Reinforcement learning-based assist-as-needed control for shoulder rehabilitation robot based on virtual biomechanical model

Abstract

This paper presents an adaptive control strategy for a shoulder rehabilitation robot, leveraging a machine learning-based Assist-as-Needed (AAN) framework to improve transparency and user participation during rehabilitation tasks. The proposed control architecture integrates a conventional impedance controller with a reinforcement learning-based torque modulation layer that actively adapts to patient effort. A biomechanical model of the shoulder joint, developed using a Virtual Biomechanical Shoulder Robot Model (VBSRM), informs the estimation of joint torques and stiffness values. Experimental evaluations involving ten healthy subjects compare the performance of the AAN mode against conventional passive control. Results demonstrate that the AAN strategy reduces joint torque and effort by up to 30%, while preserving accurate trajectory tracking and stable impedance adaptation. Stiffness profiles under AAN control exhibit smoother, user-contingent variations, indicating responsive and compliant assistance. The proposed method provides a robust and energy-efficient rehabilitation interface, paving the way for intelligent, patient-specific therapy protocols in physical human-robot interaction systems.