kamwoh/mini-carla-192x320-v3

mini-carla-192x320-v3

Action- and camera-pose-conditioned driving clips rendered offline from CARLA 0.9.16 — the v3 scale-up of `mini-carla-192x320`, built as the training corpus for `miniworld`, a minimal flow-matching world-model framework.

10.48 M frames / 145.6 hours at 192×320 (h×w), 20 Hz, across 21 training environments (7 towns × 3 camera regimes), plus a held-out-town validation split (Town07, all three regimes).

Every episode is 600 frames = 30 s, rendered at Low quality, fixed_delta_seconds = 0.05, fov 90° (pinhole intrinsics in meta.json).

Three camera regimes

Each town appears under three prefixes:

• —*`vehicle_town** — camera on a Traffic-Manager-driven car. The vehicle **blueprint is randomised per episode** (epnpz["epvehicle"]) and the camera mount is fitted to each vehicle's bounding box, so ride height and viewpoint vary. Actions are read back from vehicle.get_control()` at capture time.
• —*`pedestrian_town`** — walker-height camera following sidewalk-level paths; slow, close to geometry, frequent pedestrians.
• —*`freecam_town`** — a camera flying the lane graph with no body attached; produces motions a traffic-obeying car never does. Speeds up to ~34 m/s.

What v3 adds over v1

• —~55× more data (17,471 vs 320 episodes) with uniform spawns drawn from the lane graph (3,266 waypoints/town rather than the ~255 fixed spawn points), after v1-style spawning was measured to collapse 840 episodes onto 28 distinct paths.
• —Per-episode weather: 11 CARLA presets, near-uniformly distributed (1,493–1,646 episodes each), from ClearNoon to HardRainNoon and four sunset variants — recorded per episode in ep_weather.
• —Per-frame traffic census: scene_near (actors within radius) and scene_moving (how many are moving), captured from the world snapshot each tick. The accept policy uses it: an episode is rejected only if the ego AND the scene are both static, so stop-and-go traffic ("brake–go–brake") is kept rather than filtered as stationary footage.
• —Camera pose per frame (pose: x, y, z, pitch, yaw, roll — metres and degrees, world frame), enabling pose- and ray-map-conditioned training.
• —A held-out town as the validation split rather than held-out episodes of seen towns.

Files

Layout: train/<env>/ and val/ (the Hub caps 10,000 files per directory, so the 17,471 train episodes are split one directory per environment):

.npz fields (N = frames in the env; E = episodes):

Frame i of the env (concatenating its mp4s in episode order) pairs with row i of every per-frame array; a one-frame shift would pair frames with the wrong action, so the alignment is the dataset's core contract.

Conditioning conventions used downstream

miniworld derives from pose a per-frame 7-D relative camera trajectory (tx, ty, tz, qx, qy, qz, qw) (translation in metres, xyzw unit quaternion), expressed relative to a clip's first frame — and from that, Plücker ray maps and raxel maps (arXiv:2604.09429-style, VAE-encoded). The v3 corpus was captured to make those trainable: measured on this data, camera speed spans 0–34 m/s (p50 ≈ 2 m/s), per-frame rotation reaches ~3°/frame at p99, and only ~1 % of clips are near-pure rotation — numbers worth knowing before designing a camera encoding on top.

Known limitations

• —Low render quality; this corpus trades fidelity for scale on purpose.
• —Freecam trajectories follow the lane graph (collision-free by construction) rather than flying freely — a documented simplification of the free-camera regime in arXiv:2603.15583.
• —Episode weather is fixed within an episode; there are no same-trajectory different-weather pairs (relevant to cross-temporal training schemes).
• —Sun-angle presets only (Noon/Sunset); no night driving.

License

CC-BY-4.0. CARLA itself is MIT-licensed; assets rendered from CARLA maps.