Adaptive Admittance Control for Physical Human-Robot Interaction based on Imitation and Reinforcement Learning

Physical human-robot interaction (pHRI) is applied in various applications such as hand guiding of robots in manufacturing. In this paper, an adaptive admittance controller based on reinforcement learning is proposed to help operators perform pHRI tasks. As reinforcement learning is hard to train and has a high technology threshold, the method of applying imitation learning to get a primary deterministic model is proposed, and it is used to determine the suitable policy model for reinforcement learning. In particular, reinforcement learning obtains the optimal damping value of the admittance controller by minimizing human-robot interaction force and avoiding obstacles. The experiments are performed on a UR5 robot. It is verified that the damping of the admittance controller can adjust adaptively and improve of performance of pHRI.