Towards User-Driven Exoskeletal Assistive Control using Deep Reinforcement Learning

Movement disorders such as sarcopenia or age-related muscle weakness affect nearly one in seven adults, transforming routine tasks like carrying objects and mobility into formidable challenges. Recent research has shown promise in robotic interventions for assisting and restoring movement in individuals with such disabilities. However, dynamic control of robotic exoskeletons is an open problem. In this paper, we present a user-driven exoskeletal assistance system developed using deep reinforcement learning (RL), specifically Proximal Policy Optimization (PPO). This system leverages the observed body’s pose, including joint positions, velocities and measured electromyography signals (EMG) or muscle activations, to provide exoskeletal assistance and precisely compensate for muscle weakness. We used a human musculoskeletal model acting in a high-dimensional, physics-based simulation environment to train and evaluate the exoskeleton control system. The designed exoskeleton parameters and policies achieved 99.4% accuracy in compensating for the loss of muscle function. Overall, this system serves as a step towards developing a task-agnostic, user-driven exoskeletal assistance system that could significantly enhance daily functional capabilities for sarcopenia patients.