RoboCup robotic soccer tournament seriesAn annual competition that serves as the focus of most of the following work.
Work is under way at several labs to apply machine learning techniques to soccer-playing robots, both real and simulated.
| Jay : game learning : soccer |
RoboCup robotic soccer tournament series
An annual competition that serves as the focus of most of the following work.
The Dynamo Project at the University of British Columbia.
Peter Stone at CMU.
Asada Laboratory at Osaka University.
Microb in France.
JavaBots simulator by Tucker Balch at Georgia Tech.
The robots used range from the tabletop toy cars of The Dynamo Project to the medium-sized camera-carrying machines of Asada Lab. The investigators modify the rules of soccer to simplify the robots' jobs and to reduce expense--for example, nobody seems to be using more than four robots on a team. The robots "kick" by bumping the ball. Tasks investigated range from learning simple behaviors like "kick the ball" to dealing with high-level questions of cooperation with one's own team and recognizing opponent actions.
Teaching robots to play soccer (1995)
Peter Stone
Text. Brief abstract of a talk on robotic soccer, given April 1995.
What a great domain! Robotic soccer is extremely difficult, with vast scope for new ideas, but since it's a game it's not as hard to tell when you're making real progress. (Compare it to a vaguely-defined field like computer vision.) It integrates problems of vision, low-level control, real-time response, planning and plan recognition, and competition and cooperation all in one domain, spurring wide-ranging work. Playing well is so difficult that I imagine learning will be necessary for good performance--no one will have the time to hand-code all the behaviors the robots will need, and even then there is scope for learning the opponents' behavior.
A drawback is that robots are expensive and complicated. A lot of people who could contribute won't have the time, money, or expertise to deal with real robot hardware, or won't be interested in the important but low-level details of sensing and motor control. (They can still compete in the RoboCup simulator league.)
Standard rules, like those for the RoboCup tournament, are an excellent way to keep researchers focused on the problem, promoting progress.
interesting pages: ASCII Soccer - a soccer-like game.
online papers: "Markov games as a framework for multi-agent reinforcement learning", using a simplified, slightly soccer-like task