Professional Documents
Culture Documents
Network Weight and Time-Varying Potential Function For Obstacle Avoidance of Swarm Robots in Column Formation 2022
Network Weight and Time-Varying Potential Function For Obstacle Avoidance of Swarm Robots in Column Formation 2022
Integration
To cite this article: Shotaro Shibahara, Takuma Wakasa & Kenji Sawada (2022) Network
weight and time-varying potential function for obstacle avoidance of swarm robots in column
formation, SICE Journal of Control, Measurement, and System Integration, 15:1, 24-35, DOI:
10.1080/18824889.2022.2032542
1. Introduction
Two issues are raised in this paper. The first issue is
A multi-agent system (MAS) allows multiple agents to to alleviate the imbalance between the attractive force
share information with each other, so that the agents for staying the formation and the repulsive force for
as a whole can efficiently accomplish tasks that would avoiding obstacles. The second issue is to eliminate the
be difficult to accomplish independently [1, 2]. There stagnation of the swarm robots in front of the large
has been a growing motivation to control large swarm obstacle.
robots of more than several thousand units [3, 4]. One To address the first issue, this paper proposes a net-
of the important problems is obstacle avoidance while work weight function that changes the gravitational pull
swarm robots stay in the desired formation [5]. Typical for staying the formation between robots based on the
methods of obstacle avoidance include potential func- positions of the robots and obstacles. Swarm robots in
tions [6, 7] and model predictive control [8]. Poten- formation are forced to break the desired formation
tial function methods realize the collision avoidance to avoid obstacles. The larger the difference between
by using the repulsion force that increases with the the actual distance between robots and the desired dis-
distance between the robot and the obstacle. In this tance, the larger the gravitational pull of the forma-
case, the robot near the obstacle may show oscillatory tion control. The previous study [9] proposes a weight
behaviour. This is caused by the fact that the control function that weakens the attraction to stay in the for-
inputs for staying formation and avoiding obstacles mation as the distance between robots becomes larger
in opposite directions. In the previous study of the than the desired distance. However, when the robots are
current authors [9], the imbalance between the two close to each other, the force to stay the formation also
control inputs is alleviated by lowering the priority of increases, which causes oscillation. In other words, the
the formation control. However, the number of robots force imbalance between the robots in preserving the
assumed in [9] is much smaller than that in [1, 2]. This formation causes the oscillation. To address this issue,
paper considers the case where the number of robots is we propose a weight function that weakens the forma-
significantly increased from the number set in [9]. tion control according to the displacement between the
In this paper, we propose a method of changing the distance between robots and the desired distance. This
potential function for obstacle avoidance in addition to network weight prevents the formation control from
the formation control for the column formation prob- becoming too large even when the robots are close to
lem [10]. The objective of this paper is to reduce the each other.
oscillatory behaviour near obstacles and to recover and To address the second issue, we set a relative angle
preserve the formation after the avoidance of obstacles. based on the positional relationship between the robot
and the obstacle and introduce a weight parameter adjacent matrix A = [aij ] ∈ Rn×n is defined as follows:
to change the repulsion force of obstacle avoidance.
The stagnation in front of the large obstacle causes a 1 if (j, i) ∈ E and i = j
aij = . (1)
misalignment in the positional relationship between 0 otherwise
the robots and may make it impossible to stay in the
desired formation after avoiding obstacles. This reason The neighbourhood of agent i is the set of agent
is thought to be that the potential function for obsta- j(= i) to which agent i is adjacent. The number of edges
cle avoidance is a circularly spreading function. Refer- entering vertex i is called the input degree diin and given
ence [11] adds weights depending on the relative angle by
between the robot and the obstacle to the circularly
n
spreading potential function. The resultant range of the diin = aij , i = 1, 2, . . . n.
influence of the repulsion force changes from circularly j=1
spreading one to teardrop one and smooth obstacle The degree matrix D of graph G with the diagonal
avoidance is achieved. Motivated by this method, we element of d1in , d2in , . . . , dnin is defined as follows:
propose a new weight that changes the shape of the
potential function depending on the relative angle and D = diag d1in , d2in , . . . , dnin .
construct a time-varying repulsive potential function.
From D and A, the Graph Laplacian is given by
By changing the shape of the potential function to an
ellipse instead of a circle, swarm robots do not stagnate L = D − A.
in front of the large obstacle and can stay in the desired
formation after passing through. Also, Perron matrix with some positive number
The conference version of this paper [9] only con- ε(0 < ε < 1) is defined as follows:
siders the design of the network weight function against P = I − εL
the first issue. The function of this paper is a generalized
form of [9]. Also, the second issue is not considered in where I is the identity matrix.
