2007 IEEE International Conference on FrE10.

4
Robotics and Automation
Roma, Italy, 10-14 April 2007

Increased Mars Rover Autonomy using AI Planning,

Scheduling and Execution
Tara Estlin, Daniel Gaines, Caroline Chouinard, Rebecca Castano,
Benjamin Bornstein, Michele Judd, Issa Nesnas, and Robert Anderson

Abstract—This paper presents technology for performing navigate through unknown terrain.
autonomous commanding of a planetary rover. Through the Some autonomous capabilities for rovers have been well
use of AI planning, scheduling and execution techniques, the tested on the Mars surface. For example, the MER
OASIS autonomous science system provides capabilities for the navigation algorithm has been used to drive the rovers up to
automated generation of a rover activity plan based on science
140 meters in one sol. Future rover missions will likely
priorities, the handling of opportunistic science, including new
science targets identified by onboard data analysis software, continue to use this capability and increase allowable
other dynamic decision-making such as modifying the rover distances for autonomous driving. This increase will also
activity plan in response to problems or other state and create new requirements for other onboard autonomy
resource changes. We first describe some of the particular software that can closely monitor rover state and resources
challenges this work has begun to address and then describe (to ensure both the safety of the rover and the most efficient,
our system approach. Finally, we report on our experience
testing this software with a Mars rover prototype.
use of such resources). Further onboard software will need
to react to unforeseen opportunities during drives that could
I. INTRODUCTION be missed if the rover could not handle opportunistic science
events. Onboard data analysis is already being used on MER
T HE Mars Exploration Rovers (MER 2003) have traveled
many kilometers over the Martian soil and have
outperformed all expectations by lasting an order of
to automatically detect dust devils and other work is being
done to develop analysis software that can detect a large
variety of image features [1, 2].
magnitude longer than their original mission goal. Both This paper presents the planning, scheduling, and
rovers have now lasted over 1000 sols (or Martian days), execution component of the OASIS autonomous science
whereas their primary mission was for 90 sols. The system [1]. This component provides capabilities for initial
longevity of these vehicles will have significant effects on plan generation, plan execution and monitoring, and re-
future mission goals, such as objectives for the Mars Science planning in response to new science opportunities and/or
Laboratory rover mission which is scheduled to fly in 2009. changes in rover state or environment. This work has been
Current mission operations for MER require a team of focused on supporting Mars exploration and science goals
mission engineers to plan rover activities for each sol. and specifically on supporting the concepts of opportunistic
Though many planning activities have been significantly science and long-distance traverse. In particular, this paper
streamlined as compared to days early in the mission, the provides an overview of how the OASIS planning and
large majority of rover activities are sequenced manually execution system has been tested both through simulation
with some aid from ground software packages that help and with actual rover hardware in the JPL Mars Yard.
model rover behavior. Numerous tests have been run showing that OASIS can
Common objectives for future rover missions to Mars robustly respond and re-plan for a variety of situations,
include the handling of opportunistic science and long- including the handling of new science goals, dynamic
range, or over the horizon, driving. Though some software changes in resources levels, unknown obstacles, and
already exists onboard the MER rovers to support these autonomous targeting based on data analysis results. OASIS
goals, additional autonomous capabilities will be required. is also integrated with a large set of robotic control software
To handle newly-identified science targets, onboard software that provides Mars mission-relevant capabilities such as
will need to schedule new rover activities based on specific autonomous navigation, vision, and manipulation.
target data and current rover state. To handle long-range Past OASIS publications have given a high-level system
traverse, onboard software will need to closely monitor rover description or details on the data analysis components of
state and resources, as well as dynamically modify rover OASIS [1]. This paper focuses and provides details on the
activity start-times and orderings since rovers will have to planning and execution elements of the system.

