Research Analyzer
← Back ICRA 2026

Conditional Flow-VAE for Safety-Critical Traffic Scenario Generation

Zimu Gong, Brian Zhaoning Zhang, Chris Zhang, Kelvin Wong, Raquel Urtasun

PDF

AI summary

Key figure (auto-extracted from paper)
A conditional flow VAE successfully maps nominal driving scenes to realistic safety-critical scenarios in latent space, outperforming existing baselines in realism and diversity.
Safety-critical scenarios Flow matching Conditional VAE Autonomous vehicle testing Scenario generation Latent space transport

Problem

Safety-critical driving events are rare in real data, and current simulation methods either scale poorly due to manual curation or produce unrealistic adversarial behaviors, hindering robust autonomous vehicle testing.

Approach

The method trains a conditional VAE on mixed real and synthetic data, then uses a flow matching transformer to learn a latent-space transport map that steers nominal scenarios toward safety-critical distributions.

Key results

  • Highest near-miss generation rate among evaluated baselines
  • High distributional similarity to real-world driving statistics
  • Controllable scenario difficulty via heuristic maneuver labels
  • Superior realism and diversity compared to VAE and STRIVE baselines

Why it matters

Enables scalable, realistic stress-testing for autonomous vehicles, accelerating safe deployment and benchmarking.

Abstract

Safety-critical scenarios are essential for the devel- opment of autonomous vehicles (AVs) but are rare in real-world driving data. While simulation offers a way to generate such scenarios, manually designed test cases lack scalability, and adversarial optimization often produces unrealistic behaviors. In this work, we introduce a conditional latent flow matching approach for scalable and realistic safety-critical scenario gen- eration. Our method uses distribution matching to transform nominal scenes into safety-critical rollouts. Furthermore, we demonstrate that incorporating both simulation and real-world data enables our framework to efficiently generate diverse, data-driven scenarios. Experimental results highlight that our approach is able to more consistently and realistically generate novel safety-critical scenarios, making it a valuable tool for training and benchmarking AV systems.

Index terms

Autonomous Vehicle Navigation Learning from Demonstration Simulation and Animation

Related papers