Recent Research in Cooperative Control of Multi-Vehicle Systems

Richard M. Murray
Control and Dynamical Systems
California Institute of Technology
Submitted, ASME Journal of Dynamic Systems, Measurement, and Control
Special issue on the Analysis and Control of Multi Agent Dynamic Systems
Submitted, 2007 International Conference on Advances in
Control and Optimization of Dynamical Systems
10 September 2006
Abstract
This paper presents a survey of recent research in cooperative control of multi-vehicle sys-
tems, using a common mathematical framework to allow dierent methods to be described in
a unied way. The survey has three primary parts: an overview of current applications of
cooperative control, an summary of some of the key technical approaches that have been ex-
plored, and a description of some possible future directions for research. Specic technical areas
that are discussed include formation control, cooperative tasking, spatio-temporal planning and
consensus.
1 Introduction
Research on control of multi-vehicle systems performing cooperative tasks dates back to the late
1980s, initially beginning in the eld of mobile robotics (see [37] for a more detailed history).
Aided by the development of inexpensive and reliable wireless communications systems, research
in this area increased substantially in the 1990s. Californias Partners for Advanced Transit and
Highways (PATH) project [2] demonstrated multiple automobiles driving together in platoons
and this was quickly followed by other highway automation projects [18, 10]. In the late 1990s and
early 2000s, cooperative control of multiple aircraft, especially unmanned aerial vehicles (UAVs),
became a highly active research area in the United States [3], spurring further advances. Over the
last decade this research area has blossomed, with many new systems being proposed in application
areas ranging from military battle systems to mobile sensors networks to commercial highway and
air transportation systems.
The purpose of this article is to provide a survey of some of the recent research in cooperative
control of multi-vehicle systems. We focus on research in the last two decades, with some historical
notes on work before this period. To help focus the topics that are surveyed, we focus exclusively
on control of multi-vehicle systems that are working together to complete a shared task. Several
other surveys of the literature in cooperative control are available that complement the current
paper (see, e.g., [37]).
1
It will be helpful in the sequel to have a clear notion of some terms that will dene the object of
the survey, in particular a concise denition of cooperative, which has been used in many dierent
ways by the broad research communities interested in this topic. For the purposes of this survey,
we will consider a vehicle to be a dynamical system whose position is given by its location in three
dimensional space. We will consider a collection of N vehicles that are performing a shared task,
where the task depends on the relationship between the locations of the individual vehicles. The
vehicles are able to communicate with each in carrying out the task, with the individual vehicles
able to communicate with some subset of the other vehicles.
We assume that the dynamics of the ith vehicle can be written as
x
i
= f
i
(x
i
, u
i
) x
i
R
n
, u
i
R
m
y
i
= h
i
(x
i
) y
i
SE(3),
where x
i
is the state of the ith vehicle, u
i
is the input that controls the vehicles state and f
i
is a smooth vector eld representing its dynamics. We assume that the location of the vehicle
is given by the output y
i
SE(3), where SE(3) is the set of rigid body congurations (position
and orientation). More general formulations allowing position and velocity as part of the location
description are possible as well, but will be omitted for simplicity. We let x = (x
1
, . . . , x
N
) represent
the complete state for a collection of N vehicles.
In addition to the location of the vehicle, we will also assume that each vehicle has a discrete
state,
i
, which we dene as the role of the vehicle. The role of the vehicle will be represented
as an element of a discrete set A whose denition will depend on the specic cooperative control
problem under consideration. As indicated by the terminology, we will generally consider the role
variable
i
to represent the portion of the vehicles overall state that encodes its current actions
and its relationship with the overall task being performed. We will assume that the role of a vehicle
can change at any time and we will write a change of role as

= r(x, ),
where

indicates the new value of . We let = (

1
, . . . ,
N
) represent the roles of the collection
of N vehicles and write
i
(t) for the role of vehicle i at time t.
We assume that the vehicles are able to communicate with some set of other vehicles and
we represent the set of possible communication channels by a graph G. The nodes of the graph
represent the individual vehicles and a directed edge between two nodes represents the ability of
a vehicle to receive information from another vehicle. We write N
i
(G) to represent the neighbors
of vehicle i, that is, the set of vehicles that vehicle i is able to obtain information from (either by
explicit communication or by sensing the position of the other vehicle). In general, N
i
can depend
on the locations and roles of the vehicles, in which case we will write N
i
(x, ). The number of
neighbors of the ith vehicle is given by the number of elements of N
i
, written |N
i
|.
Given a collection of vehicles with state x and roles , we will dene a task in terms of a
performance function
J =
_
T
0
L(x, , u) dt + V (x(T), (T)),
where T is the horizon time over which the task should be accomplished, L represents the incre-
mental cost of the task and V represents the terminal cost of the task. As special cases, we can
take T = to represent innite horizon problems or take L = 0 to represent tasks in which we are
2
only interested in the nal state. We may also have constraints on the states or inputs, although
we shall generally consider such constraints to be included in the cost function (eg, via Lagrange
multipliers) for ease of presentation.
A strategy for a given task is an assignment of the inputs u
i
for each vehicle and a selection of
the roles of the vehicles. We will assume that the inputs to the vehicles dynamics are given by
control laws of the form
u
i
= (x, )
where is a smooth function. For the choice of roles, we make use of the notion of a guarded
command language (see [27]): a program is a set of commands of the form
{g
i
j
(x, ) : r
i
j
(x, )}
where g
i
j
is a guard that evaluates to either true or false and r
i
j
is a rule that denes how the role

i
should be updated if the rule evaluates to true. Thus, the role evolves according to the update
law

i
=
_
r
i
j
(x, ) g(x, ) = true
unchanged otherwise.
This update is allowed to happen asynchronously, although in practice it may be assigned by a
central agent in the system, in which case it may evolve in a more regular fashion. We write
i
to represent the overall strategy (control law and guarded commands) for the ith vehicle. =
(
1
, . . . ,
N
) is used to represent the complete strategy for the system.
Using these denitions, we can now provide a more formal description of a cooperative control
problem. We say that a task can be additively decoupled (or just decoupled) if the cost function J
can be written as
J =
N

