Investigating the Impact of Time-Lagged End-to-End Control in Autonomous Driving

End-to-end training, the strategy by which deep neural networks learn to map raw pixels from front-facing cameras directly to steering commands have recently gained attention in the autonomous driving field. Here, we investigate the possibility of extending this approach with a time-lagged procedure, by training the system to map raw pixels at time T to steering commands at time T+Lag. We are interested in evaluating such an approach for two main applications: (1) time-lagged end-to-end control towards an artificial driving instructor (ADI) that recommends future control actions to novice human drivers; (2) time-domain data augmentation to improve the performance of standard non-lagged end-to-end control. Our results show that time-lagged end-to-end training is not appropriate for time-lagged control, but using it for data augmentation leads to a smoother output in standard non-lagged end-to-end control. This suggests that time-lagged training improves the anticipatory ability of end-to-end control and augmentation reduces overfitting.