license: mit
pretty_name: block-stack
tags:
- world-models
- robotics
configs:
- config_name: train
data_files: train/data/**/*.parquet
- config_name: val
data_files: val/data/**/*.parquet
- config_name: eval_purple_play
data_files: eval_purple_play/data/**/*.parquet
- config_name: eval_purple_stack
data_files: eval_purple_stack/data/**/*.parquet
block-stack
Recorded trajectories (no simulator) packaged as LeRobot splits for world-model training.
Generated with quickdraw (data.processors).
Observation / action
- observation_vector (17): the recorded state
- observation.images.scene_left (144×192×3, video): recorded camera stream, stored as MP4, aligned 1:1 with the vector frames
- observation.images.scene_right (144×192×3, video): recorded camera stream, stored as MP4, aligned 1:1 with the vector frames
- observation.images.gripper_right_bottom (144×192×3, video): recorded camera stream, stored as MP4, aligned 1:1 with the vector frames
- observation.images.gripper_right_top (144×192×3, video): recorded camera stream, stored as MP4, aligned 1:1 with the vector frames
- action (5): the recorded actions, at 30 Hz
Splits
| split | episodes | steps/ep (mean) | frames |
|---|---|---|---|
train |
43 | 4019 | 172835 |
val |
2 | 17400 | 34799 |
eval_purple_play |
5 | 1810 | 9048 |
eval_purple_stack |
6 | 887 | 5323 |
Normalization statistics are computed on train only and applied to every split.
How this dataset was produced
See PROCESSING.md for the full account: source streams and rates, exactly which fields were kept and dropped and why, the resampling and synchronisation procedure, the train/val split rule, and the verification that images align with state rows.
Update 2026-09-24 — cube object-state features + observation.state rename
observation_vector was renamed to observation.state (lerobot-canonical; matches our other datasets).
Values are byte-identical to the old observation_vector [17]; only the field name changed.
New field: observation.objects.cube — shape [32], float32. Per-frame cube object-state, extracted by
per-episode color segmentation (hue-peak palette discovery, so only colors actually present in a split are
tracked — R/G/B for the standard task, R/B/P for the eval_purple_* splits).
Layout: 4 cube slots (canonical order R, G, B, P) x 2 cameras (cam_scene=scene_right, cam_wrist=gripper_right_top) x 4 features, flattened as [(cube j, cam c) for j in RGBP for c in (scene, wrist)] -> [f0..f3]. Per (cube,camera):
| idx | feature | meaning |
|---|---|---|
| 0 | cx / W | centroid x / image width, in [0,1] |
| 1 | cy / H | centroid y / image height, in [0,1] |
| 2 | sqrt(area / TYP) | apparent linear extent (~ 1/depth proximity cue); TYP=120 px |
| 3 | soft_vis | clip(area/TYP, 0, 1); 0 = not currently observed |
Occlusion: if a cube is not detected but was seen earlier, its position is held at last-seen while
soft_vis (and size) go to 0 -> [cx, cy, 0, 0] marks a held/occluded position. A cube color absent from a
split (P in train/val; G in the purple splits) is all-zeros every frame. So soft_vis>0 = current;
soft_vis=0 with nonzero position = carried through occlusion; all-zero = absent.
Normalization stats for observation.objects.cube (normalization_stats.json + per-split meta/stats.json) are on the train split. Cameras/action/indices and the four observation.images.* video streams are unchanged. Extraction is deterministic classical CV (no learned detector, no labels).