i=0
__
T
0
L
i
(x
i
,
i
, u
i
) dt + V
i
(x
i
(T),
i
(T))
_
.
If a task is not decoupled, we say that the task is cooperative, by which we mean that the task
performance depends on the joint locations, roles and inputs of the vehicles. (Note that we are
assuming here that all vehicles are trying to solve a common objective and hence not considering
adversarial tasks, for which a more careful notation would be required.)
We say that a strategy is centralized if
i
depends on the location or role of any vehicle that is
not a neighbor of i. A strategy is decentralized if
u
i
(x, ) = u
i
(x
i
,
i
, x
i
,
i
)
{g
i
j
(x, ) : r
i
j
(x, )} = {g
i
j
(x
i
,
i
, x
i
,
i
) : r
i
j
(x
i
,
i
, x
i
,
i
)},
where we use the shorthand x
i
and
i
to represent the location and roles of vehicle is neighbors
(hence x
i
= {x
j
1
, . . . , x
jm
i
where j
k
N

and m
i
= |N
i
|.}. We will mainly be interested in
cooperative tasks that can be solved using a decentralized strategy.
We note that the denitions used here are not the most general possible and we have ignored
some subtleties regarding the formal denition of the solution of a task (i.e., we assume existence
and uniqueness of solutions for a given strategy). These details are important and can be found
in the various papers referenced in this survey. One alternative set of denitions for cooperative
agents can be found in the work of Parker [35], which makes use of the notions of local/global goals
and control.
3
We these denitions in hand, we now proceed to consider some of the primary applications
of cooperative control of multi-vehicle systems, followed by some of the key technical results that
have been proposed in the last decade. We end the paper with a partial listing of some of the open
research directions that are currently under exploration.
2 Applications Overview
In this section we summarize some of the main applications for cooperative control of multi-vehicle
systems. This summary is based on those applications of which the author is most aware (including
the results of a result survey of future directions in control, dynamics and systems [31]), as well
as a survey of the literature (with emphasis on papers that are frequently referenced by others).
Although not comprehensive, the applications cited here demonstrate some of the key features that
must be addressed in solving cooperative control problems.
2.1 Military Systems
Modern military systems are becoming increasingly sophisticated, with a mixture of manned and
unmanned vehicles being used in complex battleeld environments, such as the one depicted in
Figure 1. Traditional solutions involve a centralized resource allocation (assignment of planes to
targets), followed by decentralized execution (each attack vehicle is responsible for a set of targets).
More modern battlespace management systems are considering the use of cooperative operation of
large collections of distributed vehicles, with location computation, global communication connec-
tions and decentralized control actions [31, 46].
Formation ight One of the simplest cooperative control problems is that of formation ight:
a set of aircraft y in a formation, specied by the relative locations of nearby aircraft. This area
has received considerable attention in the literature. Some of the earliest work in this area is that
of Parker [35], who consider the design of control laws that use a combination of local and global
knowledge to maintain a formation.
NASA has experimented with formation ight as a method for reducing drag on a collection of
aircraft [28]. The key idea is to locate the aircraft such that the tip vortices of one aircraft help
reduce the induced drag of the tailing aircraft. This task requires precise alignment of an aircraft
with the aircraft in front of it. To date, demonstrations of this concept in engineering systems have
been restricted to small numbers of aircraft. Similar formations in nature can involve many more
individuals [14].
Cooperative classication and surveillance Chandler et al. [3] dene the cooperative classi-
cation problem as the task of optimally and jointly using multiple vehicles sightings to maximize
the probability of correct target classication. More generally, we can dene the cooperative
surveillance problem as that of using a collection of vehicles to maintain a centralized or decentral-
ized description of the state of a geographic area. This description might include the current state
of features that are spatially xed (such as the number of people in a given location) or information
about entities that are moving in the region of interest (eg, locations of cars and planes in a given
region).
4
Figure 1: Battle space management scenario illustrating distributed command and control between
heterogeneous air and ground assets. Figure courtesy of DARPA.
The cooperative classication problem is one in which the performance function involves the
collection of maximal amounts of relevant information. One typically assumes that the vehicles
can communicate over some range (possibly limited by line of site, especially for ground-based
vehicles) and information shared between the vehicles can be used by the vehicles in determining
their motion.
Cooperative attack and rendezvous The rendezvous problem involves bringing a collection
of vehicles to a common location at a common time. Depending on the application, the rendezvous
time may either be xed ahead of time or determined dynamically, based on when all vehicles
reach the same area. Military applications of rendezvous include minimizing exposure to radar by
allowing aircraft to y individual paths that are locally optimized [3].
Mixed initiative systems A variant of the cooperative control problem is the mixed initiative
cooperative control problem, in which collections of autonomous vehicles and human operators (on
the ground or in vehicles) must collectively perform a task or a mission. This class of problems
adds the complexity of providing situational awareness to the operators and allow varying levels of
control of the autonomous system.
5
Figure 2: Autonomous ocean sampling network: (a) a depiction of the collection of vehicles that
were part of the summer 2003 experiment; (b) an example of using a collection of gliders for
sampling a region of interest.
2.2 Mobile Sensor Networks
A second area of application in cooperative control is networks of sensors that can be position so
as to maximize the amount of information they are able to gather. In this section we provide some
examples of the types of cooperative control applications that are being pursued in this area.
Environmental sampling The Autonomous Ocean Sampling Network (AOSN) [30], pictured
in Figure 2 is an example of an environmental sampling network. The network consists of a
collection of robotic vehicles that are used for adaptive sampling, in which the motion of the
vehicles is based on the observations taken by the vehicles. This approach allows the sensors to
be positioned in the areas in which they can do the most good, as a function of the data already
collected. Because of the distributed nature of the measurements being taken, a cooperative control
strategy is used to control the motion of the vehicles. In tests done in the summer of 2003, over 20
vehicles were controlled over 4 weeks weeks to collect data [30].
Distributed aperture observing A related application for cooperative control of multi-vehicle
systems is distributed aperture (or phased array) imaging, illustrated in Figure 3. The proposed
TechSat 21 project was sponsored by the US Air Force Research Laboratory (AFRL) and was to
have launched a collection of microsatellites that would be used to form a virtual satellite
with a single, large aperture antenna (the project was canceled in 2003). Another example of a
distributed aperture observing system is the terrestrial planet nder (TPF), being proposed by
NASA. TPF uses optical interferometry to image distance stars and to detect slight shifts in the
stars positions that indicated the presence of planets orbiting the stars [20].
2.3 Transportation Systems
Finally, the use of cooperative control in transportation systems has received considerable attention
over the last few decades.
6
Figure 3: Distributed aperture observing systems: (a) the proposed TechSat 21 concept would use
a collection of microsatellites to form the equivalent of a larger aperture imaging system; (b) the
terrestrial planet nder uses formation ying to enable optimal interferometry for detecting planets.
(a) (b)
Figure 4: Transportation systems. (a) A platoon of cars driving down the San Diego freeway as
part of the PATH project [2]. (b) The San Francisco Bay area aircraft arrival and departure routes
(courtesy of Federal Aviation Authority).
Intelligent highways Several groups around the world have begun to explore the use of dis-
tributed control for problems related to intelligent highway and transportation systems. These
problems include increased interaction between individual vehicles to provide safer operations (e.g.,
collision warning and avoidance), as well interaction between vehicles and the roadway infrastruc-
ture. These latter applications are particularly challenging since they begin to link heterogeneous
vehicles through communications systems that will experience varying bandwidths and latency
(time delays) depending on the local environment. Providing safe, reliable, and comfortable op-
eration for such systems is a major challenge that will have application in a variety of consumer,
industrial, and military products and systems.
A representative example of this class of applications is the California Partners for Advanced
Transit and Highways (PATH) project [2]. In 1997 the PATH project developed and demonstrated
a system for allowing cars to be driven automatically down a freeway at close spacing, as shown in
Figure 4a. By decreasing the spacing of cars, the density of trac on a highway can be increase
7
Figure 5: The RoboFlag playing eld [7].
without requiring additional lanes. Additional work within the PATH project has looked at a
variety of other systems for better managing trac ow [2].
Air trac control Air trac control systems are another area where methods for cooperative
control are being explored (see, e.g., [45]). As the density of air trac continues to increase,
congestion at major airports and automated collision warning systems are becoming increasingly
common. Figure 4b illustrates some of the complexity of the current air trac control networks.
Next generation air trac control systems will likely move from a human-controlled, centralized
structure within given region to a more distributed system with free ight technologies allowing
aircraft to travel in direct paths rather than staying in pre-dened air trac control corridors.
Eorts are now being made to improve the current system by developing cockpit sensors such as
augmented GPS navigation systems and data links for aircraft to aircraft communication citeatc.
2.4 Testbeds
A variety of testbeds have been developed to explore cooperative control problems in laboratory
settings. Perhaps the most well known is RoboCup, a multi-vehicle game of robot soccer. RoboCup
was initially conceived as an attempt to foster research in articial intelligence, specically that
of multiple vehicles in a highly dynamic environment [25]. The RoboCup competition is now held
annually and has competitions involving a variety of dierent physical and simulation platforms.
Most of the RoboCup competitions allow the use of centralized computation, although some teams
have made use of decentralized strategies [6].
A related game, dubbed RoboFlag has been developed at Cornell [7] and is loosely based on
Capture the Flag and Paintball. Two teams play the game, the red team and the blue team,
as depicted in Figure 5. The red teams objective is to inltrate blues territory, grab the blue ag,
and bring it back to the red home zone. At the same time, the blue teams objective is to inltrate
reds territory, grab the red ag, and bring it back to the blue home zone. The game is thus a mix
of oense and defense: secure the opponents ag, while at the same time prevent the opponent
8
from securing your ag. Sensing and communications are both limited to provide a more realistic
distributed computing environment. The game is meant to provide an example of multi-vehicle,
semi-autonomous systems operating in dynamic, uncertain, and adversarial environments. Human
operators can also be present in the system and can be used either as high level controllers or as
low level (remote) pilots. A centralized control unit may be used coordinate the vehicles, but it
must respect the communication constraints (bandwidth and latency) of the system.
Several physical testbeds have also been developed, ranging from wheeled robots such as those
sued in RoboCup, to hovercraft that provide some of the dynamics more typical of aircraft [5, 47],
to small-scale aircraft [17, 24] and helicoptersSKS03-cep,SV99-cca. These citations are far from
complete, but give an example of the range of physical testbeds that have been developed.
3 Technology Overview
In this section we provide a brief survey of some of the techniques that have been developed for
designing strategies for cooperative control tasks. We make use of the mathematical notation
dened in the introduction wherever possible. We focus primarily on the problem formulation and
the approached used in its solution, leaving the details of the proofs of stability, convergence and
optimality to the original papers.
3.1 Formation Control
Many of the applications above have as part of their solution the ability to maintain the position of
a set of vehicles relative to each other or relative to a reference. This problem is known as formation
control and has received considerable attention, both as a centralized and as a decentralized problem.
Optimization-based approaches One way to approach the formation control problem is to
formulate it as an optimization problem. If we let L
i
(x
i
, x
i
) represent the individual formation
error between the ith vehicle and its neighbors, then we can establish a cost function
L(x, , u) =

