Adapting Interaction Control in Robot-Aided Rehabilitation Using Reinforcement Learning

Adapting control strategy parameters in robot-aided rehabilitation is an important task to enhance the patients' engagement and recovery. Thus, this work investigated the feasibility of applying a Reinforcement Learning algorithm to automatize the learning process. A Q-learning algorithm dynamically adjusts the interaction stiffness taking as input the motor performance and physiological state of the user. The experiments demonstrated the feasibility of the proposed approach and the ε-greedy capability in learning an optimal policy.