Design of a Variable-Stiffness Elbow Rehabilitation Robot and Its Control: A Variable Impedance Strategy Based on LSTM Elbow Torque Prediction

Abstract

In the rehabilitation training of stroke patients, traditional rigid-driven robots struggle to effectively address spasticity, leading to suboptimal training and risks of secondary injury. This paper proposes a dual-motor driven variable stiffness elbow Joint rehabilitation robot based on a cam-spring-roller mechanism, enabling independent control of stiffness and position. We developed an LSTM-based elbow joint torque estimation model trained on a dataset generated using a musculoskeletal model incorporating Hill muscle model, achieving accurate and reliable torque prediction. A stiffness and impedance cooperative control framework is then established based on the estimated torque values: wherein the system provides higher assistive torque with lower mechanical stiffness when lower muscle strength is detected, and reduces assistive torque while increasing mechanical stiffness when higher muscle strength is identified. The framework also includes a spasticity detection model is proposed, incorporating metrics of antagonist muscle abnormal activation, normalized jerk score and velocity-activation quantification, aiming to enhance training safety and demonstrating the potential for improved patient engagement. Finally, experiments on four healthy subjects and a patient with spasticity demonstrated the feasibility and preliminary efficacy of the stiffness-variable impedance control framework and the spasticity detection model in a pilot study.