Control of a Rotary Flexible Joint Experiment based on Reinforcement Learning

This paper presents the Reinforcement Learning (RL) technique for control of a Flexible Joint Experiment. In the firs part of the paper an overview of the field of Reinforcement Learning in the frame of Machine Learning context and some basic RL strategies for control are highlighted. The design of RL controllers for linear systems is briefly presented. The RL based on two neural networks is designed for an experiment that consists of a DC motor and a rigid beam with flexible joint. The goal of the control system is that the beam tap to follow a predefined trajectory. The experimental setup is a nonlinear system, but in order to tune the controller parameters, a linear model is derived. The controller performances are evaluated using a real plant experiment. In the last sections of the paper the results obtained by numerical simulation and the main conclusions regarding the performance of the control system are presented.