Reinforcement learning based on backstepping approach for a wheeled mobile robot

Abstract

We propose to combine the backstepping method with the reinforcement learning to take advantage of each method, where learning is not performed at once, but is performed step by step following the structure of the plant to be controlled. To demonstrate the basic idea, a tracking control problem of a wheeled mobile robot (WMR) under a nonholonomic constrain is discussed. Firstly, a tracking controller at the kinematic level is obtained by reinforcement learning with the actor critic structure, and then the actual control law is obtained by applying the reinforcement learning again taking the actuator dynamics into account by using the (fictitious) kinematic control obtained at the first step