Manuscript received September 15, 2006. This work was supported by II. CHALLENGES FOR ONBOARD AUTONOMY
the NASA Mars Technology Program and NASA Intelligent Systems
Rovers that possess extended autonomy capabilities, such
Program.
All authors are with the Jet Propulsion Laboratory, California Institute of as onboard command generation and re-planning, have the
Technology, Pasadena, CA 91109 USA (email: potential for not only supporting the more efficient use of
[Link]@[Link], i.e., [Link]@[Link]).

1-4244-0602-1/07/$20.00 ©2007 IEEE. 4911

 FrE10.4

Another predominant challenge in developing autonomy

software for rover operations is robustly handling the
inherent uncertainty of rover behavior. Since rovers are
typically driving in unknown terrain, important state
knowledge, such as rover position or activity duration, is
often difficult to accurately predict. The MER rovers have
seen wheel slippage cause significant position errors, such as
the 21% error accrued when traversing Eagle crater and the
100% slippage error that occurred when Opportunity got
stuck in a Purgatory dune [3]. Rover resource levels will also
fluctuate based on how certain operations interact with the
terrain, including drives and instrument placements. Some
resources (such as power) may be affected by other
environment variables such as sun angle and atmosphere
dust opacity. The difficulty of operating in unknown terrain
is further compounded by the tight resource and state
constraints a rover faces. For instance, for much of 2006, the
MER Spirit rover had such a low power level (due to the
angle of the sun in Martian winter), that it could only
Fig. 1. View of the Dome Fuji target on the Mars perform a small number of stationary operations per sol.
surface taken by the MER Spirit rover. Onboard autonomy software should also be able to
effectively reason about plan quality and science. Over the
rover resources over longer periods but also for enabling course of a mission, a rover will be asked to perform a large
data collection opportunities that might have never been variety of science operations. Different priorities will exist
realized had the autonomy software not been onboard. New for these measurements and priorities may fluctuate
missions are being designed that will require rovers to throughout the mission. When resources are over-taxed, the
support more autonomous endeavors such as long-range rover should also be capable of making science/resource
traversals, complex science experiments, opportunistic trade-offs in an effort to produce the highest science return.
science handling, and lengthy mission durations. To address Engineering objectives may also exist that should be
these goals, autonomy software designers face a number of considered such as leaving the rover at a desirable tilt for
challenges. In this section, we consider a few key challenges acquiring solar panel energy or including observations that
for using planning and execution techniques to provide take images of rover hardware to help evaluate system state.
extended rover-autonomy capabilities.
To generate or modify a rover activity plan for carrying III. OASIS AUTONOMOUS SCIENCE SYSTEM
out a set of science goals, the onboard planning and To address the issues outlined in the previous section, we
execution software will need to reason about a rich model of have developed a planning, scheduling and execution system
resource and temporal constraints. For example, it will need that can autonomously command a planetary rover. This
to predict power consumption of variable duration activities software is part of a larger framework called OASIS
such as downlinks and traverses, keep track of available (Onboard Autonomous Science Investigation System) [1],
power levels, and ensure that generated plans do not exceed which is shown in Figure 2. OASIS provides a number of
power limitations. capabilities for performing rover autonomous science,
Sequence generation for rover missions raises a number of including image and data analysis, planning, execution and
interesting challenges regarding spatial reasoning. One of interaction with robotic control. This paper focuses on the
the dominating characteristics of rover operations is drives planning, scheduling and execution element of this
to designated waypoints and science targets through either framework. Other components will only be briefly
partially-known or unknown terrain. (Figure 1 shows a view described, but are further detailed in related publications.
of Martian terrain captured during early parts of the MER
mission.) Onboard planning and execution needs to A. OASIS Planning, Scheduling and Execution
effectively coordinate with several layers of navigation and In the OASIS framework, planning, scheduling and
other spatial reasoning software in order to generate correct execution techniques are used to provide rover-plan
and efficient rover plans. This coordination typically generation, execution, monitoring, and re-planning in
includes querying a path planner for route options, using response to problems and new science opportunities. To
position estimators to track rover progress and re-plan if provide these capabilities, OASIS closely integrates the
progress is impaired, and monitoring navigation and obstacle CASPER (Continuous Activity Scheduling, Planning,
avoidance software that may cause the rover to move off the Execution and Re-planning) planning system and the TDL
predicted route. (Task Description Language) executive system. The
following capabilities are provided:

4912
 FrE10.4

Science
Goals

Data Analysis Planning / Scheduling / Execution

Target
Novelty Optimize Repair
Signature
Rock Science
Sampling Alerts Execution

Rock Features Commands State/Resource

Updates

Visual Path
Albedo Shape Navigation Estimation
Texture Data
Planning

Rock Vision Locomotion …

Detection
Control / Functional Layer
Feature Extraction (provided by CLARAty)

Fig 2. OASIS (Onboard Autonomous Science Investigation System) framework. This diagram shows how different
decision-making capabilities interact within OASIS (shown in the yellow boxes) and how OASIS interacts with low-
level control software (shown in the gray box).
As this information is acquired, CASPER updates future-
• Automated plan generation based on an input set of plan projections. These updates may cause new conflicts
science and engineering goals
and/or opportunities to arise, requiring CASPER to replan to
• Maintenance of resource, state, and temporal constraints
accommodate the new data. A simple re-planning situation is
• Plan execution through the interaction with basic rover
shown in Figure 3. Here, a rover science activity (of drilling
control functionality
for sample collection) took more power than originally
• Monitoring of plan execution to ensure plan activities
estimated, which causes a resource conflict. CASPER fixes
are successfully executed
the conflict by deleting a low-priority science activity to
• Dynamic modification of the plan based on activity,
ensure that enough time remains to execute a later, high-
state or resource updates or new goal information
priority science activity and a required communications link.
• Plan optimization for reasoning about plan quality
To reason about science goal priorities and other plan
• Automated handling of newly identified science goals
quality measures we use the CASPER optimization
framework to search for a higher quality plan. User-defined
Planning in OASIS is provided by the CASPER
preferences are used to compute plan quality based on how
continuous planning system [4]. For initial plan generation,
well the plan satisfies these preferences. An overall plan
CASPER is given a model of rover operations and an input
score is computed based on the preference specification.
set of science and engineering goals (e.g., take a panoramic
Plan optimization works in an iterative fashion (similar to
image at a certain azimuth and elevation). CASPER
plan repair) and searches for plan modifications that could
generates a sequence of activities that satisfies as many goals
potentially improve the overall plan score.
as possible while obeying relevant resource, state and
Plan execution in OASIS is handled by the TDL execution
temporal constraints. For instance, plans may have strict
system [5]. TDL was designed to handle intermediate
limits on energy usage and science activities may be
control activities for a robotic system and to mediate
required to occur during certain time windows. Plan
between a planning system and low-level robot control
activities are grounded in time, enabling CASPER to quickly
software. It expands activities into lower-level commands,
determine plan conflicts and solutions. Conflicts occur when
executes the commands, and monitors their execution. It also
a plan constraint has been violated where this constraint
provides support for exception handling and fine-grained
could be temporal or involve a resource, state or activity
timing control such as synchronization of subtasks. TDL is
parameter. Plans are produced using an iterative repair
implemented as an extension of C++ and uses a construct
algorithm that classifies conflicts and resolves them
called a task tree to describe the tree structure that is
individually by performing one or more plan modifications.
produced when tasks are broken down into lower-level
To provide re-planning capabilities, CASPER monitors
commands. In OASIS, TDL handles expanding high-level
current rover state and the execution status of plan activities.
activities into lower-level commands (based on current state

4913
 FrE10.4

Traverse1 Traverse2 Traverse3

Comm
Drill1 Image1 Panorama1

Energy
A)
Orientation

Current Time

Traverse1 Traverse2 Traverse3

Comm
Drill1 Image1 Panorama1

B) Energy

Orientation

Current Time

Traverse1 Traverse3
Comm
Drill1 Panorama1

C) Energy

Orientation

Current Time

Fig. 3. Example re-planning scenario. A) The initial plan generated by CASPER which drives the rover to three
locations, performs a science operation at each location, and finally performs a critical end-of-day communications
link with Earth. B) The plan after partial execution. The first science activity has taken more energy than expected,
which has caused several resource conflicts with later activities. C) The plan after re-planning has occurred to resolve
the conflict. The second science operation and drive were dynamically removed so the last (higher priority) science
operation and communication activity can still be achieved.