[9]. The time varying repulsive force for swam robots is A weighted graph is a graph in which each edge has a
the contribution of this paper compared with [11]. The positive real number (weight). The weight graph is rep-
number of robots to which the proposed method can resented by (G, w) where ω : E → R+ \{0} is a weight
be applied is comparable to those in [3, 4]. This paper function. Also, the adjacency matrix A of the weighted
suppresses the oscillatory behaviour of a large number graph (G, w) is defined as follows:
of robots, which has not been solved in [9].
The organization of this paper is as follows. Section 2 ω((j, i)) if (j, i) ∈ E and i = j
aij = .
explains consensus control of MAS. Section 3 describes 0 otherwise
the problem setting of the obstacle avoidance for swarm
robots in column formation. Section 4 describes two 2.2. Consensus control
proposed methods: a method for setting the network
weight function and a method for designing a time- Consider the case that agent i(= 1, 2, . . . , n) is given by
varying potential function. In Section 5, numerical an integral system in discrete-time domain as follows:
experiments are conducted using MATLAB, and the
xi [k + 1] = xi [k] + ui [k].
summary of this paper is given in Section 6.
The control input of agent i considering the offset d
is given by
2. Preliminaries
n
This section explains the distributed optimization of ui [k] = −ε aij xi [k] − xj [k] − d . (2)
MAS [12–15] which is the basis of the proposed j=1
method.
We transform Equation (2) as follows:
n
2.1. Graph theory ui [k] = −ε aij xi [k] − xj [k] − d
j=1
The network structure for information transfer bet
ween agents is represented by graph G = (V, E) com-
n
= −ε aij xi [k] − xj [k] + bi (3)
posed of the vertex set V = {1, 2, . . . , n} and the edge j=1
set E ⊆ V × V. The edge (i, j) ∈ E indicates an infor-
mation transfer path from agent i to agent j, that is, where bi = εd nj=1 aij . Equation (3) is a consensus
agent i ∈ V is adjacent to agent j ∈ V. If all the paths are control with bias, so bi has no effect on the stability anal-
bi-directional, then G is called an undirected graph. An ysis [12]. In Equation (3), we set bi = 0. Summarizing
26 S. SHIBAHARA ET AL.
the agents and the inputs, the state-space equation of 3. Problem setting
consensus control is represented by
This paper considers a column formation problem in
x[k + 1] = Px[k]. (4) which N(= m × n) swarm robots move straight ahead
and avoid collisions on the x−y plane with c obstacles.
Since the Perron matrix P has eigenvalue 1, we get The network structure between the robots is an undi-
rected and connected graph as shown Figure 1. The
P1n = 1n . shapes of each robot and obstacle are circulars of radius
r and R, respectively, and their position coordinates are
Let k → ∞ in Equation (4), we get
their centres.
1 Let Ni be the neighbourhood set of robot i. When
lim x[k + 1] = 1n 1Tn x0 ,
k→∞ n the direction of the swarm robots is the positive direc-
tion of the x-axis, the other robots networked to
1
n
lim xi [k + 1] = x0j . the right, left, front and back of robot i are defined
k→∞ n j=1 as Ri , Li , Fi , Bi ∈ Ni (Ni = {Ri , Li , Fi , Bi }) as shown in
Figure 2.
This system proves that all agents achieve the average We assume a leaderless formation. In addition, each
consensus. robot obtains information about its surrounding envi-
ronment from sensors whose detection range is radius
2.3. Distributed optimization ρ(> 0). The adjacency matrix A = [aij ] ∈ Rn×n rep-
resenting the inter-robot connections is defined as fol-
Consider the unconstrained optimization problem: lows:
min Ji (ξ ) (5) 1 if pij ≤ ρ
ξ ∈R N aij = , j ∈ Ni (7)
i∈V 0 otherwise
where the cost function Ji (ξ ) is a real-valued func- where pij is the distance between robot i and robot j.
tion defined in RN and convex. The sub-gradient of This paper aims to make each robot follow the target
differentiable Ji (ξ ) is as follows: coordinates xg ∈ R while avoiding obstacles under the
∂Ji formation movement problem. For straight-line move-
dJi (ξ̄ ) = (ξ̄ ). ment of the column formation, xg is set far enough
∂ξ
away. In the y-direction, each robot only stays the
By using the consensus control of Section 2.2, the desired distance d ∈ R from its neighbouring robots.
system equation for solving Equation (5) is defined as Let the set of robots be V = {1, 2, . . . , n}. The dynamics
follows: of robot i ∈ V at discrete time k ∈ N is given by
xi [k + 1] = pij xj [k] − s[k]dJi (xi [k]) pi [k + 1] = pi [k] + ui [k],
j∈V T
pi [k] = pix [k] piy [k] (8)
where pij is an element of the Perron matrix P.
