Reinforcement Learning with a Supervisor for a Mobile Robot in a Real-world Environment

Abstract

This paper describes two experiments with supervised reinforcement learning (RL) on a real, mobile robot. Two types of experiments were preformed. One tests the robot's reliability in implementing a navigation task it has been taught by a supervisor. The other, in which new obstacles are placed along the previously learned path to the goal, measures the robot's robustness to changes in environment. Supervision consisted of human-guided, remote-controlled runs through a navigation task during the initial stages of reinforcement learning. The RL algorithms deployed enabled the robot to learn a path to a goal yet retain the ability to explore different solutions when confronted with a new obstacle. Experimental analysis was based on measurements of average time to reach the goal, the number of failed states encountered during an episode, and how closely the RL learner matched the supervisor's actions.