so100_black_cube_merged

Deduplicated merge of three SO-100 pick-and-place datasets, prepared for SmolVLA training.

• 130 episodes / 61,442 frames, 30 fps, robot type so_follower
• Single camera observation.images.front, 640×480, AV1
• 2 tasks

Sources

Source	Episodes kept	Notes
mahi-zade/merged_data_smolVLA	100 (all)	episodes 0–99
Mehul2112/so100_black_cube_v2	30 (all)	episodes 100–129
Mehul2112/so100_black_cube	0	fully duplicate — see below

Deduplication

Mehul2112/so100_black_cube was dropped in full. All 50 of its episodes are the same recordings as episodes 50–99 of mahi-zade/merged_data_smolVLA, verified by matching episode length together with per-episode action and observation.state means — all 50 matched, in order. Including it would have duplicated 27,371 frames (31% of the naive 180-episode / 88,813-frame concatenation).

Task distribution

Task	Episodes	Provenance
Pick up the blue cube and place it in the target location	50	mahi eps 0–49 (overhead view)
Put the black cube in box	80	mahi eps 50–99 (side view) + v2 eps 0–29 (overhead, wide)

v2's original task string "Pick the cube and drop in the box" was normalized to "Put the black cube in box" so all black-cube episodes share one instruction.

Transformations applied to so100_black_cube_v2

To make it schema-compatible with the 640×480 single-camera sources:

1. Dropped observation.images.ee — the wrist camera. The other 100 episodes have no second view, so the merged dataset is front-camera only.
2. 1280×720 → 640×480: center-cropped to 960×720 (4:3) then scaled with lanczos. Cropping rather than stretching keeps object proportions undistorted and consistent with the 4:3 sources. Frame count (12,059) and duration (401.967 s) preserved exactly, so per-episode video timestamps remain valid.

Video statistics were recomputed after the crop; the numeric (action / observation.state) stats are unchanged from the sources.

Caveats

• Mixed camera viewpoints. The three groups were recorded with different camera poses: mahi eps 0–49 overhead, mahi eps 50–99 side-on close-up, v2 overhead-wide. Since actions are absolute joint positions, the image→action mapping is viewpoint-dependent and the policy must infer viewpoint from pixels. Only the 30 v2 episodes match the current run-record-v2.sh rig, so the three groups will not contribute equally at deployment.
• The wrist camera is not present. A policy trained here is single-camera. The existing two-camera smolvla_black_cube_v2 rollout config does not apply.

Upstream defects repaired

Two pre-existing inconsistencies in mahi-zade/merged_data_smolVLA were fixed here:

• Episode metadata carried observation.images.gripper stats/video columns for episodes 0–49, but no gripper videos exist in the repo and info.json never declared the feature. Columns dropped.
• meta/episodes/file_index claimed 19 episodes lived in meta/episodes/chunk-000/file-001.parquet, which does not exist — all 100 rows are in file-000.parquet. LeRobotDataset tolerates this (it globs the directory) but aggregate_datasets reads those columns and fails. Reindexed.

Additionally, lerobot's remove_feature leaves the removed key's per-episode videos/… and stats/… columns behind in meta/episodes (it only filters the global stats.json); the orphaned observation.images.ee columns were stripped.

Usage

lerobot-train \
  --dataset.repo_id=Mehul2112/so100_black_cube_merged \
  --policy.type=smolvla \
  --policy.device=cuda