Deep reinforcement learning to assess lower extremity movement intention and assist a rehabilitation exoskeleton

This paper introduces a pioneering approach for controlling a unilateral lower extremity exoskeleton designed for rehabilitation and enhancing the quality of life for individuals with neuromuscular weakness of the lower limbs. At the core of our methodology is the integration of Long Short-Term Memory (LSTM) networks with Proximal Policy Optimization (PPO) models, utilizing a deep reinforcement learning framework to interpret and predict user movement intentions in real time. By harnessing sensor fusion that combines surface electromyography (sEMG) and Inertial Measurement Units (IMU) from sensor arrays placed around the quadriceps and gastrocnemius muscles, our system employs an adaptive nonlinear sliding mode control with Pneumatic Artificial Muscles (PAMs), thereby directing the exoskeleton's movement and positioning. The LSTM network processes temporal sequences of sensor data to capture the dynamics of human motion, while the PPO model optimizes the control policy to ensure smooth and responsive movements aligned with the user intentions. Focusing initially on basic maneuvers integral to Activities of Daily Living (ADL), our system demonstrates promising preliminary results in mimicking natural limb movements, laying the groundwork for future clinical applications. This paper specifically delves into the utilization of the LSTM-PPO framework for controlling an avatar prior to testing the exoskeleton, representing a significant step towards realizing a responsive and intuitive exoskeleton control system.