of a learning phase and a query phase. In

the learning phase, a set of random collision-free
configurations are generated over the configuration space and
connected together to form a graph. This graph is supposed to
capture the connectivity of the environment. In the query
phase, the resulted graph is used to solve any given query.
Since the query phase is basically a graph search procedure, the
focus of this research is on the learning phase and the proposed
improvements are being applied on this part of the PRM
planner. The learning phase of the PRM planner begins with an
empty graph. At each iteration, a collision-free sample is
generated and added to the set of graphs vertex ( ). Then,
connections are attempted between the sample and other
vertices in
within a ball of radius
centered at the
configuration of the sample using a local planner like straightline connection. Successful connections result in adding a new
edge to the set of edges
. This procedure continues times
and the resulted structure is an undirected graph with nodes.
B.

Algorithm
In the standard PRM algorithm, connections are attempted
between roadmap vertices that are within a fixed radius from
one another. The constant is thus a parameter of PRM. The
PRM* algorithm is similar to PRM, with the only difference
being that the connection radius is chosen as a function of ,
the size of the graph.
]

) (

(1)

Where, is the dimension of the configuration space and

denotes the volume of the obstacle-free space.
Clearly, the connection radius decreases with the number of
samples. The rate of decay is such that the average number of
connections attempted from a roadmap vertex is proportional
to
. Algorithms 1 shows the learning phase of the PRM
and PRM* planners. Note that the only difference is in
choosing the value for which is a constant in PRM and a
function of in PRM*.

Algorithm 1. PRM and PRM* learning phase

1
2
for i=0, , n do
3
a random collision-free configuration
4
U
5
6
for each
, do
7
if
is collision-free
8
9
return
III.

samples for capturing the connectivity of the configuration

space.
Before discussing the proposed algorithm a brief notation
about measuring the quality of a set of samples seems
necessary for the purpose of this research. Considering
as
the Lebesgue measure (or volume) for , one can conclude that
if a set of samples are ideally uniform, then it seems reasonable
that the number of those samples that lie in any subset should
be roughly
. The concept of discrepancy is defined
to measure how far from the ideal the samples are:
|

(2)

Where | | denotes the cardinality of a finite set . While,

discrepancy, as a measure-based criterion, provides a measure
of how uniformly the samples are distributed over the space, a
metric-based criterion, known as dispersion, provides a
measure of the largest portion of the space that contains no
samples:
(3)
Note that
denotes any metric, such as the Euclidean
distance. Intuitively, corresponds to the radius of the largest
empty ball centered inside the space. Therefore, an
experimental way to measure the dispersion of a set of samples
is to find the largest ball centered in the workspace and
containing no samples.
The concept of dispersion is used for defining the sampling
radius around each sample. This radius can be calculated as
follows:
[ (

(4)

Each time a sample is generated, a virtual circle is created

entered at the samples position and with the radius
.
Next samples are not allowed to be inside the sampling circles
of the current samples. In other words, each pair of samples
should not be closer to each other than
.
(

(5)

This procedure will make the resulted graph sparser and the
connectivity of the environment can be captured will a much
smaller set of samples. In has been shown that using this
method improves the dispersion of the generated samples [13].
Fig. 2 shows the difference between uniform sampling and the
proposed method for sampling in a plain environment. The
resulted samples from the proposed mechanism are sparse and
almost evenly distributed throughout the configuration space
without imposing substantial incensement in the running time
of the planner.

PROPOSED ALGORITHM

A variant of the PRM* planner is presented which

improves the quality of the generated samples by defining a
virtual circle around each sample and forbidding further
sampling in that area. This heuristic procedure makes the
resulted graph sparser and reduces the number of required

IEEERAS 2014

104

only different is in the sampling procedure (line 3) where the

generated sample is being checked to be out of the forbidden
regions of other samples.

(a)

Algorithm 2. The learning phase of the proposed

algorithm
1
2
for i=0, , n do
3
a random collision-free configuration
if
4
U
5
6
for each
, do
7
if
is collision-free
8
9
return

(b)

Fig. 1. The resulted samples from (a) uniform sampling and (b) proposed
approachin a plain 2D environment using 1000 nodes.

