Robust Neural Network Controllers for Lower Limb Exoskeletons: A Deep Reinforcement Learning Approach

In this study, we introduce a novel deep reinforcement learning (DRL) based approach for controlling lower limb rehabilitation exoskeletons (LLREs). Our method employs a neural network-based controller that accurately forecasts real-time commands for the exoskeleton’s actuators using only proprioceptive signals from the LLRE. This controller is trained within a sophisticated virtual simulation environment integrating a comprehensive human musculoskeletal model and an exoskeleton interaction model. To enhance adaptability, we utilized domain randomization during training to simulate diverse patient musculoskeletal conditions. We validate the effectiveness and robustness of our DRL-based LLRE controller across various neuromuscular conditions during walking, evaluating key metrics such as stability and gait symmetry. This innovative approach supports seamless deployment of trained controllers onto physical hardware through sim-to-real transfer, eliminating the need for patient-specific experimentation and parameter tuning. Our work represents a significant advancement in LLRE control methodology, promising enhanced functionality and adaptability for real-world applications.