Robot Learning for Moving between Subgoals

Abstract

In this research, our goal is that a mobile robot learns to move between subgoals without human interactions. The robot does not have knowledge on the environment: positions of subgoals and obstacles. Unfortunately, the robot learning is unavoidably influenced by the errors in a real world. We suggest how to learn to move efficiently between subgoals by resetting errors. In our system, we define the distinctive place on which the movement control of a robot changes from sensor-based to coordinates-based. The distinctive place is identified only with local information, and the cumulative errors through the movement between two subgoals are reset on distinctive places. First, the robot moves to search for subgoals. Next, it moves between subgoals repeatedly, and learns to move efficiently. We made some experiments in a real environment. In the experiments, we found out that the robot's movement was gradually improved by learning.