Batch Reinforcement Learning for Controlling a Mobile Wheeled Pendulum Robot

In this paper we present an application of Reinforcement Learning (RL) methods in the field of robot control. The main objective is to analyze the behavior of batch RL algorithms when applied to a mobile robot of the kind called Mobile Wheeled Pendulum (MWP). In this paper we focus on the common problem in classical control theory of following a reference state (e.g., position set point) and try to solve it by RL. In this case, the state space of the robot has one more dimension, in order to represent the desired variable state, while the cost function is evaluated considering the difference between the state and the reference. Within this framework some interesting aspects arise, like the ability of the RL algorithm to generalize to reference points never considered during the training phase. The performance of the learning method has been empirically analyzed and, when possible, compared to a classic control algorithm, namely linear quadratic optimal control (LQR).