Safe Robot Learning by Energy Limitation
Online robot learning has been a goal for researchers for several decades. A problem arises when learning algorithms need to explore the environment as actions cannot easily be anticipated. Because of this, safety is a major issue when using learning algorithms.
This paper presents a framework for safe robot learning by the use of region-classification and energy limitation. The main task of the framework is to ensure safety regardless of a learning algorithm’s input to a system. This is necessary to allow a learning robot to explore environments without damaging itself or its surroundings. To ensure safety, the state-space is divided into fatal, supercritical, critical and safe regions, depending on the energy of the system.
To show the adaptability of the framework it is used on two different systems; an actuated swinging pendulum and a mobile platform. In both cases obstacles with unknown locations must are avoided successfully.
Unable to display preview. Download preview PDF.
- 1.Connell, J.H., Mahadevan, S.: Introduction to Robot Learning. Springer (1993)Google Scholar
- 2.Olivier Chapelle, A.Z., Schölkopf, B.: Semi-Supervised Learning. The MIT Press (2006)Google Scholar
- 5.Kaelbling, L.P., Littman, M.L., Moore, A.W.: Reinforcement learning: A survey. Journal of Artificial Intelligence Research 4, 237–285 (1996)Google Scholar
- 6.Gillula, J.H., Tomlin, C.J.: Guaranteed safe online learning of a bounded system. In: IROS 2011, pp. 2979–2984 (September 2011)Google Scholar
- 7.Hans, A., Schneegaß, D., Schäfer, A.M., Udluft, S.: Safe exploration for reinforcement learning. In: European Symposium on Artificial Neural Network, pp. 143–148 (April 2008)Google Scholar
- 8.Fjerdingen, S.A., Kyrkjebø, E.: Safe reinforcement learning for continuous spaces through Lyapunov-constrained behavior. In: Frontiers in Artificial Intelligence and Applications, pp. 70–79 (May 2011)Google Scholar