In this paper, we assume that the step width is a where pi [k] ∈ R2 is the x−y coordinate of robot i
constant s[k] = s(> 0). For any i ∈ V, the following position, and ui [k] ∈ R2 is the control input. ui [k] is
equation holds:
sC12 C2
lim sup J(x̂i [k]) ≤ J ∗ + ,
k→∞ 2
nβ
C1 = max dJi (ξ ) , C2 = n 1 + 8 2 + ,
ξ 1−β
η
β =1− , Figure 1. Network structure (m rows and n columns).
4n2
1
k−1
x̂i [k] = xi [h] (k = 2, 3, . . .)
k−1
h=0
f
composed of the consensus control law ui [k] ∈ R2 for For simplicity, regarding the convergence of the
g
preserving formation, the control law ui [k] ∈ R2 for algorithm proposed in this paper, we do not consider
reaching the target position, uci [k] ∈ R2 for avoiding the transient state with collision avoidance, but discuss
robot collision and uoi [k] ∈ R2 for avoiding obstacle the steady state. This is because the potential func-
collision as follows: tions do not always converge to certain values except
in steady-state conditions where there are no collisions.
f g
ui [k] = ui [k] + ui [k] + uci [k] + uoi [k]. (9) That is, if uci [k] = uoi [k] = 0 in Equation (9), then
relative distance between the robots. Also, as a solution where σ > 0 is a threshold. As shown in the blue line of
to the second issue, we design the input uo [k] in (10) Figure 3(a), this weight becomes 0 when the strength of
that changes the shape of the time varying potential the connection between the robots is less than σ . The
function depending on the relative angle between the weight aij in (20) includes the weight aij in (19) as a
robot and the obstacle. special case σ → 0. This disconnection eliminates the
imbalance between the attractive force for preserving
4. Main results the formation and the repulsive force when the robots
diverge each other due to the collision avoidance. On
To consider the two issues, we introduce the variable âi the other hand, there still remains the imbalance force
that represents the detection/nondetection of obstacle when the distance between the robots becomes close.
(= 1, 2, . . . , c) by robot i as follows [16]: This phenomenon is demonstrated in Section 5.
To solve the above open problem of [9], this paper
1 if pi − po − r − R ≤ D
âi = (18) proposes a new network weight function that changes
0 otherwise its value depending on the displacement between the
robot relative distance pi − pj and the desired dis-
where D > 0 is a threshold of detection/nondetecton, po
tance d as shown in Figure 3(b). That is, the proposed
is the coordinate of obstacle , r and R are radius of
network weight function is defined as follows:
robot i and obstacle , respectively. This variable âi is
used in the design of the network weight function and
the time-varying potential function.
aij = exp − pi − pj − d if j ∈ Ni . (21)
Figure 3. Network weight function: (a) previous method [9] and (b) proposed method.
SICE JOURNAL OF CONTROL, MEASUREMENT, AND SYSTEM INTEGRATION 29
go go go
Figure 5. Angle-based potential function: (a) θi 0, (b) θi π/4, (c) θi π/2.
Figure 6. Trajectory of robot avoiding obstacle: (a) conventional and (b) proposed.
First of all, we focus on convergence in steady state enough time has passed. This result indicates that the
without collision. The convergence with respect to the all potential functions do not adversely affect the objec-
objective function is shown in Figure 7 where only V g tive function V g for tracking in a collision-free steady
is considered because V g is for tracking. The target x- state.
coordinate of the first row (i = 1, 2, . . . , 100) is set to In the following sections, the target position is set to
xg and the target position of the robots in the n th row a common value for all robots for simplicity (all xgi are
from the top is set to xg − n d. Figure 7 shows that the same) and we discuss the oscillatory behaviour and
the all patterns converge to the target position after stability of the algorithms during collision avoidance.
