A Safety Architecture for Self-Driving Systems

Shai Shalev-Shwartz, Moran Molnar, Ilai Granot, Almog Shany, Amnon Shashua

Abstract:

With the increasing presence of autonomous vehicles on the road and the imminent commercialization of fully
autonomous systems at scale, the need for transparency regarding the robust safety architectures underpinning these
systems has become paramount. Mobileye is advancing self-driving technologies across a spectrum of autonomy
levels, including Driver Assistance, Eyes-off, and fully No-driver systems. This paper describes our architecture
for Eyes-off and No-driver Self-Driving Systems (SDSs), with a strong emphasis on adherence to critical safety
principles. We advocate two primary safety goals. The first is the elimination of unreasonable risk, with an emphasis
on transparency about the boundaries distinguishing reasonable and unreasonable risks. We provide crisp definitions
for these boundaries by addressing lapses of judgment in planning, identifiable hardware failures, the elimination of
reproducible errors, and leveraging redundancy to mitigate black swan events. The second goal focuses on the greater
good: self-driving cars must reduce overall harm compared to the status quo of human-driven vehicles.