HaoweiLi/CrashFactory-Dataset

CrashFactory Dataset

Crash reports and reconstructed scenarios of the CrashFactory benchmark

📄 Paper | 🌐 Website | 💻 Code

CrashFactory turns official traffic crash records into executable, safety-critical simulation scenarios. This dataset is its benchmark: the crash reports it starts from and the complete outputs of every case.

Four reconstructed MTCF crashes from above: rear-end, head-on, angle and sideswipe.

📊 Contents

Database	Cases	Inputs	Outputs
Michigan Traffic Crash Facts (MTCF)	100, 20 per crash type	mtcf/reports.zip (5 MB)	mtcf/results.zip (1.0 GB)
NHTSA Crash Investigation Sampling System (CISS)	48	nhtsa/reports.zip (39 MB)	nhtsa/results.zip (137 MB)

Each folder also holds an index.csv with the case ID and crash type of every case, and the crash coordinates for MTCF. Seven MTCF cases additionally include a generated sensor video, see Sensor Videos below.

🚀 Download

hf download HaoweiLi/CrashFactory-Dataset --repo-type dataset --local-dir CrashFactory-Dataset

The data guide of the code shows where to unpack the files and how to reconstruct further crashes.

📁 Result Folder

results.zip holds one folder per case, named crash_<report number> for MTCF and by the CISS case number for NHTSA:

File	Content
report.yaml	Crash record in CrashFactory's unified form
map.net.xml	Road network around the crash (SUMO); also as map.osm and map.xodr
prevtraj.txt	Sparse pre-crash states of the crash participants
final.fcd.xml	Final trajectories; earlier stages are kept as *.fcd.xml
final_bev.mp4	Bird's-eye video of the final scenario
processing.log	Run log of the pipeline stages
cosmos/	Inputs and result of sensor simulation, in seven cases

🎥 Sensor Videos

Seven MTCF cases also contain a cosmos/ folder with the inputs and the result of sensor simulation:

File	Content
streetview_image_<camera>.jpg	Google Street View images around the ego vehicle, one per camera
prompt_<camera>.txt	Scene description written from each image
hdmap_*.mp4	HDMap control videos of the six cameras
front_view.mp4, front_view.txt	Front-view video generated by NVIDIA Cosmos-Drive (1280×704, 5 s at 24 fps) and the prompt it was generated from
multi_view.mp4, multi_view.txt	Six synchronized views and their prompts, in crash_2023174894

The six cameras are front, front_left, front_right, rear, rear_left and rear_right. front_view.txt is a refined version of prompt_front.txt that also describes the crash; the generated video varies noticeably with the prompt.

Top: pedestrian downtown at night, pedestrian on a residential road, bicyclist crossing. Bottom: head-on at night, angle, rear-end.

Case	Crash
crash_2023165056	Head-on at night
crash_2023167829	Rear-end
crash_2023174894	Bicyclist struck from behind
crash_2023197392	Pedestrian downtown at night
crash_2023225879	Angle at an intersection
crash_2023249707	Bicyclist crossing a side street
crash_2023298086	Pedestrian on a residential road

multi_view.mp4 holds the six views one after another, 57 frames each, in the order of the cameras above.

The six views of crash_2023174894, arranged around the vehicle.

The videos were rendered from an earlier run of these cases; in three of them the trajectories differ from final.fcd.xml by up to 2 m.

📄 License

• CrashFactory outputs: CC BY 4.0
• MTCF and NHTSA CISS reports: the terms of Michigan Traffic Crash Facts and NHTSA CISS
• Road networks: © OpenStreetMap contributors, ODbL
• Street View images: © Google

Citation

@article{li2026crashfactory,
  title  = {CrashFactory: From Crash Databases to Scalable Safety-Critical Data Synthesis for End-to-End Autonomous Driving},
  author = {Li, Haowei and Wang, Jiawei and Sun, Haowei and Yan, Xintao and Liu, Henry X.},
  year   = {2026},
  note   = {Preprint},
  url    = {https://ssrn.com/abstract=7288377}
}

Contact: haoweili@umich.edu