End to End Learning for Self-Driving Cars
Mariusz Bojarski NVIDIA Corporation Holmdel, NJ 07735、Davide Del Testa NVIDIA Corporation Holmdel, NJ 07735、Daniel Dworakowski NVIDIA Corporation Holmdel, NJ 07735、Bernhard Firner NVIDIA Corporation Holmdel, NJ 07735、Beat Flepp NVIDIA Corporation Holmdel, NJ 07735、Prasoon Goyal NVIDIA Corporation Holmdel, NJ 07735、Lawrence D. Jackel NVIDIA Corporation Holmdel, NJ 07735、Mathew Monfort NVIDIA Corporation Holmdel, NJ 07735、Urs Muller NVIDIA Corporation Holmdel, NJ 07735、Jiakai Zhang NVIDIA Corporation Holmdel, NJ 07735、Xin Zhang NVIDIA Corporation Holmdel, NJ 07735、Jake Zhao NVIDIA Corporation Holmdel, NJ 07735、Karol Zieba NVIDIA Corporation Holmdel, NJ 07735
Abstract
We trained a convolutional neural network (CNN) to map raw pixels from a single front-facing camera directly to steering commands. This end-to-end approach proved surprisingly powerful. With minimum training data from humans the system learns to drive in traffic on local roads with or without lane markings and on highways. It also operates in areas with unclear visual guidance such as in parking lots and on unpaved roads.
中文速览
英伟达的研究团队用一个卷积神经网络（CNN）直接把车载摄像头拍到的原始画面映射成方向盘转角，让汽车学会了端到端（end-to-end）自动驾驶，完全不需要人工编写车道线检测、路径规划等分步骤规则。他们收集了约72小时的人类驾驶数据，并通过模拟偏移和旋转来扩充训练样本，使网络还能学会在跑偏后自主纠正。实测结果显示，在新泽西州混合路况下自动驾驶时间占比约98%，在高速公路上行驶10英里零次人工接管，而网络激活图也直观地证明了它在没有任何显式标注的情况下自己"学会"了识别路面轮廓。这项工作的意义在于：端到端训练让系统内部各环节协同优化，省去了大量人工设计的中间步骤，为低成本、高性能的自动驾驶方案提供了一条可行路径。
原文 arXiv:1604.07316;中英对照 + 大白话阅读 https://aha.fim.ai/paper/1604.07316v1