Efficient Learning of Pre-attentive Steering in a Driving School Framework
KI - Künstliche Intelligenz
Springer Berlin Heidelberg
© Springer-Verlag Berlin Heidelberg 2014
Autonomous driving is an extremely challenging problem and existing driverless cars use non-visual sensing to palliate the limitations of machine vision approaches. This paper presents a driving school framework for learning incrementally a fast and robust steering behaviour from visual gist only. The framework is based on an autonomous steering program interfacing in real time with a racing simulator: hence the teacher is a racing program having perfect insight into its position on the road, whereas the student learns to steer from visual gist only. Experiments show that (i) such a framework allows the visual driver to drive around the track successfully after a few iterations, demonstrating that visual gist is sufficient input to drive the car successfully; and (ii) the number of training rounds required to drive around a track reduces when the student has experienced other tracks, showing that the learnt model generalises well to unseen tracks.
The final publication is available at Springer via http://dx.doi.org/10.1007/s13218-014-0340-1
Vol. 29 (1), pp. 51 - 57