DAMN: A Distributed Architecture for Mobile Navigation

Julio K. Rosenblatt
Robotics Institute
Carnegie Mellon University
Pittsburgh, PA 15213
jkr@cmu.edu

Abstract
An architecture is presented in which distributed task- to reason about the systems own intentions and goals.
achieving modules, or behaviors, cooperatively determine a Hierarchical approaches allow slower abstract reasoning at
mobile robots path by voting for each of various possible
the higher levels and faster numerical computations at the
actions. An arbiter then performs command fusion and selects
that action which best satisfies the prioritized goals of the lower levels, thus allowing varying trade-offs between
system, as expressed by these votes, without the need to responsiveness and optimality as appropriate at each level
average commands. Command fusion allows multiple goals (Payton, 1986; Albus, McCain and Lumia, 1987). While
and constraints to be considered simultaneously. Examples of such an approach provides aspects of both deliberative
implemented systems are given, and future research directions
planning and reactive control, the top-down nature of
in command fusion are discussed.
hierarchical structures tends to overly restrict the lower
Keywords: mobile robots, distributed architecture, behaviors, levels (Payton, Rosenblatt and Keirsey, 1990). In
voting, arbitration, command fusion hierarchical architectures, each layer controls the layer
beneath it and assumes that its commands will be executed
as expected. Since expectations are not always met, there is a
Introduction
need to monitor the progress of desired actions and to report
In order to function in unstructured, unknown, or dynamic failures as they occur (Simmons, Lin and Fedor, 1990). In an
environments, a mobile robot must be able to perceive its unstructured, unknown, or dynamic environment, this
surroundings and generate actions that are appropriate for approach introduces complexities and inefficiencies which
that environment and for the goals of the robotic system. could be avoided if higher level modules participated in the
Mobile robots need to combine information from several decision-making process without assuming that their
different sources. For example, the CMU Navlab vehicles commands will be strictly followed.
are equipped with sensors such as video cameras, laser range Rather than imposing a top-down structure to achieve this
finders, sonars, and inertial navigation systems, which are desired symbiosis of deliberative and reactive elements, the
variously used by subsystems that follow roads, track paths, Distributed Architecture for Mobile Navigation takes an
avoid obstacles and rough terrain, seek goals, and perform approach where multiple modules concurrently share control
teleoperation. Because the raw sensor data and the internal of the robot by sending votes to be combined rather than
representations used by these subsystems tend to vary commands to be selected and executed (Rosenblatt, 1995;
widely, especially when these modules have been developed Rosenblatt and Thorpe, 1995).
independently, combining them into one coherent system The Distributed Architecture for Mobile Navigation has
which combines all their capabilities has proven to be very been successfully used to integrate the various subsystems
difficult. mentioned above, thus providing systems that perform tasks
To function effectively, an architectural framework for such as road following, cross-country navigation, and
these sensing and reasoning processes must be imposed to teleoperation while avoiding obstacles and meeting mission
provide a structure with which the system may be developed, objectives. In addition to its use on the CMU Navlab
tested, debugged, and understood. However, the architecture vehicles, DAMN has also been used at Martin Marietta,
should serve as an aid, not a burden, in the integration of Hughes Research Labs, and the Georgia Institute of
modules that have been developed independently, so it must Technology.
not be overly restrictive. It must allow for purposeful goal-
oriented behavior yet retain the ability to respond to
potentially dangerous situations in real-time while The Distributed Architecture for
maintaining enough speed to be useful. Mobile Navigation
The earliest work in robot control architectures attempted Deliberative planning and reactive control are equally
to reason by manipulating abstract symbols using only pure important for mobile robot navigation; when used
logic (Nilsson, 1984). The limitations of this top-down AI appropriately, each complements the other and compensates
approach led to a new generation of architectures designed in for the others deficiencies. Reactive components provide
a bottom-up fashion to provide greater reactivity to the the basic capabilities which enable the robot to achieve low-
robots surroundings, but sacrificed generality and the ability level tasks without injury to itself or its environment, while
deliberative components provide the ability to achieve while an attempt is made to keep the perception and
higher-level goals and to avoid mistakes which could lead to planning components of a behavior as simple as possible
inefficiencies or even mission failure. But rather than without sacrificing dependability, they can and often do
imposing an hierarchical structure to achieve this symbiosis, maintain internal representations of the world. Brooks
the Distributed Architecture for Mobile Navigation (DAMN) (1993) has argued that the world is its own best model,
takes an approach where multiple modules concurrently but this assumes that the vehicles sensors and the
share control of the robot. In order to achieve this, a scheme algorithms which process them can not benefit from
is used where each module votes for or against various evidence combination between consecutive scenes. In
alternatives in the command space based on geometric addition, disallowing the use of internal representations
reasoning, without regard for the level of planning involved. requires that all environmental features of immediate
Figure 1 shows the organization of the DAMN interest be visible to the vehicle sensors at all times; thus
architecture, in which individual behaviors such as road adding unnecessary constraints and reducing the flexibility
following or obstacle avoidance send votes to the command of the overall vehicle system
arbitration module; these inputs are combined and the The DAMN architecture is designed to provide the basic
resulting command is sent to the vehicle controller. Each capabilities essential to any mobile robot system, or first
action-producing module, or behavior, is responsible for a level of competence in the parlance of the Subsumption
particular aspect of vehicle control or for achieving some Architecture. In DAMN, this consists of safety behaviors
particular task; it operates asynchronously and in parallel which limit turn and speed to avoid vehicle tip-over or
with other behaviors, sending its outputs to the arbiter at wheel slippage, obstacle avoidance behaviors to prevent
whatever rate is appropriate for that particular function. Each collisions, as well as various auxiliary behaviors (see
behavior is assigned a weight reflecting its relative priority DAMN Behaviors section). As new functions are needed,
in controlling the vehicle. A mode manager may also be used additional behaviors can be added to the system without
to vary these weights during the course of a mission based on any need for modification to the previously included
knowledge of which behaviors would be most relevant and behaviors, thus preserving their established functionality.
reliable in a given situation. Since both deliberative and reflexive modules are
DAMN is a behavior-based architecture similar in some needed, DAMN is designed so that behaviors can issue
regards to reactive systems such as the Subsumption votes at any rate; for example, one behavior may operate
Architecture (Brooks, 1986). In contrast to more traditional reflexively at 10 Hz, another may maintain some local
centralized AI planners that build a centralized world model information and operate at 1 Hz, while yet another module
and plan an optimal path through it, a behavior-based may plan optimal paths in a global map and issue votes at
architecture consists of specialized task-achieving modules a rate of 0.1 Hz. The use of distributed shared control
that operate independently and are responsible for only a allows multiple levels of planning to be used in decision-
very narrow portion of vehicle control, thus avoiding the making without the need for an hierarchical structure.
need for sensor fusion. A distributed architecture has several However, higher-level reasoning modules may still exert
advantages over a centralized one, including greater meta-level control within DAMN by modifying the voting
reactivity, flexibility, and robustness (Payton, Rosenblatt weights assigned to behaviors and thus controlling the
and Keirsey, 1990). However, one important distinction degree to which each behavior may influence the systems
between this system and purely reactive systems is that, decision-making process and thus the robots actions.

