Learning the track and planning ahead in a car racing controller

We propose a robust approach for learning car racing track models from sensory data for the car racing simulator TORCS. Our track recognition system is based on the combination of an advanced preprocessing step of the sensory data and a simple classifier that delivers six types of track shapes similar to the ones a human would recognize. Out of these, establishing a complete track model is straightforward. This model provides an information advantage to controller strategies, as it generally enables planning. We demonstrate how such a planning controller can be derived by a mixture of expert knowledge and a simple evolutionary learning approach and give experimental evidence that knowing not only the current conditions but also the big picture of the track is beneficial, as may be expected.