Stroke patients with hemiplegia require personalized upper-limb rehabilitation, yet designing safe and effective robot-assisted trajectories that mimic natural human movement remains a significant challenge. This paper proposes a trajectory planning and optimization method to address this need by leveraging multi-objective constrained reinforcement learning. The method involves dynamically capturing motion data from the patient's healthy limb to define personalized Activities of Daily Living (ADL). A reinforcement learning algorithm, guided by a specially designed reward-punishment function, then optimizes the trajectory with objectives for smoothness, jerk minimization, and accurate tracking of key points. The approach was validated on a 4-degree-of-freedom (4-DOF) upper limb rehabilitation robot, which successfully achieved multi-joint coordinated trajectory tracking based on the learned ADL movements. The experiments confirm the method's effectiveness in designing personalized rehabilitation trajectories that improve the continuity and smoothness of robot-assisted movements, offering a promising solution for patient-specific therapy.
Xu et al. (Thu,) studied this question.