L
i
(x
i
, x
1
) +u
i

2
R
,
where the summation over the individual formation errors gives the cumulative formation error [35]
and the nal term is a penalty on the inputs (other forms could be used).
This problem can be solved in either a centralized manner or a distributed manner. One dis-
tributed approach is the work of Dunbar et al. [11], who considers cooperative control problems
using receding horizon optimal control. For a cost function whose coupling reects the commu-
nication constraints of the vehicles, he generates distributed optimal control problems for each
subsystem and establishes that the distributed receding horizon implementation is asymptotically
stabilizing. The communication requirements between subsystems with coupling in the cost func-
tion are that each subsystem obtain the previous optimal control trajectory of those subsystems at
each receding horizon update. The key requirements for stability are that each distributed optimal
control not deviate too far from the previous optimal control, and that the receding horizon updates
happen suciently fast.
Figure 6 shows a simulation of Dunbars results. The vehicles are ying in ngertip formation,
with vehicles 2 and 3 maintaining position relative to vehicle 1 and vehicle 4 maintaining position
relative to vehicle 2. The control goal is to maintain formation around the black square, which is
9
Figure 6: Four vehicle formation using distributed receding horizon control [11].
ying along a trajectory that is not known to the individual aircraft. The localized optimization
for each vehicle uses a previous optimal path for its neighbors while constraining its own path to
stay near the previous path that it communicated to others.
Potential eld solutions Another approach to solving the formation control problem is to
consider the mechanical nature of the systems and to shape the dynamics of the formation using
potential elds. In this case, the control law for the individual vehicles has the form
u
i
= V (
i
, x
i
, x
1
)
where V is a potential function that depends on the mode of the vehicle,
i
(typically whether it
is a leader or a follower).
A representative body of research in this area is the work of Fiorelli and Leonard, who use the
concept of virtual leaders that guide the motion of the other vehicles [29, 32]. They consider
two types of potential functions: an interaction function V
I
and a potential generated by leaders,
V
h
. Each function generates a repulsive force if vehicles are very close to each other, an attractive
force if the vehicles are within some interaction range but not too close or too far and no force for
vehicles beyond a certain radius. Their resulting control law is of the form
u
i
=
N

