Content uploaded by Terry Huntsberger
Author content
All content in this area was uploaded by Terry Huntsberger on Nov 26, 2017
Content may be subject to copyright.
1998 Special Issue
BISMARC: a biologically inspired system for map-based autonomous
rover control
Terry Huntsberger*, John Rose
Intelligent Systems Laboratory, Department of Computer Science, University of South Carolina, Columbia, SC 29208, USA
Received 31 October 1997; revised 3 June 1998; accepted 3 June 1998
Abstract
As the complexity of the missions to planetary surfaces increases, so too does the need for autonomous rover systems. This need is
complicated by the power, mass and computer storage restrictions on such systems (Miller, Proceedings SPIE Conference on Cooperative
Intelligent Robotics in Space III, 1829, pp. 472-475, 1992). To address these problems, we have recently developed a system called
BISMARC (Biologically Inspired System for Map-based Autonomous Rover Control) for planetary missions involving multiple small,
lightweight surface rovers (Huntsberger, Proceedings SPIE Symposium on Sensor Fusion and Decentralized Control in Autonomous Robotic
Systems, pp. 221-227, 1997). BISMARC is capable of cooperative planetary surface retrieval operations such as a multiple cache recovery
mission to Mars. The system employs autonomous navigation techniques, behavior-based control for surface retrieval operations, and an
action selection mechanism based on a modified form of free flow hierarchy (Rosenblatt and Payton, Proceedings IEEE/INNS Joint
Conference on Neural Networks, pp. 317-324, 1989). This paper primarily describes the navigation and map-mapping subsystems of
BISMARC. They are inspired by some recent studies of London taxi drivers indicating that the right hippocampal region of the brain is
activated for path planning but not for landmark identification (Maguire et al., Journal of Neuroscience, 17, 7103-7110, 1997). We also report
the results of some experimental studies of simulated navigation in planetary environments. 䉷1998 Elsevier Science Ltd. All rights reserved.
Keywords: Robot navigation; Hippocampal maps; Wavelets; Neural networks
1. Introduction
The recent successful Mars Pathfinder mission has
demonstrated that a limited set of geological science tasks
can be accomplished remotely using small lightweight
rovers. More sophisticated analysis procedures will have
to be performed in terrestrial laboratories due to the mass
and fragility of the sampling equipment. What this entails is
a return of Martian planetary surface samples to the Earth.
Such a mission is being planned by NASA in 2005 with the
Sample Return Rover (SRR1). SRR1 will retrieve of the
cache containers that have been stockpiled by the 2001
and 2003 Long Range Science Rovers (LRSR) during
their year-long traversal of the surface, and pass the sample
to an orbiter for return to Earth.
Current prototypes of these new rovers, developed at the
Jet Propulsion Laboratory, were designed with specific mass
and power requirements. The typical mass lies between 7
and 10 kg, and the maximum power use during fast
movement (30–50 cm/s) is around 35 W, which can only
be sustained for about 6 h without recharging the batteries.
The primary sensing modalities on SRR1 include a stereo
camera pair (5 cm separation, 130 degree field of view), a
goal camera mounted on the manipulator arm (20 degree
field of view), an internal gyro and accelerometers, and a
planned sun sensor that will give global positioning infor-
mation. The current SRR1 prototype is shown in Fig. 1,
where the manipulator arm is only complete to the first
joint and the top of the rover has been removed. Since the
available power is shared between computer-related com-
ponents and drive mechanisms, the active memory capacity
is limited to 64 MB of RAM, and a processor with low
power requirements such as the PowerPC 603e. Although
the hard disk size is currently 500 MB, access is strictly
limited to essential needs due to heat dissipation. These
limitations constrain the types of algorithms that can be
used for sensing and navigation on the planetary surface.
The complicated nature of planetary environments pre-
cludes the use of most of the robotic planning systems that
are currently available. Although these environments tend to
be relatively static, there are often drastic changes in
* Corresponding author. Tel.: +1-803-777-2404; Fax: +1-803-777-3767;
E-mail: terry@cs.sc.edu
0893–6080/98/$19.00 䉷1998 Elsevier Science Ltd. All rights reserved.
PII: S0893-6080(98)00088-4
Neural Networks 11 (1998) 1497–1510
PERGAMON
Neural
Networks
altitude, wide temperature variations, and obstacles that are
highly irregular in shape. The various internal modules in
the rover control system must communicate state informa-
tion for integrated navigation. Straightforward planning, as
opposed to a reactive approach to long-range navigation in
such uncertain and harsh environments, would drain the
battery reserves within a short time into the mission. The
rover prototypes have the option of navigating around an
obstacle or riding over it since their ground clearance can be
as high as 200 mm. Pure obstacle avoidance would tend to
put more stress on the rover structure, since turns involve
differential forces. A robust control system needs to be able
to combine these possibly conflicting behaviors in order to
maximize the mission goal of cache retrieval.
Brooks developed the subsumption control architecture in
order to address some of these problems for unstructured
terrestrial environments (Brooks, 1986). Using this architec-
ture, seemingly complex behavior arising from a hierarchy
of simple, augmented finite state machines is generated.
Recent work of Parker has extended this approach and
included fault-tolerant characteristics for collections of
heterogeneous robots (Parker, 1994). Despite its notable
successes, the subsumption architecture loses internal state
information due to inhibition or subsumption operations. It
also lacks the ability to combine conflicting behaviors due to
its winner-take-all strategy for control generation, and is
relatively inflexible for highly uncertain environments
without reprogramming.
There are also a number of neural network approaches
that have been used for robot motion control. Among these
are neural dynamics (Baloch and Waxman, 1991; Gaudiano
et al., 1997; Hallam et al., 1997), operant conditioning
(Bu
¨hlmeier and Manteuffel, 1997), reinforcement learning
(Kontoravdis et al., 1992; Kro
¨se and van Dam, 1997;
Prescott and Mayhew, 1992), backpropagation (Pomerleau,
1991), and self-organization (Heikkonen and Koikkalainen,
1997). Most of these systems are potentially capable of
meeting the needs and demands for planetary environments,
but have not been tested in that domain.
In an effort to provide a comprehensive set of capabilities
suitable for autonomous planetary exploration, this paper
presents a multirover control system called BISMARC
(Biologically Inspired System for Map-based Autonomous
Rover Control) (shown in Fig. 2), which is built on previous
work that used wolf pack hunting behavior for retrieval of
spinning and tumbling satellites (Huntsberger and Hilton,
1995; Huntsberger, 1996). Local rover operations are con-
trolled, using stereo sensor and accelerometer inputs, by a
three-level system, which is a hybrid combination of
wavelet signal processing, neural networks based on the
fuzzy self-organizing feature map (FSOFM) algorithm
(Huntsberger and Ajjimarangsee, 1990), and the fine-
grained action selection hierarchical network of Rosenblatt
and Payton (1989). Our previous work has demonstrated
that the FSOFM network is a flexible framework for sensor
processing (Huntsberger and Ajjimarangsee, 1990; Hunts-
berger, 1992a, b, 1995).
BISMARC provides a type of behavior-based control
without the need to explicitly program the sensory-to-action
behavior mapping as in the original subsumption archi-
tecture of Brooks (1986). It also has the ability to adapt to
environmental cues because the free-flow hierarchy (FFH)
used for action selection is not governed by binary inhibi-
tion mechanisms. Experimental studies of 500 missions in a
simulated Martian environment using three cooperating
rovers demonstrated a 98.9% mission success rate for
retrieval of four widely separated cache containers
(Huntsberger, 1997). We are not aware of any other study
that has simulated missions of similar complexity.
