Learning Replanning Policies With Direct Policy Search

摘要

Direct policy search has been successful in learning challenging real-world robotic motor skills by learning open-loop movement primitives with high sample efficiency. These primitives can be generalized to different contexts with varying initial configurations and goals. Current state-of-the-art contextual policy search algorithms can however not adapt to changing, noisy context measurements. Yet, these are common characteristics of real-world robotic tasks. Planning a trajectory ahead based on an inaccurate context that may change during the motion often results in poor accuracy, especially with highly dynamical tasks. To adapt to updated contexts, it is sensible to learn trajectory replanning strategies. We propose a framework to learn trajectory replanning policies via contextual policy search and demonstrate that they are safe for the robot, can be learned efficiently, and outperform non-replanning policies for problems with partially observable or perturbed context.