Local movement control with neural networks in the small size league

In the RoboCup small-size league, most teams calculate the robots' positions by means of a camera that is mounted above the field as well as different kinds of artificial intelligence methods that run on an additional PC. This processing loop induces various time delays, which require forecasting routines, if more accurate behaviors are desired. This paper shows that by utilizing a combination of a neural network and local sensors, the robot is able to estimates its actual position quite accurately. This paper furthermore shows that the learning procedure is also able to compensate for slip and friction effects that cannot be observed by the local sensors.