The next section describes biological/biologically
inspired models of navigation. This is followed by a dis-
cussion of some neural network approaches to robot
navigation and control. The overall organization of
Fig. 1. SRR1 prototype in Planetary Robotics Lab of the NASA Jet Propul-
sion Laboratory in Pasadena, CA.
Fig. 2. Multilevel system organization of BISMARC for cache retrieval
operations in planetary environments. Coefficients from the wavelet detail
channels are used to generate actions with a FSOFM. An ASM then per-
forms a combination operation on the possible actions for final navigation.
1498 T. Huntsberger, J. Rose/ Neural Networks 11 (1998) 1497–1510
BISMARC, including the navigation and map-making por-
tions of the system is detailed in the next section. It should
be mentioned here that the maps used in our system are
sensory-to-action mappings, and that landmarks as we use
them refer to the actions and internal rover state associated
with obstacles, cache containers, and topological features
such as cliffs and crevasses. The results of some experi-
ments on simulated multiple cache retrieval missions is
described next, followed by a final summary section.
2. Models of navigation
A number of models of navigation are based on theories
about the role that the hippocampus plays in spatial map
formation and navigation.
Bachelder and Waxman (1994) have developed a
biologically motivated, mobile robot visual mapping and
localization system. This system uses the Seibert–Waxman
3D object learning and recognition system to learn visual
landmarks. Learning a landmark entails the generation of
10–20 aspect categories per landmark. In other words, a
landmark is learned by viewing that landmark from 10–20
different directions in 3D space. This allows the system to
recognize a landmark when viewing it from different direc-
tions. When learning a place, a panoramic view of the
environment, which captures information concerning all
visible landmarks, is recorded. In this system, places are
defined by the locations from which visible aspects of
landmarks can be viewed.
Burgess et al. (1994) describe a model of rat hippocampal
function in which firing rate maps of hippocampal place
cells are treated as approximate radial basis functions over
the 2D spatial environment. The phase of firing of individual
place cells gives an indication of the relative position of the
rat with respect to the spatial location for which the place
cell codes. In particular, place cells corresponding to
locations in front of the rat fire late in the Theta cycle (the
5–12 Hz EEG Theta rhythm). In contrast, place cells coding
for locations through which the rat has traversed fire early in
the Theta cycle. The simulation of this model assumes that
the search space of 150 ⫻150 cm is bounded by four walls.
In this model, place cell firing is not strongly modulated by
direction. Furthermore, all visual cues are evenly distributed
and distinguishable and the issue of movement or removal
of individual cues is not addressed.
One view of the function of the hippocampus in naviga-
tion is as a path integrator. This view is supported by the
ability of rats to navigate in complete darkness suggesting
that they are capable of fairly accurate dead reckoning.
Samsonovich et al. (1995) suggest a neuron-level model
which uses head-direction and ideothetic data to perform
path integration which gives the animal’s position in
space. In this paradigm, landmarks serve the additional
purpose of correcting cumulative error in path integration.
The computer model CRAWL developed by Touretzky
and Redish (1995) embodies a theory of landmark-based
navigation that takes into account behavioral as well as
neurophysiological data. Their model combines visual
input, head direction, and motor activity to give an estimate
of the current location through path integration. Place cells
allow a mapping between dead reckoned coordinates and
the perceived location of landmarks.
Blum and Abbott (1996) propose a model in which a map
of 2D space can be created in the rat hippocampus through a
mechanism of long-term potentiation of place cells. The
sequential firing of place cells during exploration results
in a pattern of long-term potentiation (LTP) in their model
which affects subsequent place cell firing. They suggest that
the location of the rat is expressed by the ensemble activity
of active place cells. Thus, the rat could navigate by moving
from its current position to the location corresponding to the
strongest place cell activity. They suggest that the coded
location would be shifted toward and forward along the
specific path that the animal is traversing if LTP occurs
while the animal is on that path. The LTP has the effect
of evolving the navigational map, extending its range.
They evaluate their ideas in a computer simulation of the
Morris maze.
In a subsequent paper, Gerstner and Abbott (1997) pre-
sent a model in which navigational maps are learned
through temporally asymmetric potentiation and depression
of interacting hippocampal place cell synapses. This model
is capable of handling multiple maps to different goals. This
is accomplished through the introduction of modulation of
receptive fields by goal location. This allows a network of
place cells to simultaneously encode maps to several differ-
ent goals. An interesting aspect of this approach is that
modulation corresponding to a goal that has not been pre-
viously learned results in a new map that interpolates
between learned goal locations to provide a path to the
new goal.
A more recent model of map-based navigation proposed
by Recce and Harris (1996) applies Marr’s theory of hippo-
campal function. In contrast to many hippocampal-based
navigation models, they suggest that many of the additional
functional components entailed in map-based navigation are
not located in the hippocampus. Specifically, in their model
an egocentric spatial map is located in the neocortex and is
continuously updated by ideothetic data so that it is possible
to have an idea of position in 2D space in the absence of
sensory cues. This ability to deduce a homing vector in the
absence of specific sensory cues has been observed in
animals, but never simulated by models that rely exclusively
on these cues. The hippocampus acting as an auto-
associative memory stores snapshots of this egocentric
map. Head direction cells are used to select the best
egocentric map rotation to match the snapshots in the
hippocampus which addresses the issue of direction without
requiring directional firing patterns in hippocampal place
cells. This model is evaluated using a mobile robot in an
enclosed internal environment.
1499
T. Huntsberger, J. Rose/ Neural Networks 11 (1998) 1497–1510
Balakrishnan et al. (1997) describe a hippocampal model
of spatial learning and navigation for a mobile robot that
represents space in a metric framework. In this model, the
environment is represented as distinct places in which the
center of each place is labeled with metric information
derived from path integration. Representing goals using
the same coordinate system makes it possible to navigate
directly to a goal by finding the vector difference between
the current position and the location of the goal. This model
combines dead-reckoning position information along with
sensory inputs to determine the current location. Thus path
integration as well as sensory cues are used to localize the
robot in 2D space. They take the approach of using Kalman
filter-based tools for analyzing their model for the fusion of
uncertain sensory data.
3. Neural network robot control
Traditional approaches to rover navigation follow a
sense/plan/act strategy that usually requires long time
delays between movements and relatively large computa-
tional resources. This technique was successfully applied to
cross-country navigation in the CMU Ambler project
(Simmons et al., 1991), the JPL Robby project (Gat et al.,
1991), and the recent CMU/NASA Ames Nomad trek across
the Atacama Desert in Chile (Whittaker et al., 1997). The
vehicles used in these projects weighed between 500 and
3000 kg and required up to 2.4 kW of power. The mass and
power requirements of long-range rover missions preclude
use of such a strategy. These concerns were addressed by
subsequent studies undertaken at JPL using reactive
methods (Gat et al., 1994; Miller et al., 1992).
The MAVIN system of Baloch and Waxman (1991) as
described in the previous section is a comprehensive neural
network approach to robotic sensing and navigation. It
incorporates an action selection mechanism within the
behavior network portion of the system (see Fig. 4 in Baloch
and Waxman, 1991). This behavior network is fed by a layer
of READ (REcurrent Associative gated Dipole) circuits
(see Grossberg and Schmajuk, 1988) that act to associate
recognized objects with actions (behaviors), and an emotion
network which is based on competitive emotional states.
