By requesting access you agree to the RoadData.AI Evaluation License (LICENSE.md): use for research and internal evaluation only, no redistribution, and no use to identify or track people or vehicles, or for surveillance or law-enforcement purposes.

RoadData.AI US Urban Driving Events (Sample)

54 human-verified, six-second dashcam event clips from commercial vehicles driving in US cities, each with synchronized IMU and vehicle-motion telemetry. This is a free evaluation sample drawn from RoadData.AI's continuously growing fleet capture. Version 0.1.

What's in the sample

Event label	Clips
School bus ahead	39
Low sun / glare	8
Transit / city bus	3
Construction / work zone	2
General urban driving	1
Moped / motor scooter	1

Region: 39 of 54 clips are from the New York City / North New Jersey metro; 11 from 10 other US metros; 4 from elsewhere in the US.
Time of day: daylight 38, dusk 9, night 7.
Weather: clear 48, overcast 5, rain 1.
School bus stop arm extended: 2 clips confirmed; for the rest the reviewer did not record it.
Video: road-facing camera, 1280×720, ~30 fps, about 6 s per clip, H.264.
Video–sensor alignment: median offset 0.05 s, 95th percentile 0.33 s (n=52).

Files

data/train/<clip_uid>.mp4      video
data/train/metadata.csv        one row per clip (fields below)
frames/<clip_uid>.jpg          trigger frame
telemetry/<clip_uid>.json      accel, gyro, speed/heading; time in seconds from clip start

Field	Meaning
`label`, `label_display`	Published event label, verified by a person
`label_source`	`human-verified`, or `human-verified (relabeled)` where the reviewer corrected the original label
`device_trigger`	On-device trigger that captured the clip
`cloud_event_type`	Label from RoadData.AI's cloud model, before human review (may disagree with `label`)
`school_bus_stop_arm_out`	`y` / `n` / blank (not recorded)
`region`, `month`	Coarse US metro and capture month
`weather`, `time_of_day`, `lighting`, `road_type`	Scene context from the cloud model (not human-verified)
`speed_mph_mean`, `video_sensor_offset_s`	Mean vehicle speed; measured video-to-sensor clock offset
`machine_caption`	Per-second scene description from the cloud model (not human-verified)

How the sample was built

Clips are captured on-device when a trigger fires (e.g. school bus ahead, low sun) and uploaded.
Cloud QA checks video integrity, camera alignment and visibility; failed clips are excluded. Faces and licence plates are blurred in the cloud with EgoBlur Gen2.
A candidate set was sampled across devices and trips (max one clip per trip per label).
Each published clip was reviewed by a person who confirmed or corrected the label and checked that faces and plates are blurred. Clips with an unsure label or an unconfirmed blur check are excluded.

Privacy

Faces and licence plates are blurred. No GPS coordinates, absolute timestamps, device, trip or vehicle identifiers are published; location is reduced to a metro name. If you find an unblurred face or plate, email abdul@roaddata.ai and delete the affected clip.

Known limitations

Small and unbalanced: 39 of 54 clips are one label. Later versions add categories.
Labels are verified by a single reviewer.
Scene context and captions come from a model and are not human-verified.
Single road-facing camera; no LiDAR, radar or vehicle CAN data.

Quickstart

from huggingface_hub import snapshot_download
import pandas as pd, json

path = snapshot_download("roaddata-ai/roaddata-us-urban-events-sample", repo_type="dataset")   # requires accepted access + HF token
meta = pd.read_csv(f"{path}/data/train/metadata.csv")
row = meta.iloc[0]
video = f"{path}/data/train/{row.file_name}"
telemetry = json.load(open(f"{path}/{row.telemetry}"))
print(row.label, len(telemetry["gyro"]), "gyro samples")

Going further

This sample is intentionally small. The full RoadData.AI catalog adds:

Scenario library: curated long-tail events at volume (VRUs, cut-ins, work zones, emergency vehicles, weather).
Continuous urban driving: hours of uninterrupted video with telemetry for world-model and simulation training.
C³O causal annotation: Context, Cognition, Control and Outcome per event, exported as OpenLABEL / OpenSCENARIO.
Targeted collection: events by category, region and conditions on request.

Licensing is non-exclusive. Contact abdul@roaddata.ai · https://roaddata.ai

Citation

@misc{roaddata_us_urban_events_sample,
  title  = {RoadData.AI US Urban Driving Events (Sample)},
  author = {{RoadData.AI}},
  year   = {2026},
  howpublished = {Hugging Face Hub},
  url    = {https://huggingface.co/datasets/roaddata-ai/roaddata-us-urban-events-sample}
}

Downloads last month: -

Total file size:

260 MB