University of Pennsylvania

ScholarlyCommons
Department of Mechanical Engineering & Applied
Departmental Papers (MEAM)
Mechanics

12-1-2001

Modeling and Control of Formations of

Nonholonomic Mobile Robots
Jaydev P. Desai
Drexel University

James P. Ostrowski
University of Pennsylvania

Vijay Kumar
University of Pennsylvania, [email protected]

Follow this and additional works at: http://repository.upenn.edu/meam_papers

Part of the Mechanical Engineering Commons

Recommended Citation
Desai, Jaydev P.; Ostrowski, James P.; and Kumar, Vijay, "Modeling and Control of Formations of Nonholonomic Mobile Robots"
(2001). Departmental Papers (MEAM). 243.
http://repository.upenn.edu/meam_papers/243

Suggested Citation:
Desai, Jaydev P., James P. Ostrowski and Vijay Kumar. (2001). Modeling and Control of Formations of Nonholonomic Mobile Robots. IEEE
Transactions on Robotics and Automation, Vol. 17(6), p. 905-908 .

©2001 IEEE. Personal use of this material is permitted. However, permission to reprint/republish this material for advertising or promotional purposes
or for creating new collective works for resale or redistribution to servers or lists, or to reuse any copyrighted component of this work in other works
must be obtained from the IEEE.
Modeling and Control of Formations of Nonholonomic Mobile Robots
Abstract
This paper addresses the control of a team of nonholonomic mobile robots navigating in a terrain with
obstacles while maintaining a desired formation and changing formations when required, using graph
theory.We model the team as a triple, (g, r, H), consisting of a group element that describes the gross position
of the lead robot, a set of shape variables g that describe the relative positions of robots, and a control graph H
that describes the behaviors of the robots in the formation. Our framework enables the representation and
enumeration of possible control graphs and the coordination of transitions between any two formations.

Keywords
Formation control of mobile robots, graph theory, nonlinear control

Disciplines
Engineering | Mechanical Engineering

Comments
Suggested Citation:
Desai, Jaydev P., James P. Ostrowski and Vijay Kumar. (2001). Modeling and Control of Formations of
Nonholonomic Mobile Robots. IEEE Transactions on Robotics and Automation, Vol. 17(6), p. 905-908 .

©2001 IEEE. Personal use of this material is permitted. However, permission to reprint/republish this
material for advertising or promotional purposes or for creating new collective works for resale or
redistribution to servers or lists, or to reuse any copyrighted component of this work in other works must be
obtained from the IEEE.

This journal article is available at ScholarlyCommons: http://repository.upenn.edu/meam_papers/243

IEEE TRANSACTIONS ON ROBOTICS AND AUTOMATION, VOL. 17, NO. 6, DECEMBER 2001 905

Modeling and Control of Formations of Nonholonomic

Mobile Robots
Jaydev P. Desai, James P. Ostrowski, and Vijay Kumar

Abstract—This paper addresses the control of a team of nonholonomic

mobile robots navigating in a terrain with obstacles while maintaining a
desired formation and changing formations when required, using graph
theory. We model the team as a triple, ( ), consisting of a group ele-
ment that describes the gross position of the lead robot, a set of shape vari-
ables that describe the relative positions of robots, and a control graph Fig. 1. Formation of robots changing shapes.
that describes the behaviors of the robots in the formation. Our framework
enables the representation and enumeration of possible control graphs and
the coordination of transitions between any two formations.
Index Terms—Formation control of mobile robots, graph theory, non-
linear control.

