An Architecture for UAV Team Control

Abstract

Recent years has seen a widespread interest in the use of Unmanned aircraft vehicles for military applications. These UAV's can be used in many applications such as surveillance, information gathering, suppression of enemy defenses, air to air combat, mapping building and facilities etc. In this paper, we present an architecture with the necessary algorithms that we have implemented to control a team of UAVs to search for targets such as SAMs, ground troops, artillery, tanks etc in a given region.

Full text

AN ARCHITECTURE FOR UAV TEAM
CONTROL
Sivakumar Rathinam ∗Marco Zennaro ∗∗
Tony Mak ∗∗∗ Ra ja Sengupta ∗∗∗∗
∗Graduate Student, CEE Systems, Berkeley
∗∗ Graduate Student, CEE Systems, Berkeley
∗∗∗ Systems Analyst, PATH, Berkeley
∗∗∗∗ Assistant Professor, CEE Systems, Berkeley
Abstract: Recent years has seen a widespread interest in the use of Unmanned
aircraft vehicles for military applications. These UAV’s can be used in many
applications such as surveillance, information gathering, suppression of enemy
defenses, air to air combat, mapping buildings and facilities etc. In this paper, we
present an architecture with the necessary algorithms that we have implemented
to control a team of UAVs to search for targets such as SAMs, ground troops,
artillery, tanks etc in a given region.
Keywords: UAV, search, architecture
1. INTRODUCTION
Unmanned Aerial Vehicles (UAV’s) has received
significant attention in recent years for military
applications {Pachter, 98} {Bortoff, 99} {Mclain,
99} {McLain, 00} {Chandler, 01} {Nygard, 01}
{Passino, 01} {Lee, 03}. The motivation behind
this interest is to realize a vision where these
vehicles cooperatively accomplish missions such
as search and attack {Murphey, 99} {Polycarpou,
01} {Beard, 02} {Jaques, 03} {Bellingham, 03}.
The reason for the widespread interest are the
many advantages that the unmanned vehicles
have over manned vehicles such as reduced human
risk, manoeuvrability, and superior cooperation.
In this paper, we primarily concentrate on the
search mission and describe a architecture with
the algorithms that we have implemented to re-
alize the same. A detailed review of the search
literature is presented in {Polycarpou, 01}. Our
contributions of this work compared to the previ-
ous work are:
•Modularity with which the different types of
sensors and the strategies are coupled in the
architecture.
•A safe flight path planning algorithm that
allows sufficient time for the sensors for basic
operations such as imaging and image pro-
cessing. Also this algorithm satisfies the the
yaw rate and sensor range constraints of the
vehicle.
1.1 Basic Approach
The primary step before developing search strate-
gies and its architecture is to first come up with
a way of representing the threats and to answer
what kind of information is shared between the ve-
hicles. To address these questions, we use a proba-
bility map to represent the threats. As the vehicles
move in the unknown regions, the sensors collect
new information (location and the type of targets)
about the environment. This new information is
updated with the probablity map using bayes rule.
The probability maps of all the different types

