Tracking Control using standalone Reinforcement Learning for a Robot Manipulator

摘要

In this paper, we demonstrate the utilization of an upper-extremity (UE) rehabilitation robot for tracking control employing a reinforcement learning (RL) agent. Our methodology incorporates the use of a Deep Deterministic Policy Gradient (DDPG) agent, an off-policy actor-critic RL algorithm, to interpret observations from the robot and generate optimal torque efforts. The principal goal was training an RL agent capable of accurately following trajectories curated for rehabilitation exercises. Our devised reward function aimed to optimize tracking accuracy while minimizing observed chattering effects. The effectiveness of our approach was evident through successful agent training and positive simulation results, showcasing precise tracking. A well-structured rehabilitation system for upper-limb patients has the potential to significantly benefit the rehabilitation industry by reducing wait times at physiotherapy centers and improving overall efficiency, facilitated by the assistance of the robot.