Reinforcement Learning with Merged Motor Learning Theory for Personalized Upper-Limb Robot Rehabilitation

This paper presents a rehabilitation-robot framework that integrates motor-learning theory with reinforcement learning (RL) to personalize upper-limb exoskeleton training. A classical two-state adaptation model and a merged learning–forgetting state are combined with an RL policy that adjusts impedance and assistance under safety constraints as a function of patient progress. In a clinically-inspired scenario (60-year-old, 3-month post-stroke, right hemiparesis) performing target-reaching with session-wise frame rotations, the approach shows improvement of the error-effort trade-off and smoothness (jerk) in simulation. It also accelerates early learning by strengthening the learning signal without increasing applied forces. This research also sketches early phase outcome prediction that uses estimated learning parameters and process-level features.