Safety Assurance for Automated Vehicles Beyond Collision Avoidance

Abstract

Each year, automotive crashes cause thousands of deaths and injuries. Autonomous safety systems have the potential to greatly reduce this tragic loss of life and improve safety, but such systems must meet existing requirements for automotive certification. Particularly, active safety systems must designed to comply with the Automotive Safety Integrity Level risk classification scheme described in the ISO 26262 standard.

In this thesis, I design a system using redundant components to independently enforce safety requirements across parallel software supervisors within an autonomous vehicle planning pipeline. I use Hamilton-Bellman-Jacobi reachability analysis to provide new guarantees for safe navigation on public roadways. I create new and extend existing safety modules to independently verify collision avoidance, obedience to traffic rules, and vehicle lane discipline. This project provides theoretical proof of safety and implements control methods within Nvidia’s DriveWorks autonomous vehicle framework.