Safety-Centric Architecture for AI-Enabled Autonomous Driving Systems

Industry and academia are rapidly moving towards the development of next-generation Autonomous Driving (AD) systems. These AD systems are expected to be equipped with novel AI algorithms to perform several complex tasks, ranging from object detection and path planning to control decisions and cloud communication. However, safety is of utmost importance in such systems because errors, including AI decision errors, e.g., object detection failure, traffic light detection error, speed limit violations, etc., can lead to catastrophic consequences. This paper presents a safety-centric architecture for building dependable AD systems. Specifically, we introduce the concept of adaptive safe state manager to identify potential safety risks in the AD system and activate respective fail-safe mechanisms to prevent safety hazards. The proposed architecture is aligned with industrial safety standards such as ISO 26262 and ISO 21448.