of the threats is used to calculate the risk map
which indicates the risk of being shot at any given
point in the desired search region. This risk map
is shared by all the vehicles and is used for path
planning. Following the path based on the par-
tially known risk map, would damage or destroy
the vehicles if they enter regions where Surface
to Air Missiles (SAMs) are present. Therefore,
we reduce the forward speed of the vehicles by
refining the nominal path, thus allowing sufficient
time for image processing and target recognition.
This refinement of the nominal path is basically
done by a safe flight design which is discussed in
the later sections.
Two strategies for navigating in an unknown re-
gion are outlined in this paper: one for reaching
a destination and an other for searching a given
area. All the ideas presented above forms the
backbone for the UAV team architecture. Ideally,
what we want is an architecture that allows hu-
man interface at each level of control and also
can automatically make its decisions if required
as to which speed mode or sensor choice to use to
suitably realize a mission. The architecture should
allow an human interface to pick and choose the
regions that is required to be searched.
The paper is organized as follows: The problem
setup which includes the capabilities of the various
sensors, threat models, assumptions on the motion
of the vehicle is presented in 2. The threat map
representation and the risk function calculation
is presented in section 3. Safe flight algorithm is
discussed in 4. Section 5 and 4 presents the two
main search strategies. The main architecture is
explained in section 6 and the properties are dis-
cussed in 6.5. Simulation results are presented in
section 7 and the paper concludes with comments
and future work.
2. PROBLEM SETUP
2.1 Sensors and their capabilities
Sensors such as Electro-optical(EO), Infrared(IR)
and Synthetic Aperture Radar Systems can be
used onboard of these vehicles to collect infor-
mation about the partially known or unknown
environments. The information that we are inter-
ested from these Surveillance and Reconnaissance
(S&R) systems are the type of threats and the
location of the threats. Each of these sensors has
its own advantages and disadvantages and cannot
be used for all applications. {Leachtenauer, 01}
proposes a guideline in the selection of a sensor
for a S&R application. This is presented in figure
1. Retrieving this useful information requires a
sufficient amount of processing time (varies for
different sensors) for performing basic operations
Heavy cloud, cover/
rain or long range? Daytime?
(illuminated)
Electro Optical
Inexpensive
High Resolution
Infrared
yes
no no
yes
Synthetic Aperture
Radar
All weather conditions
Low Resolution
Fig. 1. General guidelines for sensor selection
{Leachtenauer, 01}.
such as image formation and image processing. In
this paper we assume that if sufficient amount of
time is spent on imaging and processing an area,
then the probability of detecting a target located
in that area is 1.
2.2 Threat Model
In this work, we assume that the only threat
that can destroy the vehicles are the stationary
SAM launchers. We also assume that the range
of the search radars of the launchers are larger
than its fire control range. Therefore, as soon as
the vehicles enter their search radar zone, they
are locked for destruction and the weapons with
its tracking radars are instantly fired when the
vehicles enter the fire control range. Hence, a
vehicle is instantly destroyed when they fall in the
range of the fire control radars of the SAMs. The
only way to keep the vehicles safe is to avoid the
threat regions of all the launchers.
2.3 Assumptions on the Motion of the UAV
The vehicles are assumed to fly at a constant
height. A UAV at any time t can be specified by its
coordinates {x(t),y(t),θ(t)}. We treat each UAV
as a Dubins car (simple stick model), travelling
at a constant speed with a bound on its yaw
rate. Let vdenote the velocity of the UAV. Let
ωrepresent the bound on the yaw rate of the
UAV. The motion of the UAV is governed by the
following equations:
˙
x(t) = vcos θ(t) (1)
˙
y(t) = vsin θ(t) (2)
˙
θ(t) = Ω where Ω²[−ω, +ω] (3)
3. THREAT REPRESENTATION
The targets consists of different types of targets
such as Surface to Air Missile Launchers (SAMs),

Surface to surface missiles, tanks, etc. A set of N
targets is represented as:
T argets ={target1= (type1,(x1, y1)), ..
, ..., targetN= (typeN,(xN, yN))}(4)
Here, the targetiis of typeiand is situated at a lo-
cation (xi, yi). Let the targets be distributed over
kareas A1, A2, ..., Ak.Ntj targets of type tare
assumed to be independently and uniformly dis-
tributed over Area Aj. The information about the
targets is represented as a probability distribution
of the random variables N, typei,{xi, yi}. This
probability distribution of targets is expressed as:
Pthreat(T ar gets) =
Y
t
k
Y
j=1
Ntj
Y
i=1
ptj (type =t, (xi, yi))Ptj (Ntj )(5)
in which,
ptj (type =t, (xi, yi)) = ½|Aj|−1,(xi, yi)∈Aj
0. otherwise (6)
At the start of the mission, this information about
the targets or the probability distribution is also
referred as the Initial Preparation of the Bat-
tlefield (IPB). As the vehicles move through the
region, sensors on the vehicles gather information
about the targets. We use Bayes rule the update
the distribution with the incoming sensor infor-
mation.
3.1 Risk Function Calculation
The instantaneous risk function r(x, y, P ) at any
point (x, y) given the probability distribution P
of the targets is given by:
r(x, y, P ) =
k
X
j=1 X
t
∞
X
Ntj =0 X
n=1
Ntj




