Argoverse 2 · Pittsburgh

Real autonomous-vehicle logs from Argoverse 2 — ~16-second drives captured by a 64-beam LIDAR rig with a full perception stack. One scene per AV2 city ships (Pittsburgh, Miami, Austin, Detroit, Palo Alto, Washington DC); switch between them in the cockpit. Every LIDAR return, tracked car / pedestrian / cyclist box, the HD map, and the ego pose are georeferenced from each city’s coordinate frame (via the AV2 devkit CRS) onto a real basemap and served as spatiotemporal tiles.

The cockpit at /drive/argoverse-02678d04 composes the full stream set on one playback clock: the accumulated LIDAR cloud (colored by height band), the ego trail, tracked objects as oriented 3D boxes, the HD-map substrate (lane boundaries + centerlines + drivable areas + crosswalks), a ring-camera inset, and a telemetry gauge panel. Argoverse logs carry no CAN bus, so speed / acceleration / yaw-rate / heading are derived from the ego pose.

Built by argoverse_extract.py (driven by argoverse_batch.sh) from public sensor logs pulled with no auth from the Argoverse AWS Open Data bucket; each scene is decimated to ~190k LIDAR points and drops zero-point (occluded) GT boxes.

The same cockpit drives three other sensor sources — nuScenes Boston/Singapore, a comma.ai dashcam log and a synthetic reference drive — all reachable from the scene rail without leaving the page.

bash stt:scripts/data-generation/argoverse_batch.sh   # 1 scene per AV2 city (6 total)

The batch driver selectively pulls each log (lidar + ego + annotations + map + one ring camera) with aws s3 cp --no-sign-request, extracts, and deletes the raw log. Open the cockpit at /drive/argoverse-02678d04 (switch cities in-cockpit).

Full recipes for every dataset live in the data-generation guide.