[4Rin1] Interactive 2

[4Rin1-11] A Comparison of Deep Reinforcement Learning Methods for Driving a Small UGV

〇Takuto Fujii¹, Makoto Miwa¹, Yutaka Sasaki¹ (1.Toyota Technological Institute)

Keywords:self driving, deep reinforcement learning, simulation environment, autonomous unmanned vehicle

We aim to develop a safe and efficient autonomous driving program in a crowded dynamic space that requires no environmental map by using deep reinforcement learning.We constructed a simulation environment for a small unmanned ground vehicle (UGV) using Robot Operating System (ROS). The environment provides OpenAI Gym API to perform training and prediction using socket communication. Using the environment, we compared several deep reinforcement learning methods that produce robot control commands in discrete values based on 2D-LiDAR inputs and investigated which deep reinforcement learning method is effective for autonomous driving. Finally, in order to confirm that learning in the simulator environment can be applied to real world, we transferred one of the trained models on the real UGV and confirmed the behaviors.