I. INTRODUCTION
In this paper, we discuss the fundamental issues underlying the con-
trol and coordination of multiple autonomous robots. We formulate the
problem of modeling a formation of nonholonomic mobile robots and
develop a framework for transitioning from one formation to another.
We focus on tasks in which the robots are required to follow a trajec-
tory while maintaining a desired formation and avoiding obstacles. In Fig. 2. Notation for l 0 and l 0 l control.
a situation such as the one shown in Fig. 1, for example, it may be nec-
essary to change the formation in order to negotiate the obstacle, and
the number of robots and obstacles and quickly becomes intractable. In
then reform the original formation.
contrast, we address the problem of coordinating a number of robots,
While there are many approaches to solving such a problem [1], we
each with their own sensors and feedback controllers. One possible ap-
are interested in a method that scales with the number of robots and
proach to designing independent controllers is to use simple control
obstacles in the environment. For this reason, we pursue in the cur-
laws based on the potential field theory [6]. In Arkin’s behavior-based
rent work easily computable, decentralized feedback laws that can be
control paradigm [7], this approach is used to coordinate formations
used in conjunction with a higher level (but lower complexity) motion
of robots. It is possible to synthesize an impressive array of group be-
planner.
H
We model a team of robots in formation as a triple (g; r; ), where
haviors [7] and coordinate robots for such tasks. However, the interac-
g 2 SE (N ) represents the gross position and orientation of the lead
tion between the controllers and planners for independently controlled
robots is complex and the performance analysis of such systems is very
robot in N dimensions (N equals two or three), r is a set of shape vari-
difficult.
ables that describe the relative positions of the robots in the team, and
H is a control graph which describes the control strategy (or behavior) Another related concept for a formation of multiple mobile robots is
the concept of string stability, which has been addressed by several re-
used by each robot and the dependence of its trajectory on that of one
searchers in the context of automated highway systems (AHS) [8], [9].
or more of its neighbors. When viewed in this framework, the problem
We realize the importance of dynamic analysis, such as is done in un-
of locomotion can be broken down into three subproblems: 1) trajec-
tory planning (g ); 2) robot control; and 3) formation control (r; H).
derstanding string stability. However, our focus is on kinematic issues
for nonholonomic mobile robots. Though string stability is important,
Most previous work in motion planning has focused on obtaining the
the primary goal of this paper is to develop a new framework for mod-
path and, in some cases, designing feedback architecture and model-in-
eling a formation of mobile robots using graph theory and relating the
dependent coordination strategy. For example, see [2]–[4]. When the
changes in formation to changes in the graph structure.
actuator inputs for each robot are concurrently planned [5], the compu-
tational complexity of the planning task increases exponentially with
II. CONTROL LAWS FOR SHAPE VARIABLES
We first develop a set of decentralized control laws that allows each
Manuscript received March 17, 2001. This paper was recommended for pub-
lication by Associate Editor S. Agrawal and Editor A. De Luca upon evalua-
robot to maintain a desired position within a formation and to enable
tion of the reviewers’ comments. This work was supported by the U.S. Army changes in the shape of a team. We develop two types of feedback con-
Research Office under Grant DAAH04-96-1-0007, the Defense Advanced Re- trollers for maintaining a formation of a team of mobile robots. These
search Projects Agency under Grant N00014-92-J-1647, and the National Sci- control laws are useful for maintaining either: 1) the desired separation
ence Foundation under Grants BES92-16691, BES92-21796, IRI-9711834, and
and relative angle between the leader and the follower robot or 2) the
CMS91-57156.
J. P. Desai is with the Program for Robotics, Intelligent Sensing, and desired separation of the follower robot from its two leaders. These two
Mechatronics (PRISM) Laboratory, Mechanical Engineering and Mechanics types of controllers are shown in Fig. 2.
Department, Drexel University, Philadelphia, PA 19104 USA (e-mail: In the case of l 0 control, the state of the follower robot can be
[email protected]). written in coordinates relative to the lead robot as (l12 ; 12 ; 2 )T .
Similarly, in the case of l 0 l control, the state of the follower robot
J. P. Ostrowski and V. Kumar are with the GRASP Laboratory, University of
Pennsylvania, Philadelphia, PA 19104 USA (e-mail: [email protected];
[email protected]). can be written in terms of the two leader robots as: (l13 ; l23 ; 3 )T . In
Publisher Item Identifier S 1042-296X(01)10911-0. the l 0 control mode for two mobile robots, the aim is to maintain

