Robot-Assisted Upper Limb Rehabilitation Using Imitation Learning

Abstract

Robotic rehabilitation offers an innovative approach to enhance motor function recovery in patients with upper-limb impairment. However, the primary challenge lies in the development of adaptive and personalized therapies to meet the unique needs of patients. In response to this challenge, this paper presents a Rehabilitation Learning from Demonstration (RLfD) framework, which integrates Dynamic Movement Primitives (DMP) for learning and generalizing movements, and a Model Reference Adaptive Controller (MRAC) for real-time adaptive control. This combination enables a two-link manipulator to accurately replicate and adapt therapist demonstrations specifically designed for upper-limb rehabilitation. Unlike conventional task-specific controllers, which are limited by poor adaptability, minimal feedback, and lack of generalization, our system dynamically adjusts robotic assistance in real time based on the subject’s tracking error to optimize therapy outcomes. The objective is to minimize assistance while maximizing patient participation in the rehabilitation process. To facilitate this, the framework employs visual tracking technology to capture therapist demonstrations accurately. Once captured, the DMP component of the framework learns from these movements and generalizes them to new goals, while maintaining the original motion patterns. Our evaluations with a simulated two-link manipulator demonstrated the framework’s precise trajectory tracking, robust generalization, and adaptability to disturbances mimicking patient impairments. These tests confirmed the system’s ability to follow complex trajectories and adapt to dynamic patient motor functions. The promising results from these evaluations highlight our approach’s potential to significantly enhance adaptability and generalization in variable patient conditions, marking a substantial improvement over conventional systems.