Z
Aj
ft(|(x, y)−(xn, yn)|)|Aj|−1dxndyn




.(7)
where, the function ftis chosen as follows: ft(r) =
1 if r≤RLand ft(r) = 0 if r > RL.
4. SAFE FLIGHT DESIGN
A detailed description of a general safe flight
algorithm is presented in {Rathinam, 04}. Here,
we just present the gist of designing a path for
a vehicle required to travel from waypoint Ato
waypoint B0. As the vehicle moves from Ato B0,
let the sensor on the vehicle collect information
about the area S1S2S3S4as shown in the figure
2. If the vehicle travels straight from Ato B0,
since the forward speed of the vehicle is generally
large (200km
hr ), the sensor on the vehicle may
not have enough time to process the information.
Therefore, the vehicle may not be able to take
any evasive action in the presence of any SAM
launcher in the area S1S2S3S4. So, the vehicle
has the risk of being shot. To avoid this, we
increase the length of the path from Ato B0to
give sufficient time for the sensor to process the
information. This is done by piecing together a
series of semi circles and half ellipses as shown
in figure 3. The minor axis xand the major
axis yof the ellipses are the variables and are
chosen to satisfy all the required constraints. Let
us denote the generated path from Ato Bwith
nsuch semicircles and half ellipses be denoted by
G(A, B, n). The values of xand yare chosen to
satisfy the following constraints:
•Minimum curvature constraint: Choosing y=
√xr where r=v
ωis the radius of the semi-
cirle satisfies this constraint. Refer to the
appendix for the proof.
•Sensor timing constraint: As shown in the
figure 4, the sensor on the UAV, as it moves
from Ato B0, is required to collect and
process the information of the targets present
in the area S1S2S3S4. The constraint here is
that, the flight time required to travel from
Ato B0should be sufficient enough for the
sensor to process the collected information.
Let RLbe the firing range (km)of the SAM
launchers present in the area of interest. Let
τindicate the image processing rate per unit
area ( hr
km2) of the sensor on the UAV. Then,
2(RL+d)τ≤(πr +L(x, y))
2v(x−r)(8)
where, L(x, y) indicates the length of the half
ellipse.
•Sensor range constraint: The range of the
sensor RSshould be large enough (>length
P S2in figure 4)for the sensor on the UAV to
reach the entire area S1S2S3S4. That is,
(RL+y+ 2x)2+ (RL+ 2y)2≤R2
S(9)
5. SEARCH STRATEGIES
5.1 Strategic Search
The objective in strategic search is to find a fea-
sible path (if there exists one), devoid of threats
to a given destination. At the start of the mission,
only partial or no information is known about the
location of the launchers. {Stentz, 96}proposed a
efficient path planning algorithm for vehicles for

