Supervised-reinforcement learning for a mobile robot in a real-world environment

Abstract

This research measures how well supervised-reinforcement-learning techniques perform when applied to real-world tasks, managed as a discrete-event dynamic system (DEDS). Two types of experiments are tested. One tests the robot’s stability in implementing a task it has been taught. The other experiment includes obstacles blocking the path to the goal and measures the robot’s flexibility. The supervisor consists of human-guided remote-controlled runs through the navigation task and acts as a teacher for the initial stages of reinforcement learning. Experimental analysis is based on measurements of average time to reach the goal and the number of failed states encountered during a trial of episodes.