blazeofchi/so101-car-parking-validation

SO-101 car parking: validation

Test split with 20 scripted robot demonstrations and 15,325 frames at 20 Hz. Half the episodes push a toy car. The other half pick and place it. Each instruction names a car color and a bay color. Each scene has 2 to 4 cars.

These scenes helped during development and checkpoint selection. They are not an untouched final test set. Use the training statistics for normalization.

The preview shows the scripted teacher. It is not a result of the learned model.

Format

The data uses the native LeRobot v3.0 format:

• data/: parquet frames.
• videos/: overhead and wrist RGB videos, synchronized with the frames.
• meta/: episode metadata and statistics.
• raw/: simulator trajectories and task descriptions.
• episode_map.json: maps the raw episodes to LeRobot episode indices.

State and action each have six values. Actions are absolute joint targets in radians. Each camera frame shows the scene before its action.

Quality check

Every demonstration passed a physics check. The car is inside its bay, upright and still. The other cars moved 5 mm or less. Pick and place demonstrations also lift the car at least 25 mm. The split contains only accepted demonstrations, so it does not cover every possible layout. The start rotation of the cars does not change.

hf download blazeofchi/so101-car-parking-validation --repo-type dataset --local-dir data/parking-300/lerobot/validation

Source code · Model · Other split

Apache-2.0. The data is synthetic MuJoCo simulation. It contains no human recordings.