New Potential Functions for Multi robot path planning : SWARM or SPREAD

Sung-hwan Kim Gyungtae Lee, Inpyo Hong, Young-

Department of Information and Joo Kim, Daeyoung Kim
communication engineering Department of Computer Science
Korea Advanced Institute of Science and Technology Korea Advanced Institute of Science and Technology
Daejeon, Korea Daejeon, Korea
[email protected] {handreic, Inpyo, yjkim73, kimd}@kaist.ac.kr

Abstract— Artificial Potential Field (APF) is widely used for an placed are defined to have high potential. Therefore, it leads
autonomous robot path planning and navigation because of the robot to avoid obstacles. The potential factors are the
light complexity and elegance of results. Although it is useful in followings; repulsive high potential by obstacle and another
single robot path planning, appropriate algorithm for multi- robot, and attractive low potential by goal with lowest
robot path planning has not been proposed. Existing APFs potential. Because calculation overhead for getting potential
which can apply to multi-robot only regard robots as obstacles factors needs to only simple calculation and the number of
even if these robots are not obstacles, or focus on SWARM potential factors is little, the APF methods are attractive in
formation and moving. This paper proposes a multi-robot robot path planning especially in real-time and dynamic
support APF by defining new potential functions: SWARM or
environment application, not pre-calculating application.
SPREAD. This APF uses priority selection scheme between
However, APF methods causes local minimum [9][10].
multi robots and calculates a new potential functions based on
priority, and then solves local minimum problem in multi- Because a next plan is calculated based on the local obstacles
robot APF which is a traditional problem of APF. Therefore by adding or subtracting potential values, sometimes it
this paper enables distributed path planning in multi-robot causes local minimum problem. It means that although the
with real-time using suggested APF. All of these works are current position is not global minimum, the current position
verified its effectiveness with simulation using Player/Stage looks like the lowest potential position so the robot can’t
simulator. make next plan. To solve local minimum problems, these
works [17~19] propose potential functions which guarantees
Keywords-component; multi-robot path planning; Aritificial no local minimum occur, global minimum occurs in goal
potential field; potential function; priority selection position. To extend APF to multi-robot path planning,
complex calculation is harmful to use APF. All of existing
I. INTRODUCTION schemes [1][11][12][14] for multi robot path planning are
like this; one is centralized approach which one agent plans
For an autonomous robot path planning, Artificial all other robots path planning then spread each plan to each
Potential Field (APF) methods are widely used from network robot [11], the other is coordination approach which each
topology control [1], Robot soccer team in ROBOCUP [2], robot plans their path plan and path coordination procedure
planetary exploration application [3] and etc. The reason exists to coordinate all robots path to avoid collision between
why these methods are used in various fields has advantages robots in Time-Space configuration [12] which consider
on simplicity. Comparing methods to APF for robot path space collision with time change. To apply these APF
planning is Artificial Intelligence (AI) approaches. Generic schemes to multi robot systems with reducing calculation
algorithm [4], fuzzy control algorithm [24], and neural complexity, there are some researches [1][3][14]. Although
network [5] approaches are used for robot path planning. those researches include applying APF to multi-robot
These approaches, however, have limits in real-time systems and get results from that, it is only application for
application because of their computational complexity. specific situation; SWARM formation [3], Topology control
Compared with these AI approaches, APF requires only a [1], and perception potential sharing [14] which does not
simple calculation but powerful result and elegance output have specific procedure in general path planning using APF.
are generated in short time [6-8]. Priority selection or different goal is not considered in those
The basic idea of APF is to define potential [8] researches. Therefore, these approaches are not appropriate
throughout the fields and robot where potential is a power to apply general application which uses multi-robot path
which obstacles generate for collision avoidance in robot planning.
navigation. Lower potential position attracts the robot and To solve these limitations, we propose a multi-robot
higher potential position repulses the robot. So each robot support APF for multi robot path planning aiming on real-
makes a plan to go to lowest potential position defined as time and distributed algorithm. By adding new potential
goal position. Positions where obstacles or other robots are factors to potential functions for SWARM on group moving