information), executing commands by dispatching them to of several prioritization algorithms, which analyze the data
functional-level control software, relaying updates on by searching for items such as rocks with features that match
activity status and state/resource data to CASPER, tracking pre-known signatures of interest (as identified by scientists
plan execution progress, and relaying identified problems to on Earth), or novel rocks (i.e., outliers) that have not been
CASPER. seen in past activities. If the analysis component detects new
science opportunities of significant interest, it will generate a
B. OASIS Feature Extraction and Data Analysis
science alert that is sent to the planner, which signifies that
The OASIS Feature Extraction and Data Analysis new measurements should be taken of the identified target.
modules (shown in Figure 2) are responsible for analyzing
onboard image data from rover cameras, identifying C. CLARAty Robotic Architecture
interesting or novel parts of that data, and generating new The systems described in this paper are also integrated
science goals to gather additional measurements when new with the Coupled Layered Architecture for Robotic
science targets have been identified. Analyzed images could Autonomy (CLARAty) [6], which provides a large range of
be taken by either rover science or engineering cameras and basic robotic functionality and simplifies the integration of
could be shot during a variety of rover activities including new technologies on different robotic platforms. Through
obstacle avoidance imaging, high-resolution science CLARAty, the OASIS planning and execution system has
imaging, and imaging done for long-range route planning. been tested with several JPL rover platforms, including the
As shown in Figure 2, new science data is first processed FIDO rover, which is shown in Figure 4.
by the Feature Extraction component. Images are broken To run realistic scenarios with rover hardware, a number
down by first locating individual rocks, and second, by of supporting pieces of software were used. These
extracting a set of rock properties (or features) from each components were provided through CLARAty and could run
identified rock. Extracted rock properties (e.g., shape, on the relevant hardware platforms. This software includes
albedo, visual texture, size) are then passed to the Data the Morphin navigation system, which enables a rover to
Analysis component of the system. This component consists avoid obstacles and navigate to specified waypoints. It also

4914
 FrE10.4

Fig. 4. The FIDO rover, which was used for testing.

