Safe autonomous driving may be one of the most difficult engineering challenges that any artificial intelligence system has been asked to do since the birth of AI over sixty years ago. The difficulty is not within the task itself, but rather in the extremely small margin of allowable error given the human life at stake and the extremely large number of edge cases that have to be accounted for. In other words, we task these systems to expect the unexpected with near 100% accuracy, which is a technical challenge for machine learning methods that to date have generally been better at memorizing the expected than predicting the unexpected. In fact, the process of efficiently and automatically discovering the edge cases of driving may be the key to solving this engineering challenge. In this work, we propose and evaluate a method for discovering edge cases by monitoring the disagreement between two monocular-vision-based automated steering systems. The first is a proprietary Tesla Autopilot system equipped in the first generation of Autopilot-capable vehicles. The second is a end-to-end neural network trained on a large-scale naturalistic dataset of 420 hours or 45 million frames of autonomous driving in Tesla vehicles.
View on arXiv