Adaptive Learning-Based Haptic Motion Copying for Stroke Rehabilitation Using Gaussian Process Regression

Abstract

Upper-limb rehabilitation for stroke patients often depends on subjective assessments and intensive therapist supervision, which limits scalability and consistency. This study developed a motion-copying system integrated with a Gaussian Process Regression (GPR)-based trajectory correction framework to replicate healthy subjects' movements and improve rehabilitation accuracy. A haptic device was used to collect precise position and velocity data during three predefined trajectory patterns: Square, Sine, and Sawtooth waves. The GPR-based correction achieved mean error reductions of 91.02 ± 3.05% for the Square wave, 82.12 ± 6.24% for the Sine wave, and 69.62 ± 24.30% for the Sawtooth wave. These findings confirm the system's technical capability to deliver accurate, real-time trajectory correction for consistent and predictable motions, while identifying areas for improvement in handling irregular movements. The proposed framework provides an objective and adaptive technical tool that holds promise for reducing physiotherapist workload and potentially enhancing patient outcomes in future clinical applications.