earth37815/lead-alpamayo-pai_medium

LEAD Alpamayo PAI: medium (curated, multi-rig)

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 --incremental --curate --flush).

This is the second, larger collection: a single CARLA pass recorded LEAD's own sensors, the Alpamayo cameras and a Bench2Drive rig (for HiP-AD; not included here) at once.

Snapshot (2026-09-28): 204 of 400 routes. Collection is still running. The target is 400 routes with the scenario distribution of the full LEAD route catalogue (9,692 routes, 43 scenario types; collection/route_list_400.txt). The routes collected so far were drawn round-robin over scenarios, so plain driving (noScenarios, 14.5 % of the catalogue) is still under-represented. This repo will be updated when collection finishes.

	train	val
routes / clips (curated)	113 / 230	5 / 11
nav samples (curated)	512	64
left / right / straight	128 / 128 / 256	17 / 17 / 30
stationary samples	8.6 %	0 %
all 1 s samples before curation (*_all.json)	3,847	158

• 204 routes converted, 0 conversion errors; 400 clips over 8 chunks, 39 scenario types in train. Verified against the raw LEAD logs: 0.00 cm trajectory error, clip timeline contract met on every clip (scripts/verify_pai_output.py).
• 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 %. Right turns are the scarce class and set the curated size; annotations_{train,val}_all.json holds every 1 s sample before curation.
• Routes whose expert run had collisions, red-light or stop-sign infractions, timeouts, or < 100 % completion are excluded.
• Val is small in this snapshot (5 routes); it grows with the final version.
• No chain-of-causation (CoC) labels; the Alpamayo 1.5 nav recipe trains on trajectories only.

Train: see instruction.md in the lead2alpa repo (set PAI_DIR to this dataset). Collection provenance (route lists, camera rigs, expert config, collector and conversion logs) is in collection/.

Variants

Repo	Routes	Train / val samples	Selection
lead-alpamayo-pai_test	78 (first collection)	520 / 115	curated
lead-alpamayo-pai_medium	118 of 204 (multi-rig)	512 / 64	curated
lead-alpamayo-pai_large	238 of 399 (multi-rig, final)	1,384 / 127	curated