Still on the Way: Too Many Traffic Scenarios? Not Enough! [Editor’s Column]

Abstract

When speaking of virtual testing of driverless systems (hereinafter referred to as “systems”) in open scenarios, researchers, and engineers usually frown, scratching their heads and complaining about how time-consuming and effort-intensive it is to design valuable and useful virtual traffic scenarios (hereinafter referred to as “scenarios”). At present, although more and more virtual scenarios are coming up every once in a while, effective data are becoming fewer and fewer, the cost of model training is going higher and higher, and improvement in driving ability is getting slower and slower. Countless virtual scenarios may never cover the vastness of reality, and various long-tail corner cases will always “amaze” us in unknown and unexpected ways.