Planning and Resilient Execution of Policies For Manipulation in Contact with Actuation Uncertainty (1703.10261v2)

Abstract: We propose a method for planning motion for robots with actuation uncertainty that incorporates contact with the environment and the compliance of the robot to reliably perform manipulation tasks. Our approach consists of two stages: (1) Generating partial policies using a sampling-based motion planner that uses particle-based models of uncertainty and simulation of contact and compliance; and (2) Resilient execution that updates the planned policies to account for unexpected behavior in execution which may arise from model or environment inaccuracy. We have tested our planner and policy execution in simulated SE(2) and SE(3) environments and Baxter robot. We show that our methods efficiently generate policies to perform manipulation tasks involving significant contact and compare against several simpler methods. Additionally, we show that our policy adaptation is resilient to significant changes during execution; e.g. adding a new obstacle to the environment.