A B’
S1S2
S3
S4
x
area to be
sensed
x
KNOWN REGION UNKNOWN REGION
Fig. 2. Reason behind safe flight design.
Direction of motion
d = nx
r
2x
y
RL
A B
connecting ‘n’ semicircles and half ellipses to reach the waypoint B
Fig. 3. Connecting semicircles and half ellipses to
augment the length of the path.
r
2x
y
RL+y
A B’
RL+y
Safe distance = RL+y
P
S1S2
S3
S4
area to be
sensed
Fig. 4. Scanning area for the vehicle as it moves
forward.
reaching a destination in partially known environ-
ments. The algorithm is functionally equal to the
brute-force optimal strategy that is presented as
follows:
•Follow the minimum risk path using the
information known at the start.
•If any obstacle is identified, update the map
with the collected information and follow the
new minimum risk path to the destination
(from the current location). Repeat this step
until you reach the destination or all the
paths to the destination are blocked by ob-
stacles
The above algorithm has the completeness prop-
erty that the vehicle will find a feasible path to
the destination if there exists one.
5.2 Threat Search
The aim of the threat search is to find all the
SAMs in a given area. For this search, we use a
adaptive space filling algorithm as follows:
•Generate a space filling curve that covers the
given area
•Fly the generated curve until one of the
following happens:
·If a new threat is observed, then fly the
minimum risk path to the next destina-
tion on the curve. Repeat this process
until all the destination points in the
curve are used.
·Final destination on the curve is reached.
6. THE UAV TEAM ARCHITECTURE
6.1 Architecture overview and design guidelines
The system goal is to employ a team of UAVs to
search collaboratively an unknown environment
while avoiding obstacles and threats. The UAVs
have different capabilities (e.g. different sensors
accuracy and type, speed, endurance), that have
to be used efficiently to complete the mission.
Given the nature of the scenario we design the
system to have the following four properties: mod-
ularity, safety, resilience, scalability. The system
consists of heterogeneous components, e.g. dif-
ferent UAVs equipped with different sensors. We
want this heterogeneousness to be transparent to
the system. In order to achieve this goal we made
the design modular: we decompose the system into
logical blocks (e.g. sensor controller, UAV autopi-
lot) and we define the interfaces between them.
In this particular scenario the set of available
resources dynamically changes: resources may be
added on the run, while other may be destroyed or
damaged. The system must be able to cope with
these continuous changes. It has to be scalable, in
order to be able to operate with all the available
resources and accommodate additional resources.
It has to be resilient to failures in order to con-
tinue to operate even when some of the resources
are destroyed or damaged. The system perfor-
mances should degrade gracefully. Scalability and
resilience are addressed using a layered architec-
ture that offers different level of control. Since the
environment is unknown and potentially unsafe,
the system should try to minimize destruction or
damage to the UAVs. Moreover the output of the
available sensors is subjects to errors and long pro-
cessing delay. The sensor inaccuracy problem has

Formation
Navigation
Resource allocation/scheduling
Operation decomposer
Operation monitor
Dispatcher
Target
Distribution
Model
Distribution
Update Risk Map
Team Manager
UAV UAV
Human interface
Fig. 5. Overview of the Architecture.
been addressed by sharing the threat information
among the members of the team. The processing
delay problem has been addressed by having every
single UAV adopting a safe flight strategy (as
described in section 4). The high level structure of
the system architecture is given in figure 5. The
system consists on three main components: the
Team Manager, the UAV Managers and the Sen-
sor Information Processing Unit(Risk map, Target
distribution modul e, Prob update module ). The
UAV Managers offers a higher level of control of a
single UAV to the Team Manager: instead of flying
the UAV waypoint by waypoint, we can command
it to execute a more complex task (e.g. flying to
an area while avoiding obstacles or search an area
for threats). The Sensor Information Processing
Unit is used to share and fuse the information
gathered by different sensors from different UAVs.
The Team Manager block coordinates and mon-
itors the team members, and it offers a higher
level of control. Instead of controlling every single
UAV using tasks we can assign to the team a
mission (e.g. efficiently search an area or look for
a minimum risk corridor).
6.2 Team Manager
The Team Manager offers a higher level of control
above the single UAV task control level. It allows
the user to control a team of UAVs to collabora-
tively perform a mission, such as target search in
an unknown environment and safe corridor discov-
ery. The Team Manager components is also given
in figure 5. A mission is broken down by the Oper-
ation Decomposer into a sequence of tasks. The re-
source allocation block then allocates the available
resources to perform these tasks. The dispatcher
assigns the tasks to the allocated resources. The
Operation Monitor monitors the system status.
If some of the resources are destroyed, or new
ones become available, it prompts the Resource
Allocator to redistribute the task load among the
new set of resources. This component allows the
system to dynamically adapt to changes.
6.3 UAV Manager
The current available systems, for example the
Cloudcap Piccolo system offer a low level of con-
trol, usually a waypoint navigation control with-
out obstacle and threat avoidance. The UAV man-
ager is built on top of such systems in order to
provide a higher level of control as well as obstacle
avoidance. The structure of the UAV manager
is described in figure 6. The Sensor Manager
deals with the complexity of the sensors using
the information provided by the Dynamic Path
Planer (e.g. the direction and the speed of the
motion). It takes care of the sensor aiming and
ensures that the sweeping pattern does not leave
behind unsearched areas. The autopilot deals with
the waypoint navigation (without obstacle avoid-
ance). An example of such an Autopilot is the
just mentioned Cloudcap Piccolo system. On top
of it, we built a safe controller. The Dynamic
Path Planer (DPP) creates a minimum risk path
(using the risk map) between the current position
and the destination given by the task manager,
avoiding known obstacles and threats. The task
manager provides a task execution service to the
above layers.
If any threat is detected, the safe controller stops
the autopilot from following the assigned path and
interrupts the Dynamic Path Planner for a re-
vised nominal path. Meanwhile, the safe controller

