Robust control of three‐degree‐of‐freedom spherical actuator based on deep reinforcement learning

Abstract

Multi‐degree‐of‐freedom (multi‐DOF) spherical actuators have been developed for the fields of robotics and industrial machineries. We have proposed an outer rotor type three‐DOF spherical actuator that can realize a high torque density. Each coil input current is calculated using a torque generating equation based on the torque constant matrix. The permanent magnet type actuators have a problem with generating unexpected cogging torque due to various manufacturing errors. Manufacturing errors mainly mean differences between the ideal dimensions at the motor design stage and the actual dimensions in mass production. In this case, the actuator would exceed the limitations of classical proportional‐integral‐differential (PID) controllers. Therefore, we propose a current compensator using reinforcement learning by introducing a deep neural network that is expected to improve the robustness of spherical actuators. This current compensator was applied to uncertainty problems such as manufacturing fluctuations of cogging torque. We examined the reward, which is the main parameter of deep reinforcement learning, and reduced the control error compared to classical PID controller and simple neural network (NN) controller. © 2022 Institute of Electrical Engineers of Japan. Published by Wiley Periodicals LLC.