Abstract
This paper presents an adaptive Assist-as-Needed (AAN) control framework for a shoulder rehabilitation robot enhanced by a Virtual Biomechanical Shoulder Robot Model (VBSRM) and an online stiffness adaptation module. The proposed system adapts support levels dynamically based on user interaction and motor effort, ensuring both safety and active participation during rehabilitation training. The VBSRM is first calibrated to each user's anthropometric dimensions and used to estimate joint torques during movement. Assistance coefficient $\beta $ (t), is then dynamically updated based on the user's measured interaction force, increasing support when effort is low and reducing it when active participation improves. The AAN controller features a gradient-based online stiffness learning mechanism, enabling the system to dynamically adjust joint stiffness during movement based on tracking errors and interaction forces. This dual adaptation enhances rehabilitation training individualization. Experiments with ten healthy subjects performing an Activity of Daily Living (ADL) task task demonstrated improved torque profiles, effective stiffness modulation, and reduced trajectory tracking errors in active and passive modes. The proposed control scheme demonstrates high potential for patient-specific, responsive rehabilitation training in neurorehabilitation and post-stroke therapy.