End-to-End Deep Reinforcement Learning for Exoskeleton Control

Patient-specific control and training on lower body exoskeletons can help improve a user's gait during post-stroke rehabilitation by increasing their amount of participation and motor learning. Traditionally, adaptive control techniques have been used to provide personalization and synchronization with exoskeleton users, but they require predefined dynamics models of the user and exoskeleton. However, these models can be difficult to accurately define due to the complexity of the human-robot interaction. Most recently deep reinforcement learning techniques have shown potential to effectively learn control schemes without the need for system dynamics models. In this paper, we present for the first time an end-to-end model-free deep reinforcement learning method for an exoskeleton that can learn to follow a desired gait pattern, while considering a user's existing gait pattern and being robust to their perturbations and interactions. We demonstrate the effectiveness of our proposed method for user personalization of gait training in simulated experiments.