Abstract

As self-driving vehicle fleets expand across North America, new stories regularly report mishaps where the vehicle's perceptions systems misunderstand a situation that humans have no problem with. The immense variety of the human driving environment continues to confound our best efforts in learning-based autonomy, a situation that will get worse as we add northern climates into the mix. In this talk, I'll briefly outline the main challenges and progress made toward deploying AV fleets in winter driving conditions, and present some emerging methods for predicting the perception uncertainty associated with each driving scene a vehicle encounters.