978-1-4244-5586-7/10/$26.00 C 2010 IEEE

557 Volumn 2
and SPREAD on different priority, APF can be extended for but limited in time complexity so it only supports up to three
multi-robot path planning and distributed approaches to robots. Other approach is using a heuristic search [15][17] or
multi-robot APF erase avoid path coordination phase randomized approach [17]. However heuristic search are
between robots. limited because they cannot support in both velocity and
This paper is organized as follows; in section 2, we first total path length and randomized approach takes much times
describe the robot model and priority selection scheme to get the satisfied result. In this paper, Priority selection
applying to multi-robots with real-time and this between multi-robots is defined for both predefined and
model/scheme are used throughout the paper. In section 3, state-dependent priorities. To aim on real world, predefined
we propose new potential factors to support multi-robot APF priority is used for emergency robot. In addition to this,
issuing relationship between robots; SWARM or SPREAD state-dependent priority is defined as following. The purpose
reflected in potential functions and modifying potential of priority selection is to minimize overall robot’s path
functions for solving local minimum problems to guarantee length. SWARM robots which has same goal have same
no local minimum occurs during path planning. In section 4, priorities.
we define new potential functions with potential factors ‡ SWARM robots have same priority which is
described in section 3 and suggest methods applying to robot.
predefined.
In section 5, we verify the algorithms by simulation with
Player/Stage which is famous robotics simulator. Finally in ‡ Predefined robot has lower priority than others.
section 6, we conclude and sum up this paper and suggest ‡ Higher velocity has higher priority than lower
future work.
velocity. (Needs different constant for admitting)
II. BACKGROUND ‡ If they have same velocity, the robot which has
A. Robot Model longer distance has higher priority.
This scheme will support priority selection with
The robot model throughout this paper is two-wheel considering both velocity of robot and total path length
based robot. To emphasize with reality robot, it has a overall robot. First two conditions are needed to describe in
maximum turn-rate in unit time. It means that the robot predefined priority and conditions for SWARM potential
cannot turn in fixed position so it changes its direction during function which is requires to track leader robot predefined.
moving forward. It is a simple robot model from Third condition is needed for reducing total path length, the
Player/Stage simulator which is a simulation tool for lower priority means himself goes around higher priority one.
verification our idea described in the paper. The surplus path is generated by priority. To minimize
The robot has a two part described in Fig. 1. One part is surplus path, it needs that higher velocity one goes around
inner part of robot. Because of inexact control and measure lower velocity one. Final condition is for minimizing total
of sensor, it always tries to keep distance above than inner end time between robots by giving higher priority to the
part between other objects. The outer part means the robot which is expected to take more time to reach the goal.
SWARM distance which is used for standardized distance
when robots that have same priorities move together in same III. NEW POTENTIAL FACTORS
SWARM moving group.
In this section, we describe new potential factors to
B. Priority Selection extend APF for multi-robot path planning. In these new
For the multi-robot path planning, how to select priority potential factors, we assume multi-robot relation can be
of each object is first step to calculate potential. Priority for defined as SWARM or SPREAD using priority selection
the path planning is classified as two kinds of element; described in section II. It means that robots have a plan with
predefined priority and state-dependent priority. Predefined together in SWARM robot, SPEARD in different priority. A.
priority is needed for distinct object like police car or and B. describe each potential factor which occur for
ambulance which is not same as ordinal robot. It can be SWARM and SPREAD potential to the multi-robot.
defined by predefined robot planning policy. State-dependent A. Potential factor for SWARM
priority is a priority for competing between objects which
In SWARM which constructed same priority, robots in
has non-predefined priority. For selection of state-dependent
same group generate potentials to track leader robot which is
priority, there exists complete priority select calculation [16]
predefined to move goal together. If a robot goes farther than

Figure 1. Robot model Figure 2. Potentials by robot between same priority

x:distance between robot y: potentials

558 Volumn 2
outer part, minus potential is generated to attract other robots C. Local minimum for new potential facotrs
by leader robot to move together. If robot places shorter than Total potential function by adding original APF and
outer parts, positive potential is generated to repulse and potential factors described in above causes local minimum
prohibits collision between robots. This scheme which uses problems in using APF.
farther attractions and shorter repulsions robots between APF path planning has a problem called to local
same priority robots in same SWARM group enables move minimum which is internal problem of its structure of
them together. calculation. To solve this problem, Zhu [23] suggest solution
The potential factors generated by SWARM describes of local minimum by adding goal direction. For our new
following in Eq. (1). potential functions to solve local minimum problem, we
Ut = αρ(R, L) +
1
+ γ extend the scheme of Zhu and it is for our new potential
βρ (R,L) factors described following expression Eq. (3).
α, β, γ ∶ constant
ρ(R, L): distance bewteen robot and leader Ul = Uc ∗ ρ
r, rgoal  