Risk Map
UAV manager
Dynamic Path Planner
Safe Flight
Controller
Task Manager
Sensor Manager
Autopilot Sensor
Controller Observation Update
Destination
Nominal path
Waypoint
Fig. 6. UAV Manager
generates dummy paths that forces the vehicle to
just fly around a fixed point until it receives a
new nominal path from the DPP to follow. This
interrupt from the safe controller ensures that
irrespective of the decisions that are taken at the
higher levels of the architecture, it always controls
the safety of the vehicle.
6.4 Architecture implementation using an example
Lets consider an example of a search area mission.
The following are the break up of the functions of
each block in the proposed architecture:
•The Operation decomposer produces the fol-
lowing list of tasks: fly to the area to be
searched, search the area and fly back home.
•The Resource Allocator assigns each of these
tasks to the team of UAVs available. The
following three tasks are assigned to every
UAV: a fly to to the search area, a search
area task, where the area to be searched is
a fraction of the total area to be searched
and a fly to the base. If any of the UAVs
are destroyed or damaged, the Operation
Monitor prompts the Resource Allocator to
redistribute the areas to be searched among
the remaining UAVs.
•Each UAV manager is assigned with an area
to searched. The Task Manager produces a
set of way points to be followed based on a
adaptive space filling algorithm for the given
search area.
•Dynamic path planner generates the mini-
mum risk path (nominal path) to any desti-
nation way point specified by the task man-
ager. The safe flight controller refines the
path with the safe flight design.
•Apart from the functions that the sensor
manager performs as mentioned before, a
human interface is provided to pick the type
of sensor to be used in the mission. SAR
sensors can be first used to identify all the
main threats that can destroy the vehicle
such as SAMs. Once they are identified, then
EO sensors can be used to classify targets
such as trucks or buses accurately.
A simulation of the above search mission to find all
the SAMs is shown in figure 7. One of the other
missions that we implemented was the strategic
search mission. This mission would be useful when
the aim is to find a feasible safe path to the
destination. In this mission, the only real part that
is different from the search mission is the Task
Manager. In this case, the Task Manager has just
one destination point to reach. A simulation of
this mission is shown in figure 8. The following
subsection lists the properties of the architecture
that we presented above.
6.5 Properties of the Architecture
The interrupt that the safe controller has basi-
cally provides the completeness property of the
strategic search that if a there exists a safe path,
the components in the architecture will guide the
vehicle to find that path. The safe controller by
refining the nominal path to the destination pro-
vides sufficient time for the sensors to process the
information about the scanned region in front of
the vehicles for threats. If the sensors report any
presence of SAMs, then the entire mission is im-
mediately stopped and the interrupt is passed on
to the Dynamic path planner. The Dynamic path
planner then again probes the sensor information
processing unit for the latest risk map with the
new updates. The DPP uses this new risk map
to find a feasible nominal path and this process
repeats itself according to the strategy outlined
in section 5. Thus the completeness property of
the strategy is satisfied. Also, the presented ar-
chitecture has a property of information adequacy.
That is, the safe controller generates a path such
that the information that is gathered by flying
the refined safe path is atleast equal to the infor-
mation that would have been gathered by flying
the nominal path generated by DPP. The next
section discusses the Mission control interface that
we built for the architecture.
7. SIMULATION
An implementation of the proposed architecture
has been developed on Mixed Initiative Control
for Automata-teams (MICA) Open Experimental
Platform (OEP) Simulator. We implemented also
a Mission Control interface to interact with the
system. The interface is shown in figure 9.
The user specifies a mission (currently only a
search area mission is implemented) and its pa-
rameters (e.g. the area to be searched, some
constraints on the sensor type to use, etc).Then