MODE weights DAMN commands VEHICLE

MANAGER ARBITER CONTROLLER
votes

AVOID SEEK AVOID

OBSTACLES GOAL TIP-OVER

FOLLOW MAINTAIN
ROAD HEADING

Figure 1: Overall structure of DAMN.

 Command Fusion changed dynamically (Payton, 1986). In (Maes, 1991),
behavior selection is achieved via activation and inhibition
Centralized, blackboard, and hierarchical architectures links in what is effectively a winner-take-all network.
perform sensor fusion in order to construct a coherent world In these architectures, the satisfaction of multiple, possibly
model which is then used for planning actions. This conflicting goals is achieved by having one behaviors
approach has the advantage of being able to combine data commands completely override anothers. While this is an
from previous world models to overcome ambiguities and effective scheme for choosing among incompatible
noise inherent in the sensing process, but has the commands, it does not provide an adequate means for
disadvantage of creating a computationally expensive dealing with multiple goals that can and should be satisfied
sensory bottleneck - all sensor data must be collected and simultaneously. This inability to compromise stems from the
integrated before it can be acted upon. Furthermore, fact that priority-based arbitration of commands masks all of
information from disparate sources such as maps, sonar, and the knowledge within each individual behavior that was used
video, are generally not amenable to combination within a to reach its decision. In the process of selecting a command,
single representational framework that is suitable for behaviors may weigh several alternatives before making its
planning such dissimilar tasks as following roads and ultimate choice. In other behaviors, it may be more natural to
avoiding obstacles, and a single monolithic world model is merely rule out certain inappropriate actions rather than
also more difficult to develop, maintain, and extend. selecting a desirable one, yet their only means of expression
In contrast, behavior-based architectures do not need to is to completely suppress or inhibit the outputs of other
create a central world model; instead, the perceptual behaviors. A compromise between behaviors cannot be
processing is distributed across multiple independent achieved in such an all-or-nothing scenario; whenever one
modules. Each behavior requires only fragmentary behaviors output is inhibited by another, the information
knowledge of the world and receives exclusively that and knowledge represented by that behavior is completely
sensory data which is directly relevant to its particular lost to the system (Rosenblatt and Payton, 1989).
decision-making needs, so there is no need to fuse all
available data into a single coherent world model. Based on
this input, each behavior then produces a desired action, and DAMN Arbiters
the system must use these independently produced outputs to
In order to allow multiple considerations to affect vehicle
control the robot.
actions concurrently, DAMN uses a scheme where each
By appropriately fusing behavior commands through
behavior votes for or against each of a set of possible vehicle
arbitration, a robot control system can respond to its
actions (Rosenblatt and Payton, 1989). An arbiter then
environment without suffering the problems inherent in
performs command fusion to select the most appropriate
sensor fusion. Instead of performing sensor fusion, the
action. Although all votes must pass through the command
system must combine command inputs to determine an
arbiter before an action is taken, the function provided by the
appropriate course of action.
arbiter is fairly simple and does not represent the centralized
bottleneck of more traditional systems. While the Motor
Priority-Based Command Selection
Schema framework (Arkin, 1987) also offers a means of
In a distributed architecture, it is necessary to decide which fusing commands from multiple behaviors, it suffers from
behaviors should be controlling the vehicle at any given the well known problem of local minima in potential fields.
time. In some architectures, this is achieved by having Another, perhaps more serious problem, is that arbitration
priorities assigned to each behavior; of all the behaviors via vector addition can result in a command which is not
issuing commands, the one with the highest priority is in satisfactory to any of the contributing behaviors. DAMN
control and the rest are ignored. In the Subsumption arbiters do not average commands, but rather select the
Architecture (Brooks, 1986), this prioritization is implicit in command which has the most votes from the behaviors.
the wiring of behavior modules; a higher-level behavior can
override the output of a lower-level behavior via an Turn Arbiter
inhibition link. In the Gapps architecture (Rosenschein and
In the case of the turn arbiter, each behavior generates a vote
Kaelbling, 1986), such priorities are compiled into a
between -1 and +1 for every possible steering command,
mediator which combines inputs from the behaviors through
with negative votes being against and positive votes for a
a combinatorial logic circuit which yields the system output;
particular command option. The arbiter collects the new
as in the Subsumption Architecture, behaviors consist of
votes from each behavior that has sent them, and performs a
finite state automata. The lowest level of the hierarchical
normalized weighted sum to find the turn command with the
architecture developed at the Hughes Research Labs also
maximum vote value. In order to avoid problems with
consists of a similar distributed control mechanism; a
discretization such as biasing and bang-bang control (i.e.,
centralized blackboard is used for arbitration and
alternating between discrete values in order to achieve an
communications purposes; priorities are explicit and can be
intermediate value), the arbiter performs sub-pixel
interpolation. This is done by first convolving the votes with from the coupling of the importance that the vehicle head in
a Gaussian mask to smooth the values and then selecting the a certain direction to the speed that the vehicle travels in that
command option with the highest resulting value. A parabola direction; these are independent in general.
is then fit to that value and the ones on either side, and the An entirely separate DAMN speed arbiter with its own set
peak of the parabola is used as the command to be issued to of associated behaviors has now been developed. Thus, the
the controller. This is similar to defuzzification in Fuzzy turn behaviors can vote for turn commands without concern
Logic systems (Zadeh, 1973; Lee, 1990; Kamada, Naoi and that the absolute magnitude of their votes will affect vehicle
Goto, 1990); indeed an architecture has been implemented speed. At present each speed behavior votes for the largest
which recasts DAMN into a Fuzzy Logic framework (Yen speed possible which meets that behaviors constraints, and
and Pfluger, 1992). the arbiter simply chooses the minimum of those maxima, so
The arbitration process used in DAMN is illustrated in that all speed constraints are satisfied.
Figure 2, where: (a and b) the votes from behaviors are
received, (c) a weighted sum of those votes is computed, nd Field of Regard Arbiter
(d), the summed votes are smoothed and interpolated to
A field of regard arbiter and its associated behaviors have
produce the resulting command sent to the vehicle
also been recently implemented and used for the control of a
controller.
pair of stereo cameras on a pan/tilt platform. Behaviors vote
for different possible field of regard polygons (the camera
a) Behavior 1, weight = 0.8, desired curvature = 0.04 field of view mapped on to the ground plane); the arbiter
maintains a local map of these votes and transforms them as
the vehicle moves. At each iteration, these votes are mapped
into a pan-tilt space, and then smoothed and interpolated as
-0.125 0 +0.125 described above. Figure 3 shows the votes for field of regard
polygons generated by an behavior which votes against
b) Behavior 2, weight = 0.2, desired curvature = 0.0 looking in the direction of known obstacles since travelling
in that direction is impossible. Behaviors also vote based on
considerations such as looking toward the goal and looking
at a region contiguous to already mapped areas. The darkest
-0.125 0 +0.125 polygon in the figure corresponds to the pan and tilt angles
selected by the arbiter.
c) Weighted Sum, max vote curvature = 0.035

