Multi-Skill Learning-Based Variable Stiffness Control for Upper-Limb Rehabilitation Robot

Abstract

Learning from demonstration (LfD) is a promising approach for transferring therapist skills to rehabilitation robots, enabling the generation of human-like trajectories for activities of daily living (ADL) tasks. However, existing LfD-based rehabilitation methods mainly focus on motion reproduction and rarely consider the adaptive stiffness modulation behaviors of therapists during physical interaction, limiting their ability to realize assist-as-needed (AAN) rehabilitation. To address this limitation, this letter proposes a kernelized movement primitives (KMP)–based variable stiffness learning framework that jointly learns motion and endpoint stiffness from therapist demonstrations. The stiffness matrices are mapped into a Euclidean space via logarithmic mapping, enabling unified probabilistic learning of multi-skill features. Furthermore, a performance-based variable-stiffness impedance controller is developed to adapt robot assistance according to patient performance, thereby promoting active participation. The proposed framework was validated in a 3D water-drinking task with five post-stroke patients. Experimental results show that the proposed method enables effective multi-skill learning and generalization, while promoting active patient engagement.