Learning drivers for TORCS through imitation using supervised methods

In this paper, we apply imitation learning to develop drivers for the open racing car simulator (TORCS). Our approach can be classified as a direct method in that it applies supervised learning to learn car racing behaviors from the data collected from other drivers. In the literature, this approach is known to have led to extremely poor performance with drivers capable of completing only very small parts of a track. In this paper we show that, by using high level information about the track ahead of the car and by predicting high level actions, it is possible to develop drivers with performances that in some cases are only 15% lower than the performance of the fastest driver available in TORCS. Our experimental results suggest that our approach can be effective in developing drivers with good performance in nontrivial tracks using a very limited amount of data and computational resources. We analyze the driving behavior of the controllers developed using our approach and identify perceptual aliasing as one of the factors which can limit performance of our approach.