A LiDAR based end to end controller for robot navigation using deep neural network

Abstract

Navigation is a primary task for mobile robot to accomplish its tasks. Conventionally, these navigation methods require lots of parameters to be tuned artificially. In this paper, we propose a navigation method to learn the end-to-end control policy using convolution neural network, which directly outputs the velocity and angular rates using only the current observation and the target. Besides, a sliding window of past information is incorporated to add the memory to the controller, so that the hesitation is reduced when ambiguity occurs. To train the model, the data is generated from expert planner in the simulation environment, leading to a low cost for massive data. To validate the trained end-to-end controller, we compare our method with the expert global planner both in the training map and complex map in the simulation environment.