Obstacle Avoidance and Safeguarding for a Lunar Rover

Reid Simmons, Lars Henriksen, Lonnie Chrisman and Greg Whelan

School of Computer Science/Robotics Institute
Carnegie Mellon University
Pittsburgh, PA 15213

Abstract reliable, goal-driven navigation. The idea is to take

advantage of the human’s common sense and long-range
We are developing techniques for safeguarding the remote planning capabilities and the rover’s ability to sense and
operation of lunar rovers. This paper presents two react quickly and dependably. The idea, which we call
complementary techniques: One, based on stereo vision, “safeguarded teleoperation,” [9] is to let the humans guide
evaluates the traversability of paths the rover could follow, the rover, but have software running on-board that
and produces preferences for steering directions. The other, safeguards the vehicle by preventing dangerous movements,
based on laser proximity sensing, looks for hazards or biases the vehicle's actions towards more easily
immediately in front of the rover, commanding an traversable areas of the terrain.
emergency stop if any are detected. The stereo-based
technique provides reliable obstacle avoidance, but operates Our implementation approach is systemic and layered. By
fairly slowly, while the laser-based technique operates systemic, we mean that we are building a complete,
faster and is more conservative in its evaluations. The integrated robot system — from real-time control to user
stereo-based obstacle avoidance planner has been used to interface. By layered, we mean that higher-level system
drive a rover over ten kilometers in outdoor, natural terrain. functionality is built on top of lower layers. For example,
The laser proximity system has been tested, and is currently the lowest level — the real-time control — accepts
being integrated with the rest of the rover system. commands in the form of steering angle and velocity. Local
obstacle avoidance uses this layer to autonomously traverse
terrain. The safeguarded teleoperation combines the local
1. Introduction obstacle avoidance and user interface layers to produce
safe, reliable navigation. Advantages of the layered
approach are increased flexibility (the system can be
We are investigating techniques to help humans operate
commanded at any layer) and increased reliability (if
rovers on the Moon. Our work belongs to a larger Lunar
designed correctly, the reliabilities of each layer
Rover Initiative, which aims to conduct a lunar mission,
complement one another).
sponsored by private ventures, dedicated to returning to the
Moon before the turn of the Millennium [3]. The mission
Central to our approach is on-board software to sense and
would have rovers navigate hundreds, if not thousands, of
react to terrain features. The work reported here consists of
kilometers over several years, visiting sites of geological
two techniques: one for local obstacle avoidance based on
and historical interest. The research reported here involves
stereo vision, and one for hazard detection based on a laser
techniques for planning rover motions that are (locally) safe
proximity rangefinder.
and easily traversable.
The rationale for using two techniques based on different
Our motivation starts with the observation that teleoperation
sensing modalities is that they effectively complement one
of mobile robots is often fatiguing and disorienting for
another. Stereo-based vision provides a relatively wide,
operators. This is especially true for remote Lunar driving,
medium-range view of the terrain (three to seven meters in
in which the environment is foreign, and operators would
front of the vehicle), but is rather slow (less than 1 Hz) and
have to contend with up to a five second communications
has only fair resolution (5-10 cms). The local obstacle
time delay. While an alternative scenario is to have the
avoidance planner that uses the stereo data is rather
rover drive itself autonomously, the ability to teleoperate
sophisticated, and is used to make decisions about where the
the vehicle constitutes much of the appeal of the mission.
rover should, and should not, be heading. The laser-based
Furthermore, the current state-of-the-art is not reliable
proximity sensor provides high-resolution (under a
enough to enable the robot to make correct decisions in
centimeter) data at a fast rate (minimum of 4 Hz), but in a
every conceivable situation.
relatively narrow band close to the front of the vehicle (100-
150 cms). The hazard detection software uses simple,
Our philosophical approach is to combine the relative
conservative heuristics to determine if a variety of hazards
strengths of the human operator and the rover to produce
are present, and acts to stop the vehicle in an emergency.
Although we presume that the stereo system will keep the left and right body segments. This articulation enables all
rover out of most hazardous situations, the laser system acts four wheels to maintain ground contact even when crossing
as a short-range backup. The combination increases the uneven terrain, which increases the Ratler’s ability to
overall reliability of the navigation system, and increasing surmount terrain obstacles. The body and wheels are made
our confidence that the rover will not accidentally drive of a composite material that provides a good strength-to-
over a cliff or become stuck. weight ratio.