j=i
V
I
(y
i
y
j
)

kL
V
h
(y
i
y
k
) + f
v
i ,
where L is the set of leaders, f
v
i is a dissipative force based on the velocity of the i vehicle, and local
coordinates are used for y
i
SE(3). By appropriate choice of f
v
i they are able to show asymptotic
stability of various schooling and ocking behaviors.
10
(a) (b) (c)
Figure 7: Formation stabilization using potential functions [33]. (a) Stabilization of three vehicles
in the plane. (b) Time traces for individual positions of the vehicles. (c) Stabilization of a six
vehicle formation.
Other work on the use of potential elds includes that of Olfati-Saber [33], who uses potential
functions obtained from structural constraints of a desired formation in a way that leads to a
collision-free, distributed, and bounded state feedback law for each vehicle. Figure 7 demonstrates
some of the results of his algorithm for formation control.
String stability One issue that arises in formation control is that of string stability, in which
disturbances grow as they propagate through a system of vehicles [42]. One of the early sources of
research on this problem was in the control of vehicle platoons, in which one wanted to insure that
small disturbances at the beginning of a chain of vehicles did not get amplied as one progressed
down the chain.
For simplicity, we assume that the disturbances enter through the initial states of the vehicles.
String stability is dened in terms of an innite collection of vehicles and our goal is to nd a
control law for each of the vehicles so that given > 0 there exists a > 0 such that
sup
i
x
i
(0) < sup
i
x
i
()

< ,
where the norm is taken with respect to time. In particular, this implies that the motion of each
vehicle is bounded for all time. More general norms can also be used, as described in [42].
Using this denition, one can show that a system is string stable if the H

gain between any two

neighbors is less than one. If this is the case, then disturbances are attenuated as they pass down
the chain of vehicles. Conversely, if the dynamics and control laws for each vehicle are identical and
if the gain of the transfer function is greater than 1 at some frequency, then disturbances at that
frequency can be amplied as they propagate down the chain. These denitions can be generalized
to dierent topologies in which the neighbor sets are more complicated than a single chain.
To help compensate for string instabilities, one can make use of globally transmitted information
that allows the vehicles to pre-compensate for disturbances. In essence, one changes the topology
of the information ow from one in which each vehicle only sees the vehicle in front of it, to one in
which vehicles also have global information about the position of the lead vehicle. Figure 8 shows
the responses of a set of vehicles with dierent topologies and dierent levels of global information.
In this simulation, the lead vehicle responds to a step input at time t = 15. The variable controls
11
Figure 8: String stability results for a ve vehicle formation [23]. Each column represents a dierent
information topology, as shown in the diagram at the top of the column. The rst row of plots
corresponds to the use of purely local information, while the second two rows allow increasing
amounts of global information.
the amount of mixing between the purely local strategy ( = 0) and a purely centralized strategy
( = 1).
It is also possible to dene the performance in ways that are more structured than string
stability, for example asking whether the distances between specied sets of vehicles have certain
levels of disturbance attenuation [22, 21].
Swarms Finally, although not strictly a formation control problem, there has been a great deal
of interest in so-called swarms of vehicles. Roughly speaking, a swarm is a large collection of
vehicles that perform in a collective fashion, such as ying together in a given direction. One early
work in swarm-like behavior was that of Reynolds, who developed a set of rules that he used to
12
generate realistic motion of vehicles for animation purposes [38].
An innovative approach to understanding swarm behavior was taken by Jadbabaie, Lin and
Morse [19], who described how to achieve coordination fo groups of mobile autonomous agents
using nearest neighbor rules. The control law was quite simple, making use of a simple heading
model in which each agent updated its heading according to the rule
u
i
=
1
1 +|N
i
(t)|
_
_

i
(t) +

jN
i
(t)

j
(t)
i
(t)
_
_
where N
i
(t) is the set of vehicle that are within a radius r of vehicle i at time t. The rst term is
the average heading of the neighbors of vehicle i and hence this control essentially tells each vehicle
to steer in the same direction as its neighbors.
Jadbabaie et al. are able to demonstrate that with this control law, all vehicles will converge to
a common heading. They make use of an eventual connectivity assumption in which the vehicles
are connected together across intervals. In other words, while it may never be the case that at
a given instant of time the graph describing the interconnectivity is complete, as long as over a
suitable interval all vehicles are able to share information, the solution will converge to a common
value.
Control laws for swarms often involve using attractive and repulsive functions between nearby
vehicles. In addition to the work of Leonard et al. already described above, another representative
work in this regard is that of Olfati-Saber [40], who makes of a control input consisting of three
terms
u
i
= f
i
g
+ f
i
d
+ f
i

.
The rst term f
i
g
= V (y
i
, y
1
) is a gradient-based term where V is a potential function. The
second term f
i
d
is a damping term based on the relative velocities of neighboring vehicles and has
the form (q)(v
i
v
j
). The nal term f
i

is a navigational feedback term that takes into account

a group objective, such as moving to a given rendezvous point. Figure 9 shows a sample maneuver
in which 150 agents squeeze through an opening without collision.
3.2 Cooperative Tasking
A major element of cooperative control is deciding on the tasks that dierent vehicles will perform
to satisfy the team objective. This essentially amounts to choosing the role of the vehicles,
i
.
MILP formulations Several groups have formulated this problem as a mixed integer linear
program (MILP), in which the integer variables correspond to the role
i
.
The work of Richards et al. [39] considers the problem of designing trajectories for a group
of vehicles that collectively visit a set of waypoints within a given set of time constraints. They
minimize a cost function of the form
J =

t +
1
N

p=1
_
t
p
+
2
T

t=0
_
|u
1
(t)| +|u
2
(t)|
_
_
where t
p
is the time at which the pth vehicle completes its task and

t is the time at which the last
vehicle completes its task. This cost function thus trades o the input forces on the vehicles with
the time that the overall task is completed as well as the tasks of the individual vehicles.
13
Figure 9: A squeezing maneuver using ocking algorithms of Olfati-Saber [40].
In the MILP formulation used by Richards et al. [39], the individual assignments of waypoints
to vehicles is handled by using decision variables to constrain the problem such that each waypoint
is visited exactly once by a vehicle. This constraint can be written in the form

