KG-ASG: New framework generates safer self-driving crash scenarios with 40% fewer multi-collisions

Autonomous driving systems need rigorous safety validation, but generating realistic, high-risk accident scenarios with clear causality is notoriously difficult. Existing methods often produce ambiguous collision causes or uncontrolled multi-vehicle pileups. Enter KG-ASG (Collision-Knowledge-Guided Closed-Loop Adversarial Scenario Generation with Primary-Support Attribution), a new framework developed by Cheng Wang and colleagues that combines a structured collision knowledge base with a lightweight Collision Expert. The expert infers the target collision mode, designates a primary adversary (the main causal vehicle), and assigns support vehicles that shape the risk environment without becoming additional colliders. Hard constraints—rule-based, physical, interaction-safety, and single-collider—filter out non-executable samples, while planner-controller feedback enables failure diagnosis, candidate re-ranking, and terminal refinement.

Tested on the WOMD (Waymo Open Motion Dataset) scenarios reconstructed in MetaDrive, KG-ASG demonstrates strong adversarial effectiveness across three different controllers (IDM, Cruise, Expert). Key results include improved Valid Primary Attack rates, a significant reduction in multi-vehicle collisions, and measurable closed-loop recovery gains. The framework’s primary-support single-collider reasoning not only improves adversarial strength but also enhances interpretability and executability—crucial for safety engineers debugging AV behavior. By generating scenarios with clear causal chains and single-point failures, KG-ASG could become a standard tool for AV safety validation, helping to expose edge cases that current methods miss.