KG-ASG: Collision-Knowledge-Guided Adversarial Scenario Generation for Autonomous Driving
A novel framework named KG-ASG (Collision-Knowledge-Guided Closed-Loop Adversarial Scenario Generation with Primary-Support Attribution) has been introduced to enhance the safety validation of autonomous driving technologies. This approach overcomes the shortcomings of current methods that depend on low-level trajectory adjustments, collision-proxy optimization, or single-adversary searches, which may lead to unclear collision causes or unmanageable multi-vehicle incidents. KG-ASG develops a comprehensive collision knowledge database and trains a streamlined Collision Expert to deduce the target collision mode, principal adversary, support vehicles, and their interaction roles. The generation of multi-vehicle adversaries is conceptualized as a primary-support mechanism. This research is available on arXiv under ID 2605.18895.
Key facts
- KG-ASG is a collision-knowledge-guided closed-loop adversarial scenario generation framework.
- It uses primary-support attribution to model multi-vehicle interactions.
- A structured collision knowledge base is constructed.
- A lightweight Collision Expert infers collision mode, primary adversary, and support vehicles.
- The framework addresses ambiguous collision causes in existing methods.
- Existing methods rely on low-level trajectory perturbations or single-adversary search.
- The paper is available on arXiv under ID 2605.18895.
- The approach is designed for safety validation of autonomous driving systems.
Entities
Institutions
- arXiv