Fig. 7. Search mission
Fig. 8. Strategic search
he/she monitors the mission progress. The user
interface displays the location of the UAVs and
of the detected objects (in the current appli-
cations trucks, SAM and busses) in real-time.
When an object is detected but cannot be clas-
sified (because, for example, SAR sensors can-
not distinguish between truck and busses since
they belong to the same visually similar objects
group), the user can interactively dispatch a new
UAV equipped with EO sensors for classification.
Again, the results are displayed in real time on
the user interface.
8. CONCLUSION
An architecture with the algorithms and its imple-
mentation for controlling a team of vehicles for the
search mission was presented in the paper. Some
of the assumptions (which also are the drawbacks)
of the current work are as follows:
•Sensors are perfect or the probability of de-
tecting a SAM launcher is 1 if sufficient
amount of time is spent.
•A simple stick model is used for the kine-
matics of the UAV and only 2-D motion is
considered.
•The sensor control onboard UAV’s are capa-
ble of scanning a specified part of an area
irrespective of the direction of the UAV.
The following are the future directions of this
current work
•The resource allocation during cooperation
between the vehicles has not been solved.

Fig. 9. Mission Control user interface
Also, the question as to how to redistribute
the resources when the vehicles are shot has
not been answered
•The information structure in the presented
architecture was centralized. In other words,
the information about the entire region or the
risk map is known to all the vehicles. So when
large number of vehicles are operating, the
aim now is to make this information structure
decentralized.
9. ACKNOWLEDGEMENTS
The authors thank Prof Pravin Varaiya, Univer-
sity of California, Berkeley and Prof Joao Sousa,
Universidade do Porto, Portugal for their useful
comments. The research was supported in part by
the Boeing Mica Program - grant #F33615-01-C-
3150 and by ONR AINS Program - grant #25833.
10. APPENDIX
Lemma:
Assume x ≥y. If x ≥v
ωand if y ≥pxv
ω, the yaw
rate constraint is always satisfied.
Proof: The speed of a UAV is a constant and is
always along the direction of the path. Let the
ellipse be represented in a parametric form by
(X(t),Y(t)), where X(t) = xcos t,Y(t) = ysin t
and t varies in the interval [0,2π]. Also let the
radius of curvature and the angular velocity at
any point be denoted by ρ(t) and ˙
Φ(t). Since the
velocity vector of the vehicle is always tangent
to the path of the vehicle and its magnitude is
a constant, the following claim is true:
max
∀t∈[0,2π]
˙
Φ(t)∼min
∀t∈[0,2π]ρ(t) (10)
The radius of curvature is given by,
ρ(t) = (X02+Y02)3
2
X0Y00 −Y0X00 (11)
where X’ = dx
dt and Y’ = dy
dt . Hence by substituting
the parametric forms for X and Y, we have,
ρ(t) = (x2sin t2+y2cos t2)3
2
xy (12)
It is trivial to see that the radius of curvature is
the same at all the points if x = y. ρ(t) is minimum
at t ={0, π}if x >y. The minimum value for x
≥y is equal is equal to y2
x. Hence for the yaw
rate condition to be satisfied, x ≥v
ωand y2
a≥v
ω.
Hence proved.
11. REFERENCES
Beard R.W., T.W. McLain, M. Goodrich, and
E.P. Anderson, “Coordinated target assignment
and interceptfor unmanned air vehicles”, in IEEE
Transactions On Robotics and Automation, 18:911
922, 2002.
Bellingham J., M. Tillerson, A. Richards, and J.
How, “Multi-task allocation and path planning
for cooperative UAVs”, in Cooperative Control:
Models, Applications and Algorithms, Editors - S.
Butenko, R. Murphey, and P.M. Pardalos, Kluwer
Academic Publishers, pages 2341, 2003.
Bortoff S.A., “Path-planning for unmanned air
vehicles ”, Summer report, AFRL, 1999.
Chandler P., S. Rasmussen and M. Pachter, “UAV
cooperative control,” in Proceedings of American
Control Conference, Arlington, VA, June 25-27
2001.
Dubins L.E., “On curves of minimal length with
a constraint on average curvature, and with pre-
scribed initial and terminal positions and tan-
gents”, in American Journal of Mathematics, vol
79, Issue 3, July 1957, pp. 487-516.
Hwang Y.K., N. Ahuja, Gross Motion Planning -
A Survey, ACM Computing Surveys 24, no. 3, pp.
219-291, September 1992.
Jacques D.R., “Search, classification and attack
decisions for cooperative wide area search muni-
tions”, in Cooperative Control: Models, Applica-
tions and Algorithms, Editors - S. Butenko, R.
Murphey, and P.M. Pardalos, Kluwer Academic
Publishers, pp 7593, 2003.
Latombe J.C., Robot Motion Planning, Kluwer
Academic Publishers, Norwell, Mass. 1991.
Laumond J.-P., Robot motion planning and con-
trol, Springer-Verlag, 1998.