The emotion network is influenced by the READ circuits
as well as internal drive inputs such as battery strength and
other sensory modalities.
The NETMORC system of Zalama et al. and others
(Zalama et al., 1995; Gaudiano et al., 1996a, b, 1997) uses
a VAM (Vector Associative Map) (Gaudiano and
Grossberg, 1991) to perform unsupervised learning of the
mapping of wheel velocities to displacements coupled with
the DIRECT (DIrection-to-Rotation Effector Control
Transform; Bullock et al., 1993) model to learn the mapping
of a desired target angle/distance to wheel velocities that
provide the necessary movements for goal satisfaction. The
advantage of using such a system in uncertain environments
lies in its real-time error correction properties. This was
exhibited in the experimental studies, where the robot’s
performance was relatively stable to perturbations of the
sensory input, wheel characteristics such as radius and
slippage, and noise in the weights connecting the maps
(Gaudiano et al., 1996a).
An operant conditioning model was used to extend the
NETMORC system to include obstacle avoidance
(Gaudiano et al., 1996a). Since the NETMORC system
generates the necessary wheel velocities for navigation to
a target, the generation of fictitious targets triggered by
detection of an obstacle is used for avoidance. The trigger-
ing process is modeled based on the operant conditioning
model of Grossberg (Grossberg, 1971; Grossberg and
Levine, 1987). This model has three cell populations: (1)
Short Term Memory (STM) modeled as a recurrent com-
petitive field (Grossberg, 1982) stimulated by sensory inputs
(sonar), (2) polyvalent cells driven by inputs from STM and
a drive node, and (3) cells connected to a angular velocity
map. Robust obstacle avoidance is exhibited by the system
due to the learning of the temporal connection between
angular velocity patterns and collisions.
The vector field approach used by Tani and Fukumura
(1994) requires knowledge of the goal location in order to
construct mappings from the sensory inputs into maneuver-
ing outputs. Although training is required for interpretation
of the sensory inputs, the use of local minimum in the
potential profile derived from the range sensors is particu-
larly effective for obstacle avoidance. A conservative
steering equation is used that minimizes the possibility of
collisions due to the finite size of the mobile robot. Their
experiments demonstrated that knowledge of temporal
patterns in the sensory inputs eliminates potential
ambiguities in the state space.
The correlation matrix approach of Fukushima et al.
(1997) for chained recall of previously learned maps utilizes
a shifting stimulus pattern technique based on cross-
correlation between the current sensed pattern and the
sum of all patterns previously learned in the matrix. Recent
studies of London taxi drivers support this framework,
where landmark information and spatial maps in path plan-
ning activated different areas of the brain (Maguire et al.,
1997).
4. BISMARC system organization
The complicated nature of a multiple cache recovery task
using multiple rovers involves system organization at many
levels. For such operations, we designed the three level
BISMARC system (shown in Fig. 2) which uses a hybrid
mix of neural networks and behavior-based approaches. The
first level performs a wavelet transform on the rover’s stereo
image pair, the second level inputs these processed images
into an action generation navigation network, and then to a
third level action selection mechanism (ASM) network
1500 T. Huntsberger, J. Rose/ Neural Networks 11 (1998) 1497–1510
modeled after that of Rosenblatt and Payton (1989). Some
examples of the other external inputs would include internal
temperature sensing, relative time of day, sun sensor
positioning, and communications with other rovers.
BISMARC uses a FSOFM to learn landmarks (obstacles
and goals). In the operational mode, this network generates
membership values to the classes of visual input that the
system has previously seen. When coupled with onboard
rover components such as accelerometers and dead reckon-
ing inputs, an egocentric map of the environment is built
using the FSOFM response as an index. This would be
equivalent to navigation based on a statement such as
‘approximately five steps past the big oak tree’, which
does not indicate which big oak tree, but would allow
localization after an oak tree is sensed. In the same manner,
a human would have trouble distinguishing between naviga-
tion in a real environment and in an accurate Hollywood
movie set of the same environment. Since unsupervised
operation or training a system in a planetary environment
such as Mars would be costly and potentially dangerous to
the rover, BISMARC offers a compromise solution.
The three-level self-organizing system of Heikkonen and
Koikkalainen (1997) is closest to BISMARC in structure.
They use Gabor filtered visual input fed into the first level,
followed by clustered feature extraction in the second level,
and action generation in the third level. Lack of a specific
action selection mechanism in their system is the most
notable difference between the two. Our system replaces
the first two levels of their system with a single FSOFM
by using the wavelet transform as a preprocessing level.
This was done in order to eliminate the large storage
space needs and coverage problems of the frequency
space by the Gabor wavelet (Lee, 1996). We use the FFH
of Rosenblatt and Payton (1989) for action selection, which
had been successfully fielded for an autonomous land
vehicle (ALV). The FFH used in our study includes both
unidirectional and multidirectional sensors, temporal
penalties for actions (Sutton, 1988), ordinary and multi-
directional nodes, and both additive and multiplicative
processing of inputs to a node.
This type of FFH was recently shown to be optimal within
the multiple objective decision-making (MODM) formal-
ism, which produces an action through the maximization
of a global objective function that includes all possible
actions (Pirjanian and Christensen, 1997). Additive pro-
cessing of inputs to a node can combine conflicting
behaviors for maximum motion efficiency (shown in
Fig. 3). The rover would move south if the Avoid Obstacle
action generation were the only node that is activated since
it has a high negative strength as represented by the large
shaded circle in the north direction. Instead its combination
with the Approach Goal activation, which contains a small
positive value, results in motion towards the cache container
perceived to the west.
Although the rover(s) that place the caches have some
sense of position, they will not generally have access to a
detailed map. This would be the case even with known
landmarks due to uncertainty in accuracy of localization
using triangulation (Sutherland and Thompson, 1994).
Field trials with the current generation of Lightweight
Survivable Rovers (LSR) and SRR1 at the Jet Propulsion
Laboratory have indicated that there is a great amount of
slippage in the drive wheels during traversal of Martian-like
terrain. This precludes the use of purely dead-reckoning
techniques for navigation and obstacle avoidance. This
point was recently addressed by Baumgartner and Skaar
using an extended Kalman filter analysis of staged visual
cues in the environment to refine the dead reckoning posi-
tion estimates (Baumgartner and Skaar, 1994). Their system
accurately localized a rover in the laboratory to within 1 cm
using only a single camera and a series of previously defined
reference patterns. Additional work on this problem by
Roumeliotis and Bekey indicates that the extended Kalman
filter analysis can be further improved by an order of
Fig. 3. Combination of conflicting behaviors where pure obstacle avoidance action generation leads to inefficient motion away from the cache container.
Shaded circle indicates a negative value and relative size of circles indicate strength.
1501
T. Huntsberger, J. Rose/ Neural Networks 11 (1998) 1497–1510
magnitude with the inclusion of a sun sensor input
(Roumeliotis and Bekey, 1997). The ASM used in
BISMARC combines the information from the sun sensor
and the stereo cameras on the scout rover to extract visual
cues from the environment, coupled with a FSOFM to per-
form the kinematic adjustments. The output of the FSOFM
serves as a behavior based index into a sensory-to-action
map as shown in Fig. 4, where a relatively high level of
sensory-to-action detail is only maintained at obstacles
and goals.