Both techniques have been implemented and are being Sensors on the Ratler include wheel encoders, turn-rate
tested on a prototype wheeled rover. In particular, in one gyro, a compass, a roll inclinometer, and two pitch
experiment the stereo-based local obstacle avoidance inclinometers (one for each body segment). There is a color
system was used to drive the rover safely over ten camera for teleoperation, and we have added a camera mast
kilometers in outdoor, natural terrain. We are currently and four black-and-white cameras for stereo vision, and an
integrating the laser-based hazard detection subsystem, and Accuity laser proximity rangefinder.
will test the complete system by traveling autonomously
over greater distances and rougher terrain. Figure 2 presents a block diagram of the overall navigation
software system. The real-time controller handles servoing
of the motors, collecting and processing of the internal
sensor signals (encoders, compass, inclinometers), and
dead-reckoning calculations [4]. It runs on a 286 and a 486
CPU board, connected by an STD bus, and communicates
with the rest of the system via serial link. The laser
subsystem (Section 4) also runs on-board, on another 486
board.

The controller module (Figure 2) transforms higher-level

commands (steering angle and velocity) into the lower level
commands (individual wheel velocities) used by the real-
time controller, and transforms raw sensor signals into more
familiar units (radians and meters). The stereo and obstacle
avoidance planner modules work together, taking pairs of
images and producing recommendations on which paths are
best for the rover to traverse. The arbiter module combines
information from the planner and user interface subsystems
to select paths that satisfy both user preferences and vehicle
safety [7, 8]. Each module is a separate process, running
concurrently, and communicating with one another via
Ethernet, using the message passing protocol of the Task
Control Architecture [10]. Currently, these modules run
off-board, on two Sparc 10 workstations, but we are in the
process of porting them (except for the user interface) to run
on-board, on two Pentium processors running Linux.

Figure 1. The Ratler Rover

2. The Rover And Its Navigation System

While we await the completion of our new lunar rover [1],
we are using a vehicle designed and built by Sandia
National Laboratories [5] as a testbed to develop the remote
driving techniques needed for a lunar mission. The Ratler
(Robotic All-Terrain Lunar Exploration Rover) is a battery-
powered, four-wheeled, skid-steered vehicle, about 1.2
meters long and wide, with 50 cm diameter wheels (Figure
1). The Ratler is articulated, with a passive axle between the
 The Ranger algorithm works by analyzing the paths the
vehicle would traverse along the terrain for a number of
different steering angles, and choosing the one that
evaluates as the safest. It merges individual stereo-
produced elevation maps to create a 25 cm resolution grid
map up to seven meters in front of the rover. Map merging
is necessary because the limited fields of view of the
cameras do not allow a single image to view sufficient
terrain.