includes a position estimation algorithm, which integrates Fig. 5. Image taken in response to a science alert on the
JPL FIDO rover. In this test, science alerts were
IMU (Inertial Measuring Unit) measurements with wheel
generated for “high albedo” (or light-colored) rocks.
odometry to estimate rover position and attitude (roll, pitch
and heading). Other used algorithms include mobility and The first three levels of response (calling Earth, taking
stereo processing as well as control functions for mast additional measurements from the rover current’s position,
pan/tilt and camera operation. and driving closer to the object of interest to take
measurements) have been implemented and significantly
IV. OPPORTUNISTIC TRAVERSE SCIENCE tested with hardware. A sample image taken in response to
a science alert during testing is shown in Figure 5.
To handle opportunistic science during rover traverses, we
enabled the OASIS planner, CASPER, to recognize and V. AUTONOMOUS TARGETING
respond to science alerts, which as described above, are new
science opportunities detected by the onboard OASIS data Another objective for future rover missions is to provide
analysis software. For example, if a rock is detected in autonomous targeting for measurements of science targets
navigation imagery and the rock has a previously unseen that cannot easily be selected in advance. A number of
color or texture, a science alert can be generated to take remote sensing instruments for rovers have a very narrow
additional measurements of that rock. field-of-view (FOV) and thus require very specific targets
CASPER uses its optimization framework to decide how for correct sampling. Examples of such instruments include
to respond to science alerts. Because it may not be possible the MER mission Mini-Thermal Emission Spectrometer (or
to accommodate all alerts (due to other constraints or mini-TES) and the ChemCam spectrometer, which performs
activities), a science alert is represented as an optional goal, Laser-Induced Breakdown Spectroscopy and is planned to
which indicates its achievement is not mandatory but may be flown as part of the Mars Science Laboratory (MSL)
improve the plan’s optimization score if included in the plan. 2009 rover mission. Targeting of these instruments by
Before attempting to handle a science alert, CASPER mission personnel requires a lengthy planning process, since
protects the current plan by saving a copy before images of the targeted areas will need to be first analyzed on
optimization. If the quality has not increased after a certain Earth.
time limit, the previous plan is restored. If CASPER can One scenario suggested for the MSL ChemCam
handle a new science alert (e.g., by adding additional science instrument is to autonomously select and shoot a number of
measurements) without causing other negative affects (e.g., targets selected by onboard analysis software. New
resource over-subscriptions or the deletion of higher-priority measurements would then be autonomously scheduled
science goals) then the new plan that accommodates the onboard the rover. To address this scenario, we applied our
science alert is used. planning and execution system to both execute a set of
Science alerts can have different levels of reaction in
images from which targets could be selected and to
OASIS. The most basic reaction is to adjust the rover plan so
incorporate new science measurements based on results from
that the rover holds at the current position and the flagged
data analysis. Similar techniques could be applied as were
data is sent back to Earth at the next communication
opportunity. The next level of reaction is to collect used to support traverse science, where measurements were
additional data at the current site before transmitting back to only added if they did not disrupt other high priority
Earth. Further steps include having the rover alter its path to activities or require an excess of resources. The OASIS
get closer to objects of interest before taking additional planner can also handle these new measurement requests at
measurements and taking a close-contact measurement. different frequencies; in other words, requests could either
be sent as available (in an iterative fashion as image data is

4915
 FrE10.4

analyzed) or they could be sent as one batch set.

VI. ADDITIONAL PLAN IMPROVEMENTS AND

FAULT HANDLING
Another key area where plan optimization techniques
were used to improve the rover plan was when extra time or
resources became dynamically available during plan
execution. Since traverse times and resource usage are
difficult to predict, it is often the case that a rover operation
