Host: The Japan Society of Mechanical Engineers
Name : [in Japanese]
Date : June 28, 2023 - July 01, 2023
Autonomous navigation systems of mobile robots using deep reinforcement learning with camera images have been studied. However, there are differences among images obtained during training in simulation and obtained in actual environments. The differences make it difficult for robots to be navigated in the real world using action models trained in the simulation. To reduce the differences, semantic segmentation is applied to the navigation system. The semantic segmentation can simplify camera images such as floors or walls. The experimental results show that the proposed method can obtain action models for mobile robots that can be autonomously navigated to a specified destinations with segmentation images as input. A ROS-based navigation system to apply the obtained models to the actual robot has been built and the experiments in the real world have been conducted.
