Self-Directed Cooperative Planetary Rovers

Sponsored by NASA
Aerospace Technology Enterprise

Shlomo Zilberstein, PI
Co-Investigators: Eric Hansen, Victor Lesser, and Rich Washington
Collaborators: Francois Charpillet, and Abdel-Illah Mouaddib
Research Assistants: Raphen Becker, Daniel Bernstein, Max Horstmann, and Zhengzhu Feng

Planetary rovers are unmanned vehicles equipped with cameras and a variety of scientific sensors. They have proved to be a cost-effective mechanism in space exploration and will continue to play a major role in future NASA missions. Recent rover missions, such as Sojourner's Mars exploration, have suffered from total reliance on ground-based commanding and employed on-board autonomy only to safely follow uplinked commands. The inherent uncertainty that characterizes exploration of new environments and the limited communication bandwidth and time delays increase the risk of execution failures and rover downtime.

This project focuses on the question of how to best utilize the rover's resources in the face of the above difficulties. Our approach is to equip the rovers with pre-compiled control polices for making fast decisions on such issues as: how to best perform a given task given a set of alternatives; when the quality of the result is satisfactory; how to react to failure; how many times to retry to perform a certain operation; and how to best allocate limited resources to the entire set of activities over a certain window of operation.

To achieve these goals we are developing and evaluating several fundamental technologies, focused on the basic need to carefully manage the limited computational resources, power, and communication capabilities of the rover. First, we are enriching the rover plan language to describe different methods to achieve each sub-goal and the associated costs and expected quality. We are also providing the system with a model of the uncertainty about the outcome of actions and the resources they consume. Off-line reinforcement learning algorithms are used to construct control policies to choose among the alternative plan options at run-time. These choices are sensitive to the importance of the task, what has been achieved so far, the remaining workload, and the available resources. Second, we are examining the viability of a multi-rover design and its implications on system complexity, autonomy, reliability, and scientific return. Finally, we are building an execution architecture that exploits preliminary on-board data interpretation in order to assess the quality of the collected data and provide feedback to the planner.

Access the Private Project Archive via Postdoc

Related Publications

shlomo@cs.umass.edu