1042–296X/01$10.00 © 2001 IEEE

906 IEEE TRANSACTIONS ON ROBOTICS AND AUTOMATION, VOL. 17, NO. 6, DECEMBER 2001

d d
a desired length l12 and a desired relative angle 12 between the two
robots. The kinematic equations for the follower robot having l 0
control is given by

l_12 = v2 cos 1 0 v1 cos 12 + d!2 sin 1

_ 12 = 1 (v1 sin 12 0 v2 sin 1 + d!2 cos 1 0 l12 !1)
l12
_2 = !2

where 1 = 1 + 12 0 2 and vi ; !i (i = 1; 2) are the linear and

angular velocities at the center of the axle of each robot. Fig. 3. Examples of formations.
In l 0 l control, this requires regulating the desired lengths, l13
d
and
l23 , of the third robot from its two leaders. The kinematic equations for
d
and Nji 2 f1; 1 1 1 ; ng denotes the index of the robot in the formation
the follower robot is expressed as with l 0 control, i.e., RN . The 2 appears in each i because R2
always has l 0 control.
l_13 = v3 cos 1 0 v1 cos 13 + d!3 sin 1 The set fg uniquely determines a corresponding set, f g, defined
l_23 = v3 cos 2 0 v2 cos 23 + d!3 sin 2 as the collections of the sets with the remaining robot indices. We
_3 = !3 denote these by  1 ;  2 ; . . . ;  where
 i = fL1i ; L2i ; . . . ; Lni g
for the third robot, where i = i + i3 0 3 (i = 1; 2). Using
input/output linearization, we can guarantee exponential convergence and Lji denotes the index of the robot in the formation with l 0 l control,
for the controlled variables and derive the control law for the inputs ! i.e., RL .
and v for the follower robot in l 0 and l 0 l control [10]. Theorem 1: Given n robots in the formation with n  1 robots
having l 0 control and the remaining nl = (n 0 n 0 1) robots
having l 0 l control, with nl > 0, there are exactly 0(n ) control
III. ENUMERATION AND TRANSITIONS IN FORMATIONS graphs where 0(n ) is given by
In our formulation, a team of n robots has one designated lead robot n n
labeled R1 that directly or indirectly controls all other (follower) robots 0(n ) = 1 (n 0 1)! (Lij 0 2): (1)
in the formation. Within the formation, the follower robots depend on
2 i=1 j =1
other robots for their motion. Thus there are many leaders that “lead” Proof: a) Given that R1 is the lead robot in any formation, and R2
other follower robots, but there is a unique lead robot, R1 . For example, by constraint B always has l 0 control, the sets fg and f g give
in Fig. 3, R1 is the lead robot and R2 , R3 , R4 , and R5 are all leaders all possible combinations of formations with n robots having l 0
each with one or more followers. and nl robots having l 0 l controllers. Let us consider one such possible
In our work, we define control graphs to be labeled digraphs with combination, i for robots with l 0 control and  i the corresponding
each vertex having a uniquely assigned integer number and subject to set for robots with l 0 l control. Based on constraint B, each number Lji
the following three constraints. L 01
has 2 possible choices for the index of the leader robots. Thus, the
total number of l 0 l formations possible for these nl robots is given by
Constraint A: Every vertex (robot) except one has at least one in-
coming edge. There is one vertex (robot) with no incoming edges, and
at least one outgoing edge, which is labeled R1 . This is the lead robot n
Lji 0 1 = (Lji 0 1)(Lji 0 2)
n

