Torque Map Identification and Control of Spherical Actuators Using Deep Reinforcement Learning Based on One Simple Magnetic Pole Model

Abstract

　In recent years, multi-degree-of-freedom spherical actuators have been expected to be utilized to reduce the size and weight of multi-degree-of-freedom drive systems used in product manufacturing. To control spherical actuators, data on the torque characteristics in all positions, called a torque map, is required. However, the torque maps on actual machines are difficult to be measured with high accuracy due to variations in permanent magnet characteristics and friction, leading to reduced control accuracy. In this paper, we propose a torque map identification method using deep reinforcement learning. By using the proposed method, it is possible to create a torque map that includes nonlinear disturbances without measuring the torque map, and to construct a highly accurate control system with excellent stability and computational cost. In this paper, a mathematical model of a spherical actuator is used for verification. Control was performed using the identified torque map, and the effectiveness of the proposed method was verified from its accuracy.