Reinforcement learning-enhanced adaptive PD control for Bowden cable-driven soft exoskeletons

Purpose This paper aims to address nonlinear challenges in Bowden cable-driven soft exoskeletons, including asymmetric friction, hysteresis and transmission backlash, to enhance joint angle tracking accuracy and motion smoothness. Design/methodology/approach Propose a hybrid control framework integrating proximal policy optimization (PPO) reinforcement learning with classical PD control. The agent dynamically outputs time-varying derivative gain adjustment Δkd and reference angle compensation θc. A lightweight cubic Hermite interpolation model characterizes cable bending geometry for time-varying friction estimation. Findings Experiments on a wearable knee exoskeleton platform show maximum tracking error reduced from 3.2° to 0.7° (73% RMS error reduction). Ablation studies confirm θc dominates error suppression, while Δkd optimizes dynamic response. Originality/value This study develops a structured compensation framework tailored for Bowden cable-driven soft exoskeletons. The PPO agent outputs a derivative-gain correction term and a reference compensation term to improve transient response and reduce tracking bias. Together with a lightweight cable-bending geometric model, the method provides a practical solution for nonlinear compensation in flexible-drive exoskeletons.