Ranger then projects the rover’s state (position, roll, pitch,

yaw) as it travels along a path. The projection is based on a
desired steering angle, the vehicle dynamics, and the
underlying terrain. The vehicle’s current pose, its dynamics
and the steering angle are used to determine the position
Figure 2. Navigation System Block Diagram and yaw of the vehicle at the next time step. The height of
3. Stereo-Based Obstacle Avoidance the terrain under the wheels is then used to determine the
roll and pitch of the vehicle at that point.
3.1 Stereo Vision
Once a projection of the vehicle along a path has been
The local obstacle avoidance planner uses stereo-based computed, the vehicle state at each point in time is
terrain elevation data to determine safe paths for the rover evaluated. Four criteria are used to determine the
to travel. The stereo module takes its input from black-and- “goodness” of a path: roll, pitches of the left and right body
white CCD cameras, mounted on a motion-averaging mast segments of Ratler, and number of known terrain points the
(Figure 1). The camera images are first rectified to ensure vehicle crosses along the path. If any of the criteria exceed
that the scan lines of the image are the epipolar lines [6]. a given threshold of safety (such as excessive roll or pitch),
The best disparity match within a given window is then the whole path is given a very low evaluation. Otherwise,
computed using a normalized correlation. Disparity the criteria are normalized to the range [0..1] and are
resolution is increased by interpolating the correlation combined using a linear weighted function. This
values of the two closest disparities. Various heuristics are determines the overall merit of choosing that steering angle
employed to minimize outlier values (caused by false stereo for the rover. These evaluations are then combined with the
matches), for example, by eliminating low-textured areas user's preferences to determine the overall best command,
using lower bounds on the acceptable correlation values and which is then sent to the rover to be executed. The cycle
variance in pixel intensity [4, 8]. time for this process is about 1-2 seconds, with the stereo
computations taking up about 75% of the time.
The output of the stereo subsystem are sets of (x, y, z)
triples, given in the camera coordinate frame, along with the 3.3 The Morphin Planning Algorithm
pose of the robot at the time the images were acquired.
Using the pose information, the obstacle avoidance planner While the Ranger algorithm has worked well for high-speed
transforms the (x, y, z) values into world coordinates to navigation of Humvees, it is not entirely well-suited to the
form a (non-uniformly distributed) terrain elevation map. much smaller, and slower, lunar rover. As is often the case
in robotics, the problems are mainly attributable to an
To make the stereo computation tractable, the planner abundance of noise, particularly in the stereo-produced
requests only a small segment of the stereo image (about terrain maps and the dead-reckoning. The main effect of
2%), at reduced resolution (every fifth row and column). the noise is to make it difficult to cleanly merge terrain
Experiments show that this is sufficient to reliably detect maps acquired from separate images, which is required by
features on the order of 20 cm high. the Ranger algorithm since it uses only a small percentage
of each image. Map merging often produced artifacts in the
3.2 The Ranger Planning Algorithm map, such as crevasses and ridges, which the rover would
refuse to cross. This is less of an issue with the Humvees,
Our first local obstacle avoidance planner was an adaptation since they can cross much taller obstacles. We tried several
of a planner, called Ranger, that was developed at CMU for merging techniques in an attempt to minimize the artifacts,
ARPA’s Unmanned Ground Vehicle (UGV) program for but none was robust enough to yield consistent driving
cross-country navigation [2]. This planner enabled the results.
rover to travel up to a kilometer in mild terrain [4, 8, 9].
Another effect of noisy terrain data is that, because of the The former indicates the roughness of the overall area,
rover’s relatively short wheelbase, small deviations in while the latter indicates if the patch is bumpy/spiky.
perceived terrain elevation under the wheels produced Finally, two factors are used to assess the confidence in the
relatively large changes in estimated roll and pitch. For evaluations: the number of stereo-generated terrain points in
example, a 20 cm “spike” in the terrain map (not a patch and the spatial distribution of these points (based on
uncommon) produces a 13 degree change in pitch, given a an entropy-like measure), which is used to ensure that the
90 cm wheelbase. Thus, it is often difficult to distinguish points are representative of the patch as a whole.
noise from steep bumps. This same problem makes it
difficult to reliably determine whether high-centering might Morphin then projects the path of the rover over the terrain
occur, since the clearance of the rover is not much more patches. Unlike Ranger, which uses a discrete numerical
than the noise in the map. Finally, the Ranger algorithm simulation to project paths, Morphin uses closed form
presumes that the rover can track the path exactly, and does solutions to calculate the intersections between arcs of a
not account for uncertainty in the execution of commands or circle and the terrain patches. Morphin then sums the
for uncertainty in the vehicle dynamics models used to traversability metrics of the intersecting terrain patches,
project paths. weighted by the length of the intersection between the arc
and terrain patch. For each patch, Morphin determines roll,
To address these problems, we modified parts of the Ranger pitch, roughness, and confidence in the data. The pitch of
algorithm, creating an algorithm called Morphin (a “power” the vehicle is easily calculated as the slope of the line along
Ranger). In contrast to the path-based approach of Ranger, the plane in the direction of the current heading (yaw). A
Morphin is area-based: it analyzes patches of terrain to similar calculation yields the vehicle roll. The roughness
determine the traversability of each patch, and evaluates the and confidence measures are calculated as described above.
traversability of a path by determining the set of patches it If there are overlapping patches from different images, only
travels through. As such, it is more akin to the terrain the one associated with the most recently acquired image is
navigation planner of [11]. used (given the dead-reckoning uncertainty of the rover, we
find this to be much more effective than trying to combine
overlapping evaluations in some way). Then, as with
Ranger, the criteria are combined using a linearly weighted
function.

