"Curse of rarity" for autonomous vehicles

Achieving the human-level safety performance for autonomous vehicles (AVs) remains a challenge. One critical bottleneck is the so-called "long-tail challenge", which usually refers to the problem that AVs should be able to handle seemingly endless low-probability safety-critical driving scenarios, even though millions of testing miles have been accumulated on public roads. However, there is neither rigorous definition nor analysis of properties of such problems, which hinders the progress of addressing them. In this paper, we systematically analyze the "long-tail challenge" and propose the concept of "curse of rarity" (CoR) for AVs. We conclude that the compounding effects of the CoR on top of the "curse of dimensionality" (CoD) are the root cause of the "long-tail challenge", because of the rarity of safety-critical events in high dimensionality of driving environments. We discuss the CoR in various aspects of AV development including perception, prediction, and planning, as well as validation and verification. Based on these analyses and discussions, we propose potential solutions to address the CoR in order to accelerate AV development and deployment.