By potential generated by Eq. (1), potential factors make ²

r, rgoal  distance between robot and its goal

robots moves together. As you can see in Fig. 2, the domain
By multiplying goal position factors to potential field
of potential between SWARM forces and navigation
generated by SPREAD potential factors made by priority, it
potential are different. It leads robot not to conflict on
guarantees that robot does not stop in local minimum and
navigation by SWARM forces between robots to move
goes to goal without stop by potential.
tighter for formation SWARM.
B. Potential factor for SPREAD IV. NEW POTENTIAL FUCTION

If robot paths generated by potential field navigation By original APF schemes and new potential factors
have a common point in Time-space configuration, it needs described in III, we finally describe procedure of getting
path coordination for avoiding collision. In this paper, we potential in path planning using SWARM or SPREAD
propose potential factors for SPREAD in different swarm scheme described by new potential factors. Potentials are
groups. The basic idea of this factor is higher priority robot basically generated by goal attraction and obstacle repulsion
turn around lower priority robot with ellipse motion. This with multiply goal distance factor for solving local minimum.
idea describes as expression following in Eq. (2). In addition to this, potential factors for SWARM or
SPREAD are added to potential. For robot swarm Eq. (1) are
Uc = ∑p i <p r P if ρ(i, r) <   used for tracking leader robot by another robot in same group.


P: constant for priority speard In addition to this, Eq. (2) is used for priority factors incur
pi : priority of other robot pr : priority of robot robot goes to spread without obstacle repulsion which induce
ρ(i, r) ∶ distance between other robot and self high priority robot can keep their going to the goal without
By giving a potential to the low priority robot, lower consider around lower priority robot. All of calculations are
priority robot keep its path and higher priority robot will turn executed in distributed. Therefore, all of potential field can
around lower priority robot with ellipse motion with be different in each robot.
preventing collision between robots. Expected result by this The expressions and forces to manipulate robots are
potential can see in Fig. 3. In figure it represents difference described in following expression Eq. (4).
between potential APF with obstacles in other robot and U = Kρ(R, G)n + ∑Ni=1 μ(
1
)n ∗ ρ
r, rgoal  + Ut + Ul (4)
potential factors for SPREAD. The left figure show that two ρ(R,O i )
robot repulse together even if it is not necessary to repulse by Kρ(R, G)n  attract by goal
gray line robot. In right figure which apply potential factors ρ(R, G)n distance between robot and goal
it shows expected result that gray line turn around black line ρ(R, Oi ): distance bewteen robot and obstacle
and black line goes first to end up planning. Force in Eq. (5) gathered by differential to potential
functions will lead robots navigation to the goal.
F = −∇U (5)
1 n 1
F = K1 ρ(R, G) + K2 ∑Ni=1 μ( ) + + K3
ρ(R,Oi ) (βρ (R,L))2

K1 , K 2 , K 3 ∶ constant generated by potential function

This navigation by forces can be seen as path planning by

APF. In simulation they use forces expression because it is
definitely result of potential calculation, also it is hard to
differentiate function in real time to get the force in point
place. The validation of these new potential functions verify
Figure 3.expected result using potential factors for spread by simulation in section V.

