Complete Safety Analysis of Known and Unknown Scenarios in Autonomous Vehicles Based on STPA Loss Scenarios

Artificial intelligence turns out an increasingly important role in autonomous driving system (ADS), especially for world model perceptions and ego vehicle trajectory planning in an ADS, contributing to the safety for the occupies and surrounding traffics. The performance of an ADS depends on the level of absence for functional insufficiency and performance limitation of the components and algorithms including AI in known and unknown scenarios. In this paper, we propose using System Theoretic Process Analysis, STPA, to characterize those known and unknown scenarios for SAE automation Levels 3 and 4. A key challenge of STPA is the identification of an appropriate dynamic control structure that is efficient for the purpose at hand. An ideal control structure should be able to include all causes of failure. What “all” really implies here is one of the central challenges. One of the implications of “all” is that a safety analysis based on STPA control structures should have a predictive power, meaning that all root causes that could violate safety goals should be covered. We propose driving environment model to structure an ADS including its relations to the user, the environment, and all other traffic actors. More specifically, we show how an analysis based on this model can be used to identify safe and unsafe control actions (UCAs) in known and unknown scenarios.