Python Deep Learning - Second Edition

Training

ChauffeurNet was trained with 30 million expert driving examples, using imitation supervised learning. The middle-level, top-down input allows to use different sources of training data with ease. On one hand, it can be generated from real-world driving with a fusion between the vehicle sensor inputs (cameras, lidar) and mapping data such as streets, traffic lights, traffic signs, and so on. On the other hand, we can generate images of the same format with a simulated environment. As we mentioned in section Imitiation driving policy, this allows us to simulate situations that occur rarely in the real world, such as emergency braking or even crashes. To help the agent learn about such situations, the authors of the paper explicitly ...