takes less time or power than expected. For instance, a
traverse could take much less time than expected due to
benign terrain. For these cases, the planning optimization
framework was used to dynamically add additional science
goals to the plan that could not be fit in the original plan due
to time and resource constraints. This capability enables the
scenario where scientists on the ground specify a number of
prioritized science goals, but not all of them may be
achievable due to limited rover resources. However, some
goals (that were not included in the initial plan based on
resource estimates) may eventually be fit into the plan if
resource usage is lower than predicted.
On the converse side, rover traverse activities and some
science activities may take longer or more resources than Fig. 6. Sample plan used in testing. Green lines show the
expected. Just as terrain could be more benign than planned path of the rover; blue lines shown the real path;
originally estimated, it could also be significantly more and pink lines show the path that is currently executing.
difficult. In this scenario, the OASIS planning and execution
can dynamically adjust the plan to ensure that resources are typically prioritized and constraints on time, power or
not over-subscribed during plan execution. The CASPER memory would limit the number of science targets that could
planner can often delete a low priority science activity be handled. A large focus of these tests was to improve
during plan execution in order to ensure that higher priority system robustness and flexibility in a realistic environment.
science and engineering activities will be correctly executed. Towards that goal we used a variety of target locations and
Further, it can also be detected that an activity is behind consistently selected new science targets and/or new science
schedule during execution. If it falls significantly behind target combinations that had not been previously tested.
schedule, the TDL executive can fail the activity and request A primary scenario element was dynamically identifying
that CASPER adjust the plan to better accommodate the and handling opportunistic science events. For these tests,
current situation. If new obstacles have been detected, we used onboard data analysis software to generate science
CASPER may re-order (or delete) science targets to ensure
alerts based on a target rock signature. Various types of
more high priority targets are correctly sampled.
signatures were used, but they typically corresponded to a
combination of target albedo (brightness) levels, shape
VII. SYSTEM TESTING WITH HARDWARE AND SIMULATION
characteristics, and size estimations. If rocks were identified
To evaluate our system we have performed a large series in rover camera imagery that had a high score for these
of tests both in simulation and using rover hardware. These features, then a science alert was created and sent to the
tests covered a wide range of scenarios that included the planner. Science alerts often happened during rover traverses
handling of multiple, prioritized science targets, limited time to new locations, but they were also used for testing
and resources, opportunistic science events, resource usage autonomous targeting at the end of a traverse. If a science
uncertainty causing under or over-subscriptions of power alert was detected, the planner attempted to modify the plan
and memory, large variations in traverse time, and so an additional image of the rock of interest would be
unexpected obstacles blocking the rover’s path. acquired. A sample image that was taken in response to a
Our testing scenarios often consisted of a random number science alert was shown in Figure 5.
of science targets specified at certain locations. A map was Other important scenario elements included adding or
used that would represent a sample mission-site location deleting ground-specified science targets based on resource
where data would be gathered using multiple instruments at under- or over-subscriptions. For instance, in some tests, the
a number of locations. Figure 6 shows a sample scenario that rover covered distances more quickly than estimated and the
was run as part of these tests. This particular map is of the planner was able to add additional science targets that could
JPL Mars Yard. The pre-specified science targets (shown in not be fit into the original plan. Conversely, in other tests,
Figure 6 as the larger circles) represented targets that would the rover used more energy than expected during traverses or
be communicated by scientists on Earth. These targets were science activities, which often caused a power over-

