Fast robot motor skill acquisition based on bayesian inspired policy improvement

Abstract

Learning from demonstration with the reinforcement learning (LfDRL) framework has been successfully applied to acquire the skill of robot movement. However, the optimization process of LfDRL usually converges slowly on the condition that new task is considerable different from imitation task. We in this paper proposes a ProMPs-Bayesian-PI2 algorithms to expedite the transfer process. The main ideas is adding new heuristic information to guide optimization search other than random search from the stats of imitation learning. Specifically, we use the result of Bayesian estimation as the heuristic information to guide the PI2 when it random search. Finally, we verify this method by UR5 and compare it with the traditional method of ProMPs-PI2. The experimental results show that this method is feasible and effective.