This approach and the connection radius used in the PRM*

planner are utilized in the proposed planner. The procedure of
the sampling and connection is illustrated in the following
figure. After a sample was generated, a circular area around it
with the radius of
will be considered as forbidden and
further sampling cannot take place in this area. At the same
time, another circular area will be considered with the radius
of
and current sample will only be connected to the
samples inside this region. Its noteworthy that the following
statement always remains true.

IV.

RESULTS

For testing the performance of the proposed algorithm and

also to compare it with the original PRM and the PRM*
planners, four different environments are designed and the
proposed algorithm is simulated to solve these problems using
MatLab on a computer with Core i5 processor and 8 GB of
memory. Based on the random behavior of the algorithm, a
large number of experiments (1000) are performed to present a
more realistic characterization of the performance of the
proposed planner. One instance of the results in each
environment is shown in Fig. 4.

(3)
Fig. 2 illustrates the concepts of neighborhood and
forbidden regions. As can be seen, there is no samples inside
the forbidden area and connections only happen inside the
neighborhood region.

Fig. 2. The illustrations of sampling and neighborhood ranges in the proposed

PRM. Sampling is forbidden in the dark area while connections between
neighbors are made only in the neighborhood region.

Fig. 3. Performance of the proposed planner in 4 test environments. Start and

goal configurations of the circular robot are shown by black circles. Note that
these are actual environments not the configuration spaces.

These two radiuses together improve the performance of

the planner in terms of path length and runtime. The procedure
of the proposed algorithm is presented in Algorithm 2. The

Note that actual environments are shown in this figure with

a circular robot with radius equal to 0.25 meter. The robot
moves from the start configuration which is depicted with

IEEERAS 2014

105

yellow color and reaches the goal position as shown in green

color in Fig. 3. Simulation result shows that the proposed
algorithm generates safe and almost optimal paths. More
details about the simulation studies are presented in Table I. As
can be seen in Fig. 3, the resulted graphs are spars and include
significantly less nodes comparing to the original PRM and the
PRM* planners.
TABLE I.

SIMULATION RESULTS IN TEST ENVIRONMENTS.

VALUES ARE AVERAGED OVER 1000 ITERATIONS THROUGH DIFFERENT
PROBLEMS.
Graph
Size

PL

RT

PL

RT

PL

RT

10
100
200
300
400
500
600
700
800
900
1000

19.26
17.47
15.98
14.81
13.96
13.03
12.29
11.52
11.03
10.62
10.49

4.10
6.23
9.10
13.33
19.28
27.67
39.43
55.84
78.83
110.92
155.72

18.05
15.75
14.07
12.93
12.01
11.29
10.49
10.04
9.78
9.64
9.61

2.74
4.37
6.66
9.67
13.99
20.11
28.72
40.45
57.07
80.47
113.06

17.84
13.87
11.89
10.93
10.43
10.06
9.76
9.61
9.51
9.47
9.47

2.72
4.29
6.34
9.29
13.49
19.39
27.54
38.96
55.16
77.49
108.98

According to Table I, the proposed planner provides almost

optimal solutions with shorter paths than PRM and PRM*
planners. More important, as shown in Fig. 4, these short paths
are provided using smaller graphs than PRM* graph. For
instance, for getting a path with the length of 10, PRM*
requires at least 750 nodes while the proposed algorithm only
needs 500 nodes. Surprisingly, the runtime of the proposed
algorithm is similar to that of the PRM* algorithm as can be
seen in Fig. 4(b).

(a)

(b)
Fig. 4. Comparison results of the proposed algorithm, PRM and PRM* in
terms of (a) path length and (b) runtime. Averaged path lengths and runtimes
over test problems are being used for evaluations.

As an extension to this work, the proposed procedure can

be applied on the RRT* algorithm which is an asymptotic
optimal extension of the original RRT planner. Furthermore,
other improvement techniques for the quality of the set of
samples may be applied and analyzed to see the effect of
choosing different techniques in the sampling phase.

