Model-based Reinforcement Learning Variable Impedance Control for 1-DOF PAM Elbow Exokeleton

Providing effective physical human-robot interactive control is one of the most important considerations in designing an exoskeleton robot. The purpose of the study is to develop a variable impedance control method for providing a reliable approach to different users to operate the exoskeleton by using interactive torque and a spring-damper model. The proposed method based on a model-based reinforcement learning structure and model predictive control frame, designed a neural network dynamic model to replace the state-action value function and using a linear estimate model to fit the action generated from the agent. Analysis and experiment results indicate that the designed method has less interactive torque and tracking error than the traditional impedance control method. Meanwhile, the results also illustrate the changing process of impedance parameters. The proposed control method sufficiently applies to the 1-DOF PAM elbow exoskeleton system.