Pedestrian Avoidance Navigation of Mobile Robots Using Monocular Camera in Urban Environments Based on Deep Reinforcement Learning

This study aims to acquire RL models for mobile robots to avoid pedestrians in unknown urban environments using deep reinforcement learning. Training is performed in urban environments in a virtual world created by the Unity game engine. The mobile robot runs using only monocular camera images and positional information, and does not use any external sensors other than the monocular camera, such as a LiDAR. Not only the current step, but also the last 5 steps are utilized for input states in learning. This is so that the robot can learn and predict the pedestrians’ movement. Since this is a challenging learning task for the robots, curriculum learning, in which the difficulty of the learning task increases step by step, is adapted. As a result of the learning, the robot acquired the ability to reach the goal while avoiding pedestrians, even in an unknown urban environment including pedestrians with unknown color appearances.