-0.125 0 +0.125

d) Smoothed and Interpolated, peak curvature=0.033

Figure 3: Field of regard voting and arbitration

-0.125 0 +0.125 DAMN Behaviors

Figure 2: Command fusion process Within the framework of DAMN, behaviors must be
defined to provide the task-specific knowledge for
controlling the vehicle. Each behavior runs completely
Speed Arbiter independently and asynchronously, providing votes to the
The emphasis in the research thus far has been in command arbiter each at its own rate and according to its own time
fusion for the control of vehicle steering; until recently the constraints. The arbiter periodically sums all the latest votes
commanded speed was decided in a very simplistic fashion from each behavior and issues commands to the vehicle
based upon the commanded turn radius. The user-specified controller.
maximum vehicle speed was multiplied by the normalized
weighted sum for the chosen turn radius; the result was the Safety Behaviors
speed command issued. This is similar to the method used in A basic need for any mobile robot system is the ability to
Motor Schemas where the magnitude of the resultant vector avoid situations hazardous to itself or to other objects in its
determines the speed of the vehicle. Both methods suffer environment. Therefore, an important part of DAMN is its
first level of competence (Brooks, 1986), which consists where is the ratio of the distance between the vehicles
of behaviors designed to provide vehicle safety. In contrast center of gravity (c.g.) and the wheels to the c.g. height, g is
to priority-based architectures which only allow one the acceleration due to gravity, and is the vehicle roll with
behavior to be effective at any given moment, the structure respect to the gravity vector. Likewise, for a given vehicle
of DAMN and its arbitration scheme allow the function of turn curvature, the maximum velocity is:
max = Min ( g cos ) + g sin -
these safety behaviors to be preserved as additional levels of 12
competence are added. -------------------------------------------------------------

Similar constraints can be imposed on vehicle turn radius
Obstacle Avoidance The most important behavior in the
and speed in order to avoid tire slippage. The limit on
context of vehicle safety is the Obstacle Avoidance behavior.
curvature for slippage is:
In order to decide in which directions the vehicle may safely
( g cos ) ( g sin )
travel, this behavior receives a list of current obstacles in max = -------------------------------------------------------------
-
vehicle-centered coordinates and evaluates each of the 2
possible command options, as illustrated in Figure 4. The where is the dynamic coefficient of friction between the
source of these obstacles may be intraversable regions of tire and the terrain, and for speed:
max = (-------------------------------------------------------------
g cos ) ( g sin )-
terrain determined by range image processing or stereo 12
vision, by sonar detection of objects above the ground plane,

or any other means of obstacle detection as appropriate to the
Two behaviors, Limit Turn and Limit Speed send votes to
current task and environment (Daily et al, 1986; Langer,
the arbiter that implement these constraints, voting against
Rosenblatt and Hebert, 1994).
commands that violate them.
STRAIGHT
SOFT AHEAD SOFT Road Following
LEFT RIGHT
Once vehicle safety has been assured by the obstacle
HARD HARD avoidance and dynamic constraint behaviors, it is desirable
LEFT RIGHT to add additional behaviors that provide the system with the
ability to achieve the tasks for which it is intended, such as
road following; one of the behaviors that have been
implemented within DAMN to provide this function is
Figure 4: Arc evaluation in Obstacle Avoidance behavior ALVINN.
The ALVINN road following system is an artificial neural
If a trajectory is completely free of any obstacles, such as network that is trained, using backpropagation, to associate
the Straight Ahead or Hard Right turns shown in Figure 4, preprocessed low resolution input images with the
then the obstacle avoidance behavior votes for travelling appropriate output steering commands (Pomerleau, 1992). In
along that arc. If an obstacle lies in the path of a trajectory, the case of ALVINN, creating a behavior that independently
the behavior votes against that arc, with the magnitude of the evaluated each arc was relatively straightforward. The units
penalty proportional to the distance from the obstacle. Thus, of the neural networks output layer each represent an
the Obstacle Avoidance behavior votes more strongly evaluation of a particular turn command, with the layer
against those turns that would result in an immediate impact trained to produce Gaussian curves centered about those
(Hard Left in the figure) and votes less strongly against those turns that would follow the road ahead. These units are
turns which would only result in a collision after travelling simply resampled to the DAMN voting command space,
several meters (Soft Right). In order to avoid bringing the using a Gaussian of the appropriate width. This process is
vehicle unnecessarily close to an obstacle, the behavior also illustrated in Figure 5.
votes against those arcs that result in a near miss (Soft Left),
although the evaluation is not as unfavorable as for those DAMN TURN COMMANDS
trajectories leading to a direct collision.

