Human-Exoskeleton Interaction Simulation Framework via Deep Reinforcement Learning

Abstract

Designing and controlling lower-limb exoskeletons is a complex and costly process that requires extensive prototyping, human testing, and iterative refinement. A key challenge in active exoskeleton development is accurately perceiving human-robot interaction (HRI) and delivering practical assistance through advanced control strategies. Simulating HRI in controlled environments offers a powerful alternative, enabling efficient experiment design and assistive device development while reducing time, costs, and reliance on physical testing. This paper presents a simulation framework based on deep reinforcement learning (DRL) to investigate the interaction dynamics between human locomotion and powered lower-limb exoskeletons, utilizing the OpenSim physics-based simulator. A musculoskeletal model, integrated with an exoskeleton, functions as an agent that generates muscle forces derived from kinematics, ground reaction forces, and muscle data. The DRL architecture enables the agent to learn natural walking motion through training with experimental data in the simulator. Results demonstrate the model's ability to simulate locomotion dynamics and provide insights into several factors, including muscle activity, muscle forces, human-exoskeleton interaction, and gait patterns.