Robot Reinforcement Learning using Crowdsourced Rewards

This work is inspired by the idea that robots should possess the ability to learn new skills with minimum guidance from a dedicated human teacher and they should enhance their abilities by utilizing external resources such as the Internet. In this paper we describe our approach for achieving robot learning by combining basic skill transfer through physical demonstrations, reinforcement learning and web-based crowdsource rewards obtained via remote laboratory. We demonstrate how a robot that is taught very basic motion skills through kinesthetic teaching (i.e. lift, push, etc.), can learn to manipulate several objects through reinforcement learning by taking advantage of the feedback from multiple non-expert crowd workers. Although reinforcement learning is commonly intractable for real robot applications due to high exploration times, we demonstrate that by transferring the training to the crowd, it is possible to achieve robot learning while minimizing the effort of a dedicated teacher. This paper describes the system architecture, as well as the integration of our application to a current remote laboratory framework (Robot Management System). Experimental results with a real robot performing a time consuming learning task show a successful learning performance using crowdsouced feedback.


Demo Videos


Related Publications

Christian I. Penaloza, Sonia Chernova, Yasushi Mae and Tatsuo Arai: "Robot Reinforcement Learning using Crowdsourced Rewards", IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS 2013) Cloud Robotics Workshop. Tokyo, Japan. Nov. 3, 2013.