A proposal for an online imitation method of path-tracking behavior by end-to-end learning of vision and action

- Analysis of unknown obstacle avoidance behavior and consideration for real-world application -

Abstract

We have proposed an end-to-end imitation learning method for path-following using camera images to imitate behavior generated by LiDAR-based navigation. Experiments have confirmed that this method enables path-following using visual information. However, these experiments were limited to static environments with predefined obstacles. Observations suggest that the robot may have unintentionally learned to avoid unknown obstacles. To investigate this, we analyzed its behavior under varying obstacles and environmental conditions. Our results indicate that the method may be effective even in environments with unknown obstacles. This insight highlights its potential for real-world applications.