Functional Safety in AI-controlled Vehicles: If Not ISO 26262, Then What?

Since its establishment in 2011, the ISO 26262 international functional safety standard has rapidly emerged as the definitive guideline for automotive engineers looking to optimize the safety of electrical and/or electronic (E/E) automotive systems. But because ISO 26262 implies strict adherence to analyzing the architecture and its effect on safety, the shift toward machine learning for critical driving decisions in self-driving cars threatens to break the standard's direct link between functional safety and the requirements for how these new concepts should be fulfilled. After presenting a short history of automotive functional safety standards up to ISO 26262, this session outlines the standard's deficiencies relative to AI-controlled vehicles. In this session we'll discuss the requirements and challenges that automated driving presents to existing standards bodies which increasingly have to accommodate a more autonomous future. Whereas ISO 26262 specifies requirements for eliminating safety hazards in the presence of an E/E system fault, this discussion suggests that new standards for the AI era must address Safety of the Intended Function (SOTIF) helping to validate that advanced automotive functionalities are engineered into the vehicle to avoid safety hazards even in the absence of a fault.