Host: The Japan Society of Mechanical Engineers
Name : [in Japanese]
Date : June 28, 2023 - July 01, 2023
The purpose of this study is to make a robot to acquire path-following behavior with fewer learning steps by end-to-end learning using camera images as input. Our past studies have proposed an online imitation method of path-following behavior of vision and action. We added a function to select a path so that the robot can move in any direction, and verified the effectiveness of the function in simulations. However, the number of learning steps increased significantly when we added the path selection function, and it was desired to reduce the number of learning steps for experiments in a real environment. Therefore, we propose two methods to reduce the number of learning steps and verify their effectiveness through experiments. Furthermore, we verify the validity of the methods through experiments using a real robot.
