RL in real, distributed, and cooperative Environments

Within the DFG 'Schwerpunktprogramm' (SPP 1125: 'Kooperierende Teams mobiler Roboter in dynamischen Umgebungen') we are investigating Reinforcement Learning methods for real robots. Especially, we are interested in two research questions:

1. to develop learning methods, that are fast and robust enough to be applicable directly to real robots.
2. to investigate algorithms for the learning of team behaviour in distributed multi-agent systems.

To prove the practical relevance of our learning algorithms, we maintain a team in the simulation league of RoboCup, the Brainstormers. Having realized a growing part of decision making components by neural networks (trained on the basis of reinforcement learning), this team has performed quite well winning many national and international titles.

In 2003, we started to build a MidSize team of real autonomous soccer agents, as a testbed for reinforcement learning on real systems. This team has won many national and international titles, as well. At the German Open 2005, as part of a collaboration with the University of Freiburg, we demonstrated two teams of three humanoid robots (RoboSapiens) each playing a match of soccer.

Start: 2000 End: 2006

This project was funded for six full years by the DFG as part of the 'Schwerpunktprogramm' 1125. A follow-up project (DFG Transferprojekt) ist still running.

Researchers working on this project:

• Prof. Dr. Martin Riedmiller
• Dr. Thomas Gabel
• Dr. Roland Hafner
• Dr. Sascha Lange
• Dr. Martin Lauer
• Dr. Artur Merke

For more information on this research project, please contact Martin Riedmiller.