i = :
in the formation.
j =1
2 j =1
2
Constraint B: Every directed edge in the digraph goes from a lower
vertex label to a higher vertex label. Similarly for each index Nji , in the set i denoting l 0 control for the
Constraint C: The number of incoming edges for any vertex Ri robot, there are (Nji 0 1) choices of a leader. Thus, the total number of
(i > 1) is less than or equal to an integer p  dim (SE(N)), that l 0 formations possible for these n robots is
describes the number of output variables for a given robot. p = 2 for n
Hilare type robots and we will work with this p for the rest of this paper.
A convenient method for representing control graphs is through an
9i = (Nji 0 1):
n 2 n adjacency matrix [11]. It can be shown that if there are n vertices j =1

in a control graph, there are exactly M(n) = n!(n 0 1)!=2n01 distinct Finally, combining the above two results, we obtain the total number
control graphs (based on the constraints stated above) [11], [12]. 2i of formations possible for a given set i and the corresponding set
It is possible to classify control graphs based on the number of l 0 i
and l 0 l controllers. We first introduce some notation. Since we will
need to count the number of possible graphs, we will need to investi- 2i = 9i 1
i :
gate permutations of the robot’s indices. Given n robots with l 0
Since this computation is true for all the sets 1 ; 2 ; . . . ;  2 fg
controllers, let fg represent the set of all possible orderings of n of
(and the corresponding sets  1 ;  2 ; . . . ;  2 f g), the total number
the integers (robot indices) f1; . . . ; ng. Thus, there are possible sets,
of control graphs for n robots with n robots having l 0 control and
1 ; 2 ; . . . ;  , where i and are given by all other robots having l 0 l control is given by

i = fN1i ; N2i ; . . . ; Nni g = f2; N2i ; . . . ; Nni g n n

0(n ) = 9i 1
i = 1 (n 0 1)! (Lij 0 2):
2
= n02 = (n 0 2)! i=1 l i=1 j =1
n 01 (n 0 1)!(n 0 n 0 1)!
IEEE TRANSACTIONS ON ROBOTICS AND AUTOMATION, VOL. 17, NO. 6, DECEMBER 2001 907

Fig. 4. Enumeration of transitions in control laws for robot R .

Given the total number of robots in the formation and n , we can

easily compute all possible sets i and the corresponding set  i . For
n = 3, there are only two control graph with two l 0 controllers and
one control graph for one l 0 (i.e., R2 ) and R3 having l 0 l control.
So there are totally three control graphs.
WhileTheorem 1 allowsthe classification ofcontrolgraphs,itdoes not
give us any information about the equivalence classes of control graphs.
On the other hand, since control graphs are derived from digraphs, it is
productive to enumerate the possible digraphs for an n robot formation
and establish equivalence classes of digraphs. An upper bound on the Fig. 5. Initial and final control graph for Example 1.
digraphs for n robots is given by the polynomial [11], [12]
(2n03)
Qn (x) = ak xk : (2)
k=(n01)
Based on (2), there are (n 0 1) equivalence classes of digraphs, where
the k th equivalence class has ak members, with ak given by Polya’s
theorem [13]. A more detailed description of the mathematical tools
for enumeration of control graphs can be found in [11] and [12].
In summary, the procedure for enumerating control graphs involves
the following two steps.
Fig. 6. Formation change for three robots in the presence of sensory noise.
1) Use Polya’s Theorem to enumerate all digraphs of order n, i.e.,
dn (x) = 1 + a1 x + a2 x2 + 1 1 1 + an(n01) xn(n01) : Fig. 4 is an exhaustive list of all possible transitions in the control graph
Equation (2) provides a tighter bound on allowable digraphs, i.e., for the j th robot, Rj (j 6= 1).
Qn (x):
Qn (x) = a(n01) x(n01) + an xn + 1 1 1 + a(2n03) x(2n03) :
IV. SIMULATION RESULTS
In this section, we will demonstrate the use of the proposed control
2) Choose any control graph G, enumerated by Qn (x), and label laws in the presence of obstacles and show how we can transition from
the vertices arbitrarily. Based on this labeling, construct its ad- one formation to another based on the transitions enumerated in Fig. 4.
jacency matrix. By trial and error, obtain the control graphs for Example 1: In this example, we examine the effect of making
which the adjacency matrix is upper triangular. changes in the control graph. We take the example of three robots
As the exponent of x in the expression for Qn (x) varies from (n 0 1) moving originally in a triangular formation and transitioning to a
to (2n 0 3), there is a spectrum of control graphs ranging from pure straight line formation (see Fig. 5). Fig. 6 shows the path of the robots,
l 0 control to pure l 0 l control with a medley of combined l 0 and R2 and R3 , in the presence of sensory noise, where the noise is
l 0 l control falling in between the two exponents (except for n = 3 modeled as a uniformly distributed random noise. It is seen that, even
which only has pure l 0 and pure l 0 l control). in the presence of noise, the formation converges to the desired shape
The transition from one control graph to another is modeled by a while maintaining the constraints.
transition matrix. We define the transition matrix, T , as the difference Example 2: In this example, the task is to transition from a trian-
between the final (F) and the initial (I) adjacency matrices. The appear- gular formation to a rectangular formation. Fig. 7 illustrates the initial
ance of (01) in the transition matrix denotes that the edge connecting and final control graphs. The change in controller for each robot is il-
the vertices representing the robots in the formation needs to be broken lustrated through the various transition laws in Fig. 4. Fig. 8 illustrates
to achieve the transition. Similarly, the appearance of a (+1) denotes the path of the various robots during the transition from the triangular
the addition of an edge (i.e., communication setup between the robots). formation to a rectangular formation while avoiding the obstacles.
908 IEEE TRANSACTIONS ON ROBOTICS AND AUTOMATION, VOL. 17, NO. 6, DECEMBER 2001

