Improving robot plans for information gathering tasks through execution monitoring

Recent advances in navigation and control of robots has increasingly led to systems where the actions are deterministic and the challenge is to collect information about the world using noisy sensors. Examples include search and rescue, Mars rover planning and robotic monitoring tasks. However, theoretical results show that in general these problems are as hard as solving partially observable Markov decision problems (POMDPs). We propose an approach where we build plans assuming both the actions and the observations are reliable, then monitor the execution of the plan and use a value of information calculation to add information gathering actions on-line. We describe two variants: one using a classical contingency planner to generate the initial plan, and the other using a Markov decision problem planner. We show how in both cases the addition of execution monitoring can considerably improve overall performance with lower computational cost than solving the original POMDP.