SAFETY ENHANCEMENT OF AI-ENABLED AUTONOMOUS VEHICLE ARCHITECTURES VIA ASSURANCE CASE DESIGN ANALYSIS

Abstract

The exponential advancements in Artificial Intelligence (AI) capabilities in recent years have sparked rapid integration of this technology within the automotive industry via the addition of numerous autonomous vehicle control features. Despite these features being touted for the safety improvements they provide, AI-enabled feature development remains largely focused on performance with less research being done on corresponding safety systems. This work makes use of an existing Connected & Automated Vehicle (CAV) architecture designed and developed for the EcoCAR EV challenge - a collegiate vehicle engineering competition sponsored by the United States of America (USA) Department of Energy (DOE), General Motors (GM), and Mathworks. The CAV architecture makes use of AI-enabled perception algorithms without appropriate mitigation strategies required in safety-critical systems to cover hazards inherently introduced by AI as this was beyond the scope of the competition. An assurance case is used to argue the safety of the vehicle within the context of the competition, and to identify perceived points of failure from a system safety perspective. A modified CAV architecture inspired by the Perception Simplex architecture [Bansal et al., 2024] is proposed which introduces a parallel perception safety layer using the same sensor data and comprised exclusively of deterministic algorithms. The safety layer compares its output to that of the high-performance AI-powered layer and overrides vehicle control commands if discrepancies are found. This proposed architecture was designed to meet the constraints of the EcoCAR competition. A second assurance case was constructed to outline the assumptions and supporting evidence required to demonstrate the safety of the proposed architecture. This thesis aims to document the feasibility of rendering existing AI-enabled systems fault-tolerant, the usefulness of assurance cases as tools to guide the iterative design process of complex safety critical systems, and the practicality of using assurance cases to justify the safety of these systems throughout their development timeline.