Robot Learning through Crowd-Based Games
Andrew Peter Wolff, Scott P. Cornman · Digital WPI · 2014
The field of robot learning from demonstration focuses on algorithms that enable a robot to learn new abilities from examples provided by a human teacher. We develop a system that enables users to remotely control a KUKA youBot robot to play a game of Whack-A-Mole through a common web browser. We collect data on how users play and utilize a decision tree learning algorithm to teach the robot to play autonomously. We find that filtering data by score does not significantly improve learned robot performance over using all data. Using A/B testing techniques, we find that motivational game features noticeably improve the quantity and quality of collected data.