Leachtenauer J.C. and R.G. Driggers, Surveil-
lance and Reconnaissance Imaging Systems -
Modelling and Performance Prediction, Massachusetts:
Artech House, 2001.
Lee, Huang, Vaughn, Xiao, Hedrick, Zennaro,
Sengupta, Startegies of Path-Planning for a UAV
to Track a Ground Vehicle, AINS Conference 2003
Lumelsky V. and A. Stepanov , ”Dynamic path
planning for a mobile automaton with limited
information on the environment,” in IEEE Trans-
actions on Automatic Control, vol. 31, no. 11, pp.
1058-1063, 1986.
Mclain T.W., “Coordinated control of unmanned
aircraft vehicles”, Summer report, AFRL, 1999.
McLain T.,“Cooperative rendezvous of multiple
unmanned air vehicles,” in AIAA GNC 2000,
Denver, CO, Aug 2000, Paper no. AIAA 2000-
4369.
Murphey R.A., “An approximate algorithm for
a weapon target assignment stochastic program,
in Approximation and Complexity in Numerical
Optimization: Continuous and Discrete Problems,
P.M. Pardalos, Ed., Boston: Kluwer Academic
Publishers, pp. 1-2, 1999.
Nygard K.E., P.R. Chandler and M. Pachter,“Dynamic
network flow optimization models for air vehicle
resource allocation,” in American Control Confer-
ence 2001, Arlington, VA.
Pachter M. and P. Chandler, “Challenges of au-
tonomous control”, IEEE Control Systems Maga-
zine, pp. 92-97, April 1998.
Passino K.M., M.M. Polycarpou, D. Jacques and
M. Pachter, “Distributed cooperation and control
for autonomous air vehicles”, in The Workshop for
Cooperative Control and Optimization, Gainsville,
FL, December 3-5 2000.
Polycarpou M.M., Y. Yang, and K. Passino, A co-
operative search framework for distributed agents,
in Proc.IEEE ISIC2001, pp 16, 2001.
Rao N., S. Hareti, W. Shi, and S. Iyengar, Robot
navigation in unknown terrains : Introductory
survey of non-heuristic algorithms, Technical Re-
port ORNL/TM-12410, Oak Ridge National Lab-
oratory, 1993.
Rathinam S. and R. Sengupta, “A Safe Flight
Algorithm for Unmanned Aerial Vehicles”, to be
presented in IEEE Aerospace Conference, Mon-
tana, 2004.
Stentz A.,“Map-based Strategies for Robot Nav-
igation in Unknown Environments”, in Planning
with Incomplete Information for Robot Problems:
Papers from the 1996 AAAI Spring Symposium,
AAAI Press, Menlo Park, California, pp. 10-116,
1996.