The rovers in the simulated missions were of two types:
scout and retrieval. The navigation system on the scout
rover is responsible for locating the cache containers and
‘marking’ obstacles that it encounters in its path to each of
the containers using the sensory-to-action information gen-
erated by the ASM level in BISMARC. These obstacles
need to be avoided both by the scout rover, and the retrieval
rovers on their way to each cache with their augmented
knowledge of the terrain. This process requires an effective
representation of the terrain, including obstacles and the
cache container locations. The use of stereo cameras for
characterization of the terrain offers the possibility of
real-time rover obstacle avoidance (Matthies, 1992).
Many biological systems use anchor points (landmarks)
for navigation in spatially extended environments. This was
recently demonstrated in a study of London taxi drivers
where the right hippocampus was activated for route
planning, but not for landmark identification (Maguire
et al., 1997). The medial parietal cortex was activated for
both tasks, indicating that landmark identification is an
important portion of the egocentric sequential aspects of
the route planning task (Maguire et al., 1997). In other
words, navigation between two waypoints that are not in a
direct line involves the sequential reference to landmarks
along the way. This is the primary motivation for the
sensory-to-action maps that are built for navigation in
BISMARC.
We have developed a landmark identification subsystem
in BISMARC based on the FSOFM, which generates
behaviors associated with wavelet-encoded visual input. It
should be stressed here once again that landmarks in
BISMARC are an encoded action representation, and are
not meant in the traditional sense (i.e. the Statue of Liberty).
This subsystem is similar to the approach taken by Mataric
´
(1992a, b, 1997) except that her studies used a behavior
chain to characterize a path through the indoor environment.
The output of the FSOFM is used as a local index into a
crude map of the terrain (5 m resolution) that would be
obtained with a camera as the rover landed. The indexing
process is shown in Fig. 4.
Navigation with such a low resolution map would be
extremely dangerous for the rover, since the rover size is
on the order of 50 cm. BISMARC uses a 5 cm resolution for
the local maps, which would correspond to a frequency of
10 samples/s when moving at full speed. The crude map
used in BISMARC is a metric map of the terrain with a
4 bit label at each grid point indicating the possible terrain
character (texture measure of roughness). Due to the unique
landing mechanism used in the Sojourner mission, acquisi-
tion of this type of visual 2D map was not possible. Future
NASA mission plans include a more conventional landing
scenario. Links between the landmark points are encoded
using the odometry data, which is used for rough dead
reckoning analysis of the connecting path. Although
wheel slippage and other effects make these data unreliable,
significant information compression is achieved by not
storing details of the uninteresting (at least to these rovers)
portions of the path.
Since the area that is traversed is on the order of square
kilometers, a detailed occupancy grid approach like that of
Moravec and Elfes (Elfes, 1987; Moravec and Elfes, 1985)
is not feasible due to storage limitations on the rovers. For
example, our simulation studies covered a 1 km ⫻1km
area at a local resolution of 5 cm. This would require a
grid of 20,000 ⫻20,000 places for full coverage, or about
400 MB if a single byte is used for each grid point. Exten-
sive access of the hard disk would be required for such a
map to be kept, which would lead to possible internal tem-
perature problems for the rover. BISMARC only encodes
the local details of the surface that are relevant for naviga-
tion, which include the relative height of the obstacles,
action taken immediately prior to and after encountering
the obstacles (direction of travel and velocity), acceler-
ometer state (to encode surface slope), and a detailed
360 degree sweep of the area at any cache container. This
representation uses four bytes for obstacles and twelve bytes
for the cache. This information is only stored for obstacles
and goals (cache containers).
A vision preprocessing level uses wavelet-based
algorithms to decompose the images generated by each of
the stereo cameras. The wavelet decomposition provides
Fig. 4. Sensory-to-action index into local maps. These local maps are only
obtained at obstacles and goals and are in an order of magnitude better
spatial resolution than the global map.
1502 T. Huntsberger, J. Rose/ Neural Networks 11 (1998) 1497–1510
information about the scale, location, and orientation of
features in an image. Because the wavelet decomposition
contains information about the local frequency content of an
image, it can represent visually important features (such as
edges) more compactly than many of the other transforms
commonly used in image processing (Mallat and Zhong,
1992). Our previous work with wavelet-processed images
has shown their utility for sensor fusion (Huntsberger and
Jawerth, 1993), morphological image processing
(Huntsberger and Jawerth, 1995), motion analysis
(Huntsberger et al., 1994), face processing (Huntsberger
et al., 1998), image enhancement (Hilton et al., 1993), and
texture processing (Espinal et al., 1998).
After the wavelet transform is performed on each of the
stereo images, a vector is formed using the multiresolution
information from the two highest levels of the wavelet
horizontal and vertical detail channels. This vector of length
40,960 elements is the input to the FSOFM, the output being
any of six action states: go forward, backup, turn right, turn
left, stop, or pick either direction in turn. This sensory-to-
action mapping approach is similar to that used by
Pomerleau in the CMU ALVINN road following system
(Pomerleau, 1991), although in that system specific features
such as edges of roads were identified. BISMARC is only
encoding the raw stereo visual information without any
attempt to label individual features or objects beyond the
desired action associated with the input pattern. Although
the FSOFM is an unsupervised network, it can be trained by
presenting it with samples and labeling the output nodes that
correspond to each of the generated actions. The weights are
then clamped for all subsequent runs. This approach was
successfully used for the inverse kinematics study of a
seven degree-of-freedom robot arm (Soltys and
Huntsberger, 1993).
The network is trained with a set of wavelet-processed
images that simulate the types of terrain seen on the Martian
surface during the recent Sojourner mission. Some of the
stereo pairs that were used to train the navigation FSOFM
are shown in Fig. 5. A total of 500 stereo pairs were used for
the training session, which took 637 epochs to converge.
This should be compared to a backpropagation implementa-
tion, which typically takes hundreds of thousands of epochs
to converge. Recall of the trained images was 100%.
An advantage of using the FSOFM for the action genera-
tion level lies in the membership values that are generated at
the output nodes. The sum of these values is normalized to
one, and the relative size of the membership values gives a
ranking of the actions that are possible. For example, a set of
stereo images shown in Fig. 6 that were not in the training
set were shown to the system after training. This set of
images contains a cache container that is partially hidden
by rocks on the right-hand side of the field of view. The
Fig. 5. Samples of stereo image pairs used to train the navigation FSOFM; generated action is given beneath each pair. Sample stereo pairs generated using
VistaPro with USGS Martian Digital Elevation Maps.
1503
T. Huntsberger, J. Rose/ Neural Networks 11 (1998) 1497–1510
ranking of the membership values indicated the highest
actions were to turn left (0.49), goal (0.39), with the rest
of the memberships distributed over the other four states.
The obstacle avoidance behavior of the system is invoked
due to it having the highest membership value, but the
system still senses that the goal is close as it moves. Move-
ment far from an obstacle is automatically generated, and
the rover will only turn when its field of view has the vector
of wavelet coefficients that indicate looming of an obstacle
large enough to need avoidance behavior. The velocity of
the rover is selected within the ASM dependent on the
FSOFM output, it is slower during turns (3 cm/s), when
approaching the goal (1 cm/s), andinrockyterrain(2 cm/s).
