Reinforcement Learning-based Control for an Upper Limb Rehabilitation Robot

The concept and development of a robot-assisted rehabilitation program using an actor-critic-based deep reinforcement learning controller are detailed in this paper. This research aims to provide a therapy approach that employs a variety of technologies to decrease human participation to give intelligent robot-assisted therapy sessions. In this paper, a Reinforcement Learning (RL) agent utilizing the Deep Deterministic Policy Gradient (DDPG) algorithm was trained to tune a PID controller's parameters to precisely control the robot to track any set of therapeutic exercise trajectories. The reward function was formulated to obtain fast convergence as well as accurate tracking. The proposed approach was designed such that the robot's joint angle is always within an allowable range to ensure the safety of the patient and the effectiveness of the therapy. The RL-based control approach is simulated on a 2 degree-of-freedom (DOF) robot named iTbot, and the simulation results show efficient training and good tracking performance which highlight the effectiveness of the control strategy.