Towards Safer Rehabilitation: Improving Gait Trajectory Tracking for Lower Limb Exoskeletons Using Offline Reinforcement Learning.

The application of online reinforcement learning (RL) in lower limb exoskeleton control has the potential to improve gait rehabilitation for individuals with impaired mobility. However, online RL approaches require real-time exploration and pose safety risks during training as suboptimal policies can lead to unstable or unsafe actions being executed by the exoskeleton. This study explores the application of offline RL methods, including Implicit Q-Learning (IQL), Twin Delayed Deep Deterministic Policy Gradient with Behavior Cloning (TD3+BC), and Revisited Behavior Regularized Actor-Critic (ReBRAC), for trajectory control of lower limb exoskeletons using a pre-collected dataset. The transition involved generating a diverse and representative dataset using online RL methods like Proximal Policy Optimization (PPO), which was then utilized to optimize offline RL models with advanced hyperparameter tuning via Optuna. Our results demonstrate improved gait trajectory tracking over a PPO baseline, with our TD3+BC model achieving the best performance. These findings highlight the potential of offline RL to enhance exoskeleton trajectory control while minimizing safety risks inherent in online approaches.