Fig. 5 shows the required graph size for having different

levels of optimality. As you can see, for instance, in order to
get 80% optimality, the proposed planner requires 300 nodes
while PRM and PRM* planners need 850 and 580 nodes
respectively.

V.

DISCUSSION

In this paper, a new extension of the PRM motion planning

algorithm was proposed which incorporates the adaptive
neighborhood radius of the PRM* algorithm and the variable
sampling radius of low-dispersion approaches in order to
reduce the cost of the generated solutions in terms of path
length and runtime. Each time the algorithm iterates, a new
sample is generated which is required to be collision-free and
also far enough from other samples. A sampling radius was
defined to define the forbidden region around each member of
the graph. Afterwards, a neighborhood radius is defined and
connection will be made between different samples if their
distance is less than this radius. The resulted algorithm is able
to provide asymptotic optimal solutions with surprisingly fewer
samples.

Fig. 5. Required nodes for achieving different levels of optimality. The

proposed algotihm is able to achive greater optimality eith smaller graphs.

REFERENCES
[1]

[2]

[3]

[4]

IEEERAS 2014

106

K. I. Tsianos, I. A. Sucan, and L. E. Kavraki, Sampling-based robot

motion planning: Towards realistic applications, Comput. Sci. Rev.,
vol. 1, no. 1, pp. 2-11, 2007.
P. Cheng, G. Pappas, and V. Kumar, Decidability of motion planning
with differential constraints, in: IEEE International Conference on
Robotics and Automation, 2007, pp. 18261831.
L. E. Kavraki, P. Svestka, J. C. Latombe, and M. H. Overmars,
Probabilistic roadmaps for path planning in high-dimensionla
configuration spaces, IEEE Trans. Robot. Autom., vol. 12, no. 4, pp.
566-580, 1996.
S. M. LaValle, and J. J. Kuffner, Randomized Kinodynamic planning,
Int. J. Robot. Res., vol. 20, no. 5, pp. 378-400, 2001.

[5]
[6]
[7]

[8]

[9]

M. Elbanhawi, and M. Simic,"Sampling-Based Robot Motion Planning:

A Review," Access, IEEE, vol. 2, pp. 56-77, 2014.
S. Karaman, and E. Frazzoli, Sampling-based algorithms for optimal
motion planning, Int. J. Robot. Res., vol. 30, no. 7, pp. 846-494, 2011.
Y. Kuwata, J. Teo, G. Fiore, S. Karaman, E. Frazzoli, and J.P. How,
Real-time motion planning with applications to autonomous urban
driving, IEEE Trans. Cont. Syst., vol. 17, no. 5, pp. 1105-1118, 2009.
C. Urmson and R. Simmons, Approaches for heuristically biasing RRT
growth, In Proceedings of the IEEE/RSJ International Conference on
Robotics and Systems (IROS), 2003.
D. Ferguson and A. Stentz, Anytime RRTs, In Proceedings of the
IEEE/RSJ International Conference on Intelligent Robots and Systems
(IROS), 2006.

IEEERAS 2014

[10] L. Jaillet, J. Cortes, and T .Simeon, Sampling-based path planning on

configuration-space costmaps, IEEE Trans. Robot., vol. 26no. 4, pp.
635-646, 2010.
[11] D. Berenson, T. Simeon, and S. Srinivasa, Addressing cost-space
chasms in manipulation planning, In IEEE Conference on Robotics and
Automation (ICRA), 2011.
[12] M. Likhachev and D. Ferguson, Planning long dynamically-feasible
maneuvers for autonomous vehicles, Int. J. Robot. Res., vol. 28, no. 8,
pp. 933-945, 2009.
[13] W. Khaksar, T. S. Hong, M. Khaksar, and O. Motlagh, A low
dispersion probabilistic roadmaps (LD-PRM) algorithm for fast and
efficient sampling-based motion planning, Int. J. Adv. Robot. Syst.,
vol. 10, no. 397, pp. 1-10, 2013.

107