ROADDATA.AI · FEED SPECIFICATION
Cosmos Data Card schema Specification · September 2026

Road imagery with 10 Hz ego-state,
time-aligned to the frame

Forward road video paired with a GNSS+IMU pose stream — position, velocity, bearing, three-axis acceleration and rate, and a full 3×3 rotation matrix per timestep. Every record carries both a wall clock and a monotonic clock, so the telemetry stays aligned to the video across NTP corrections, and the whole delivery is built to the NVIDIA Cosmos Data Card schema.

Video — per segment

Road
Resolution1280 × 720
Frame rate30 fps
CodecHEVC (H.265)
Video bitrate~2.0 Mbps
Audionone
Segment length120 s
File size~30 MB
Per device-hour~918 MB

MP4 container — 3,600 frames per road segment.

Anonymised before delivery

Faces in the road footage are detected and blurred before delivery. The blur is burned into the delivered frames — not a viewer-side overlay, and not reversible by whoever receives the data. Device and driver identifiers are replaced with synthetic ones, so a delivered segment carries no route back to a vehicle or a person.

Ego-state — per segment

Delivered at
10 Hz
100 Hz available

Position, velocity, attitude — one sidecar per video segment, same time window. The IMU is sampled at 100 Hz and the solution decimated to 10 Hz for delivery; the full 100 Hz stream is available on request.

Uploads on its own

Three event types leave the vehicle without anyone asking, carrying the same imagery and ego-state as everything else. They upload over the unit's own LTE connection while the vehicle is still on the road — no dock, no card to collect, no return to base. Out of coverage, the event is held on the device and re-driven when the connection comes back, so nothing is lost to a dead zone.

EventWhat triggers itWhat arrives
Accident Impact detected in the inertial stream, above a threshold set to clear ordinary braking Road video either side of the moment, the inertial trace at full 100 Hz through the impact, and a structured event report — the long tail nobody can stage on a test track
VRU encounters A hard brake, swerve or sharp acceleration, or a rare actor seen in frame A five-second clip centred on the moment, tagged with what else was in the scene — real conflict with pedestrians and cyclists, driver reaction attached
Toll crossings Entering or leaving a known gantry or landmark, by position — every pass, not a sample A clip and ego-state track per crossing. The same surveyed, never-moving structure traversed hundreds of times across vehicles, lanes, light and weather — a standing reference for localisation and map validation, where the ground truth is the geometry itself

The gnss_imu record

Verbatim, unedited:

{ "type":"gnss_imu", "ts":1788481158892, "monoNs":1847746744501,
  "lat":37.3319364, "lon":-122.0553775, "alt":56.7,
  "spd":30.782, "brg":140.666, "acc":3.79,
  "gps":true, "dr_age_s":0.4,        // GPS-anchored, 0.4 s since last fix
  "ax":11.7244, "ay":-0.972, "az":-1.0008,  // m/s², device frame
  "gx":-0.02142, "gy":0.10156, "gz":-0.0318, // rad/s
  "R":[-0.036876,0.281996,0.958733, 0.030165,-0.958635,0.283037,
        0.998864,0.039140,0.026805] }  // device → navigation, row-major