t=0
T
N

p=1
K
pi
b
ipt
= 1 for all waypoints i
14
Figure 10: Resource allocation using mixed integer linear programming (MILP) [39].
where K
pi
is the suitability of vehicle p to visit waypoint i and b
ipt
is 1 if vehicle p visits waypoint
i and time t and zero otherwise.
Figure 10 shows an example of the allocation problem applied to set set of 6 vehicles. The sce-
nario includes 12 waypoints that must all be visited, along with a region of no y zones (obstacles).
An approximate method described in [39] is used to solve the problem in 27 seconds on a standard
PC.
A similar approach has been developed independently by Earl and DAndrea [12], in which
the MILP formulation is used to solve a subproblem of the RoboFlag example in Section 2.4.
Specically, they solve the problem of guarding a defense zone from attackers that are trying to
enter it. They formulate the problem in discrete time to be consistent with the MILP framework;
for simplicity we will use a single time discretization here and re-use t as the discrete time.
The objective function is given by
J =
T

t=0
(t) +
T

t=0
|u(t)|
where (t) is a binary variable that takes on the value 1 if and only if one of the attackers is in the
defense zone at time t. This function must be minimized while also constraining the position of the
defending robots so that they avoid collisions with each other and stay outside of the defense zone.
In addition to the dynamics of the vehicles, a complete description of the problem also requires
that we dene the dynamics of the attacking robots. We do this using a describe variable
i
for
15
Figure 11: The ALLIANCE architecture [36].
each attacker that describes whether an attacker is active or inactive. An attacker is active initially
and becomes inactive if it either enters the defense region or is intercepted by a defending robot
(modeled by a defending robot getting within a certain distance of the attacking robot). We assume
that if an attacking robot is active, it moves toward the defense zone in a straight line.
Note that in both of these formulations, the assignment is handled implicitly: the problem does
not explicitly assign a given defender to specic attacker, but rather relies on the optimization to
choose motions of the group of defenders such that no attackers enter the defense region.
Assignment protocols Another approach to the cooperative tasking problem has been to de-
velop protocols that are used to decide on who is assigned to what task. By protocol we mean a
set of rules that are used to determine the individual roles (assignments) of each vehicle. One seeks
to prove that this protocol results in all tasks being assigned to a vehicle, even in the presence of
changing environmental conditions or failures.
One of the early approaches to distributed task allocation was the ALLIANCE software archi-
tecture developed by Parker [36]. Their approach made us of behavior sets that were activated
under certain conditions. Each behavior could itself inhibit other behaviors, so that it was possible
for a single behavior set to control the motion of the robot. Figure 11 illustrates this architecture.
The activation of a behavior set is controlled through motivational behaviors. Each motivation
behavior responds to some set of inputs, including external sensors, inter-robot communications,
inhibitory feedback from other behaviors, and internal motivations. The two internal motivations,
robot impatience and robot acquiescence, allow the robot to progress when other robots fail to
complete a task or when the robot itself fails to accomplish a task. These motivational behaviors
can be viewed in the context of the guarded command framework discussed in Section 1.
A related approach has been taken by Klavins [26], who constructed a language for describing
and verifying protocols for cooperative control. The computation and control language (CCL) uses
the guarded command formalism to specify sets of actions for a collection of robots. Figure 12 gives
an example of how a distributed area denial task can be solved in CCL. In this example, drawn from
the RoboFlag game, 6 defensive robots are trying to protect a defense zone for an incoming set of
robots, which descend vertically at a xed speed. The defending robots must move underneath the
incoming robots, but are not allowed to run into each other. The defenders are randomly assigned
16
Figure 12: The RoboFlag Drill.
incoming robots and are allowed to talk to their neighbors and switch assignments under a given
protocol. A protocol was developed in [26] that is able to provably solve this problem, including
insuring that no two robots collide and that all defensive robots eventually end up assigned to an
incoming robot with now crossing of assignments. Extensions to this approach for observability
and controllability have also been developed [9, 8].
3.3 Spatio-Temporal Planning
A broad collection of technological developments can be described under the heading of spatio-
temporal planning, in which the paths of the robots and their locations with respect to time are
to be specied and controlled. In this section we consider two typical spatio-temporal planning
problems: rendezvous and coverage.
Rendezvous The rendezvous problem is a specic cooperative task in which one wants to have
a number of individual vehicles meet at a common point at a common time. The key element in
the rendezvous problem is that all agents should arrive at the same time, so that if one vehicle is
delayed the other vehicles should adjust their trajectories so that they do not arrive early.
Bhattacharya et al. [1, 43] formulated the rendezvous problem by dening a rendezvous region
R around the rendezvous point (take as the origin) and letting be the ratio of the maximum
and minimum distances of the vehicles at the time that one of them enters the rendezvous point.
Letting be the radius of the rendezvous region and t
a
the time in which the time at which the
rst vehicle enters the region, they dene as
=
max(x
i
(t
a
))

.
The goal can then be dened as nding control laws such that from all initial conditions,

des
1.
The case of perfect rendezvous corresponds to = 1, in which case all vehicles must reach the
rendezvous region at precisely the same time.
17
(a) (b)
Figure 13: (a) Denition of the rendezvous problem for two scalar agents. (b) Solution via con-
struction of invariant cones.
This problem can be solved using a Lyapunov-based approach that uses feedback to create an
invariant cones in the phase space [43, 1], as illustrated in Figure 13. The problem denition is
shown in Figure 13a, which shows the phase space for two scalar vehicles. To achieve rendezvous,
these vehicles must reach x = 0 at approximately the same time, without either of the individual
vehicles coming near x = 0 before that time. This creates a set of forbidden regions in the phase
space. By proper choice of control law, it is possible to render certain cones as invariant, as shown
in Figure 13b. The resulting trajectories satisfy the rendezvous problem. The feedback in this case
is centralized, requiring each vehicle to communicate its position to nearby vehicles.
Coverage The coverage control problem refers to the use of a collection of vehicles to provide
sensor coverage for a given geographic area. It is thus one approach to the cooperative surveillance
problem. Given a set of N vehicles, we wish to allocate each vehicle to a region in which it is
responsible for providing senor information. The centralized version of this problem is referred to
as the locational optimization problem and there is a large literature describing dierent approaches
(see [4] for a survey). We focus here on the decentralized solution proposed by Cortes et al. [4].
The approach taken by Cortes et al. is to partition a region Q into a set of polytopes W =
{W
1
, . . . , W
N
} that cover Q. Each polytope is assigned to a specic vehicle to each region and we
let f
i
: R
+
R
+
represent the sensing performance of a vehicle based on its distance from a given
point, with f small representing good performance. We then form the coverage control problem as
choosing the locations of each vehicle such that we minimize
L =
n