While the path projection approach of Ranger (numerical

simulation) produces higher fidelity paths (since dynamics
and the effects of moving on uneven terrain can be taken
into account), Morphin's geometrical approach is much
more efficient, and is adequate for the task since the rover's
dead-reckoning is not accurate enough to warrant a high
fidelity approach. In fact, we are extending Morphin to
explicitly deal with the uncertainty in the rover's heading:
for each nominal steering angle, we project a number of
Figure 3. Local Obstacle Avoidance Planner
paths (currently five) that differ slightly in the steering
angle. The evaluation for each of these paths is weighted
Each local terrain map, produced from a single stereo pair,
by the probability of the rover following that path (under an
is analyzed independently. The terrain is divided into
assumption of Gaussian distribution from the nominal
overlapping patches, each 125 cm on a side, with patches
steering angle).
offset by 25 cm from one another. Thus, each terrain point
contributes to 25 patches (Figure 3).
3.4 Performance
To determine traversability, a plane is fit to each patch
using least-squared error. To avoid redundant computation, To evaluate the strengths and weaknesses of the stereo-
statistics (e.g., sum of X, sum of XY) are collected for based approach, we performed extensive field trials. The
smaller 25 cm squared patches and then aggregated to test site (Figure 4) consists of soil, crumbled asphalt, loose
determine the plane parameters for each 125 cm squared gravel, scree, and some grassy vegetation. Obstacles to
patch. The plane parameters are used in determining the rover passage include soil mounds, depressions, cliffs at the
vehicle roll and pitch (see below), and the residual to the river bank, building walls, metal pipes, cement blocks,
plane fit is used to estimate the roughness of the terrain. railroad ties, trees, and bushes.
Two “roughness” measures are computed: one based on the
residual in fitting the plane to the whole patch, and one
based on the residual of each small (25 cm squared patch).
 front of the rover. The scanner can produce data at various
rates, depending on the number of samples per scan and the
required precision. In the runtime configuration, a scan is
available every 25-50 msecs.

An on-board computer collects the range and angle readings

Figure 4. Terrain For Rover Navigation Experiments
and tests them for validity. The data is then linearized and
In one particular experiment, the rover traveled more than
transformed to obtain an array of (x, y, z) triples of the
10 km over a three-day period. During the experiment, the
terrain with respect to the rover’s local coordinate frame
rover operated autonomously over 98% of the time,
(i.e., this transformation does not adjust for the angular
successfully avoiding discrete obstacles, while averaging a
inclination of the vehicle). The resulting laser data are
speed of 15 cm/sec. This is an order of magnitude farther
processed, looking for evidence of depressions/drop-offs
than we were able to traverse with the Ranger algorithm,
and obstacles that might lead to the vehicle being stuck
and needed about one-third the amount of teleoperated
when attempting to drive over them. When such hazards are
control. This experiment demonstrated the superiority of
detected, the subsystem issues an emergency stop command
the Morphin algorithm for our rover. Morphin addresses
to the vehicle and notifies the local obstacle avoidance
the problem of noisy data by aggregating independent data
planner (Morphin) of the hazard, so that it can incorporate
points into an overall statistic, thus dramatically lessening
that information into its planning.
the impact of any single point. While this aggregation can
sometimes cause the rover to behave more conservatively
As a baseline configuration the position information
than would otherwise be warranted, in our application it is available on the rover is not incorporated into the detection
better to be too conservative than to allow the rover to head
of the hazards. This frees the laser safeguarding system
into danger.
from dependence on the controller module maintaining
accurate dead-reckoning information, and hence makes it
less dependent on sensor failures (encoders, compass).
4. Laser-Based Hazard Detection Even if all other navigation systems should fail the rover
can still be controlled safely in teleoperated mode by the
While the stereo-based planner is fairly reliable, there are laser safeguarding. Besides being robust, this configuration
several hazards that it has trouble detecting. The major involves less processing, which leads to a faster update rate.
weakness is that the stereo vision often cannot detect While incorporating pose information is an option, and
depressions/craters, reporting them as unknown areas. In might possibly produce better hazard detection, doing so is
addition, the limited resolution of the stereo, combined with non-trivial since the desired cycle and reaction times of the
the large look-ahead distance (three to seven meters) means laser subsystem are considerably smaller than the inherent
that small obstacles (on the order of 10-20 cms) may be time constants of the inclinometers. Our approach instead
overlooked. These can cause problems if the rover tries to aims at identifying statistics of the scans that are insensitive
straddle them, which can cause high-centering (hitting the to sudden (and unknown) inclinations. By using these, we
bottom of the vehicle). have found this baseline configuration to be sufficient for
all but the most extreme rover configurations.
To detect such hazards, we have developed a hazard
detection technique that uses a high-resolution, laser
proximity sensor. The requirements for this subsystem are 4.2 Data acquisition
that it must be very robust in detecting hazards and have
very good response time. These requirements have driven The first step in processing a laser scan is to determine the
the design and implementation of the laser-based integrity of the laser system and to perform self-diagnostics,
safeguarding system. if necessary. The next step is to remove invalid data and to
determine if the spatial density of the remaining data is high
enough to reliably calculate the hazard metrics.
4.1 System configuration
These calculations use a number of laser ranger sensor
The sensor, an Acuity 3000-LIR laser ranger, sends a beam signals: absolute encoder, incremental encoder, range,
towards a rotating mirror projecting a plane of infrared laser temperature, data out of range, buffer overflow, intensity of
light at a 45 degree angle to the ground. It is able to image reflected laser light, and ambient light. First the motor
the ground with a resolution of under a centimeter in all subsystem is checked through a test of correct motion of the
three dimensions at a range of about 100-150 cm in front of mirror. This is done using three measures:
the rover. The effective field of view is limited by the • Is zero pulse captured? (absolute encoder)
effective angle of incidence and is, in practice, about 90 • Full cycle loaded? +/- 45 degrees in front of the vehicle
degrees producing a 4 m long laser line on the ground in captured.
 • Is motor spinning? (incremental encoder)