SICE JOURNAL OF CONTROL, MEASUREMENT, AND SYSTEM INTEGRATION 31
5.1. Trajectories the previous study [9]. In pattern (d), we use the pro-
posed network weight function (21) and a potential
The results of numerical experiments are shown in
function with weights that adjust the repulsion depend-
Figures 8 and 9. The solid black line represents the
ing on the angle [11]. Pattern (d) of Figures 8 and 9
trajectory of the swarm robots, and the gray circles rep-
shows that the proposed network weight function (21)
resent the obstacles. Pattern (a) is the power imbalance
reduces the oscillatory behaviours near obstacles. How-
between maintaining formation and avoiding obstacles,
ever, the swarm robots are not able to recover to the
Patterns (b)(c) are the force imbalance between robots
desired formation after avoiding obstacles. The cause of
in formation control, Pattern (d) is the non-recovery
this is thought to be the stagnation of the swarm robots
of the formation after avoiding obstacles, Pattern (e)
in front of the large obstacle. In pattern (e), which uses
is recover formation after avoiding obstacles. In pat-
the proposed potential function (25), the swarm robots
tern (a), oscillatory behaviours are observed near the
are able to recover to the desired formation without
obstacles. Patterns (b) and (c), to which the method of
stagnating in front of the obstacle.
the previous study [9] is applied, also show oscillatory
Figure 10 shows the x-coordinate of the robots
behaviours. This is because the repulsion force becomes
(i = 1, 2, . . . , 100) at the head of the formation as the
too large due to the network weight functions (19)
ordinate and the number of steps k as the abscissa.
and (20) when the robots are too close to each other.
These figures indicate the degree of convergence to the
As a result of increasing the total number of robots,
target x-coordinate xg . Since the variation of Pattern
the oscillation could not be reduced by the method of
(e) is small against Patterns (a)–(c), it is considered
Figure 9. Trajectories (Enlarge): (a) conventional; (b) previous 1; (c) previous 2; (d) proposed 1; (e) proposed 1 + proposed 2.
that the imbalance between the attraction of the for- numbers because the coefficient matrices are symmet-
mation control and the repulsion of obstacle avoid- ric. In Pattern (c), the coefficient matrix is sometimes
ance has been alleviated by the proposed method. Also, not symmetric because of the network disconnections.
comparing Pattern (e) with Pattern (d), we see that In Figure 11, we compare the maximum eigenvalues
the time-varying repulsive potential function elimi- for each pattern. This figure shows that the maximum
nates the stagnation in front of obstacles during the eigenvalues of Patterns (a)–(c) are oscillatory and the
formation movement. robots divergence is large. In Patterns (d) and (e),
the maximum eigenvalues are not oscillatory and the
robots divergence is small. Pattern (c) has the highest
5.2. Stability analysis robots divergence among all patterns, and Pattern (e)
has the lowest. The reason of the worst result due to Pat-
To analyse the proposed method, we approximate tern (c) is that we increased the total number of robots
the entire robots behaviour to the form p[k + 1] = compared to the problem setting in the previous study
A[k]p[k], and perform eigenvalue analysis from the [9]. In the previous study [9], the oscillatory behaviour
coefficient matrix A[k]. Specifically, for each step, we is not completely suppressed, and it is believed that
obtain a linear approximation system and its eigen- the increase in the total number of robots caused the
values. When the maximum eigenvalue exceeds 1, it oscillation to propagate. The proposed method in this
means that the robots leave the desired position and study can be applied even when the total number of
diverge due to obstacle avoidance. For Patterns (a), robots is increased, and this method is considered to be
(b), (d), and (e), the eigenvalues are obtained as real
SICE JOURNAL OF CONTROL, MEASUREMENT, AND SYSTEM INTEGRATION 33
Figure 10. x-coordinate of robots (i = 1, 2, . . . , 100): (a) conventional; (b) previous 1; (c) previous 2; (d) proposed 1; (e) proposed 1
+ proposed 2.
Figure 11. Comparison of the maximum value of the eigenvalues: (a) overall and (b) enlarge.
Table 2. Comparison of stability for each gain. Takuma Wakasa received the B.E. and
ko ρ
M. E degrees in mechanical engineer-
ing from the University of Electro-
1000 1200 1400 1600 5 15 Communications, Tokyo, Japan, in 2019
(a) 1.28 1.30 1.32 1.64 1.68 4.22 and 2021. His research interests include
(b) 122 60.6 30.1 131 44.3 177 the control theory of multi-agent sys-
(c) 20.9 48.7 80.2 52.9 48.5 319 tems.
