Gradual Acquisition of Feed-Forward Control in Repetitive Motions by State-Independent Reinforcement Learning

Human motor control is characterized by its adaptability to new dynamics. As a result of adaptation, humans can achieve motor control with less computational effort while maintaining achievement of the task. In this paper, we hypothesize that such adaptation can be modeled by acquisition process of a feed-forward control sequence. Based on the hypothesis, we propose a state-independent reinforcement learning model of feed-forward control generation and feedback control reduction. A gradual learning strategy is presented on the basis of state-independent and time-dependent reinforcement learning to improve learning efficiency for repetitive tracking control tasks. The proposed motor learning model was validated in simulation of 2-DOF manipulator tracking control task, where the robot could obtain a state-unaware control sequence under unknown dynamics and external force condition.