There are a number of factors beyond the visual sensory
input that influence navigation of the rover. These include
the health of the rover (internal/external temperature,
battery power levels, accelerometers/gyro), time of day,
and homogeneity of the terrain. For example, if the rover
is approaching a highly uneven portion of the terrain rather
late in the day, the control decision may be made to halt and
wait for the next morning in order to recharge the batteries
and to have enough light for visual sensing. Action selection
in the presence of such conflicting behaviors is done in
BISMARC using the FFH of Rosenblatt and Payton
(1989).
Possible evidence for a type of FFH behavior in nature is
found in geese when offered corn from a human hand. Fear
of the human coupled with the desire for the corn leads to a
quivering of the neck due to antagonistic muscle responses
(Lorenz, 1981). A FFH is a directed graph of action and
stimulus nodes that are combined using predetermined
rules. These rules may include addition, multiplication, or
more complicated means of combination. The Find Cache
system for BISMARC is shown in Fig. 7. Action nodes are
drawn as rectangles, stimulus nodes as ellipses, and those
with multidirectional characteristics are indicated using 8
directional bins. The combination rules are additive for a
small filled rectangle above the node, multiplicative for a
small filled triangle, and the following rule is used for plain
rectangular nodes (Tyrrell, 1993):
Aj¼
maxi(Pþ
ij )þaX
Nþ
i¼1(Pþ
ij )
1þaþ
mini(P¹
ij )þbX
N¹
i¼1(P¹
ij )
1þb,
where A
j
is the activation strength, Pþ
ij and P¹
ij are the
positive and negative preferences from node ifor node j,
N
þ
and N
¹
are the numbers of such preferences for node j,
and aand bare constants. This more sophisticated combi-
nation rule was developed by Tyrrell (1993) to guarantee the
proper transfer of goal and motivational behavior to lower
levels of the FFH.
Since the scout rovers in our simulations need to find all
of the cache containers, penalty nodes were added to prevent
a total mission failure in the event that one or more of the
containers are impossible to approach. This scenario would
have occurred during our simulation runs (Mission 329
shown in Fig. 10), leading to a circling behavior around
the cache container until the rover’s batteries were
exhausted. Tyrrell (1993) introduced the temporal penalty
(T-circle in Fig. 7) to inhibit action that will take an inordi-
nate amount of time to complete. The temporal penalty is
derived using the assigned value raised to the power of the
elapsed time during the current action. Since it is being used
with a multiplicative combination rule and is less than 1.0 to
start, the action activation will decline with increasing time.
This type of node is similar to the impatience parameter
used by Parker (1994) or the temporal discount factors of
Sutton (1988). Temporal penalty nodes increased the like-
lihood of satisfying the overall mission goal of maximizing
the number of cache containers returned.
The large motivational input of 4.0 to the Find Cache
node is mediated through its multiplicative combination
with the internal stimuli of Night and Good Health. The
Night stimuli is derived from the sun sensor input, and is
equal to 1.0 at sunrise and decreases to 0.0 at sunset. The
Good Health stimuli is a product of the battery level scaled
between 0.0 and 1.0 and a triangular function of the internal
temperature of the rover. This triangular function starts at
Fig. 6. Membership values returned from FSOFM with ambiguous input.
Although the system was not trained with this stereo pair, it was able to
generalize to the appropriate action of turn left to avoid the obstacle.
Fig. 7. Find Cache system for BISMARC. Connections from other sub-
systems are not shown. Notation for symbols is that of Tyrrell (1993). See
text for a detailed discussion.
1504 T. Huntsberger, J. Rose/ Neural Networks 11 (1998) 1497–1510
0.0 for ¹50 ⬚C, peaks at 1.0 for 55 ⬚C, and drops to 0.0 at
100 ⬚C. The multidirectional stimuli that feed the Move
Actions output node are directly derived from the output
nodes of the FSOFM.
While the design of BISMARC has been biologically
inspired, the constraints placed on its design, which are a
consequence of the environment in which it must operate,
set it apart from the models described in the previous two
sections. Any autonomous navigation model for a planetary
rover is a priori expected to operate in the open in 3D space.
Furthermore, such a remote system does not have the luxury
of being able to learn to navigate or learn a map of the
environment in situ by bumping into obstacles or careening
over a precipice. The rover must avoid self-inflicted damage
at all cost due to the paucity of rover repair facilities and
qualified technicians on remote planets. Prior knowledge of
the exact goal position in the environment will not be known
due to the uncertainties in navigation mentioned in the pre-
vious section. Additionally, time, computation and power
constraints also place limits on map learning.
The sensory input to BISMARC is also quite different
from the systems that were detailed in the preceding
sections. Since it operates in 3D space, it also uses
accelerometer data to indicate whether its vertical orienta-
tion is level, downslope or upslope. In addition to ordinary
visual information, the visual information in its field of view
is qualitatively characterized as left-field occluded,
right-field occluded, total occlusion, goal, clear, and
centered.
The direct integration of the obstacle avoidance behavior
into BISMARC’s route planning strategy is a necessary
component of the design. Since obstacles are used as land-
marks for sensory/action map making, detailed information
about size, relative height, etc., are important for the
retrieval rovers that will follow the scout. The necessary
clearance for a rover to navigate around an obstacle is
built into the training sets, and looming is used to differenti-
ate between small obstacles that can be driven over and the
larger ones that require course changes. Most of the systems
reviewed below in the context of obstacle avoidance are
different from BISMARC in the sense that they are actively
mapping the environment using obstacles as something to
avoid, rather than as significant features for landmark-based
navigation.
The MAVIN system (Bachelder and Waxman, 1994) uses
2D information for landmark recognition and has been
operated in an enclosed environment. However, there are
no constraints imposed on the MAVIN system which
preclude its operation in an open 2D environment.
The 2D models of Burgess et al. (1994), Samsonovich et
al. (1995), Blum and Abbott (1996), Gerstner and Abbott
(1997), and Balakrishnan et al. (1997) have been simulated
in enclosed environments. The lack of published data show-
ing the performance of these models in controlled open
environments makes it difficult to draw objective conclu-
sions regarding potential performance in arbitrary open
environments. This is particularly difficult to evaluate in
the cases of the 2D models of Burgess et al. (1994), Blum
and Abbott (1996), Gerstner and Abbott (1997) in which
explicit use is made of the boundaries of the closed
environment in the simulations. The approach of allowing
the simulated rat to ‘bounce’ off of walls has the effect of
redirecting the focus of exploration toward the interior of
the enclosed environment. Nevertheless, the fact that these
systems, as well as those by Touretzky and Redish (1995)
and Recce and Harris (1996), have only been tested or simu-
lated in enclosed environments should not be interpreted to
mean that such systems are only capable of operating in an
enclosed environment.
Burgess et al. (1994) state that a consequence of the sen-
sory input entorhinal cell model that they employ is that the
goal location must be inside a convex hull of the set of cues
in order for the navigation to be successful. This could pose
a problem in an arbitrary open environment.
In the case of the 2D model of Samsonovich et al. (1995),
the connections between place cells in the model are pre-
wired and fixed. In addition, the description of the model
does not address obstacle avoidance (Samsonovich et al.,
1995). At a minimum, navigation in an open/unbounded
environment would require the ability to avoid
obstacles unless the map is prelearned and the environment
is static.
The 2D models of Burgess et al. (1994), Blum and Abbott
(1996), Gerstner and Abbott (1997) address obstacle avoid-
ance by allowing the simulated rat to ‘bounce’ off of the
enclosing walls or any obstacles. While such ‘bouncing’
may not be unreasonable in the simulation of a model, it
is clearly unacceptable in the case of a navigational system
for an actual rover. Since our goal is the latter, we have been
forced to take a different tack in addressing obstacle
avoidance.