(d) 1.61 1.61 1.61 1.61 1.69 1.63
(e) 1.09 1.10 1.11 1.12 1.63 1.26
Kenji Sawada received his Ph.D. degree
in engineering in 2009 from Osaka
University. He is an Associate Pro-
fessor in Info-Powered Energy Sys-
to the distance between robots. The repulsion poten-
tem Research Center, The University of
tial function is a method to change the shape accord- Electro-Communications, Japan. He is
ing to the angle between the direction of motion and also an advisor of Control System Secu-
the obstacle. From the results of the numerical exper- rity Center since 2016. He received Out-
iments, the oscillatory behaviour of the swarm robots standing Paper Awards from FA Foundation (2015 and 2019),
during obstacle avoidance is reduced compared to the Fluid Power Technology Promotion foundation (2018), and
JSME (2018). His research interests include the control the-
conventional method, and the desired formation can ory of cyber-physical systems and control system security. He
be preserved after passing through the obstacle. Future is a member of IEICE, ISCIE, IEEJ, JSME, IEEE.
work is a generalization of the parameter selection and
consideration of the limited sensor range.
References
[1] Tang Q, Yu F, Xu Z, et al. Swarm robots search for
multiple targets. IEEE Access. 2020;8:92814–92826.
Disclosure statement [2] Khan MSA, Ahmed T, Uddin MF. Multi-robot search
No potential conflict of interest was reported by the algorithm using timed random switching of exploration
approaches. 2020 IEEE Region 10 Symposium (TEN-
author(s).
SYMP); 2020. p. 868–871.
[3] Macktoobian M, Gillet D, Kneib J-P. Astrobotics: swarm
robotics for astrophysical studies. IEEE Robot Automat
Funding Mag. 2021;28(3):92–101.
This paper is funded by JSPS KAKENHI Grant Number [4] Izumi S, Shiomoto Y, Xin X. Mass game simulator: an
JP20K20314, JP19K04444, JP19H02163, and JP17H06293. entertainment application of multiagent control. IEEE
Access. 2020;9:4129–4140.
[5] Ping S, Kejie L, Xiaobing H, et al. Formation and
obstacle-avoidance control for mobile swarm robots
Notes on contributors based on artificial potential field. IEEE International
Shotaro Shibahara received the B.E. Conference on Robotics and Biomimetics (ROBIO);
degree in mechanical engineering from 2009. p. 2273–2277.
the University of Electro- Communica- [6] Yang S, Li T, Shi Q, et al. Artificial potential-based for-
tions, Tokyo, Japan, in 2021. He is cur- mation control with collision and obstacle avoidance for
rently pursuing the M.E. degree at the second-order multi-agent systems. International Con-
University of Electro-Communications. ference on Information, Cybernetics, and Computa-
His research interests include the control tional Social Systems (ICCSS); 2020. p. 58–63.
theory of multi-agent systems. [7] Lanejin X, Xu W. The null-space-based behavioral con-
trol for a swarm of robots tracking a target region
SICE JOURNAL OF CONTROL, MEASUREMENT, AND SYSTEM INTEGRATION 35
in obstacle environments. Chinese Control Conference [12] Olfati-Saber R, Alex Fax J, Murray RM. Consensus and
(CCC); 2019. p. 5574–5578. cooperation in networked multi-Agent systems. Proc
[8] Huang D, Yuan Q, Li X. Decentralized flocking of multi- IEEE. 2007;95(1):215–233.
agent system based on MPC with obstacle/collision [13] Nedić A, Ozdaglar A. Distributed subgradient meth-
avoidance. Chinese Control Conference (CCC); 2019. ods for multi-Agent optimization. IEEE Trans Automat
p. 5587–5592. Contr. 2009;59(1):48–61.
[9] Shibahara S, Wakasa T, Sawada K. Network weight func- [14] Nedić A, Ozdaglar A. Cooperative distributed multi-
tion for obstacle avoidance of swarm robots in col- agent optimization. Convex Optimization in Signal Pro-
umn formation. SICE Annual Conference 2021; 2021. cessing and Communications. 2009. p. 340–386.
p. 259–262. [15] Azuma S-i., Nagahara M, Ishii H, et al. Control of multi-
[10] Henry J, Shuma HPH, Komura T. Interactive forma- Agent systems. Japan: Corona Publishing Co. Ltd.; 2015.
tion control in complex environments. IEEE Trans Vis [16] Dang AD, Horm J. Formation control of autonomous
Comput Graph. 2014;20(2):211–222. robots following desired formation during tracking a
[11] Kim DH, Lee H-W, Shin S, et al. Local path planning moving target. International Conference on Cybernet-
based on new repulsive potential function with angle ics (CYBCONF); 2015. p. 160–165.
distributions. International Conference on Information
Technology and Applications (ICITA’05); 2005.