4916
 FrE10.4

subscription, where enough power was not being preserved instruments, such as spectrometers, were not readily
for later plan activities. The planner resolved this situation available and thus not directly incorporated into hardware
by deleting some lower priority science targets. Unexpected tests. However different types of measurements were
energy drops during a traverse could also be handled by the included when testing in simulation.
executive, which detects the shortfall and stops the current
C. Testing and Demonstration Summary
traverse if there is not enough energy to complete it. In all
cases, the planning and execution system attempts to Testing in simulation and with real hardware provided
preserve as many high priority science targets as possible important steps in the evaluation of our system. Many bugs
while still adhering to required constraints. were caught early through simulated testing, but others did
not surface until significant runs had been performed on
A. Testing in Simulation rover hardware. Furthermore, running with hardware often
Since testing with rover hardware can be an expensive and allowed a perspective that was difficult to attain through
time-intensive process, we ran a large number of tests in simulated testing. For example, the effects of rover slippage
simulation using a relatively simple simulator. This were much easier to visualize and test when using hardware.
simulator could execute rover sequence commands and A number of live demonstrations of our system have been
simulate their effects at a coarse level of granularity. For performed, including a several hour long demonstration
instance the simulator handled items such as rover position which showed the system successfully handling a random
changes and energy usage over straight-line movements, but combination of science targets and science alerts (that had
did not simulate obstacle avoidance or rover kinematics. not been used in previous testing) and resulted in over 40
Another capability that was used in simulation was meters of autonomous driving. This demonstration consisted
triggering multiple science alerts at pre-set or random times. of several runs that showed scenario elements such as
This capability helped in evaluating the planner’s capacity to handling new science alerts, dynamically adding new
correctly handle different opportunistic science scenarios. ground-specified science when time became available, and
To easily run and evaluate large numbers of tests, we also deleting low priority science targets in a later run where
invested in a testing infrastructure, which allowed tests to be more power was used than originally estimated. The OASIS
run offline and statistics automatically gathered, including software operated correctly in all tested cases.
information such as number of plan conflicts found and Another live demonstration showed a combination of
resolved, plan generation and re-planning time, number of traverse science and autonomous targeting. Autonomous
goals satisfied, overall plan traverse distance and plan targeting was supported by adding a targeting subplan that
optimization scores. This testing infrastructure also could starts with a FIDO navigation camera panorama that is
create mpeg movies of runs that showed plan changes using analyzed online for new targets. Autonomous, new
snapshots of a plan visualization tool. measurements were then scheduled and taken by the high-
resolution FIDO panoramic cameras, which were being used
B. Testing with Hardware
as an example limited FOV instrument.
In addition to testing in simulation, a large number of tests
have been run in the JPL Mars Yard using different rover VIII. RELATED WORK
hardware platforms. For the past few years, the FIDO rover
A number of planning and executive systems have been
(shown in Figure 4) was used for the majority of tests.
used for robotic applications. One approach directed towards
FIDO is an advanced technology prototype rover similar to
rover command generation used a Contingent
the MER rovers. FIDO’s mobility sub-system consists of a
Planner/Scheduler (CPS) that was developed to schedule
six-wheel rocker-bogie suspension capable of traversing
rover science operations using a Contingent Rover Language
over obstacles up to 30 cm in height.
(CRL) [7]. CRL allows both temporal flexibility and
All OASIS software has been designed to run onboard the
contingency branches in command sequences. Contingent
rover, however during testing, only functional-level
sequences are produced by the CPS planner and then
CLARAty modules, such as navigation and vision, and the
interpreted by an executive, which executes the final plan by
OASIS rockfinding software were run onboard FIDO. Other
choosing sequence branches based on current conditions. As
modules, including the planning and execution module and
compared to OASIS, only the executive is onboard the
the analysis module, were run on offboard workstations that
rover; planning is a ground-based operation and does not
communicated with the rovers using Wireless Ethernet,
involve re-planning. Since only a limited number of
since a port of these components to the onboard operating
contingencies can be anticipated and incorporated into the
system (VxWorks) was not complete.
plan, CRL does not provide as much flexibility as OASIS
Tests in the Mars Yard typically consisted of 20-50 meter
when adjusting the sequence in response to unexpected
runs over a 100 square meter area with a range of obstacles
events.
that caused deviations in the rover’s path. Science
The LAAS-CNRS lab robotic control architecture [8] also
measurements using rover hardware were images from one
uses onboard planning and execution to create initial plans
of three sets of cameras on the rover (hazard cameras,
and to provide re-planning capabilities. However, as
navigation cameras and panoramic cameras). Other

