An Offline Reinforcement Learning-Based Auto-Tuning Framework for Continuous Impedance Control in Powered Prostheses

A continuous impedance control framework has recently gained attention in powered prosthesis control for its ability to mimic human behavior by smoothly and continuously modulating joint impedance. While recent advancements in tuning this framework have enabled more personalized control for walking, the manual tuning process remains time-intensive and empirical, limiting its scalability and adaptability. To address this challenge, we propose a reinforcement learning (RL)-driven framework for offline tuning, focusing on stiffness parameters that play a critical role in shaping knee kinematics during walking. A Gaussian Process Regression (GPR)-based Continuous Stiffness-Kinematics (CSK) model was developed to predict knee kinematics based on stiffness parameters, trained and validated using experimental data from four participants. Leveraging the Deep Deterministic Policy Gradient (DDPG) algorithm, we optimized stiffness parameters in offline, significantly reducing the need for manual adjustments. Validation results confirmed the reliability of the CSK model in capturing the relationship between stiffness parameters and knee kinematics, as well as the feasibility of the RL framework for offline tuning. This study demonstrates the potential of RL-driven tuning to enhance prosthesis control for walking within a continuous impedance control framework. Future work will focus on integrating offline-trained policies into an online human-in-the-loop framework, enabling adaptive and personalized real-time control while extending functionality to a broader range of ambulation tasks.