Abstract
Future NASA missions that include in situ explorations on distant planets require remote vehicles to operate with a high degree of autonomy. The planning system deployed on such vehicles needs to address unexpected events or outcomes from planned actions. We propose to develop and investigate an agent-based normative reasoning system for planning and plan execution in the context of planetary rovers. The system incorporates robotic vision and sensor fusion subsystems for a realistic interface between the planning system and the environment.
Our agent-oriented approach is driven by the research hypothesis that highly distributed planning based on normative reasoning can be used to produce coherent behavior in highly dynamic environments. Our approach provides a sound theoretical framework for addressing three important issues in autonomous planning for planetary exploration: dealing with large planning spaces, handling unforeseen circumstances, and ranking candidate plans. First, the large planning space is dealt with by using normative reasoning, i.e., probabilistic reasoning and decision theory. The approach supports pruning of the planning space while adhering to the principle of maximum expected utility. Second, the problems presented by large planning spaces and handling unforeseen circumstances are addressed by factoring the space into smaller, much simpler planning spaces and representing them in terms of decision networks. Each planning space is the responsibility of one of numerous miniplanners. Third, the ranking issue is handled by the decision networks as they introduce a formal mechanism for determining relative ranking of plans. The system is implemented with a collection of agents, yielding fault tolerance, computational efficiency, and ease of maintenance.
The choice of the agent encapsulated decision network paradigm requires that we also develop a corresponding run-time environment. Due to limited computational resources, the run-time environment must be small and extremely efficient in order to support the numerous planning agents effectively.
Major deliverable components of the proposed system are:
· A small efficient run-time environment to evaluate decision networks,
· Distributed planning engines, and
· Agent implementations of models for rover vision and sensors,
· An evaluation report.
Our effort will
· Demonstrate that the distributed autonomous agent approach is viable and provides the benefits specified,
· Demonstrate that the approach is compatible with existing NASA architectures for spacecraft control and navigation.
The proposed research specifically targets the areas of autonomy, automated planning and scheduling, control architecture and executive, and navigation and control in the context of planetary rovers. These are priority research areas for the AI Group at the NASA’s Jet Propulsion Laboratory. The proposed research complements current efforts at JPL and will include a subcontract for the active participation of a JPL researcher.
There are significant benefits to the state of South Carolina from the proposed research. First, it will contribute to the priority research area of information technology by developing the infrastructure for normative reasoning and distributed agent technology. The development of efficient algorithms will make it possible to transition this technology from academia to industry. Second, the majority of the requested funding is for graduate students and a postdoctoral associate. The benefit to South Carolina is the development of a trained workforce in cutting edge technology. Finally, this proposal research will foster a long-term collaboration between researchers at the University of South Carolina and NASA.