Research map / Image generation / Autoregressive and latent generation
DiffScene: Diffusion-Based Safety-Critical Scenario Generation for Autonomous Vehicles
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes DiffScene, a diffusion model over adversarial SV trajectories guided at each reverse step by safety, functionality, and constraint objectives, to generate realistic safety-critical driving scenarios in CARLA.
Problem
Generating safety-critical driving scenarios that are rare, realistic, and transferable for efficient safety evaluation and improvement of autonomous vehicles.
Approach
Use a diffusion model trained on benign trajectories and, at each reverse diffusion step, apply gradient guidance from a weighted combination of safety (distance/collision), functionality (route incompletion), and constraint (speed naturalness) objectives to produce adversarial but natural SV trajectories in a single reverse pass.
Key results
- DiffScene achieves collision rates of 0.85, 0.87, and 0.79 across the three scenario settings, the highest among all baselines (best baseline ST 0.72, 0.73, 0.59 respectively per the table's comparisons).
- Under 5 naturalness distances, DiffScene trajectories are closest to benign trajectories (e.g., Fréchet 1.04 vs 1.12 for AT in S1).
- Fine-tuning SAC on DiffScene scenarios reduces collision rate to 0.26 versus 0.90 for the un-finetuned SAC in Table 3.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use DiffScene-generated scenarios instead of the strongest baseline scenarios | Scenario generation for AV evaluation · Crossing Negotiation (S1) | Collision Rate: 0.72 ± 0.11 -> 0.85 ± 0.08 | Helped |
| Use DiffScene-generated scenarios instead of the strongest baseline scenarios | Scenario generation for AV evaluation · Red-light Running (S2) | Collision Rate: 0.73 ± 0.08 -> 0.87 ± 0.10 | Helped |
| Use DiffScene-generated scenarios instead of the strongest baseline scenarios | Scenario generation for AV evaluation · Red-light Running (S3) | Collision Rate: 0.68 ± 0.16 -> 0.79 ± 0.15 | Helped |
| Fine-tune SAC with DiffScene-generated rather than Learning-to-collide-generated scenarios | SAC · Crossing Negotiation (S1), standard test set | Collision Rate: 0.32 -> 0.26 | Helped |
| Fine-tune SAC with DiffScene-generated rather than Adversarial Trajectory Optimization-generated scenarios | SAC · Crossing Negotiation (S1), standard test set | Incomplete Route: 0.13 -> 0.11 | Helped |
| Increase the number of adversarial optimization steps during each denoising step (ablation) | DiffScene · Safety-critical scenario generation | Collision Rate | Helped |
| Increase the number of adversarial optimization steps during each denoising step (ablation) | DiffScene · Safety-critical scenario generation | Trajectory similarity (SSPD) | Hurt |
Surprising findings
- Adversarially generated scenarios reduced collision rate of a fine-tuned agent most with DiffScene, but baselines like LC and AT were effective at specific metrics, showing metric-dependent rankings.
Limitations
- Speed satisfaction is hard to achieve due to vehicle acceleration limits
- Only the naturalness constraint is used in experiments; other constraint combinations left for future work
- Trade-off between safety-criticality and naturalness controlled by number of optimization steps
Experimental setup
Datasets: CARLA benign trajectory dataset (6,995 trajectories) NHTSA pre-crash scenario settings
Benchmarks: CARLA safety-critical scenarios across 3 NHTSA pre-crash settings (S1-S3)
Metrics: collision rate incomplete route speed satisfaction SSPD Fréchet distance DTW Wasserstein distance KL divergence
Baselines: Adversarial RL (AR) Carla Scenario Generator (CS) Learning-to-collide (LC) AdvSim (AS) Adversarial Trajectory Optimization (AT) STRIVE (ST)
Builds on: Diffuser Classifier guidance / CTG STRIVE SafeBench
Closest papers by meaning
- Causal Composition Diffusion Model for Closed-loop Traffic Generation (CVPR 2025 · similarity 0.87)
- Hybrid-Driving: An Autonomous Driving Decision Framework Integrating Large Language Models, Knowledge Graphs and Driving Rules (AAAI 2025 · similarity 0.87)
- SceneDiffuser: Efficient and Controllable Driving Simulation Initialization and Rollout (NeurIPS 2024 · similarity 0.87)
- NAVSIM: Data-Driven Non-Reactive Autonomous Vehicle Simulation and Benchmarking (NeurIPS 2024 · similarity 0.86)
- Scenario Dreamer: Vectorized Latent Diffusion for Generating Driving Simulation Environments (CVPR 2025 · similarity 0.86)
- Diffusion-ES: Gradient-free Planning with Diffusion for Autonomous and Instruction-guided Driving (CVPR 2024 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.