Recovery From Plan Failures in Partially Observable Environments

The paper discusses a distributed approach to multi-agent plan execution and control, where a team of agents may perform actions concurrently in a partially observable environment. Each agent is able to supervise the execution of the actions it is responsible for by means of an on-line monitoring step and to perform agent diagnosis when a failure in the execution of an action has been detected. The emphasis of the paper is on the mechanism for synthesizing a recovery plan in presence of an action failure. One contribution of the paper concerns an in depth analysis of the characteristics the recovery plan has to satisfy. The most stringent requirement regards that the recovery plan has to be conformant, as the partial observability of the system allows just to estimate the status of the agent and the action effects may be non deterministic. The paper proposes an approach for synthesizing such a conformant recovery plan based on the adoption of symbolic methods. In particular, the recovery planning is implemented in terms of operations on Ordered Binary Decision Diagrams used for encoding both the belief states and the action models.