[6] O. Khatib, “Real-time obstacle avoidance for manipulators and mobile

robots,” Int. J. Robot. Res., vol. 5, no. 1, pp. 90–99, 1986.
[7] T. Balch and R. C. Arkin, “Behavior-based formation control for multi-
robot teams,” IEEE Trans. Robot. Automat., vol. 14, pp. 926–939, Dec.
1999.
[8] D. Swaroop and J. K. Hedrick, “String stability of interconnected sys-
tems,” IEEE Trans. Automat. Contr., vol. 41, pp. 349–357, Mar. 1996.
[9] D. Yanakiev and I. Kanellakopoulos, “A simplified framework for string
stability in AHS,” in Preprints of the 13th IFAC World Congress, vol. Q,
1996, pp. 177–182.
[10] J. P. Desai, J. P. Ostrowski, and V. Kumar, “Controlling formations of
multiple mobile robots,” in Proc. 1998 IEEE Int. Conf. Robotics and
Automation, vol. 4, Leuven, Belgium, May 1998, pp. 2864–2869.
[11] J. P. Desai, “Motion planning and control of cooperative robotic
systems,” Ph.D. dissertation, Univ. of Pennsylvania, Philadelphia, Oct.
Fig. 7. Initial and final control graph for Example 2. 1998.
[12] J. P. Desai, V. Kumar, and J. P. Ostrowski, “Control of changes in
formation for a team of mobile robots,” in Proc. 1999 IEEE Int. Conf.
on Robotics and Automation, vol. 2, Detroit, MA, May 1999, pp.
1556–1561.
[13] N. Deo, Graph Theory With Applications to Engineering and Computer
Science. Englewood Cliffs, NJ: Prentice-Hall, 1974.