4917
 FrE10.4

compared to CASPER, the IxTeT planner uses a partial ACKNOWLEDGMENT

order CSP-based planning approach, which can require This work was performed by the Jet Propulsion
larger amounts of time for re-planning since a valid plan Laboratory, California Institute of Technology, under
must be found at every search step. Further this system has contract with the National Aeronautics and Space
not addressed the handling of opportunistic science and Administration.
interacting with an onboard data analysis system.
Other similar approaches include Atlantis [9] and 3T [10], REFERENCES
which used a deliberative planner and an executive on top of [1] R. Castano, T. Estlin, D. Gaines, A. Castano, C. Chouinard, B.
a set of reactive controllers. These approaches only use a Bornstein, R. C. Anderson, S. Chien, Al. Fukanga and M. Judd,
batch planning approach in a limited fashion do not provide “Opportunistic Rover Science: Finding and Reacting to Rocks,
online re-planning or support for opportunistic science. Clouds, and Dust Devils,” Proceedings of the 2006 IEEE Aerospace
Conference. Big Sky, Montana, March 2006
The Autonomous Sciencecraft Experiment (ASE) [11] has [2] D. Thompson, T. Smith and D. Wettergreen, “Data Mining During
demonstrated the capability of planning and data analysis Rover Traverse: From Images to Geological Signatures,” Proceedings
systems to coordinate the behavior of the EO-1 Earth of the Eighth Int’l Symposium on AI, Robotics and Automation in
Space, Sept 2005.
orbiting satellite. The Remote Agent Experiment (RAX) [3] R. Li, K. Di, et al., “Incremental Bundle Adjustment Techniques
[12] demonstrated the ability of an AI planning and Using Networked Overhead and Ground Imagery for Long-Range
execution system to generate and execute plans onboard the Autonomous Mars Location,” Proceedings of the Eighth Int’l
Symposium on AI, Robotics and Automation in Space, Sept 2005.
NASA Deep Space One (DS1) spacecraft. RAX, however, [4] S. Chien, R. Knight, A. Stechert, R. Sherwood, and G. Rabideau,
used a batch approach to planning and could not “Using Iterative Repair to Improve the Responsiveness of Planning
dynamically re-plan. Further, since RAX and ASE were and Scheduling,” Proceedings of the Fifth Int’l Conference on AI
Planning and Scheduling, Breckenridge, CO, April 2000.
applied to spacecraft, neither handle the large uncertainty [5] R. Simmons and D. Apfelbaum, “A Task Description Language for
inherent in surface navigation and science. Robot Control,” Proceedings of the Intelligent Robots and Systems
Conference, Vancouver, CA, October 1998.
[6] I.A. Nesnas, A. Wright, M. Bajracharya, R. Simmons, T. Estlin, Won
IX. OTHER CURRENT AND FUTURE WORK Soo Kim, "CLARAty: An Architecture for Reusable Robotic
Towards the goal of enabling our planning and execution Software," SPIE Aerosense Conference, Orlando, Florida, April 2003.
[7] J. Bresina, K. Golden, D. Smith, and R. Washington, “Increased
system to better reason how well science measurements Flexibility and Robustness of Mars Rovers,” Proceedings of the
characterize the surrounding environment, we are in the International Symposium, on AI, Robotics and Automation for Space,
process of extending OASIS to evaluate how well a set of Noordwijk, The Netherlands, June 1999.
observations spatially covers a specified area of terrain [13]. [8] R. Alami, R. Chautila, S. Fleury, M. Ghallab, and F. Ingrand, “An
Architecture for Autonomy,” International Journal of Robotics
Scientific observations can be ranked through our algorithm Research, 17(4) April, 1998.
by their ability to improve coverage of a large geological [9] E. Gat, "Integrating planning and reacting in a heterogeneous asyn-
feature or area. This quality measure is one of many that chronous architecture for mobile robots," in SIGART Bulletin 2, 1991.
[10] R. Bonasso, R. Firby, E. Gat, D. Kortenkamp, D. Miller, and M.
could be considered in future rover missions, but it has Slack, “Experiences with an Architecture for Intelligent, Reactive
important significance in enabling our system to balance the Agents,” Journal of Experimental and Theoretical Artificial
science goals of detailed study vs. broad coverage. Intelligence Research, 9(1), 1997.
[11] S. Chien, R. Sherwood, et al.,. “Using Autonomy Flight Software To
In future work, we plan to extend our capabilities for Improve Science Return on Earth Observing One,” Journal of
opportunistic science handling to include adding Aerospace Computing, Information, and Comm-unication 2005.
observations for different types of science instruments and [12] A. Jonsson, P. Morris, N. Muscettola, K. Rajan, and B. Smith,
“Planning in Interplanetary Space: Theory and Practice, “ Proceedings
performing close-contact measurements for high priority of the Fifth International Conference on Artificial Intelligence Planning
alerts. We also intend to extend our system to handle area Systems, Breckenridge, CO, April 2000.
surveying, where all rocks (or other terrain features) within a [13] D. Gaines, T. Estlin and C. Chouinard, “Spatial Coverage Planning
and Optimization for a Planetary Exploration Rover,” Proceedings of
certain area would be properly examined and catalogued. the Fifth International Workshop on Planning and Scheduling for
Space, Baltimore, MD, Oct 2006.
X. CONCLUSION
This paper discussed a number of challenges for using
planning, scheduling and execution techniques to provide
autonomous rover capabilities for future NASA missions.
We described our system approach and explained how it
provides capabilities for sequence generation, execution,
monitoring, re-planning, sequence optimization,
opportunistic science handling, and autonomous targeting.
Through a series of tests in simulation and on rover
hardware, we have demonstrated our system’s ability to
robustly respond to unexpected problems and to take
advantage of unforeseen opportunities, thus achieving high
utilization of rover resources.

4918