Vehicle Dynamics Another vital aspect of vehicle safety is

insuring that the commanded speed and turn stay within the
dynamic constraints of the vehicle as it travels over varying
OUTPUT LAYER
terrain conditions. The most important of these constraints is
the one that insures that the vehicle will not tip over. Given a HIDDEN UNITS
velocity of magnitude , the maximum positive and negative
curvatures to avoid tip-over would be: INPUT LAYER
ALVINN
( g cos ) + g sin
max = -------------------------------------------------------------
- Figure 5: Resampling of ALVINN output layer
2
Goal-Directed Behaviors The Internalized Plans (Payton, 1990) approach uses a
detailed map to perform an A* search from the goal(s) back
Another important level of functionality that should be toward the start point to create a Gradient Field towards
present in any general purpose robotic system is the ability to the goal. The type and slope of the terrain, among other
reach certain destinations using whatever global information factors, is used to estimate the cost of traversal between grid
is available. While the low-level behaviors operate at a high cells. During run-time, the grid cell containing the current
rate to ensure safety and to provide functions such as road vehicle location is identified, and the Gradient Field pointers
following and cross-country navigation, high-level are followed forward to the point G in Figure 6; the desired
behaviors are free to process map-based or symbolic heading to reach the goal is that from the current location S
information at a slower rate, and periodically issue votes to to G, and a series of votes with its peak at that value is sent
the arbiter that guide the robot towards the current goal. to the turn arbiter.
The D* planner (Stentz, 1993) also creates a grid with
Subgoals The Goal Seeking behavior is one way to provide backpointers that represent information on how best to
this capability. This simple behavior directs the vehicle reach the goal from any location in the map. The map may
toward a series of goal points specified in global coordinates
initially contain no information, but is created incrementally
either by the user (Langer, Rosenblatt and Hebert, 1994) or
as new information becomes available during the execution
by a map-based planner (Keirsey, Payton and Rosenblatt,
of a mission, and the arc traversal costs and backpointers are
1988). The desired turn radius is transformed into a series of
updated to reflect this new knowledge. The resulting global
votes by applying a Gaussian whose peak is at the desired
plan is integrated into DAMN as a behavior by determining,
turn radius and which tapers off as the difference between
for each possible turn command, the weight w of reaching
this turn radius and a prospective turn command increases. A
the goal from a point along that arc a fixed distance ahead
goal is considered satisfied once the vehicle enters a circle
(the squares designated collectively as S in Figure 7). If
centered at the goal location; then the next goal is pursued.
wmax and wmin are the maximum and minimum values of w,
Because of errors in goal placement and accumulated errors
then the vote for each turn command is determined as:
in vehicle positioning, a goal point may not be reachable. For
( w max w )/ ( w max w min ) . In the case that a point S
this reason, an ellipse is defined with the current goal and the
is not represented on the grid, or if the goal cannot be
subsequent goal as foci; if the vehicle enters this ellipse, the
reached from it, then the vote for that arc is set to -1.
current goal is abandoned and the next one becomes the
current goal instead, thus allowing progress to continue. G
Dynamic Programming Some more sophisticated map-
based planning techniques have also been integrated and
used within the DAMN framework. These planners use {S}
dynamic programming techniques based on the A* search
algorithm (Nilsson, 1980) to determine an optimal global
path. However, an important point is that they do not hand a S
plan down to a lower level planner for execution, but rather Figure 7: D* behavior estimates goal distance for each arc
maintain an internal representation that allows them to
participate directly in the control of the vehicle based on its Teleoperation
current state. A* yields a set of pointers within the map grid
that point toward the goal, as depicted by the small arrows in Teleoperation is another possible mode in which a robotic
Figure 6. During execution, this grid may be indexed by the system may need to operate. The STRIPE teleoperation
current vehicle position to yield a path towards the goal system (Kay and Thorpe, 1993) provides a graphical user
which is optimal based on the information available in the interface allowing a human operator to designate waypoints
map at that time. for the vehicle by selecting points on a video image and
projecting them on to the surface on which the vehicle is
G travelling. STRIPE then fits a spline to these points and uses
pure pursuit to track the path. When used in isolation, it
simply sends a steering command to the controller; when
S G used as a DAMN behavior, it sends a set of votes
representing a Gaussian centered on the desired command.
This allows the dynamic constraints and obstacle avoidance
behaviors to be used in conjunction with STRIPE so that the
Figure 6: Following Gradient Field to determine safety of the vehicle is still assured.
intermediate goal heading
Auxiliary Behaviors soft right, and allowing the other turns as acceptable, with
soft left being the most favorable.
Various other auxiliary behaviors that do not achieve a
particular task but issue votes for secondary considerations HARD SOFT STRAIGHT SOFT HARD
may also be run. These include the Drive Straight behavior, LEFT LEFT AHEAD RIGHT RIGHT
which simply favors going in whatever direction the vehicle
is already heading at any given instant, in order to avoid
sudden and unnecessary turns; and the Maintain Turn
behavior, which votes against turning in directions opposite
to the currently commanded turn, and which helps to avoid
unnecessary oscillations in steering, the Follow Heading
behavior which tries to keep the vehicle pointed in a constant
OBSTACLE GOAL
direction, as well as various behaviors which allow user AVOIDANCE SEEKING
input to affect the choice of vehicle turn and speed BEHAVIOR BEHAVIOR
commands. Figure 8: Command fusion in DAMN

Because avoiding obstacles is more important than taking

Combining Behavior Votes
the shortest path to the goal, the obstacle avoidance behavior
The voting strengths, or weights, of each behavior are is assigned a higher weight than the goal seeking behavior,
specified by the user, and are then normalized by the arbiter as indicated by the thicker arrows in the diagram. The arbiter
so that their sum equals 1. Because only the relative values then computes a weighted sum of the votes it has received
are important, and because the magnitude of each behaviors from each behavior, and the command choice with the
votes vary according to their importance, DAMN is fairly highest value is selected and issued to the vehicle controller.
insensitive to the values of these weights and the system In this case a soft left turn would be executed, since its
performs well without a need to tweak these parameters. For weighted sum is the greatest, thus avoiding any obstacles
example, the Obstacle Avoidance behavior has been run in while still more or less moving toward the goal. The
conjunction with the Seek Goal behaviors with relative favorableness of the selected turn command may also used to
weights of 0.75 and 0.25, respectively, and with weights of determine vehicle speed, so that, for example, the vehicle
0.9 and 0.1, and in both cases has successfully reached goals would slow down if a command is issued which will take the
while avoiding obstacles. vehicle too far from the path to the goal point. Another
The vote weights of each behavior can also be modified by possibility is to have a separate speed arbiter that would
messages sent to the arbiter from a mode manager module. It receive commands from behaviors that, given the current
can reconfigure the weights according to whatever top-down vehicle radius, determine the maximum speed that would
planning considerations it may have, and potentially could satisfy their objectives.
use bottom-up information about the effectiveness and
relevance of a behavior (Payton et al., 1993). Different
modes of operation that exclude some behaviors can be Experimental Results
constructed by setting the weights those behaviors to 0. A For the experiment described here, the system was
Mode Manager was developed at the Hughes Research Labs configured with two behaviors, Obstacle Avoidance and
to be used with DAMN for this purpose, and at CMU Goal Seeking, which were described in the DAMN
Annotated Maps were integrated with DAMN to provide this Behaviors section. The Turn Arbiter combined the votes
capability (Thorpe et al, 1991). from these two behaviors and issued a new steering
As a simple example to illustrate the manner in which command every 100ms. The goal points were on average
votes are issued and arbitrated within DAMN, consider the 100 meters apart and the Goal Seeking behavior switched
case in Figure 8 where two behaviors are active, one goals whenever it came within eight meters of the current
responsible for obstacle avoidance and the other for goal target goal point. The weights were 0.8 and 0.2 for the
seeking (only five turn options are shown for simplicity). Obstacle Avoidance and Goal Seeking behaviors,
The magnitude of a vote is indicated by the size of a circle, respectively. The obstacle weight was large relative to the
with a large unfilled circle representing a vote of +1, a large goal weight to reflect the fact that avoiding obstacles is
striped circle a value of -1, and a small circle a value near 0. generally considered to be more important than approaching
Thus, the goal-seeking behavior is voting most strongly in the goal at any given moment. The maximum vehicle speed
favor proceeding straight and less favorably for a soft left was 2 meters/second in this experiment. A more detailed
turn, and voting against hard left or any right turns; the description of the experimental system can be found in
obstacle avoidance behavior is voting against a hard left or Langer, Rosenblatt and Hebert (1994).
 0 15 30
enhanced
arbiter
arbiter
obstacle
avoidance
goal
seeking
turn radius
command

