Reinforcement learning based control design for mobile robot motion control

Abstract

The reinforcement learning (RL) based methods show people an alternative way to solve multiple problems in robot motion control. RL based algorithms have the ability to autonomously learn the law of controller through the interaction with environments, especially with the combination with the neuronal network, the deep RL based methods attended its’ ability in continuous state-space and action-space control problems instead of solving nonlinear kinematic equations compared with the traditional method. In this thesis, we study three advanced deep Reinforcement Learning algorithms and achieve the simulation on the Minitaur robot model and Pybulet physics engine to control the motion. Furthermore, we discuss the performance of each algorithm considering the best result and overall result from multiple epochs of simulations. Finally, we assess the advantages and disadvantages of those reinforcement learning algorithms via statistical analysis based on the average reward from the simulations.