Acquisition of Autonomous Navigation Models for Mobile Robots by Deep Reinforcement Learning in Automatically Generated Urban Simulation Environments Including Pedestrians

Abstract

This paper presents an acquisition method of autonomous navigation models for mobile robots by deep reinforcement learning in the game engine Unity. In this study, the mobile robots are required to reach goals in virtual urban environments while avoiding pedestrians using only a monocular camera and positional information, without using any 3D range data from laser sensors. The navigation models are trained by performing reinforcement learning based on Proximal Policy Optimization and input images to predict the pedestrians’ movements including last five frames. As a result of curriculum learning, the robot was able to reach goals at a rate of more than 80% in both of the navigations on the sidewalk and in crosswalks.