E
vehicle path D
A BC obstacles

Figure 9: Vehicle path and corresponding votes over time, and vote distributions at five specified locations.
 Figure 9 illustrates the way the DAMN arbiter combined votes are still at a maximum for a left turn because of the
outputs from the behaviors. The first section of the figure greater weight of the obstacle avoidance behaviors votes,
shows the path of the vehicle as it travels around a set of and so the arbiter continues to steer the vehicle well clear of
obstacle points, and the distribution of votes for each of the the obstacles.
behaviors and for the arbiter along that path. The horizontal By the time the vehicle has reached point D, it is just
axis is time, with the total length of the sequence being 30 passing the obstacles, so that the obstacle avoidance
seconds. For each module, the vertical axis is the set of turn behavior is now only disallowing extreme right or left turns
choices from a hard left to a hard right turn. The votes were because of field of view constraints. The goal behavior is
sampled every 200ms for display purposes. The votes are now able to have more influence over the chosen turn
displayed as shaded pixels, with the brightest pixels direction, and a right turn is now executed so that the desired
corresponding to the highest possible vote of 1.0 and the vehicle heading is restored. Note that between points C and
black pixels corresponding to the lowest possible vote of - D, there is a transition period where the turn radius chosen
1.0. In addition to the raw votes from the arbiter, the figure by the arbiter is neither the hard left favored by the obstacle
shows an enhanced display of the arbiters distribution of avoidance behavior nor the medium right favored by the goal
votes, obtained by normalizing the pixel values using the behavior. Instead, as turns other than hard left become
actual minimum and maximum values at each time slot acceptable to the obstacle avoidance behavior, and as the
instead of the extreme possible values. The enhanced arbiter distribution of votes from the goal behavior shift further to
display is included to facilitate the visual interpretation of the right, the vehicle is commanded to make ever softer left
the vote distribution only and was not used for driving the turns, and eventually turns to the right. As can be seen in the
vehicle. In addition to the vote distribution, Figure 9 also graph of the commanded turn radius in Figure 9, rather than
includes a graph of the commanded turn radius over time; abruptly switching modes from avoiding obstacles to
the commanded turning radius is derived from the maximum following a heading, the transition proceeds smoothly and
of the vote distribution of the arbiter, as described in the steadily as the situation gradually changes.
Turn Arbiter section. Finally, at point E, the vehicle has completely cleared the
Five points of interest along the vehicle path, labeled as A obstacles and they are no longer in the local map, so that the
through E, are shown in greater detail by the five graphs that votes from the obstacle avoidance behavior are mostly +1.
follow in Figure 9. Each graph depicts the votes issued for The goal behavior now dominates completely and a right
each turn choice by the obstacle avoidance and goal turn is commanded. Through the interaction of the avoid
behaviors, as well as the weighted sum of these votes as obstacles and goal behaviors, the vehicle was thus able to
computed by the arbiter. The horizontal axis of each graph successfully circumvent an obstacle and regain its desired
shows the possible turn radius choices encoded from -8 heading. This simple experiment shows that the arbiter
meters for hard left to +8 meters for hard right (curvature and combines the preferred commands from different behaviors
turn radius are multiplicative inverses). The sequence of five in an efficient and natural way.
points selected is typical of a path around a single obstacle.
At point A, the obstacle is first reported and the obstacle
avoidance behavior generates high votes for turning left to Future Work
go around the obstacle, and inhibits right turns with negative DAMN has proven to be very effective as an architecture
votes, as shown by the solid line in the graph. At the same which greatly facilitates the integration of a wide variety of
time, the goal behaviors vote distribution is relatively flat different vehicle navigation subsystems; however, as
around the straight direction since the vehicle is currently DAMN is used for increasingly complex tasks where many
headed in the desired direction; this is shown by the dashed behaviors may be issuing votes concurrently, there will be a
line. Because of the small relative weight of the goal greater need to have the semantics of the voting process
behavior, the combined vote distribution in the arbiter, defined more carefully. By explicitly representing and
shown as a thicker solid line, is dominated by the votes reasoning about uncertainty within the decision-making
received from the obstacle avoidance behavior; a left turn is processes, a system can be created whose effects are well-
therefore commanded. At point B, the obstacle is still close defined and well-behaved.
to the vehicle and the votes distributions are similar to the We have decided to pursue utility theory as a means of
ones at A, thus maintaining the vehicle to the left of the defining the semantics of the voting and arbitration
obstacle. processes. Because we are attempting to decide which
At point C, the obstacle avoidance behavior is still voting among a set of possible actions to take, it is natural to make
in favor of a sharp left turn, but the votes for the softer left judgments on the usefulness of each action based on its
turns is now not as low as it was at A or B, since the vehicle consequences. If we assign a utility measure U(c) for each
is now clearing the obstacles. At the same time, the goal possible consequence of an action, then the expected utility
behavior is starting to shift its votes towards turning right in for an action a is:
order to bring the vehicle back to the target. The summed
 U(a) = U(c) P(c a, e) levels of competence have been implemented in DAMN thus
far, as shown in Figure 10. These levels of competence are
c
where P(c|a,e) is the probability distribution of the convenient for describing the incremental manner in which
consequence configuration c, conditioned upon selecting the systems capabilities evolve; however, it is important to
action a and observing evidence e (Pearl, 1988). Thus, if we note that all behaviors co-exist at the same level of planning.
can define these utilities and probabilities, we can then apply The importance of a behaviors decisions is reflected by the
the Maximum Expected Utility (MEU) criterion to select the weighting factor for its votes, and is in no way affected by
optimal action based on our current information. the level of competence in which it is described.
If utility theory is to be applied to the problem of evidence
combination and action selection for mobile navigation
tasks, a means for defining the utility functions must be GOALS
provided. Behaviors are defined in order to achieve some
task, so it is fair to assume that there must be at least an Subgoals
implicit measure of goodness or utility with respect to that Gradient Fields
task. For example, an obstacle avoidance behaviors task is ACTION D*
to maximize distance to obstacles, so that the distance from
the vehicle to the nearest obstacle could be used as a measure Road Following
of goodness. Likewise, proximity to the current goal could Cross-Country
be used for the goal-based behaviors, as proximity to the SAFETY Teleoperation
center of a road (or lane) could be used by road following