The CRAWL model by Touretzky and Redish (1995) is
also a 2D space simulation, although the authors indicate
that portions of the model have been implemented on a
mobile robot in an enclosed environment. They describe
computer simulations of a split landmark array task. Their
model is trained on the environment by depositing the simu-
lated animal at successive random spots. Place units are
recruited as necessary to learn the environment with respect
to randomly chosen visible landmarks. Training terminates
after 10 random spots have been visited without requiring
the recruitment of additional place units. A consequence of
this training procedure is that the issue of obstacle avoid-
ance does not arise. Although they do not present results for
an open environment, there does not appear to be anything
in their model that would preclude such experiments assum-
ing that something other than random site selection were
used to train the model.
Recce and Harris (1996) evaluated their model using an
actual mobile robot, ARNE in an enclosed 2D environment.
Irregular shaped rooms were used in their experiments to
reduce the similarity of features that would be observed with
1505
T. Huntsberger, J. Rose/ Neural Networks 11 (1998) 1497–1510
the robot’s sonar sensor. However, the rooms were large
enough so that the robot was not able to detect all walls
from one location. The exploration strategy that is described
is a simple wall-following strategy. Thus the boundaries of
the enclosed environment play an important role in the
results that are presented. However, there does not appear
to be any intrinsic limitation in their architecture which
would preclude the extension of their model to address an
open environment given an exploration strategy not predi-
cated on an enclosed environment.
5. Experimental study
The interface to BISMARC (shown in Fig. 8) has display
output for relevant mission parameters, including the cur-
rent internal temperature and battery level for the rover, a
bird’s-eye-view of the study area, the current stereo view
from the rover, the current top two actions that have been
activated, the Martian time of day, and the total accrued
mission time. The path that has been followed by the
rover up to this point in the study mission is shown as a
white line on the bird’s-eye-view display.
The mission control strategy is as follows:
•Deployment of all rovers at lander staging site.
•Scout rover heads toward first cache site, sending back
coarse path information such as major obstacles and path
lengths between them. It is followed by a single retrieval
rover using the follow behavior of Mataric
´(1992a, b). In
addition wavelet-compressed images of selected views
to augment the map and of the cache site are broadcast to
the retrieval rovers when the scout rover arrives. This
is done to safeguard the mission objectives in the
event that the retrieval rover does not make it to the
cache site.
•After deployment of a beacon, the scout rover starts
toward the next potential cache site location, once
again broadcasting compressed versions of its path. If
another retrieval rover is available, it is dispatched on an
intersection path with the scout.
•Each retrieval rover broadcasts a message indicating
successful acquisition of the cache and another when it
returns to the lander staging site. The robot impatience
behavior (Parker, 1994) will be invoked if the waiting
retrieval rovers do not receive this message within a
predetermined period of time.
•After all known caches are retrieved, the rovers attempt
to return to the lander staging site.
This control strategy was coded as a FFH, and run under
BISMARC.
We ran 500 trials using a subset of the heightfield infor-
mation returned from the Mars Pathfinder mission (shown in
Fig. 9). The area encompassed about 1 km ⫻1 km with a
grid decomposition resolution of 5 cm at the detailed map
level. The number of detailed maps will vary with the rug-
gedness of the terrain, since the maps are only generated for
obstacles and cache containers. Each trial had different
starting positions and the placement of 4 cache containers
was randomized within the area. It was assumed that each
cache placement site was known to within a 200 m radius.
Three rovers were deployed for each simulated mission: a
scout and two retrieval rovers. The bandwidth of the
Fig. 8. Interface to BISMARC. Bird’s-eye-view of study area displayed in left subwindow (grayscale is usedto indicate relative elevation), stereo view from
rover displayed in upper right subwindow. Important mission parameters such as internal temperature and battery level shown as dynamic sliders.
1506 T. Huntsberger, J. Rose/ Neural Networks 11 (1998) 1497–1510
communication channel between the rovers is 1 Mbit/s,
which is the same as the modem installed in the current
SRR1 prototype at the Jet Propulsion Laboratory. The top
speed on the rovers was set at 30 cm/s, which is consistent
with the SRR1. In order to simulate wheel slippage, we set a
5% loss of traction when climbing over a rock or traversing
rocky terrain. The battery lifetime was set at one week on all
of the rovers and the timestep size for the simulations was
fixed at 0.1 s. All of the rovers were forced to sleep during
the night hours of the simulations, since there were no
infrared sensors on any of the rovers, and navigation at
night would be dangerous.
A total mission success is defined as the return of all four
of the cache containers to the landing site. The mission
success rate for the 500 simulated missions was 98.9%.
Only one of the missions was a total failure, in that none
Fig. 9. 1 km ⫻1 km study zone used in the experimental studies. Grayscale is used to indicate relative elevation, with a highest elevation change in the study
area of 500 m.
Fig. 10. Paths taken by the scout rovers in four of the multiple cache retrieval missions. Cache container positions are marked by circles and rover start position
is marked with a cross. Filled circle in Mission 329 is a missed cache container. Grayscale is used to indicate relative elevation.
1507
T. Huntsberger, J. Rose/ Neural Networks 11 (1998) 1497–1510
of the cache containers were recovered due to the untimely
death of the scout rover from a fall down a steep incline. The
paths taken for four of the missions are shown in Fig. 10,
where the white line in each subwindow is the path and the
cache container positions are marked with white circles. An
example of the temporal penalty effect on rover navigation
is seen in Mission 329 in Fig. 10, where one of the cache
containers (shown as a filled circle) was unable to be
acquired due to it being totally surrounded by rocky terrain.
It is unclear whether this scenario would ever occur in an
actual mission, since the cache container was supposedly
put in place by the 2001 or 2003 rovers. The question raised
is how those rovers got into this area of the planetary surface
in the first place. Evaluation of the optimal path for each
mission is complex due to the obstacle avoidance options in
such a complicated 3D terrain, and only a complete path
planning analysis would give this information. In general,
the scout rovers localized the cache containers well within
the mission time constraints and with a minimal amount of
structural damage. Most of the additional path length can be
traced to the Explore behavior in the Find Cache ASM
of Fig. 7, which was motivated using random search
directions. Current NASA plans call for a beacon guidance
system for the initial 200 m approach phase, followed
by visual guidance to the cache container when within
10 m.
6. Conclusions
This paper introduced a hybrid system for planetary rover
navigation and control called BISMARC that uses neural
networks coupled with behavior-based control for the coor-
dination of multiple rovers during the navigation to and
subsequent retrieval of multiple cache containers. Out of a
total of 2000 cache containers, 1978 were successfully
retrieved over 500 simulated missions. These results indi-
cate that the sensory-to-action map representation used by
BISMARC and shared between the rovers was an effective
representation of the study area for retrieval operations. The
majority of the failures were due to expiration of the battery
lifetime on the rovers. The average completion time over the
500 missions was 4.7 days.
We are currently extending the navigation algorithms to
include better search strategies in order to minimize the
chance of loss of power leading to mission failure. An
investigation is also underway to determine the length of
the optimal vector of wavelet coefficients in order to reduce
the FSOFM processing time. It is anticipated that field
studies at the Jet Propulsion Laboratory will be used to
test the navigation portion of the system for SRR1 in real
terrain. In addition, we are evaluating incorporation of more
fault tolerance into the third level network of BISMARC to
maintain a reduced mission capability in the event of
internal sensor failure or partial structural damage to the
rover.