559 Volumn 2
 V. SIMULATION
In this section, we demonstrate the new potential
functions for multi-robot path planning. We use Player/Stage
simulator [21][22] which is famous server-client based
simulator in robot field. Simulation goes on SWARM
formation result and compares the results between multi-
robot path planning cases using origin APF which regards
other robots as obstacle and multi-robot path planning cases
using new potential functions. Finally, we show the result of
SWARM or SPEARD together for multi-robot path planning
A. SWARM formation moving result
For the same priority group they construct swarm Figure 5. APF vs new potential functions in total path length
formation and move together centered on priority leader. In
this simulation, two robot swarm, three robot swarm, and B. APF vs new potential functions in total path length
swarm robot with obstacle avoidance for multiple robots. In this simulation, we compare APF and new potential
Fig. 4 show the result of simulation and the setting functions in total path length. In this simulation, robot1 has a
parameter of robot model are inner part 0.5 and outer part 1.5. higher priority than robot2. Left figure represent simulation
The black line means the leader robot and another sign is of using APF with another robot which has potential same as
trails of other robot to track the leader robot. First figure obstacle. Therefore, even if velocity of two robots is
describes the cases of two robot swarm. Leader robot (0, 0), different which means faster robot can go front of slower
and robot2 in same group (-2, 1) are placed. In starting phase, robot, two robots keep the distance and occur unessential
robot2 goes fast because distance between leader robot and moving. Right figure represent simulation result of using
robot2 is long. Then, robot2 keep the distance with leader new potential functions. In this case, high priority robot goes
robot and goes to goal safety. Oscillation occurs because first rather than low priority one (which means lower priority
sensing and command ordering have a time difference in robot keeps their path, and high priority robot goes first
simulator. Anyway, purpose of swarm moving is satisfied. rather than low priority robot).
Second figure, Leader robot is at (0, 0), robot1 is at (-2, 2),
and robot2 is at (-3, -1) It shows success result in three robot Total path length Moving step
cases in reach to the goal without collision. APF 19.8 124
Last figure in Fig. 4 simulate swarm moving with New potential
15.9 91
obstacle avoidance. They show the result on obstacle function
avoidance with keeping distance on leader robot. Lined trail Table 1.Comparing APF and new potential functions
are path of leader robot.
The results of simulation in comparing new potential
functions with original APF sum up in table 1. Total path
length means total length of two robots. Get the each path
length of two robots and sum two lengths. It shows new
potential function shows good result in total path length.
About 20% reducing are occurs in total path length. Moving
step means the simulation scale time steps in simulator it
matches and propositional on real time scale. In robot model
defined in section II, turn overhead is required to turn the
robot in real world and player/stage simulator. Therefore
moving step is more accurate performance simulation metric
for reality. This means time required to start to goal planning.
As you can see, summation on moving step of two robots is
reducing using new potential functions about 25%. Because
new potential function hopes that no essential turning are
caused by another robot, this performance improve result are
better than total path length.
C. Result for SWARM or SPREAD
In this simulation, we apply total potential functions:
SWARM factors or SPEARD factors on between robots.
Robot1 and robot2 are same group with group leader 1. And,
Figure 4. SWARM formation and moving result Robot3 and robot4 are same group with group leader 3.In
addition to this the velocity of robot1 is higher than robot3,
therefore, priority of robot1 is higher than priority of robot3.

