単眼カメラを用いた事前の地図作成無しで自律走行する移動ロボット

抄録

This paper introduces an autonomous mobile robot system based on a monocular camera that travels to destinations without any map building in advance. The mobile robot system selects velocity commands based on action policy trained by deep reinforcement learning. The input state of the trained action policy is depth information acquired from struct2depth using the monocular camera. Also, localization of the robot is performed by online visual SLAM using the monocular camera. This paper describes the system integrating velocity command selection, depth estimation, and localization. The proposed system configuration and preliminary experiments are provided in the paper.