Acknowledgements
The authors would like to thank the reviewers for their
extremely thorough reading of the manuscript and sugges-
tions which greatly improved its readability. The first author
would like to thank Dr. Paul Schenker of the Jet Propulsion
Laboratory for his support under the ASEE NASA Summer
Faculty Fellowship Program during the summer of 1997.
This work was also supported in part under ARO grant
DAAH03-96-C-0034, and ONR grant N00014-97-1-1163.
References
Bachelder I.A., & Waxman A.M. (1994). Mobile robot visual mapping and
localization: a view-based neurocomputational architecture that
emulates hippocampal place learning. Neural Networks,7 (6) 7,
1083–1100.
Balakrishnan, K., Bousquet, O., & Honavar, V. (1997). Spatial learning and
localization in animals: a computational model and its implications for
mobile robots. Technical Report 9720. Ames, IA: Iowa State
University.
Baloch A.A., & Waxman A.M. (1991). Visual learning, adaptive expecta-
tions, and behavioral conditioning of the mobile robot MAVIN. Neural
Networks,4, 271–302.
Baumgartner E.T., & Skaar S.B. (1994). An autonomous vision-based
mobile robot. IEEE Transactions on Automatic Control,39 (4),493–
502.
Blum K.I., & Abbott L.F. (1996). A model of spatial map formation in the
hippocampus of the rat. Neural Comp.,8, 85–93.
Brooks R.A. (1986). A robust layered control system for a mobile robot.
IEEE Journal of Robotics and Automation,2 (1), 14–23.
Bu
¨hlmeier, A., & Manteuffel, G. (1997). Operant conditioning in robots. In
O. Omidvar & P. van der Smagt (Eds.), Neural systems for robotics. San
Diego, CA: Academic Press.
Bullock D., Grossberg S., & Guenther F. (1993). A self-organizing neural
network model for redundant sensory–motor control, motor equiva-
lence, and tool use. Journal of Cognitive Neuroscience,5 (4), 408– 435.
Burgess N., Recce M., & O’Keefe J. (1994). A model of hippocampal
function. Neural Networks,7 (6) 7, 1065–1082.
Elfes A. (1987). Sonar-based real-world mapping and navigation. IEEE
Journal of Robotics and Automation,RA3 (3), 249–265.
Espinal F., Huntsberger T., Jawerth B., & Kubota T. (1998). Wavelet-based
fractal signature analysis for automatic target recognition. Optical
Engineering,37 (1), 166–174.
Fukushima K., Yamaguchi Y., & Okada M. (1997). Neural network models
of spatial memory: associative recall of maps. Neural Networks,10 (7),
971–980.
Gat E., Desai R., Ivlev R., Loch J., & Miller D.P. (1994). Behavior control
for robotic exploration of planetary surfaces. IEEE Journal of Robotics
and Automation,10 (4), 490–503.
Gat, E., Slack, M., Miller, D., & Firby, R. (1991). Path planning and
execution monitoring for a planetary rover. In Iyengar & Elfes (Eds.),
Autonomous mobile robots: Control planning and architectures. IEEE
Computer Society Press.
Gaudiano P., & Grossberg S. (1991). Vector associative maps: unsuper-
vised real-time error-based learning and control of movement
trajectories. Neural Networks,4 (2), 147–183.
Gaudiano, P., Guenther, F.H., & Zalama, E. (1997). The neural dynamics
approach to sensory–motor control. In O. Omidvar & P. van der Smagt
(Eds.), Neural systems for robotics. San Diego, CA: Academic Press.
Gaudiano, P., Zalama, E., Chang, C., & Coronado, J.L. (1996a). A model of
operant conditioning for adaptive obstacle avoidance. Technical
Report. Boston, MA: Boston University.
1508 T. Huntsberger, J. Rose/ Neural Networks 11 (1998) 1497–1510
Gaudiano, P., Zalama, E., & Coronado, J.L. (1996b). An unsupervised
neural network for low-level control of a wheeled mobile robot: noise
resistance, stability, and hardware implementation. IEEE Trans.
Systems, Man, Cybernetics., 26, 485–496.
Gerstner W., & Abbott L.F. (1997). Learning navigational maps through
potentiation and modulation of hippocampal place cells. J. Computa-
tional Neurosci.,4, 79–94.
Grossberg S. (1971). On the dynamics of operant conditioning. Journal of
Theoretical Biology,33, 225–255.
Grossberg, S. (1982). In S. Grossberg (Ed.), Studies of mind and brain:
Neural principles of learning, perception, development, cognition, and
motor control. Boston, MA: Reidel.
Grossberg S., & Levine D. (1987). Neural dynamics of attentionally
modulated Pavlovian conditioning: blocking, interstimulus interval,
and secondary reinforcement. Applied Optics,26, 5015–5030.
Grossberg, S., & Schmajuk, N.A. (1988). Neural dynamics of attentionally
modulated Pavlovian conditioning: conditioned reinforcement, inhibi-
tion, and opponent processing. In S. Grossberg (Ed.), Neural networks
and natural intelligence. Cambridge, MA: MIT Press/Bradford Books.
Hallam, B., Hallam, J., & Hayes, G. (1997). A dynamic net for robot
control. In O. Omidvar & P. van der Smagt (Eds.), Neural systems
for robotics. San Diego, CA: Academic Press.
Heikkonen, J., & Koikkalainen, P. (1997). Self-organization and auto-
nomous robots. In O. Omidvar & P. van der Smagt (Eds.), Neural
systems for robotics. San Diego, CA: Academic Press.
Hilton M., Jawerth B., & Huntsberger T.L. (1993). Local enhancement of
compressed images. Journal of Mathematical Imaging and Vision,3,
39–49.
Huntsberger, T., & Hilton, M. (1995). Distributed multisensor integration
in a co-operative multirobot system. In P. S. Schenker & G.T. McKee
(Eds.), Proc. SPIE Symposium on Sensor Fusion and Networked
Robotics (Vol. 2589, pp. 134–141). Philadelphia, PA.
Huntsberger, T.L. (1992a). Data fusion: a neural networks implementation.
In M.A. Abidi & R.C. Gonzalez (Eds.), Data fusion in robotics and
machine intelligence. Orlando, FL: Academic Press.
Huntsberger, T.L. (1992b). Sensor fusion in a dynamic environment. In P.S.
Schenker (Ed.), Proc. SPIE Symposium on Sensor Fusion V (Vol. 1828,
pp. 175–182). Boston, MA.
Huntsberger T.L. (1995). Biologically motivated cross-modality sensory
fusion system for automatic target recognition. Neural Networks,8
(7) 8, 1215–1226.
Huntsberger, T.L. (1996). Hybrid teleoperated/cooperative multirobot
system for complex retrieval operations. In P.S. Schenker & G.T.
McKee (Eds), Proc. SPIE Symposium on Sensor Fusion and Distributed
Robotic Agents (pp. 11–17). Boston, MA.
Huntsberger, T.L. (1997). Autonomous multirover system for complex
planetary retrieval operations. In P.S. Schenker & G.T. McKee
(Eds.), Proc. SPIE Symposium on Sensor Fusion and Decentralized
Control in Autonomous Robotic Systems (pp. 221–227). Pittsburgh, PA.
