End-to-End Discrete Motion Planner based on Deep Neural Network for Autonomous Mobile Robots

Abstract

For autonomous navigation of mobile robots, collision avoidance with obstacles is an essential capability. Previously, we found that operators controlled a mobile robot with discrete motion commands, such as straight, right, and left, through a joystick. In this paper, therefore, we propose an end-to-end discrete motion planner. This motion planner is based on a deep neural network. For network training, we adopt a learning-from-demonstration approach. Through the developed teaching system, an operator controls the robot so as to move forward while avoiding collisions with obstacles. In doing so, the robot is allowed to record sensor data and discrete motion commands given by operators as inputs and outputs. In addition to the sensor data, we use a direction angle between the robot and goal destination as the input. All of these input and output data are the training data sets for the deep neural network. In the navigation experiments, we show that the robot based on the end-to-end discrete motion planner is able to move toward the goal destination while avoiding collisions with obstacles.