Robustness of autonomous driving systems: extending the function failure modes taxonomy

Abstract

While the development of autonomous driving systems (ADS) offers significant potential safety benefits, this is conditional upon addressing the challenges associated with increased complexity and multidisciplinarity, including the introduction of artificial intelligence (AI) embedded components. This increases the importance of assuring the robustness of these systems during their operations where possible hazards can arise from both the malfunctioning behavior of safety-related in-vehicle systems and functional insufficiencies or performance limitations of a system that is free from faults. Robustness is underpinned by the systematic consideration of failure modes which indicate how a system can potentially fail to deliver the intended function/performance and associated requirements. The introduction of AI components has made the identification for failures of ADSs challenging due to increased use of sensors and wireless communications which provide the backbone of core software functions of perception, planning and control. Current literature on failure modes mostly focuses on “physical” systems, and do not readily support the analysis of safety-critical systems which include AI components. This paper discusses the applicability of current failure mode taxonomies and their effectiveness in supporting the identification of failure modes within the context of design methodologies (such as FMEA) that underpin an effective approach to robustness of multidisciplinary autonomous driving systems.