Cascade Control with Reinforcement Learning for an Industrial Exoskeleton

Abstract

This paper presents an adaptive control framework for an industrial lower-limb exoskeleton based on cascade control combined with reinforcement learning (RL). The multilevel hierarchical structure of the cascade controller effectively decouples the fast inner-loop current and torque dynamics from the slower outer-loop kinematic tracking tasks. This decomposition significantly enhances overall system bandwidth while mitigating undesirable dynamic coupling between adjacent joints during complex gait cycles. The proposed approach aims to ensure safe and intuitive physical human-robot interaction by dynamically compensating for parametric uncertainties, while simultaneously improving adaptability to heterogeneous individual biomechanical characteristics and non-stationary task conditions. A hybrid actuation architecture with electric drives and gravity compensation is considered. Cascade parameters are adjusted online using policy-gradient-based reinforcement learning to optimize interaction forces, trajectory tracking accuracy, and energy efficiency. Experimental results obtained on a laboratory prototype demonstrate improved stability, comfort, and reduced muscular load compared to fixed-parameter cascade control.