Efficient Learning and Control of String- Type Artificial Muscle Driven Robotic Systems

Jiyue Tao, Sunil Kumar Rajendran, Yunsong Zhang, Feitian Zhang, Dexin Zhao, Tongsheng Shen

Abstract: This paper investigates the learning-based control of robotic systems driven by string-type artificial muscles. Due to the highly nonlinear dynamics of the actuators and the complicated mechanical structure, it is typically very challenging to design traditional model-based controllers that exhibit desired control performances. With the rapid development of machine learning techniques, deep reinforcement learning (DRL) algorithms have been utilized to control a variety of complex robotic systems. However, these DRL algorithms usually require a huge amount of training data and iterations to converge, which is generally unacceptable for the robotic systems of interest. Therefore, this paper designs an efficient learning-based control algorithm, aiming to improve the training efficiency for robotic systems driven by string-type artificial muscle actuators. Specially, two training improvements are proposed including imitation learning and data augmentation. This paper applies the proposed learning methods to three popular DRL algorithms and tests the control performances in three case studies using three string-type artificial muscle-driven robots, including a parallel robotic wrist, a two degrees-of-freedom (DOF) robotic eye and a robotic finger. Simulation results show that the proposed learning-based control methods significantly accelerate the convergence speed and improve the data efficiency in all the three case studies.