A Hybrid Planning Approach for Accompanying Information-gathering in Plan Execution Monitoring

Autonomous robots that execute plans in real-world environments may easily get failed due to unexpected environment dynamics. For robust plan execution, a great number of plan execution monitoring works has been proposed to estimate plan execution effects and make recovery plans for any detected plan failures. However, most works have assumed the full observability of environment states and accessibility of complete information about environment dynamics, which shows limitations when executing plans under environment dynamics and partial observability. To efficiently and effectively monitor robot plans in dynamic and partially observable environments, this paper first proposes an accompanying action policy that specifies an interaction pattern between information-gathering plan and the task plan to estimate and monitor task plan execution. To provide a concrete implementation of the above pattern, we first identify the task achievement and execution monitoring as two separated goals in a robot task, and then propose a hybrid planning approach that integrates two planners to solve the goals. For a task-achievement goal, we utilize a classical planning framework ROSPlan to efficiently compute the global task plan. While executing each action of the plan, we cast the execution monitoring problem as a Partially Observable Markov Decision Process (POMDP), which plans the information-gathering plans for monitors the action execution effects and recover failures when necessary. We further implement a typical robotic service task to verify the efficiency and effectiveness of our approach. By comparing with both the baseline plan-invariant execution monitoring approach and a full POMDP planning approach, the hybrid planning approach is efficient in task planning and execution monitoring and effectively recovering plan failures in unknown environments.