earth37815/lead-alpamayo-curated-pai
LEAD Alpamayo-rig curated PAI CARLA 0.9.16 expert driving from LEAD, re-collected with the four NVIDIA Alpamayo 1.5 input cameras added, and converted to the NVIDIA PhysicalAI-AV (PAI) on-disk format. It is read unchanged by alpamayo-recipes Alpamayo 1.5 nav SFT (PAIDatasetWithNav). It was built with lead2alpa (lead2pai --curate). train val routes / clips 72 / 229 6 / 24 nav samples 520 115 left / right / straight 136 / 136 / 248 31 / 31 / 53 stationary samples… See the full description on the dataset page: https://huggingface.co/datasets/earth37815/lead-alpamayo-curated-pai.
LEAD Alpamayo-rig curated PAI
CARLA 0.9.16 expert driving from LEAD, re-collected with the four NVIDIA Alpamayo 1.5 input cameras added, and converted to the NVIDIA PhysicalAI-AV (PAI) on-disk format. It is read unchanged by alpamayo-recipes Alpamayo 1.5 nav SFT (PAIDatasetWithNav). It was built with lead2alpa (lead2pai --curate).
- 20 s clips (≥ 17.5 s), timeline from t = 0 at the clip start, egomotion at 20 Hz (rear axle).
- Cameras
camera_cross_left_120fov,camera_front_wide_120fov,camera_cross_right_120fov,camera_front_tele_30fov: rendered natively at the PAI Hyperion 8 mount poses, 10 Hz, 960×540, H.264, per-frame timestamps. CARLA renders pinhole images; the calibration stores f-theta polynomials fitted to them. annotations_{train,val}.json:{clip_id, t0_relative, nav_text}samples ("Turn left in 35m", "Continue straight"). Every turn label is confirmed by the path the expert drove; left/right turns are balanced; straight ≤ turns; stationary ≤ 10 %.annotations_{train,val}_all.jsonholds every 1 s sample before curation.- Routes whose expert run had collisions, red-light or stop-sign infractions, timeouts, or < 100 % completion are excluded.
- No chain-of-causation (CoC) labels; the Alpamayo 1.5 nav recipe trains on trajectories only.
Train: see instruction.md in the lead2alpa repo. Collection provenance (route lists, camera rig, expert config, logs) is in collection/.