modules. Vehicle Dynamics
We will also be investigating command fusion using a Obstacle Avoidance
local map for turn and speed arbitration, as was done for the
field of regard arbiter. This will allow the arbiter to ensure Auxiliary
consistency in the voting process by transforming votes as
the vehicle moves. This will also allow behaviors to vote on Figure 10: Evolutionary system development in DAMN
where to travel rather than how to steer; however,
The safety behaviors are used as a first level of
complications due to the non-holonomic constraints of the
competence upon which other levels can be added. The
vehicle will need to be addressed.
second level of competence, Action, has been implemented
Another issue that needs to be addressed is the coupling
using road following, cross-country navigation, and
between turn and speed commands. Currently, the turn and
teleoperation behaviors; they have all been run together with
speed arbiters operate independently. Turn behaviors can use
the obstacle avoidance behavior to provide various forms of
vehicle speed as one of their inputs and vice versa; however,
generating purposeful action while maintaining safety
there is currently no mechanism which allows behaviors to
(Thorpe et al, 1991). The third level of competence is
vote for a combination of turn and speed. Research is
comprised of the various map-based goal-seeking behaviors.
ongoing to determine a command representation which
Cross-country behaviors have been combined with goal-
allows both coupled and independent voting, and is still
oriented behaviors to produce directed off-road navigation
efficient enough for real-time purposes.
(Keirsey, Payton and Rosenblatt, 1988; Stentz, 1993).
The scenario described in the Experimental Results
Conclusion section shows that a reactive architecture is capable of
navigating cross-country terrain while pursuing mission
DAMN is designed so that various behaviors can be easily objectives such as goal locations. The reactive nature of the
added or removed from the system, depending on the current system is possible because the perception module simply
task at hand. Although the modules described in the DAMN produces lists of binary obstacles as needed by the
Behaviors section all use very different paradigms and navigation system. If the output of the perception system
representations, it has been relatively straightforward to were instead a detailed geometric model of the terrain, then
integrate each and every one of them into the framework of planning would require significantly more computation, thus
DAMN. Sensor fusion is not necessary since the command precluding the use of a reactive approach.
fusion process in DAMN preserves the information that is The current design of DAMN, with the arbiter taking
critical to decision-making, yielding the capability to inputs from a set of specialized behaviors, is flexible enough
concurrently satisfy multiple objectives without the need for to allow for a variety of other behaviors to be combined or
centralized bottlenecks. All of the behaviors described have substituted. In addition to its use on the CMU Navlab
been used in conjunction with each other in various vehicles, DAMN has also been used at Martin Marietta (now
configurations, yielding systems that were more capable Lockheed Martin), Hughes Research Labs, and Georgia
than they would have been otherwise. Conceptually, three Institute of Technology. Like other behavior-based
architectures, it avoids sensory and decision-making not be reconciled, then one of the two must be chosen. Meta-
bottlenecks and is therefore able to respond in real-time to level control may be exerted so that behaviors with mutually
external events; however, it imposes no constraints on the exclusive goals do not operate simultaneously; however,
nature of the information or the processing within a plans are not used in a top-down fashion but rather as a
behavior, only on the interface between the behavior and the source of advice, so that the flexibility of the reactive level is
command arbiter. Furthermore, the behaviors are not subject preserved.
to any timing constraints; each behavior operates The hypothesis proposed here is that centralized
asynchronously and in parallel. arbitration of votes from distributed, independent decision-
Non-reactive behaviors may use plans to achieve goals making processes provides coherent, rational, goal-directed
and coordinate with other agents; thus, like centralized or behavior while preserving real-time responsiveness to its
hierarchical architectures, DAMN is able to assimilate and immediate physical environment.
use high-level information. Finally, unlike architectures with
2. Interfaces-- How can human expertise be easily
prioritized modules, DAMNs vote arbitration scheme
brought into an agents decisions? Will the agent need to
allows multiple goals and constraints to be fulfilled
translate natural language internally before it can interact
simultaneously, thus providing goal-oriented satisficing
with the world?
behavior without sacrificing real-time reactiveness.
Natural language is not needed for human-computer
interaction, and in many domains would actually be more
Acknowledgments difficult for an operator to use than other modes of
The UGV project is supported by ARPA under contracts interaction. For example geometric information or mission
DACA76-89-C-0014 and DAAE07-90-C-R059, and by the goals can be more easily specified via a Graphical User
National Science Foundation under NSF Contract BCS- Interface (GUI), and direct control of the robot can be better
9120655. Julio Rosenblatt is supported by a Hughes effected through the use of a joystick, for example.
Research Fellowship. The author would like to acknowledge In DAMN, the user may specify which behaviors are to be
the cooperation in vehicle experimentation of Martial Hebert active and what their relative weights should be, either a
and Dirk Langer, the support and guidance of Dave Payton priori or during run-time via a GUI, thus providing meta-
and Chuck Thorpe, as well as the shared efforts and level control. Behaviors that support teleoperation have also
technical support of the Navlab group at Carnegie Mellon been implemented, so that the user may specify the robot
University. path via waypoints in a video image, by using a joystick, or
by typing simple keystroke commands. As with all other
behaviors, the users input is expressed as votes which are
Appendix 1: Responses to issues raised then combined with the votes of other behaviors and
1. Coordination-- How should the agent arbitrate/ arbitrated.
coordinate/cooperate its behaviors and actions? Is there a 3. Representation-- How much internal representation of
need for central behavior coordination? knowledge and skills is needed? How should the agent
Centralized architectures provide the ability to coordinate, in organize and represent its internal knowledge and skills? Is
a coherent fashion, multiple goals and constraints within a more than one representational formalism needed?
complex environment, while decentralized architectures The DAMN system only requires votes from each behavior,
offer the advantages of reactivity, flexibility, and robustness. so that each module is free to use whatever representation
The DAMN architecture takes the position that some and paradigm best serves for its particular task. However, the
centralization is needed, but the right level must be chosen selection of which behaviors should be active at any given
so that it does not create a bottleneck, and the interfaces must time is currently done in a fairly rudimentary way; an
be defined so as to avoid being overly restrictive. explicit representation of each behaviors skills would allow