Fig. 8. Formation change for six robots in the presence of sensory noise.
Multisensor Fusion for Simultaneous Localization and
Map Building
V. CONCLUSION
In this paper, we have studied strategies for controlling formations of J. A. Castellanos, J. Neira, and J. D. Tardós
mobile robots using methods from nonlinear control theory and graph
theory. We have focused on decomposing the problem of controlling a
Abstract—This paper describes how multisensor fusion increases both
formation of nonholonomic mobile robots into: 1) controlling a single reliability and precision of the environmental observations used for the si-
lead robot and 2) controlling other follower robots in the team. We used multaneous localization and map-building problem for mobile robots. Mul-
the terms l 0 and l 0 l control to reflect whether the control laws tisensor fusion is performed at the level of landmarks, which represent sets
are based on tracking the position and orientation of the robot relative of related and possibly correlated sensor observations. The work empha-
sizes the idea of partial redundancy due to the different nature of the infor-
to a leader, or the position relative to two leaders, respectively. We also
mation provided by different sensors. Experimentation with a mobile robot
defined the concept of a transition matrix, which governs the addition equipped with a multisensor system composed of a 2-D laser rangefinder
and deletion of edges in the control graph and hence the change in the and a charge coupled device camera is reported.
communication protocol. Based on this, we presented an exhaustive Index Terms—Correlation, landmark, mobile robot, multisensor fusion,
list of all possible transitions that can occur within the robots in the simultaneous localization and map building.
formation and the corresponding transition matrix column.
There are several important issues that need to be addressed in future
research in this area, including: 1) how to choose a control graph and I. INTRODUCTION
the desired shape based on the constraints in the environment; 2) how Reliable and accurate sensing of the environment of a mobile robot
to plan changes in (g; r; H) depending on sensor constraints; 3) how is an important task both in localizing the robot and in building a de-
to allow formations to be split into sub-formations, leading to multiple tailed map of such an environment. One of the fundamental ideas to
lead robots; and 4d) though the transition matrix gives us the informa- achieve this reliability is the use of redundancy, that is, to combine en-
tion needed to change formations, it is not clear if there is an optimal way vironmental information obtained by several sensors [1]–[3]. Dealing
for carrying out these changes, rather than the sequential algorithm pre- with redundancy requires both the availability of a systematic descrip-
sented here. Some of these topics are the focus of our present research. tion of uncertain geometric information and a consistent multisensor
fusion mechanism [4].
REFERENCES Different approaches to the simultaneous localization and
[1] J. C. Latombe, Robot Motion Planning. Boston, MA: Kluwer, 1991. map-building (SLAM) problem for mobile robots have been reported
[2] R. W. Beard, J. Lawton, and F. Y. Hadaegh, “A feedback architecture for in the literature after the seminal paper of Smith et al. [5] and the early
formation control,” in Proc. American Control Conf., vol. 6, 2000, pp.
4087–4091.
[3] J.-P. Laumond, P. E. Jacobs, M. Taix, and R. M. Murray, “A motion Manuscript received October 12, 2000; revised June 6, 2001. This paper
planner for nonholonomic mobile robots,” IEEE Trans. Robot. Automat., was recommended for publication by Associate Editor N. Xi and Editor
vol. 10, pp. 577–593, Oct. 1994. S. Hutchinson upon evaluation of the reviewers’ comments. This work
[4] M. Egerstedt and H. Xiaoming, “Formation constrained multi-agent was supported by the Spanish Dirección General de Investigación Projects
control,” in IEEE Int. Conf. Robotics and Automation, vol. 4, 2001, pp. DPI2000-1265 and DPI2000-1272.
3961–3966. The authors are with the Departamento de Informática e Ingeniería de
[5] J. P. Desai and V. Kumar, “Nonholonomic motion planning for multiple Sistemas, Universidad de Zaragoza, E-50015 Zaragoza, Spain (e-mail:
mobile manipulators,” in Proc. 1997 IEEE Int. Conf. on Robotics and [email protected]; [email protected]; [email protected]).
Automation, vol. 4, Albuquerque, NM, Apr. 1997, pp. 3409–3414. Publisher Item Identifier S 1042-296X(01)10904-3.

1042–296X/01$10.00 © 2001 IEEE