다중 센서를 활용한 강화학습 기반 모바일 로봇 자율주행

Abstract

This paper presents a method to reduce the collision rate and the step over rate of reinforcement learning-based autonomous navigation by using two interdependent sensors, a Light Detection and Ranging (LiDAR) and a camera. The LiDAR is specialized in measuring distance and angles to objects using laser scanning, while the camera have the advantage of identifying the morphological features of the objects. Therefore, the proposed method utilizes both distance and morphological feature by designing the state of the reinforcement learning algorithm using the information obtained by LiDAR and camera. This research improves success rate by leveraging the synergic characteristics of these low-cost sensors, without the need for complex calibration. This method is demonstrated by using the Trutlebot3 Burger model in the ROS Gazebo environment by tracking the reward per episode and three metrics: success rate, collision rate and step over rate.