Enhancing Robot Arm Trajectory Tracking through Reinforcement Learning based Compensation

This paper proposes a reinforcement learning (RL) based compensator for minimising the trajectory tracking deviation of a robot arm under external disturbances. The compensator adaptively incorporates the weighted joint position errors into the reference joint position values, with the weighting factor provided by the RL action output. The effectiveness of the method has been verified on a simulated robot model with realistic physical properties in Gazebo simulator. The focus of this study is on evaluating the linear tracking performance of the robot under both sinusoidal and constant forces applied to the its end-effector. The experiment results show that the proposed method can reduce the mean tool center point (TCP) trajectory deviation over 44%. The result also demonstrate the feasibility of compensation without requiring the TCP position information, using only the robot joint information, offering a more cost-efficient solution.