i=1
_
W
i
f(q y
i
)(q)dq, (1)
where (q) is a distribution density function that represents the importance of a given area.
It can be shown that if the location of the vehicles are xed, the optimal decomposition of the
18
Figure 14: Coverage control applied to a polygonal region with Gaussian density function around
the point in the upper right [4].
space Q is a Voronoi decomposition where
W
i
= {q Q|q y
i
q y
j
, j = i}.
This decomposition corresponds to each vehicle being responsible for the points that are closest
to it. This decomposition also introduces a natural graph of neighbors, with two vehicles being
neighbors if their Voronoi partitions share an edge.
If we let C
V
i represent the centroids of the Voronoi partition, then it turns out that the control
law
u
i
= k(y
i
C
V
i )
converges asymptotically to a set of critical points for the cost function, and hence provides (locally)
optimal coverage. A key element of this approach is that the only communication required is
with the nearest neighbors of the vehicle (since this is what is needed to determine the Voronoi
decomposition). Figure 14 illustrates the coverage algorithm applied to a region with (q) being a
Gaussian around the point in the upper right portion of the region.
The above formulation assumes that the collection of vehicles that is available is sucient to
cover the entire region of interest. A slightly dierent problem occurs when there is not enough
sensor range to simultaneous view all portions of the environment that are of interest. In this case,
one must selectively cover dierent regions of space and change those regions over time (so that no
region goes unviewed forever). Several groups have considered this problem [15, 34, 44]
3.4 Consensus algorithms
As a nal technology in cooperative control, we briey describe the problem of consensus. The
consensus problem is to have a group of vehicles (or more general agents) reach a common assess-
ment or decision based on distributed information and a communications protocols. Many of the
decentralized problems listed above, especially those involving assignment, can be thought of as
special cases of consensus.
The consensus problem has been formulated as a coordinated control problem by Fax [13] and
Olfati-Saber [41]. A particularly simple solution to the consensus problem is to let the behavor of
each agent be governed by the rst order dierential equation
x
i
=
1
|N
i
|
|N
i
|

j=1
(x
i
x
j
),
19
where x
i
R is the internal state of the agent. For this system, one can show that if the information
ow is bidirectional (if agent i is a neighbor of agent j, then j is a neighbor of i), the states of
the individual vehicles asymptotically converage to the average of the initial state values for any
connected graph G.
If G is not bidirectional (so that there are asymmetries in the information available to each
agent), then the interaction above does not necessarily lead to average consensus. We dene a
graph to be balanced if the in-degree and out-degree of all nodes are equal. In the case of balanced
graphs, one can once again show that any connected graph solves the average consensus problem
using the interaction rules above [41]. Furthermore, even if the connections are changing as a
function of time, it can be shown that the average consensus is still reached.
When the behavior of the individual agents is more complicated, we can still pose the problem
in a similar manner. Supposed that each agents dynamics are governed by
x
i
= Ax
i
+ Bu
i
y
i
= Cx
i
(2)
Fax [13] considers a control law in which each system attempts to stabilize itself relative to its neigh-
bors. This is accomplished by constructing an error for each system that is a weighted combination
of the relative outputs of the neighbors:
e
i
=

jN
i

ij
(y
j
y
i
) (3)
where
ij
is the relative weight. For simplicity, we consider uniform weighting here, so that
ij
=
1/|N
i
| where N
i
is the number of neighbors of node i. The results are easily extended to the more
general case.
Given the error (3), we apply a compensator that attempts to stabilize the overall system. For
simplicity, we assume here that the stabilizer is given by a constant gain
u
i
= Ke
i
, (4)
with K R
mm
representing the compensation (gain) matrix. In practice, one can use a dynamic
compensator to improve performance, but for analysis purposes we can just assume these dynamics
are included in the system dynamics (2).
The interconnectedness of the system, represented by the neighbor sets N

can be studied using

tools from graph theory. In particular, for the the case of uniform weighting of the errors, it turns
out that the combined error vector e R
Nm
can be written as
e = (

L I)x (5)
where represents the Kronecker product and

L is the weighted Laplacian associated with the
(directed) graph that models the neighbors of each node. The weighted Laplacian is a standard
object in graph theory and can be dened as

L = D
1
(D A)
where D is a diagonal matrix whose entries are the out-degree of each node and A is the adjacency
matrix for the graph (see [13] for more detail). Using this framework, Fax showed the following:
20
1
2 3
4
5 6
1.5 1 0.5 0
1
0.8
0.6
0.4
0.2
0
0.2
0.4
0.6
0.8
1
Real
I
m
a
g
Figure 15: Interpretation of Theorem 1. The left gure shows the graph representation of the
interconnected system and the right gure shows the corresponding Nyquist test. The addition of
the dashed line to the graph moves the negative, inverse eigenvalues of

L from the positions marked
by circles to those marked by crosses.
Theorem 1. A local controller K stabilizes the formation dynamics in equation (2) with error (5)
and gain K if and only if it stabilizes the set of N systems given by
x = Ax + B
i
(Ky)
y = Cx
(6)
where
i
are the eigenvalues of the weighted graph Laplacian

L.
This theorem has a very natural interpretation in in terms of the Nyquist plot of dynamical
system. In the standard Nyquist criterion, one checks for stability of a feedback system by plotting
the open loop frequency response of the system in the complex plane and checking for net encir-
clements of the 1 point. The conditions in Theorem 1 correspond to replacing the 1 point with
1/
i
for each eigenvalue
i
of