The zero pulse of the absolute encoder synchronizes the

4.3 Hazard detection metrics
angles captured by the incremental encoder. If this pulse is
missed, the absolute orientation of the sweep is unknown Since the laser line hits the ground fairly close to the vehicle
and the data is of no value. Both the capture of the (100-150 cm), detection must be made quickly in order to
synchronizing zero pulse and a successful acquisition of a react in time. For this reason, we have chosen to define
full cycle depends on the speed of the mirror. If the mirror simple heuristic metrics for each type of hazard that we
is spinning too fast, the zero pulse may be missed and, if want the laser to detect. These metrics are defined in terms
spinning too slowly, a full range may not be available of a single scan of the proximity sensor, so that no
within the number of samples recorded. As the mirror information needs to be saved between scans (increasing
system has relatively slow dynamics, the system is designed robustness and decreasing computation).
so that the zero pulse or full angle measures have to fire a
number of times before the spinning of the motor is tested. When designing the metrics two approaches were
This avoids erroneous fault detection during start-up and considered. One approach evaluates whether the elevation
temporary disturbances. of the surface in front of the rover (represented in the
rover’s local coordinate frame) exceeds the capability of the
In addition to determining whether the mirror is spinning rover. While this approach is fairly general and
correctly, a check is made of the motor temperature and computationally very simple, it has the problem that the
whether there are internal errors (e.g., buffer overflow) on apparent elevation of the terrain in front of the rover is a
the SCSI interface board (which indicates that samples have function of both the actual terrain height and the rover’s
been lost). Finally the system assesses whether the density current inclination (e.g., if the front wheels of the rover are
of reliable data is sufficient. A common problem is that the on small rocks, the elevation of the terrain one meter in
laser beam hits a terrain point which does not reflect enough front of the rover appears lower than it actually is). Thus,
light to make an accurate range estimation. This can be due while true hazards will be detected reliably and quickly,
to the angle of incidence, non-diffuse reflection, or a low there are situations where potential hazards will be detected
reflectance of the object being measured (dark surface). erroneously, and the vehicle will be stopped unnecessarily.
This results in an unreliable datum, which can confound
subsequent processing. A dependable way to detect zones of The other approach involves identifying signatures of
unreliable data is high variance between adjacent range different landscape formations that are invariant to the
readings. motions that occur when driving over minor obstacles. For
example, when obliquely approaching a downward slope,
All checks, except for the variance in range estimates, are the range measurements will gradually increase starting at
very fast as their input are direct sensor signals, which are the point where the laser line intersects the beginning of the
more or less dedicated for integrity analysis purposes. Only slope, forming an “elbow bend.” This characteristic shape is
the high-variance test needs a non-trivial amount of evident regardless of whether the front of the rover is
computation to determine status. In any event, data elevated by a rock, and so is less likely to detect hazards
acquisition is fast: including integrity checking and data erroneously. However, in the signature approach it is
testing, it can be done in about 180 msecs (including 35 difficult to quantify the danger a profile constitutes to the
msecs for the laser to generate range data). vehicle. For example, when approaching a minor
downwards slope from different angles, the shift in range
When a problem occurs, corrective action is necessary. For varies and so the steepness of the slope cannot be known.
some of the very low level problems, like mirror motion, Thus, it is difficult to quantify what constitutes a real
appropriate actions can be directly associated with the hazard. In addition, in the signature approach much more
problem. In the case of mirror motion problems, new scans processing has to be performed, as the number of possible
are commanded to see if the problem was just a result of landscape feature signatures is relatively large compared to
spurious unfavorable conditions. For other problems, such the number of rover limitations (see Table 1).
as high temperature, different actions can be taken
involving other systems of the rover (like applying extra Table 1. Hazards to be detected
cooling, shutdown or seeking shade). Since other Rover limitation Landscape danger Importance
subsystems may also be affected by these kind of problems, Positive elevation Small, medium Less important.
in most cases the laser subsystem will just discard the data (step) and large rocks
as invalid, and leave it to other systems to correct the Stereo is reliable
problem. Step in landscape
(broken rock
surface)
Boulders
Negative Ditch Important.
elevation (ditch)
Craters Stereo has poor
performance
Step in landscape
(broken rock
here. Front
surface) View
Stuck on belly Objects on cross Equally good
slopes

