Preemptive Detection of Unsafe Motion Liable for Hazard

摘要

Establishing a safety standard for autonomous vehicles operating in open and dynamic environment is a challenge. As collisions are inevitable in over-constrained situations, we focus on deciding the liability for a hazard. Our insight is that hazards caused by malfunctions of autonomous vehicles result from loss of functional integrity. Design defects may leave it unnoticed, or the real-world may make integrity-preserving motion infeasible. Guarantee of functional integrity in an observable way at run-time is indispensable for revealing defects by using formal root-cause analysis, and for supporting safety claims by dismissing unreasonable doubts about design defects. From a practitical standpoint, we attempt to formalize a verification problem that consists of a novel criterion for determining liability for hazard, a safety claim comprised of confirmed observable states, and assumptions underlying the safety claim. We propose a run-time scheme of monitoring events that may lead to violations of the assumptions and a precursor to root-causes leading to loss of functional integrity and consequent hazards. We formulate a means of preemptively detecting unsafe motions liable to be hazardous as satisfiability problem within the framework of an adversarial motion planning subject to assumptions on maneuverability of movers. A numerical study shows that the run-time scheme using non-linear programming (NLP) encoding is viable in a real-world setting.