L. This interpretation is illustrated in Figure 15. The results can
easily be extended to consider weightings that are nonuniform.
Theorem 1 illustrates how the dynamics of the system, as represented by equation (2), interacts
with the information ow of the system, as represented by the graph Laplacian. In particular,
we see that it is the eigenvalues of the Laplacian that are critical for determining stability of the
overall system. Additional results in this framework allow tuning of the information ow (considered
as both sensed and communicated signals) to improve the transient response of the system [13].
Extensions in a stochastic setting [16, 41] allow analysis of interconnected systems whose dynamics
are not identical and where the graph topology changes over time.
4 Future Directions
While there has been substantial work in cooperative control over the past decade, there are still
many open problems that remain to be solved. In this section we provide a brief review of some
21
of the future opportunities in cooperative control. The topics listed here are not intended to be
exhaustive, but rather to be indicative of the classes of problems which remain open. Many of these
are drawn from the recent report on future directions in control, dynamics and systems [31].
4.1 Integrated control, communications and computer science
By its very nature, cooperative control involves the integration of communications and (distributed)
computing systems with feedback control. In many applications, the traditional separation of
computing, communications and control is no longer valid and new methods that integrate advances
from the dierent disciplines are needed. Research research in hybrid systems, in which continuous
and logical domains are integrated, are a step in the right direction but these techniques often
ignore issues associated with distributed computing and communication channels. Theories that
dene fundamental limits such as real-time computational complexity and performance limits of
feedback systems with rate limited channels are needed.
4.2 Verication and validation
Prescribed safety and reliability is a signicant challenge for current mission-critical systems. Re-
quirements, design, and test coverage and their quantication all signicantly impact overall system
quality, but software test coverage is especially signicant to development costs. For certain current
systems, verication and validation (V&V) can comprise over 50% of total development costs. This
percentage will be even higher using current V&V strategies on emerging autonomous systems.
Although traditional certication practices have historically produced suciently safe and reliable
systems, they will not be cost eective for next-generation autonomous systems due to inherent
size and complexity increases from added functionality.
New methods in high condence software combined with advances in systems engineering and
the use of feedback for active management of uncertainty provide new possibilities for fundamental
research aimed at addressing these issues. These methods move beyond formal methods in computer
science to incorporate dynamics and feedback as part of the system specication.
4.3 Higher levels of decision making
The research surveyed in this paper has focused on cooperative control problems that can be
formalated as optimization problems over some cost function. Many autonomous systems must
make decisions for which an underlying set of continuous and discrete variables may not provide
an appropriate level of abstract for decision making. Cooperative systems that must reason about
the complex interactions between the groups dynamics and the environment in which they operate
may require dierent levels of representation of their task and their dynamics. Techniques from
articial intelligence that allow identication of strategies and tactics that can be coded as lower-
level optimization-based problems are needed.
4.4 Networked control systems
Modern control theory is largely based on the abstraction that information (signals) are transmit-
ted along perfect communication channels and that computation is either instantaneous (continuous
time) or periodic (discrete time). This abstraction has served the eld well for 50 years and has
led to many success stories in a wide variety of applications.
22
Estimation/
Sensor Fusion
Optimization-
Based Control
Process 1
Sensing
Estimation/
Sensor Fusion
Optimization-
Based Control
External Environment
Process 2
Sensing
Figure 16: Control architecture for a networked control system.
Future applications of control will be much more information-rich than those of the past and will
involve networked communications, distributed computing, and higher levels of logic and decision-
making, as described above. New theory, algorithms, and demonstrations must be developed in
which the basic input/output signals are data packets that may arrive at variable times, not nec-
essarily in order, and sometimes not at all. Networks between sensors, actuation, and computation
must be taken into account, and algorithms must address the tradeo between accuracy and com-
putation time. Progress will require signicantly more interaction between information theory,
computer science, and control than ever before.
An emerging architecture for networked control systems is shown in Figure 16. This architecture
separates the traditional elements of sensing, estimation, control, and actuation for a given system
across a network and also allows sharing of information between systems. Careful decisions need
to be made on how the individual components in this architecture are implemented and how the
communications across the networked elements is managed. This architecture can be used to model
either a single system (using either half of the diagram) or multiple systems that interact through
the network.
The opportunity for networked control systems is the ability to quickly add functionality to
systems by incorporating new sensors and algorithms into an existing system. For cooperative
control systems, the ability to make use of a virtual sensor located on another vehicle without
having to redesign the control system from scartch is an example of the types of functionality
one would like to achieve. Similarly, new software modules that add functionality should provide
plug and play compatibility so that they can be integrated into systems quickly and reliably. This
network centric approach to control will require substantially better frameworks for implementing
complex, distributed control systems than currently exist today.
23
5 Conclusions
In this survey we have described some of the driving applications of cooperative control, surveyed
some of the relevant technology that has been developed over the past decade and provided some
possible directions for future study. Given the large and growing literature in this area, many
interesting results have not been included in an attempt to capture some of the key areas are
interest.
What is clear is that many of the basic problems of cooperative control have been explored
and a wealth of results are available demonstrating the potential of such systems. To transition
these research results to applciations will require additional eort in the integration of control,
communications and computer science; decision-making at higher levels of abstraction; verication
and validation of distributed embedded systems; and an extensible architecture for networked
control systems implementation.
The author would like to thank the US Air Force Oce of Scientic Research for their past and
continuing support of research in this area.
References
[1] R. Bhattacharya, A. Tiwari, J. Fung, and R. M. Murray. Cone invariance and rendezvous of
multiple agents. In IEEE Transactions on Automatic Control, 2006. Submitted.
[2] California Partners for Advanced Transit and Highways. http://www.path.berkeley.edu,
2006.
[3] P. R. Chandler, M. Pachter, and S. Rasmussen. Uav cooperative control. In Proc. American
Control Conference, pages 5055, 2001.
[4] J. Cortes, S. Martnez, T. Karatas, and F. Bullo. Coverage control for mobile sensing networks.
In Proc. IEEE International Conference on Robotics and Automation, pages 13271332, 2002.
[5] L. Cremean, W. Dunbar, D. van Gogh, J. Hickey, E. Klavins, J. Meltzer, and R. M. Murray.
The caltech multi-vehicle wireless testbed. In Proc. IEEE Control and Decision Conference,
2002.
[6] R. DAndrea. Cornell RoboCup. http://robocup.mae.cornell.edu, 2005.
[7] R. DAndrea and R. M. Murray. The RoboFlag competition. In Proc. American Control
Conference, 2003.
[8] D. Del Vecchio. Discrete dynamic feedback for a class of hybrid systems on a lattice. In IEEE
International Symposium on Computer-Aided Control Systems Design, 2006.
[9] D. Del Vecchio, R. M. Murray, and Erik Klavins. Discrete state estimators for systems on a
lattice. Automatica, 2006.
[10] Department of Transportation. National transportation libraryintelligent transporatation
systems. http://ntl.bts.gov/ref/biblio/its/its references.html.
24
[11] W. B. Dunbar and R. M. Murray. Distributed receding horizon control for multi-vehicle
formation stabilization. Automatica, 42(4):549558, 2006.
[12] M. G. Earl and R. DAndrea. Multi-vehicle cooperative control using mixed integer linear
programming. Technical Report arXiv:cs.RO/0501092, http://arXiv.org, 2005.
[13] J. A. Fax and R. M. Murray. Information ow and cooperative control of vehicle formations.
IEEE Transactions on Automatic Control, 2004.
[14] F. B. Gill. Ornithology. W. H. Freeman, second edition, 1994.
[15] V. Gupta, T. H. Chung, B. Hassibi, and R. M. Murray. On a stochastic sensor selection
algorithm with applications in sensor scheduling and sensor coverage. Automatica, 2004. In
revision.
[16] V. Gupta, R. M. Murray, and B. Hassibi. Stability analysis of stochastically varying formations
of dynamic agents. In Proc. IEEE Control and Decision Conference, 2003.
[17] G. Homann, D. G. Rajnarayan, S. L. Waslander, D. Dostal, J. S. Jang, and C. J. Tomlin.
The Stanford testbed of autonomous rotorcraft for multi-agent control (STARMAC). In AIAA
Digital Avionics Systems Conference, 2004.
[18] Intelligent Transportation Society of America. http://www.itsa.org.
[19] A. Jadbabaie, J. Lin, and A. S. Morse. Coordination of grups of mobile autonomous agents
using nearest neighbor rules. IEEE Transactions on Automatic Control, 48(6):9881001, 2003.
[20] Jet Propulsion Laboratory. Terrestrial planet nder. http://tpf.jpl.nasa.gov.
[21] Z. Jin. Coordinated Control of Networked Multi-Agent Systems. PhD thesis, California Institute
of Technology, Electrical Engineering, 2006.
[22] Z. Jin and R. M. Murray. Stability and performance analysis with double-graph model of
vehicle formations. In Proc. American Control Conference, 2003.
[23] Z. Jin and R. M. Murray. Double-graph control strategy of multi-vehicle formations. In Proc.
IEEE Control and Decision Conference, 2004.
[24] E. King, Y. Kuwata, M. Alighanbari, and J. How. Coordination and control experiments for
uav teams. Advances in the Astronautical Sciences, 118:145155, 2004.
[25] H. Kitano, M. Asada, Y. Kuniyoshi, I. Noda, and E. Osawa. Robocup: The robot world cup
initiative. In Autonomous Agents 97, pages 340347, 1997.
[26] E. Klavins. A formal model of a multi-robot control and communication task. Proc. IEEE
Control and Decision Conference, 2003.
[27] E. Klavins and R. M. Murray. Distributed computation for cooperative control. IEEE Perva-
sive Computing, 2003. In revision.
[28] E. Lavretsky. F/a-18 autonomous formation ight control systems design. In aiaaGNC, pages
AIAA 20024757, 2002.
25
[29] N. E. Leonard and E. Fiorelli. Virtual leaders, articial potentials and coordinated control of
groups. In Proc. IEEE Control and Decision Conference, pages 29682973, 2001.
[30] Monterey Bay Aquarium Research Institute. Autonomous ocean sampling network. http:
//www.mbari.org/aosn, 2006.
[31] R. M. Murray, editor. Control in an Information Rich World: Report of the Panel on Future
Direcitons in Control, Dynamics and Systems. SIAM, 2003. Available at http://www.cds.
caltech.edu/

