Autonomous control of four-wheeled vehicles by reinforcement learning

Abstract

This paper discusses autonomous control of four-wheeled vehicles using a relatively rigorous nonlinear model with three degree of freedom, e.g. the forward velocity is not constant. Since it is not easy to design a tracking controller analytically with such complicated models, a learning controller via the actor critic method in reinforcement learning with neural networks for a tracking problem at velocity level is proposed, employing the idea of input-output linearization. The effectiveness including the generalization ability of the proposed controller is shown through numerical simulation.