An Examination on Motion Planning for Mobile Robot using Deep Q-learning with RPLidar Sensor

Abstract

In this paper, we investigate about deep Q-learning for autonomous motion planning of mobile robot in indoor environment. Additionally, the proposed method utilizes an RPLidar sensor to deal with training issue. The deep Q-learning is adopted in this work, which is the combination between Q-learning and deep neural network. We propose a neural network model to learn the Q-function in order to solve the planning task, the proposed model is implemented in a simulation environment by using gazebo robot simulation. Our propose studies a new approach on motion planning for mobile robot. The new approach based on the deep Q-learning, which is start-of-art for robotic application. The approach does not based on a model of environment to navigation robot toward a given goal, therefore, it is independent with the scale of environment. As a result, the planning time is able to be real time to execute. Our model can directly map the range finder sensor data to a motion action such as the angular velocity, the result in section 3 points out our model well execute on a static indoor environment and successfully learn to reach a goal after some thousand training episodes. The research achieves two improvements. Firstly, the motion planning task is able to be real time planning comparing with the previous research. Secondly, the proposed model help robot reach to a given goal without the model of map.

Keywords: Navigation, Obstacle avoidance, Deep Q-learning, RPLidar ,Deep neural network