murray/cdspanel.
[32] P. Ogren, E. Fiorelli, and N. E. Leonard. Cooperative control of mobile sensor networks:
Adaptive gradient climbing in a distributed environment. IEEE Transactions on Automatic
Control, 49(8):12921302, 2004.
[33] R. Olfati-Saber and R. M. Murray. Distributed cooperative control of multiple vehicle forma-
tions using structural potential functions. In Proc. IFAC World Congress, 2002.
[34] J. Ousingsawat and M. E. Campbell. Estabilishing trajectories for multi-vehicle reconnaissance.
In AIAA Conference on Guidance, Navigation, and Control, pages AIAA 20045224, 2004.
[35] L. E. Parker. Designing control laws for cooperative agent teams. In Proc. IEEE International
Conference on Robotics and Automation, pages 582587, 1993.
[36] L. E. Parker. ALLIANCE: An architecture for fault tolerant multirobot cooperation. IEEE
Transactions on Robotics and Automation, 14(2):220240, 1998.
[37] L. E. Parker. Current state of the art in distributed autonomous mobile robotics. In Interna-
tional Symposium on Distributed Autonomous Robotic Systems (DARS), 2000.
[38] C. W. Reynolds. Herds, and schools: A distributed behavioral model. Computer Graphics
(SIGGRAPH 87 Conference Proceedings), 21(4):2534, 1987.
[39] A. Richards, J. Bellingham, M. Tillerson, and J. How. Co-ordination and control of multiple
UAVs. In AIAA Conference on Guidance, Navigation, and Control, 2002.
[40] R. Olfati Saber. Flocking for multi-agent dynamic systems: Algorithms and theory. IEEE
Transactions on Automatic Control, 51(3):401420, 2006.
[41] R. Olfati Saber and R. M. Murray. Consensus problems in networks of agents with switching
topology and time-delays. IEEE Transactions on Automatic Control, 2004.
[42] D. Swaroop and J. K. Hedrick. String stability of interconnected systems. IEEE Transactions
on Automatic Control, 41(3):349357, 1996.
[43] A. Tiwari, J. Fung, J. M. Carson III, R. Bhattacharya, and R. M. Murray. A framework
for Lyapunov certicates for multi-vehicle rendezvous problems. In Proc. American Control
Conference, 2004.
[44] A. Tiwari, M. Jun, D. E. Jecoat, and R. M. Murray. The dynamic sensor coverage problem.
In Proc. IFAC World Congress, 2005.
26
[45] C. Tomlin, G. J. Pappas, and S. Sastry. Conict resolution for air trac management: A study
in multiagent hybrid systems. IEEE Transactions on Automatic Control, 43(4):509521, 1998.
[46] United States Department of Defense. Network centric warfare. Report to Congress, July
2001. Available at http://www.dod.mil/nii/NCW.
[47] V. Vladimerou, A. Stubbs, J. Rubel, A. Fulford, and G.E. Dullerud. A hovercraft testbed for
decentralized and cooperative control. In Proc. American Control Conference, 2004.
27