Reinforcement Learning for a Continuous DC Motor Controller

Abstract

Electric motors control is a very knows topic in research and most of the activities are drawn towards classic PI or model predictive control methods. Implementing reinforcement learning techniques in the field of motor control depends on fidelity environments, types of considered motors and modeled power electronics used for control. Training an RL speed controller means finding an optimal control policy by offline training using an environment, before implementing it in a real-world scenario. Different environments and techniques have been developed for training RL controllers, most of them being extensions of Open AI gym environments. This paper presents a trained RL speed controller, developed through Reinforcement learning techniques, specifically TD3 RL algorithm, applied to permanently excited dc motors. In this work, the open-source Python package gym-electric-motor (GEM) [1] is used for environment setup, and pytorch framework for developing the controller.