Long-term Autonomous Mobile Manipulation under Uncertainty

We describe a hierarchical belief space planning framework to achieve robust long-term autonomous mobile manipulation behavior under uncertainty. The approach relies on condensing belief distributions across different abstractions to simultaneously suppress uncertainty and mitigate risk at run-time. We evaluate this system in an experimental domain that requires a robot to monitor and clean a dynamic unstructured environment, executing hundreds of physical actions without failures that require human intervention. Results indicate that this framework provides a sound basis for cognitive robots in uncertain and dynamic environments.