Agentic data science with the Data Agent Kit
cloud.google.com·6h ago
AI Research
arXiv CS.AI·Sep 3·5 min read
CrashDiffuser: VLM-Guided Collision Intent Reasoning for Fine-Grained Safety-Critical Traffic Scenario Generation
TL;DR
CrashDiffuser is a novel framework for generating safety-critical traffic scenarios that allows precise control over collision contact regions. It achieves a target-collision rate of 50.33% in one attempt, improving to 67.98% after three attempts.
✦ Why It Matters
Implement CrashDiffuser to improve the testing of autonomous driving systems with precise collision scenario generation.
Key Takeaways
- Utilize CrashDiffuser for generating specific collision scenarios.
- Achieve high target-collision rates with iterative attempts.
- Control collision contact regions for better testing accuracy.
- Leverage VLM for enhanced scene context understanding.
How It Works
CrashDiffuser employs a closed-loop VLM-guided diffusion framework that separates collision reasoning from trajectory generation. It initializes by extracting scene-level context and, during each replanning step, predicts structured actions that dictate vehicle speed, turning, and collision stages.
This structured intent informs a diffusion model to create realistic trajectories that adapt to the evolving behavior of the target vehicle.