A Multi-Channel Reinforcement Learning Framework for Robotic Mirror Therapy

Abstract

In the letter, a robotic framework is proposed for hemiparesis rehabilitation. Mirror therapy is applied to transfer therapeutic training from the patient's function limb (FL) to the impaired limb (IL). The IL mimics the action prescribed by the FL with the assistance of the wearable robot, stimulating and strengthening the injured muscles through repetitive exercise. A master-slave robotic system is presented to implement the mirror therapy. Especially, the reinforcement learning is involved in the human-robot interaction control to enhance the rehabilitation efficacy and guarantee safety. Multi-channel sensed information, including the motion trajectory, muscle activation and the user's emotion, are incorporated in the learning algorithm. The muscle activation is expressed via the skin surface electromyography (EMG) signals, and the emotion is shown as the facial expression. The reinforcement learning approach is realized by the normalized advantage functions (NAF) algorithm. Then, a lower extremity rehabilitation robot with magnetorheological (MR) actuators is specially developed. The clinical experiments are carried out using the robot to verify the performance of the framework.