Self-Supervision Based Learning to Move Obstacles for Path Planning

Abstract

Navigation among movable obstacles (NAMO) has been a research problem for many robotics researchers. The work presented so far includes moving obstacles out of the way and to plan shorter paths. However this interaction with obstacles is based on either hit and try where a robot tries to move obstacles one by one, or obstacles are labeled as movable and immovable. Our work is based on learning if an obstacle is movable or immov able, and also finding out what is the best possible angle to push an obstacle to move it. This has been done autonomously without human supervision, where robot will keep interacting with various obstacles and will eventually learn how to move obstacles out of the way. The developed algorithm is more efficient, time and energy saving than the previously presented approaches.