This is an old revision of the document!
End-to-end imitation systems can suffer a domain shift between the off-line training experience and the on-line behavior [35]. This problem, however, can be partially addressed in practice by data augmentation [8, 12]. Nonetheless, in spite of the early and recent successes of behavior cloning for end-to-end driving [32, 23, 10, 8, 12], it has not yet proved to scale to the full spectrum of driving behaviors, such as reacting to multiple dynamic objects.