Another problem with the interpretation approach is that the

set of features may not cover all possible landscapes Side
encountered. Hence, safe operation would not be View
guaranteed. To ensure safety (at the cost of sometimes
stopping erroneously), it was decided to employ the direct
method based on the capabilities of the rover. Three hazard Figure 5. Physical metrics and their corresponding hazards
types are considered:
• Maximum traversable step (curb-like, head on) As an example, Figure 6 shows the interpretation of a
• Maximum traversable ditch (curb-like, head on) typical scene. The elevation profile is inclined to the left,
• Belly clearance relatively flat, and shows a small mound at y=-1m. The two
dashed lines indicate the step and ditch thresholds. For
As the metrics are defined in terms of a single scan, no y>0.5 the step metric has detected a hazard (denoted by
information is available about the transition from the “o”s). No belly hazard is detected.
surface currently under the rover to the scanned surface at
the laser line. The transition must therefore be treated as a
worst case, which is a step-like transition at the laser line.
Also, since the laser subsystem does not know the current
vehicle steer angle, to be safe it must analyze the complete
laser line. For the step and ditch metrics, this translates into
defining a simple upper and lower threshold (respectively)
directly on the 3D elevation profile (Figure 5). The
thresholded data is spatially filtered to prevent spurious
signals from firing the metric. A median filtering is used,
which is quite fast since it operates in the binary domain.

The belly hazard metric first estimates the slope by linear

regression and then equalizes the elevation profile
accordingly, yielding a level elevation profile centered
around zero elevation. Based on the minimum and
maximum elevation in this compensated profile, the most
favorable levels of a positive and a negative threshold is
computed (difference between the two levels is the body Figure 6. Hazard detection. The elevation
clearance minus a margin). The compensated elevation profile is seen in the direction of travel from the vehicle.
profile is then tested for exceeding the elevation band The laser is positioned at y=0, z=1.
defined by the two threshold levels and this output is
filtered spatially as for the step and ditch metrics. Since this
metric is more computationally expensive than the other 4.4 Performance
two, it is processed last (and only if the other two do not
fire). We have implemented the integrity checks and hazard
detection metrics described above, and are currently
running experiments to characterize their performance.
Preliminary indications are that laser proximity
safeguarding will be a very valuable supplement to the
overall navigation system. In terms of missed hazards, the
performance is excellent. Some false detections are
encountered, mainly due to specular reflecting surfaces and
small angles of incidence. This is, however, very dependent
on the scene used for testing and the pose of the laser
scanner. Some of these problems can thus be overcome by
placing the sensor in a more favorable location.
Acknowledgments

