Runtime Safety Assurance of Autonomous Vehicles

The paper deals with runtime safety assurance of autonomous vehicles (AVs), First, we closely examine the current state of AV technology, highlighting the incredible progress made by industry leaders in ensuring minimal risks in their vehicles. Our analysis of reported disengagements in AVs reveals interesting insights, showing that most disengagements are initiated by human test drivers rather than by the AV systems themselves and that they often occur on streets and freeways. Next, to ensure AVs' safety, we explored various safety architectures, including 1oo2, 1oo2D, and 2oo3, and thoroughly evaluated their reliability. We also introduce the concept of runtime safety assurance (RTA), a system that closely monitors the AV's state and employs emergency measures through a backup controller if the safety conditions are breached. Such a safety architecture - having separate safety and performance loops - has several benefits, which are highlighted in this paper. Furthermore, we emphasize the crucial role of robust sensing and perception systems in AVs, showing that multi-modal sensing and sensor fusion can effectively enhance the system reliability. We validated the reliability of these systems through extensive software testing and comparisons with ground-truth data. Finally, we explore fault-detection techniques for AVs by employing simulations to detect and handle faults in the control system.