Demonstration-based Therapist Skill Transfer Learning Framework for Upper-Limb Rehabilitation Exoskeleton

Upper limb rehabilitation exoskeletons offer a promising technological solution for neuromuscular function recovery. However, their clinical application remains challenging due to difficulties in effectively integrating therapists’ professional expertise. This paper proposes a demonstration-based therapist rehabilitation skill transfer learning paradigm, where general rehabilitation assisting policy is learned from demonstration through simulation-based learning, and ultimately deployed onto a real exoskeleton system. The learning framework consists of two components: a motion imitation module that simulates muscle-actuated movements using a two-level policy, and a reinforcement learning-based exoskeleton controller that adapts assistance based on patient states. These components work jointly within a human-exoskeleton coupled system to reproduce therapist-guided movements and force patterns. Demonstration data, including video and surface electromyography, are processed to generate biomechanically consistent reference motions and torque labels. The trained controller is validated through both single-joint and multi-joint rehabilitation tasks. Experimental results show that the system can accurately replicate therapist-like motion and joint torques, and demonstrates strong adaptability across subjects and task types.