Research map / Image generation / Autoregressive and latent generation
Causal Composition Diffusion Model for Closed-loop Traffic Generation
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 CCDiff, a causal composition diffusion model that identifies a decision causal graph among traffic agents, ranks agents by causal importance, and applies masked and interventional guidance to generate controllable yet realistic closed-loop safety-critical traffic on nuScenes.
Problem
Closed-loop traffic simulation for autonomous driving must be both controllable (user-specified safety-critical objectives such as collisions) and realistic, but these objectives conflict, especially in long-tail safety-critical scenarios.
Approach
Formulates closed-loop scenario generation as a constrained factored MDP, learns a decision causal graph from TTC-masked spatial attention, ranks agents by causal importance, and applies guidance only to top-ranked agents (masked classifier guidance) plus a do-intervention style classifier-free guidance over causal parents, so realism is preserved while controllability is increased.
Key results
- In K-agent generation CCDiff reaches realism 0.97 at K=2 and controllability 0.74 at Full agents, with best or second best in 10 of 12 metrics versus baselines.
- In long-horizon generation with 5 controllable agents at T=4s, CCDiff reaches CS 0.89 versus CTG 0.78 with comparable realism.
- Causal ranking outperforms distance-based and human ranking, with human/distance ranking losing roughly 5-10% in collision rate.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use CCDiff instead of CTG | closed-loop traffic generation · nuScenes, K=5 controllable agents | CS: 0.41 -> 0.56 | Helped |
| Use CCDiff instead of CTG | closed-loop traffic generation · nuScenes, K=5 controllable agents | RS: 0.84 -> 0.88 | Helped |
| Use CCDiff instead of CTG | closed-loop traffic generation · nuScenes, full agent set | CS: 0.56 -> 0.74 | Helped |
| Use CCDiff instead of CTG | closed-loop traffic generation · nuScenes, full agent set | RS: 0.77 -> 0.76 | No clear effect |
| Use CCDiff instead of CTG | closed-loop traffic generation · nuScenes, 5 controllable agents, T=5s | CS: 0.67 -> 0.78 | Helped |
| Use CCDiff instead of CTG | closed-loop traffic generation · nuScenes, 5 controllable agents, T=5s | RS: 0.53 -> 0.49 | Hurt |
| Remove causal sparsity constraints from the scene encoder (ablation) | CCDiff · nuScenes, multi-agent, K=5 | SCR | Hurt |
| Replace factorized guidance with whole-state-space guidance (ablation) | CCDiff · nuScenes, multi-agent generation | SCR | Hurt |
| Replace causal ranking with distance-based ranking (ablation) | CCDiff · nuScenes, multi-agent generation | SCR | Hurt |
| Replace causal ranking with human ranking (ablation) | CCDiff · nuScenes, multi-agent generation | SCR | Hurt |
Surprising findings
- STRIVE achieves high controllability but the poorest realism score among baselines.
- TrafficSim achieves the best controllability in long-horizon generation yet has second-worst realism there.
Limitations
- Causal reasoning pipeline relies on hyperparameter tuning and is hard to directly evaluate.
- Evaluated only on nuScenes with rasterized maps and closed-loop simulation via tbsim.
- Realism is quantified with proxy metrics (off-road rate, FDE, comfort distance) rather than a holistic realism measure.
Experimental setup
Datasets: nuScenes traffic behavior simulation (tbsim)
Benchmarks: nuScenes closed-loop simulation
Metrics: scenario collision rate (controllability score) off-road rate final displacement error comfort distance generational distance inverted generational distance
Baselines: SimNet TrafficSim STRIVE BITS CTG
Builds on: classifier-free guidance CTG DDPM TrafficSim BITS
Closest papers by meaning
- ChainFlow-VLA: Causal Flow Planning with Vision-Language Models (NeurIPS 2026 · similarity 0.88)
- SceneDiffuser: Efficient and Controllable Driving Simulation Initialization and Rollout (NeurIPS 2024 · similarity 0.88)
- DiffusionDrive: Truncated Diffusion Model for End-to-End Autonomous Driving (CVPR 2025 · similarity 0.87)
- DiffScene: Diffusion-Based Safety-Critical Scenario Generation for Autonomous Vehicles (AAAI 2025 · similarity 0.87)
- Diffusion-ES: Gradient-free Planning with Diffusion for Autonomous and Instruction-guided Driving (CVPR 2024 · similarity 0.87)
- RLFTSim: Realistic and Controllable Multi-Agent Traffic Simulation via Reinforcement Learning Fine-Tuning (CVPR 2026 · 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.