Autonomy 2.0: Why is self-driving always 5 years away?

Despite the numerous successes of AI over the past decade in image recog1 nition, decision-making, NLP and image synthesis, self-driving technology has not 2 yet followed the same trend. In this paper, we study the history, composition, and 3 development bottlenecks of the modern self-driving stacks adopted across the industry. 4 We argue that the slow progress is caused by taking approaches that require too much 5 hand-engineering, an over-reliance on road testing, and high fleet deployment costs. We 6 observe that the classical stack has several bottlenecks that preclude the necessary scale 7 needed to capture the long tail of rare events. To resolve these problems, we outline the 8 principles of Autonomy 2.0, an ML-first approach to self-driving, as a viable alternative 9 to the currently adopted state-of-the-art. This approach is based on (i) a fully differ10 entiable AV stack trainable from human demonstrations, (ii) closed-loop data-driven 11 reactive simulation, and (iii) large-scale, low-cost data collections as critical solutions to12 wards scalability issues. We outline the general architecture, survey promising works in 13 this direction and propose key challenges to be addressed by the community in the future. 14