SafeGen: Goal-Conditioned Video Diffusion of Safety-Critical Scenarios for VLM-Based Autonomous Driving

2026-07-23 · Source: cs.CV updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Robotics & Autonomous Systems, Cybersecurity & Data Privacy · Depth: Expert, extended

Summary

SafeGen is a goal-conditioned diffusion framework designed to generate realistic safety-critical video scenarios for evaluating vision-language model-based autonomous driving systems (VLMADs). It addresses the sim-to-real gap of simulator-based methods by formulating scenario generation as a diffusion process guided by a catastrophic end-state, such as a pedestrian collision. The framework employs Context Grounded End State Reasoning to identify latent vulnerabilities in benign driving contexts and infer structured end-state specifications. Subsequently, End State Conditioned Video Evolution grounds these semantic threats into physically plausible visual dynamics, instantiating high-risk agents via depth-aware geometric projection and boundary-conditioned diffusion. Experiments across 3 VLMADs show SafeGen increases the Judge Overall Score by 24.25% on average compared to baselines. Fine-tuning a VLMAD with SafeGen-generated data also improves real-world driving performance by an average of 15.9%. The system generates 850 scenarios at 1280x720 resolution, 10 FPS, with 81 frames.

Key takeaway

For AI scientists and machine learning engineers developing autonomous driving systems, SafeGen offers a critical tool to rigorously evaluate and enhance VLMAD safety. You should integrate SafeGen-generated safety-critical scenarios into your testing pipelines to expose inherent reasoning flaws and improve model robustness against rare, high-impact events. Fine-tuning your VLMADs with this synthetic data can significantly boost real-world performance by 15.9% on average.

Key insights

SafeGen generates realistic safety-critical videos for VLMAD evaluation by goal-conditioning diffusion on catastrophic end-states.

Principles

• Scenario generation can be a goal-conditioned diffusion process.
• Leverage VLMs to infer latent vulnerabilities from benign contexts.
• Anchor catastrophic end-states to guide temporal video evolution.

Method

SafeGen uses Context Grounded End State Reasoning to define a catastrophic end-state specification from benign video, then End State Conditioned Video Evolution synthesizes the adversarial last frame and interpolates intermediate dynamics via boundary-conditioned diffusion.

In practice

• Generate diverse VRU interactions for VLMAD stress testing.
• Fine-tune VLMADs with synthetic data to improve real-world safety.
• Evaluate VLMADs using Judge Overall Score (JOS) for reasoning flaws.

Topics

• Autonomous Driving Safety
• Vision-Language Models
• Video Diffusion Models
• Scenario Generation
• Vulnerable Road Users
• Safety-Critical Evaluation

Code references

• JoFrc/SafeGen

Best for: Computer Vision Engineer, Research Scientist, AI Scientist, AI Security Engineer, Machine Learning Engineer

Related on AIssential

• SafeGen: Goal-Conditioned Video Diffusion of Safety-Critical Scenarios for VLM-Based Autonomous Driving — Goal-conditioned diffusion effectively generates safety-critical scenarios for VLMAD evaluation by leveraging catastrophic end-states.
• RiskFlow: Fast and Faithful Safety-Critical Traffic Scenario Generation — RiskFlow generates realistic safety-critical traffic scenarios rapidly by transforming action sequences in a single forward pass, avoiding iterative denoising.
• World Engine: Towards the Era of Post-Training for Autonomous Driving — World Engine enables safe autonomous driving by post-training models on synthesized, safety-critical scenarios, overcoming real-world data scarcity and risks.
• Learning of Robot Safety Policies via Adversarial Synthetic Scenarios — Adversarial gamification efficiently discovers robot safety edge cases by pitting a hazard-creating Red Team against a policy-refining Blue Team.
• A knowledge-augmented dataset of high-risk driving scenarios with LLM annotations for autonomous driving — K-Risk uses LLM-augmented data to bridge structured trajectories with semantic reasoning for autonomous driving safety.
• Proposal-Conditioned Latent Diffusion for Closed-Loop Traffic Scenario Generation — A new diffusion framework improves closed-loop traffic simulation efficiency and control for autonomous vehicles.

See Counsel's argued verdicts on the open AI decisions leaders are weighing →

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.CV updates on arXiv.org.