560 Volumn 2
As you can see in Fig. 6, robot1 goes turn around robot3 [4] Gerke, M., “Genetic path planning for mobile robots,” American
because potential generated by priority factors. They Control Conference, IEEE, 1999
SPREAD in the result. Robot2 swarms with robot1 and [5] C. C. Chang and K. T. Song, “A neural network predictor for mobile
robot navigation among moving obstacles,” the 6th International
robot4 swarms with robot3. They make swarm formation and Symposium on. Artif. Neural Networks, Tainan, Dec. 1994
moving together.. [6] O. Khatib, “Real-time obstacle avoidance for manipulators and
mobile robots,” Int. J. Robot. Res., 1986.
[7] D. E. Koditschek, “Exact robot navigation by means of potential
functions: Some topological consideration,” Int. Conf. robot. Autom.,
IEEE, 1987
[8] E. Rimon and D. Koditschek, “Exact robot navigation using artificial
potential functions,” IEEE Trans. Robot. Autom., vol. 8, no. 5, pp.
501–518, Oct. 1992.
[9] Y. Koren and J. Borenstein, “Potential field methods and their
inherent limitations for mobile robot navigation,” Int. Conf. Robot.
Autom., IEEE, 1991
[10] B. H. Krogh, ̌A generalized potential field approach to obstacle
avoidance,̍ Robot. Res., 1984
[11] Warren, C.W., “ltiple robot path coordination using artificial potential
fields,” International Conference on Robotics and Automation, IEEE,
1990
Figure 6. Result for SWARM or SPREAD [12] M. Erdmann and T. Lozano-P&ez, “On multiple moving obstacles.”
Leader1 group turns around leader3 group, because International Conference on Robotics and Automation, IEEE, 1986.
robot1 has higher priority over robot 3. It keeps their path [13] Barnes, L., Fields, M.A., Valavanis, K., “Unmanned ground vehicle
and all robots in that group go to their goal fast. The swarm formation control using potential fields,” Control &
Automation, IEEE, 2007.
predefined information is goal of leader 1 group (6, 6) and
goal of leader3 group (5, 3). The result shows same as [14] J.L. Baxter, E.K. Burke, J.M. Garibaldi, M. Norman, “The effect of
potential field sharing in multi-agent systems,” The 3rd International
expected result in swarm and spread. Conference on Autonomous Robots and Agents, IEEE, 2006
[15] S.J. Buckley, “Fast motion planning for multiple moving robots,”
VI. CONCLUSION International Conference on Robotics and Automation (ICRA89),
This paper proposes a multi-robot support APF by IEEE, 1989.
defining new potential functions: SWARM or SPREAD. [16] K. Azarm, G. Schmidt, “A decentralized approach for the conflict-
This algorithm calculates a new potential functions for free motion of multiple mobile robots,” IEEE/RSJ International
Conference on Intelligent Robots and Systems (IROS), IEEE, 1996.
swarm between robot, spread based on priority, and then
[17] Maren Bennewitsz, Wolfram Burgard, Sebastian Thrun, “Finding and
solves local minimum problem. To prove empirically, we optimizing solvable priority schemes for decoupled path planning
have used Player/Stage simulator. As a result, simulation techiniques for teams of mobile robots,” Robotics and Autonomous
shows the result of swarm moving and performance System 41, 2002, 89-99
improvement in path planning by using this algorithm. [18] P. Vadakkepat, K. Chen Tan, and W. Ming-Liang, “Evolutionary
Therefore, this paper enables distributed path planning in artificial potential fields and their application in real time robot path
multi-robot with real-time using suggested algorithm. In planning,” Congress on Evolutionary Computation, IEEE Press, 2000.
future, the work will be expanded to include APF-based path [19] S. S. Ge and Y. J. Cui, “New potential functions for mobile robot path
planning in more reality environment; planning based on planning,” IEEE Trans. Robot. Autom., vol. 16, no. 5, pp. 609–615,
Oct. 2000.
inexact perception, dynamic environment, and APF applying
[20] De-Shuang Huang, Laurent Heutte and Marco Loog, “Artificial
to various robot models. Potential Field Based Path Planning for Mobile Robots Using Virtual
Water-Flow Method”, Communications in Computer and Information
ACKNOWLEDGMENT Science vol. 2, pp. 588-595, Aug 2008
This work was supported by Korea Science & [21] B. P. Gerkey, R. T. Vaughan, K. Støy, A. Howard, G. S. Sukhatme,
Engineering Foundation through the NRL Program. and M. J. Matari´c. “player: A robot device server for distributed
control.” IEEE/RSJ Intl. Conf. on Intelligent Robots and Systems
(IROS01), IEEE, 2001.
REFERENCES
[22] R. T. Vaughan. “Stage: a multiple robot simulator.” Technical Report
[1] A. Howard, M. Matari´c, and G. Sukhatme., “Mobile sensor network IRIS-00-393, Institute for Robotics and Intelligent Systems,
deployment using potential fields: A distributed, scalable solution to University of Southern California, 2000
the area coverage problem,” the 6th International Symposium on [23] Qidan Zhu, Yongjie Yan, and Zhuoyi Xing, “Robot Path Planning
Distributed Autonomous Robotics Systems (DARS02), 2002.1 Based on Artificial Potential Field Approach with Simulated
[2] S. Johansson and A. Saffiotti., “An electric field approach to Annealing,” Sixth International Conference on Intelligent Systems
autonomous robot control,” Lecture notes in artificial Design and Applications, IEEE Computer Society , 2006
intelligence(RoboCup 2001), Springer Verlag, 2002
[24] G. T. Zoumponos, and N. A. Aspragathos, “Fuzzy logic path planning
[3] M. Massari, G. Giardini, and F. Bernelli-Zazzera., “Autonomous for the robotic placement of fabrics on a work table,” Robotics and
navigation system for planetary exploration rover based on artificial Computer-Integrated Manufacturing, Pages 174-186, 2008
potential fields,” the 6th Conference on Dynamics and Control of
Systems and Structures Space (DCSSS), 2004.

561 Volumn 2