Rather than imposing an hierarchical structure or using a for more dynamic and flexible responses to unforeseen
prioritized behavior-based systems to effect a traded control situations, and may also facilitate learning.
system, the Distributed Architecture for Mobile Navigation
takes a shared control approach where several modules 4. Structural-- How should the computational capabilities
concurrently have some responsibility for control of the of an agent be divided, structured, and interconnected? How
robot. In order to achieve this, a common interface is much does each level/component of an agent architecture
established so that modules can communicate their have to know about the other levels/components?
intentions without regard for the level of planning involved. The capabilities of an agent should be divided up as finely as
Votes from all behaviors are used in determining what the is practical among task-achieving behaviors which operate
next action should be, so that compromises are made when asynchronously. They should be completely modular and
possible; however, if two behaviors suggest actions that can independent, so that new capabilities may be added in an
evolutionary fashion without a need to disrupt or modify for other behaviors such as goal-seeking.
existing functionality. Another possibility for a normalized utility measure is
Ideally, higher level components that exert meta-level analogous to the one used by the D* behavior described in
control should only need to know which skills are provided the Goal-Directed Behaviors section. After computing gb(c)
by the lower levels, without any knowledge of how those for each possible consequence c, we determine the
skills are implemented whatsoever. maximum and minimum values gmax and gmin, and use the
following measure:
5. Performance-- What types of performance goals and
metrics can realistically be used for agents operating in U b(c) = ( g b(c) g min ) ( g max g min )
dynamic, uncertain, and even actively hostile environments?
While this measure has the desirable property of being
Objective quantitative evaluation of mobile robot systems
bounded by the interval [0,1], it has the potentially
has historically proven to be very difficult, and the task is
undesirable property that the range of utility values will
doubly difficult when attempting to compare entire
always completely span that interval. Thus, for example, if
architectures because there are so many degrees of freedom
all potential actions have exactly the same utility except for
in the design and implementation of a mobile robot system.
one that has a slightly higher value, then the normalized
The most straightforward means of evaluation is to attempt
utility values will be 1 for that one action and 0 for all the
to define measures of the quality of the path which a vehicle
rest.
has traversed under autonomous control. The measures I
suggest within the body of the paper are the paths utility,
Path Smoothness Another general measure of the quality of
smoothness, and accuracy:
the path is the smoothness with which it is controlled.
Integrating the squared derivative of curvature along the
Utility Measures One means of evaluating architectures is
vehicles actual path provides a measure of smoothness that
to use measures of goodness like those used to define the
can be applied to the architecture (Kamada, Naoi and Goto,
utility of each action for the various behaviors. Although
1990), and the derivative of acceleration, or jerk, provides a
evaluating the performance based on the utility measures
measure of smoothness in the vehicles speed as well as
defined in the Future Work section would appear to be
steering. Integrating the squared vehicle curvature along the
circular, they are the best measures available. Furthermore,
vehicles path may also be useful as a measure of
their evaluation is not vacuous given that uncertainty exists
smoothness, which reflects consistency in decision-making
in all aspects of a system: in the declarative knowledge,
and the ability to anticipate events.
sensed information, procedural knowledge, and in the effects
of its actions and the actions of other agents; thus the
Path Accuracy The accuracy of the path, i.e. the extent to
expected utility will in general not agree with the actual
which the commanded path matches the actual path taken by
utility that can be measured for those actions that are taken.
the vehicle, can also provide an important means of
Let the goodness value for behavior b at a state s be
evaluating an architecture and its planning subsystems. Path
computed by the function gb(s). Then one possible measure
accuracy can be measured by integrating the squared error
for the utility, with respect to behavior b, of consequence c is
between estimated and actual vehicle poses. If the system
the change in this measure of goodness with respect to the
commands trajectories which are not physically realizable,
current state s0:
the actual path taken may deviate significantly. Likewise, if
U b(c) = g b = g b(c) g b(s 0) large latencies exist in the system and are not adequately
compensated for, the commanded path will only begin to be
However, the value of this measure can be arbitrarily executed well after the system intended it.
large, so it must be normalized and assigned a consistent
6. Psychology-- Why should we build agents that mimic
semantic meaning for all behaviors. One means of
anthropomorphic functionalities? How far can/should we
normalization would be to divide by the larger of the two
draw metaphoric similarities to human/animal psychology?
goodness measures:
How much should memory organization depend on human/
U b(c) = g b Max(g b(c), g b(s 0)) animal psychology?
If the goal of the system is to produce an efficient means of
This would bound the absolute value of Ub(c) to be no
controlling artificially constructed agents, as is the case with
greater than 1. It would also have the effect that, for the same
DAMN, then existing systems, i.e. humans and other
gb, the utility of an action would be greater when the animals, should serve merely as an inspiration, never as a
goodness measure is small than when it is large. This
constraint. Furthermore, good software engineering practices
appears to be desirable for behaviors such as obstacle
dictates that robotic systems must evolve in a more orderly
avoidance whose importance should grow as the situation
fashion than their biological counterparts.
worsens; however, the suitability of this measure is less clear
7. Simulation-- What, if any, role can advanced There currently is no learning within the DAMN
simulation technology play in developing and verifying architecture. The most important set of parameters in
modules and/or systems? Can we have standard virtual DAMN is the behavior weights. While setting these weights
components/test environments that everybody trusts and can for a small number of behaviors has been simple, doing so
play a role in comparing systems to each other? How far can for larger systems with more complex interactions could be
development of modules profitably proceed before they problematic. Reinforcement learning would seem to be a
should be grounded in a working system? natural means of acquiring information regarding the utility
of various actions in different circumstances, and the
Simulation technology can be a very important tool for
relevance and usefulness of each behavior in particular
developing systems which are eventually realized in a
situations might also be learned through techniques such as
physical implementation, but the former must not replace the
genetic algorithms. The mode manager changes behavior
latter. In my experiences in developing navigation systems,
weights according to a precomputed schedule, but adaptive
simulation has often played a key role in their success.
