Datasets:

---

nvidia

/

PhysicalAI-Autonomous-Vehicles

Like 1.02k

Follow

NVIDIA 68.2k

Summary

Follow-up to the 26.09 calibration fix (#43). For 25 clips (15 recording sessions, one vehicle) no reliable offline LiDAR-camera solve could be produced, so their offline camera calibration is removed rather than left misaligned. The clips keep their images, LiDAR, online calibration and all other features. metadata/feature_presence.parquet is recomputed from the published files so it is exact.

What changes

• calibration/camera_intrinsics.offline/*.parquet (24 chunks): the 7 camera rows of each affected clip are removed (175 rows total).
• calibration/sensor_extrinsics.offline/*.parquet (24 chunks): the 8 rows (7 cameras + lidar_top_360fov) of each affected clip are removed (200 rows total).
• metadata/feature_presence.parquet: 25 clips change (25 affected clips plus 0 clips whose published flags did not match the files). Per column:
  • camera_intrinsics.offline: 25 True->False, 0 False->True
  • sensor_extrinsics.offline: 25 True->False, 0 False->True
    The audit found 0 mismatches between the published flags and the files for the audited features, so apart from the affected clips the table is unchanged.
• README.md: the 26.09 Version History row is updated to describe both changes.

What does not change

• Every other row of the touched chunk files is byte-for-byte equal; schema, dtypes, index order, compression and Arrow metadata are identical to the base files.
• lidar_intrinsics.offline, online calibration (camera_intrinsics, sensor_extrinsics, vehicle_dimensions), clip_index.parquet, features.csv, video, LiDAR, radar and labels are untouched.
• Unaudited feature_presence columns (radar) are unchanged.

Validation

• Presence was recomputed at base commit 2778e29cd2f3 for camera_intrinsics.offline, sensor_extrinsics.offline, lidar_intrinsics.offline, camera_intrinsics, sensor_extrinsics, vehicle_dimensions, egomotion, egomotion.offline, obstacle.offline, camera_cross_left_120fov, camera_cross_right_120fov, camera_front_tele_30fov, camera_front_wide_120fov, camera_rear_left_70fov, camera_rear_right_70fov, camera_rear_tele_30fov, lidar_top_360fov by listing the clip ids in every chunk parquet and the member files in every chunk zip (LiDAR zips via remote central-directory reads).
• Affected clips by split: train 16 / val 8 / test 1.

After merge

Move tag 26.09 to the merge commit (the Version History row is updated in place, no new version).