Autonomous Navigation of a Mobile Robot with a Monocular Camera Using Deep Reinforcement Learning and Semantic Image Segmentation

摘要

Navigation systems for mobile robots using deep reinforcement learning have the potential to achieve autonomous movement without relying on precise environmental maps. This study focuses on autonomous navigation of mobile robots using monocular camera images as input through deep reinforcement learning. Typically, RL models are trained in simulation environments. However, there are gaps between simulation and real-world environments, making it challenging to apply trained RL models to actual robots. This challenge is particularly pronounced when using images as input. Semantic segmentation is employed to address this issue, as it can simplify complex RGB images into segmented images, thereby reducing the gaps between environments. In this study, a RL model to reach their destinations and a semantic segmentation model are acquired for mobile robot navigation. These models are applied to a ROS-based autonomous navigation system. Real-world experiments confirm the successful application of the learned models in actual environments.