techniques such as those in (Maes, 1991) are necessary for
Simulation provides a means to develop and extensively test
dealing with a wider variety of situations and for fully
the system with minimal resources and without risking
autonomous control.
physical damage to a robot or its surroundings. Subsequent
testing on a real robot then provides not only a means of
validating the system, but also a means of discovering the References
flaws in the simulation so that its fidelity may be improved
for future use. Albus, J., McCain, H. and Lumia, R. (1987) NASA/NBS
Any attempt to standardize the virtual components of a Standard Reference Model for Telerobot Control System
robotic system would be premature at this point, as there is Architecture (NASREM), NBS Tech. Note 1235,
still much debate as to what form the decomposition of the Gaithersburg, MD.
architecture should take. Also, different needs in terms of Arkin, R. (1987) Motor Schema Based Navigation for a
complexity, responsiveness, and optimality will require Mobile Robot: An Approach to Programming by Behavior.
different architectures which make different trade-offs. Proceedings of the International Conference on Robotics
Standardized test environments, however, are something and Automation, Atlanta, GA, 1993.
that is sorely needed so that architectures may be compared
to each other in a systematic way. Again, different Brooks, R. (1986) A Robust Layered Control System for a
environments with different demands in complexity, Mobile Robot. IEEE Journal of Robotics and Automation
responsiveness, and optimality will be needed to highlight vol. RA-2, no. 1, pp. 14-23, April 1986.
the relative strengths and weaknesses of each architecture in Brooks, R. (1993) Intelligence Without Reason. Proceedings
various domains. While a real robot provides the best of the International Joint Conference on Artificial
testbed, simulation can also play an important role because Intelligence, Chambery, France, August 1993.
of its accessibility, its low cost, and because it provides the
ability to record every aspect of a trial for analysis and for Daily, M., Harris, J., Keirsey, D., Olin, K., Payton, D., Reiser,
reproducing results. K., Rosenblatt, J., Tseng, D. and Wong, V. (1988)
Autonomous Cross-Country Navigation with the ALV.
8. Learning-- How can a given architecture support IEEE Conference on Robotics and Automation,
learning? How can knowledge and skills be moved between Philadelphia, PA, April, 1988; (also appears in DARPA
different layers of an agent architecture? Knowledge Based Planning Workshop, December, 1987
For the sake of generality, the higher levels of a robot pp 20-1 to 20-10).
architecture tend to use knowledge-based systems containing Kamada, H., Naoi, S. and Goto, T. (1990) A Compact
explicit representations; thus, given a sufficient domain Navigation System Using Image Processing and Fuzzy
model, these levels may be amenable to techniques such as Control. Proceedings of IEEE Southeastcon, New Orleans,
explanation-based learning. For the sake of efficiency, the April 1-4, 1990
lower levels of an architecture tend to use compiled routines
and implicit representations, and therefore the various forms Kay, J., Thorpe, C. (1993) STRIPE Supervised Telerobotics
of learning via gradient descent are more applicable. Using Incremental Polygonal Earth Geometry.
Knowledge learned at the higher levels may be compiled via Proceedings of Intelligent Autonomous Systems
mechanisms such as chunking, and eventually migrated to Conference, Pittsburgh, PA, February 15-18, 1993, IOS
the lower levels for repeated use. Press.
Keirsey, D.M., Payton, D.W. and Rosenblatt, J.K. (1988) Rosenblatt, J. (1995) DAMN: A Distributed Architecture for
Autonomous Navigation in Cross-Country Terrain. Mobile Navigation. Proceedings of AAAI Spring
Proceedings of Image Understanding Workshop, Symposium on Lessons Learned from Implemented
Cambridge, MA, April, 1988. Software Architectures for Physical Agents, H. Hexmoor
and D. Kortenkamp (Eds.), Menlo Park, CA: AAAI Press.
Langer, D., Rosenblatt, J. and. Hebert, M. (1994) A
Behavior-Based System For Off-Road Navigation. IEEE Rosenblatt, J. and Payton, D. (1989) A Fine-Grained
Journal of Robotics and Automation, vol. 10, no. 6, pp. Alternative to the Subsumption Architecture for Mobile
776-782, December 1994; (also appears as An Integrated Robot Control. Proceedings of the IEEE/INNS
System For Autonomous Off-Road Navigation. International Joint Conference on Neural Networks,
Proceedings of the IEEE International Conference on Washington DC, vol. 2, pp. 317-324, June 1989; (also
Robotics and Automation, San Diego, May 1994). appears in 1989 AAAI Spring Symposium Series).
Lee, C. (1990) Fuzzy Logic in Control Systems: Fuzzy Logic Rosenblatt, J. and Thorpe, C. (1995) Combining Multiple
Controller -- Parts I and II. IEEE Transactions on Systems, Goals in a Behavior-Based Architecture. Proceedings of
Man and Cybernetics, Vol 20 No 2, March/April 1990. International Conference on Intelligent Robots and
Systems (IROS), Pittsburgh, PA, August 7-9, 1995.
Maes, P. (1991) A Bottom-Up Mechanism for Action
Selection in an Artificial Creature. Adaptive Behavior: Rosenschein, S. and Kaelbling, L. (1986) The Synthesis of
From Animals to Animats, edited by S. Wilson and J-A Digital Machines with Provable Epistemic Properties.
Meijer, MIT Press. Proceedings of Theoretical Aspects of Reasoning about
Knowledge. pp 83-98.
Nilsson, N. (1980) Principles of Artificial Intelligence. Tioga
Pub. Co., Palo Alto, Calif. Simmons, R., Lin, L.J., Fedor, C. (1990) Autonomous Task
Control for Mobile Robots. Proceedings of IEEE
Nilsson, N. (1984) Shakey the Robot, SRI Tech. Note 323,
Symposium on Intelligent Control, Philadelphia, PA,
Menlo Park, Calif.
September 1990.
Payton, D. (1986) An Architecture for Reflexive
Stentz, A. (1993) Optimal and Efficient Path Planning for
Autonomous Vehicle Control. IEEE International
Unknown and Dynamic Environments. Carnegie-Mellon
Conference on Robotics and Automation, San Francisco,
Technical Report CMU-RI-TR-93-20, Pittsburgh, PA.
CA, April 7-10, 1986, pp. 1838-1845.
Stentz, A. and Hebert, M. (1994) A Complete Navigation
Payton, D. (1990) Internalized Plans: A Representation for
System for Goal Acquisition in Unknown Environments.
Action Resources, Robotics and Autonomous Systems,
Carnegie-Mellon Technical Report CMU-RI-TR-94-7,
6(1), 1990, pp. 89-103. (Also in Designing Autonomous
Pittsburgh, PA.
Agents, ed. Pattie Maes, MIT Press, Cambridge, Mass.
1991, pp. 89-103.) Thorpe, C., Amidi, O., Gowdy, J., Hebert, M. and Pomerleau,
D. (1991) Integrating Position Measurement and Image
Payton, D., Rosenblatt, J. and Keirsey, D. (1990) Plan Guided
Understanding for Autonomous Vehicle Navigation.
Reaction. IEEE Transactions on Systems Man and
Proceedings of Workshop on High Precision Navigation,
Cybernetics, 20(6), pp. 1370-1382.
Springer-Verlag Publisher.
Payton, D., Keirsey, D., Kimble, D., Krozel, J. and
Yen, J. and Pfluger, N. (1992) A Fuzzy Logic Based Robot
Rosenblatt, J. (1993) Do Whatever Works: A Robust
Navigation System. Proceedings of AAAI Fall Symposium
Approach to Fault-Tolerant Autonomous Control. Journal
on Intelligent Applications of Artificial Intelligence to
of Applied Intelligence, Volume 3, pp 226-250.
Real-World Autonomous Mobile Robots, Boston, MA,
Pearl, J. (1988) Probabilistic Reasoning in Intelligent 1992.
Systems: Networks of Plausible Inference. Morgan
Zadeh, L. (1973) Outline of a New Approach to the Analysis
Kaufmann Publishers.
of Complex Systems and Decision Processes. IEEE
Pomerleau, D. (1992) Neural Network Perception for Mobile Transactions on Systems, Man and Cybernetics, Vol 3 No
Robot Guidance, Ph.D. dissertation, Carnegie-Mellon 1, January 1973.
Technical Report CMU-CS-92-115, Pittsburgh, PA.