Huntsberger, T.L., and Ajjimarangsee, P. (1990). Parallel self-organizing
feature maps for unsupervised pattern recognition. International
Journal of General Systems,16, 357– 372. Also reprinted in J.C. Bezdek
& S.K. Pal (Eds.), Fuzzy models for pattern recognition (pp. 483–495).
Piscataway, NJ: IEEE Press.
Huntsberger, T.L., & Jawerth, B. (1993). Wavelet based sensor fusion. In
P.S. Schenker (Ed.), Proc. SPIE Symposium on Sensor Fusion VI (Vol.
2059, pp. 488–498). Boston, MA.
Huntsberger, T.L., & Jawerth, B. (1995). Morphological image analysis
within the wavelet representation. In Proc. SPIE Symposium on Aero-
space/Defense Sensing and Dual-Use Photonics (pp. 1066–1072).
Orlando, FL.
Huntsberger, T.L., Jawerth, B., & Chang, J. (1994). Wavelet analysis of
fused EPI sequences. In P.S. Schenker (Ed.), Proc. SPIE Symposium on
Sensor Fusion VII (pp. 157–163). Boston, MA.
Huntsberger, T.L., Rose, J., & Ramaka, A. (1998). Fuzzy-Face: a hybrid
wavelet/fuzzy self-organizing feature map system for face processing.
Journal of Biological Systems. To appear.
Kontoravdis, D., Likas, A., & Stafylopatis, A. (1992). Collision-free move-
ment of an autonomous vehicle using reinforcement learning. In B.
Nuemann (Ed.), ECAI 92 (pp. 666–670). New York: Wiley.
Kro
¨se, B., & van Dam, J. (1997). Neural vehicles. In O. Omidvar & P. van
der Smagt (Eds.), Neural systems for robotics. San Diego, CA:
Academic Press.
Lee T.S. (1996). Image representation using 2D Gabor wavelets. IEEE
Trans. PAMI,18 (10), 959–971.
Lorenz, K. (1981). Foundations of ethology. Berlin: Springer-Verlag.
Maguire E.A., Frackowiak R.S.J., & Frith C.D. (1997). Recalling routes
around London: activation of the right hippocampus in taxi drivers. J. of
Neuroscience,17 (18), 7103–7110.
Mallat S., & Zhong S. (1992). Characterization of signals from multiscale
edges. IEEE Trans. PAMI,14, 710–732.
Mataric
´M.J. (1992). Integration of representation into goal-driven
behavior-based robots. IEEE Transactions on Robotics and
Automation,8 (3), 304–312.
Mataric
´, M.J. (1992b). Distributed approaches to behavior control. In P.S.
Schenker (Ed.), Proc. SPIE Sensor Fusion V (Vol. 1828, pp. 373–382).
Boston, MA.
Mataric
´, M.J. (1997). Behavior-based control: examples from navigation,
learning, and group behavior. Technical Report. Waltham, MA:
Brandeis University.
Matthies L. (1992). Stereo vision for planetary rovers: stochastic modeling
to near real-time implementation. Intern. J. Computer Vision,8 (1),71–
91.
Miller, D.P. (1992). Reducing software mass through behavior control. In
Proc. SPIE Conf. Cooperative Intelligent Robotics in Space III (Vol.
1829, pp. 472–475). Boston, MA.
Miller, D.P., Desai, R.S., Gat, E., Loch, J., & Ivlev, R. (1992). Reactive
navigation through rough terrain: experimental results. In Proc. Tenth
National Conf. on AI (pp. 823–828). San Jose, CA.
Moravec, H.P., & Elfes, A. (1985). High resolution maps from wide angle
sonar. In IEEE Conference on Robotics and Automation (pp. 116– 121).
Washington, DC.
Parker, L.E. (1994). Heterogeneous multi-robot cooperation. Technical
Report 1865. Cambridge, MA: MIT.
Pirjanian, P., & Christensen, H.I. (1997). Behavior coordination using
multiple-objective decision making. In P.S. Schenker & G.T. McKee
(Eds.), Proc. SPIE Symposium on Sensor Fusion and Decentralized
Control in Autonomous Robotic Systems (pp. 78–89). Pittsburgh, PA.
Pomerleau D.A. (1991). Efficient training of artificial neural networks for
autonomous navigation. Neural Computation,3, 88–97.
Prescott, T., & Mayhew, J. (1992). Adaptive local navigation. In A. Blake
& A. Yuille (Eds.), Active vision. Cambridge, MA: MIT Press.
Recce H., & Harris K.D. (1996). Memory for places: a navigational model
in support of Marr’s theory of hippocampal function. Hippocampus,6
(6), 735–748.
Rosenblatt, J.K., & Payton, D.W. (1989). A fine-grained alternative to the
subsumption architecture for mobile robot control. In Proc. IEEE/INNS
Joint Conf. on Neural Networks (pp. 317–324). Washington, DC.
Roumeliotis, S.I., & Bekey, G.A. (1997). An extended Kalman filter for
frequent local and infrequent global sensor data fusion. In P.S. Schenker
& G.T. McKee (Eds.), Proc. SPIE Symposium on Sensor Fusion and
Decentralized Control in Autonomous Robotic Systems (pp. 11–22).
Pittsburgh, PA.
Samsonovich, A., McNaughton, B.L., Gothard, K., & Skaggs, W.E. (1995).
Hippocampus as a path integrator. Poster presented at the XXVth
Annual Society for Neuroscience Meeting, San Diego, CA.
Simmons, R., Krotkov, E., & Bares, J. (1991). A six-legged rover for
planetary exploration. In Proc. AIAA Computing in Aerospace 8 (pp.
739–747).
Soltys, M., & Huntsberger, T.L. (1993). Inverse kinematic analysis of
articulated figures using the fuzzy self-organizing feature map. In
C.H. Dagli, L.I. Burke, B.R. Fernandez & J. Ghosh (Eds.), Intelligent
engineering systems through artificial neural networks (Vol. 3, pp.
541–546). ASME Press.
1509
T. Huntsberger, J. Rose/ Neural Networks 11 (1998) 1497–1510
Sutherland K.T., & Thompson W.B. (1994). Localizing in unstructured
environments: dealing with errors. IEEE Transactions on Robotics
and Automation,10 (6), 740–754.
Sutton R.S. (1988). Learning to predict by the method of temporal
differences. Machine Learning,3, 9–44.
Tani J., & Fukumura N. (1994). Learning goal-directed sensory-based
navigation of a mobile robot. Neural Networks,7 (1), 553–564.
Touretzky, D.S., & Redish, A.D. (1995). Landmark arrays and the hippo-
campal cognitive map. In L. Niklasson & M. Boden (Eds.), Current
Trends in Connectionism—Proceedings of the 1995 Swedish
Conference on Connectionism. Hillsdale, NJ: Erlbaum.
Tyrrell T. (1993). The use of hierarchies for action selection. Journal of
Adaptive Behavior,1 (4).
Whittaker, R., Bapna, D., Maimone, M., & Rollins, E. (1997). Atacama
desert trek: a planetary analog field experiment. In Proc. Intern.
Sympos. on Artificial Intelligence and Automation for Space. Tokyo,
Japan.
Zalama E., Gaudiano P., & Coronado J.L. (1995). A real-time,
unsupervised neural network for the low-level control of a mobile
robot in a nonstationary environment. Neural Networks,8 (1), 103–123
1510 T. Huntsberger, J. Rose/ Neural Networks 11 (1998) 1497–1510