The cycle time is currently about 4 Hz on a 66Mhz 486 in Many thanks to other members of the Lunar Rover
the test configuration. This includes no effort for optimizing Navigation team, in particular Eric Krotkov, Martial
the algorithms, relatively dense sampling, and a high range Hebert, Fabio Cozman, Richard Goodwin and Sven Koenig.
precision, which requires more time by the laser (currently This work has been partially supported by NASA, under
15% of the total time). It is expected that the speed can be grants NAGW-3863 and NAGW-1175.
increased considerably without significant loss of detection
reliability by streamlining code and reducing range References
precision to more realistic values.
[1] L. Katragadda. J/ Murphy, W. Whittaker and D. Gump.
An obvious extension would be to incorporate information Commercial lunar rover mission for science and
about vehicle movements and the planned path. This would edutainment, In Proc. AIAA Forum on Advanced
enable the metrics to evaluate each point on the elevation Developments in Space Robotics (this volume), Madison,
profile in accordance with the specific parts of the rover that WI, August 1996.
will be passing that point. However, the add-on should be
kept apart from the base-line system so that there always is [2] A. Kelly. An intelligent predictive control approach to
a working backup if the required information from the rover the high speed cross country autonomous navigation
should become unavailable. problem. Technical Report CMU-CS-TR-95-33, Carnegie
Mellon University, 1995.

5. Conclusions [3] E. Krotkov, J. Bares, L. Katragadda, R. Simmons, and

R. Whittaker. Lunar rover technology demonstrations with
This paper has presented an integrated approach to Dante and Ratler. In Proc. Intl. Symp. Artificial
safeguarded navigation for lunar rovers. The key idea is to Intelligence, Robotics, and Automation for Space, Jet
combine multiple techniques, using different sensing Propulsion Laboratory, Pasadena, California, October 1994.
modalities, to increase the reliability of the overall system.
In particular, we presented two complementary techniques: [4] E. Krotkov, M. Hebert, and R. Simmons. Stereo
stereo-based vision for obstacle avoidance and laser-based perception and dead reckoning for a prototype lunar rover.
hazard detection. Autonomous Robots, 2(4):313-331, December 1995.

The stereo-based approach allows the rover to actively [5] J. Purvis and P. Klarer. RATLER: Robotic all terrain
change its steering angle to avoid obstacles in the mid-range lunar exploration rover. In Proc. Sixth Annual Space
(3-7 meters in front of the rover), but it is rather slow, and Operations, Applications and Research Symposium,
can miss certain types of hazards, such as depressions. The Johnson Space Center, Houston TX, 1992.
laser-based subsystem is much simpler and faster, analyzing
high-resolution data immediately in front of the rover (100- [6] L. Robert, M. Buffa and M. Hebert. Weakly-calibrated
150 cms), but can only command the robot to stop. stereo perception for rover navigation. In Proc. Image
Understanding Workshop, 1994.
We anticipate that the combination of these two techniques
should give very good performance over wide range of [7] J. Rosenblatt and C. Thorpe. Combining multiple goals
terrain. Having tested the stereo and obstacle avoidance in a behavior-based architecture. In Proc. Intl. Conf.
planner extensively (in one experiment autonomously Intelligent Robots and Systems (IROS), 136-141, Pittsburgh,
traversing 10 kilometers of natural terrain), we know the PA, August 1995.
types of terrain where its performance is weak, and have
designed the laser proximity system specifically to address [8] R. Simmons, E. Krotkov, L. Chrisman, F. Cozman, R.
those weaknesses. Goodwin, M. Hebert, L. Katragadda, S. Koenig, G.
Krishnaswamy, Y. Shinoda, W. Whittaker, and P. Klarer.
We are currently working to integrate the laser-based Experience with rover navigation for lunar-like terrains. In
hazard detection component into our complete navigation Proc. Intl. Conf. Intelligent Robots and Systems (IROS),
system, and to fine tune the various metrics. The next step 441-446, Pittsburgh, PA, August 1995.
is to demonstrate that the integrated system will enable the
rover to navigate extremely reliably in rougher and more [9] R. Simmons, E. Krotkov, L. Chrisman, F. Cozman, R.
varied terrain. Goodwin, M. Hebert, G. Heredia, S. Koenig, P. Muir, Y.
Shinoda, and W. Whittaker. Mixed-mode control of
navigation for a lunar rover. In Proc. SSI/Princeton Space
Manufacturing Conference, Princeton, New Jersey, May
1995.

[10] R. Simmons. Structured control for autonomous robots.

IEEE Transactions on Robotics and Automation, 10:1, Feb.
1994.

[11] A. Stentz. The NAVLAB system for mobile robot

navigation. PhD thesis, Carnegie Mellon University, 1989.