Each caption states what the robot did with exact quantities — "moves down 38 cm while rotating 35°, then closes the gripper to pick up the blue ring" — with the numbers taken from the robot's own action log rather than estimated by a model.

Why the numbers are trustworthy

The measurements never pass through a model. Distances, rotations, directions, frame ranges and gripper events are computed deterministically from observation.state.cartesian_position and observation.state.gripper_position, and the caption is assembled from them by template.

A VLM (Qwen3.8-Flash-Next-FP8) is used for exactly one thing the action log cannot supply: naming the object each motion acts on, read off the wrist and exterior cameras. It returns one label per numbered point and never writes prose, so it cannot drop a point, invent an action, or alter a measurement. A malformed answer leaves that point unnamed and falls back to the template.

That is the point of the design: DROID's instructions could only name 41.9% of grasps and 32.1% of releases, leaving 108,367 target slots empty. The teacher filled 173,793 of 173,895 — 99.9% — leaving 102 blank.

Files

path	records	contents
`captions/shard-{00..07}.jsonl`	74,591	the teacher captions
`annotations/all.jsonl.gz`	74,591	ground truth: the scorable motion points behind each caption

Caption record

{"episode_index": 31,
 "instruction": "Put the glue stick in the cup",
 "draft_caption": "Put the glue stick in the cup. Moves down 46 cm while rotating 30°, then closes the gripper ...",
 "teacher_caption": "Put the glue stick in the cup. Moves down 46 cm while rotating 30° toward the glue stick, then closes the gripper ...",
 "subtasks": ["put the glue stick in the cup"],
 "point_labels": [0, 0, 0, 0],
 "targets": ["the glue stick", "the glue stick", "the cup", "the cup"],
 "parse_ok": true, "mode": "annotate"}

draft_caption is the template before the teacher named anything, so the pair is usable as an ablation. targets aligns one-to-one with the points array of the matching record in annotations/all.jsonl.gz; point_labels gives each point's sub-task index.

26 records have teacher_caption: null (25 unparseable, 1 truncated). They keep draft_caption and parse_ok: false. Filter on parse_ok unless you want the template as a fallback.

Ground-truth record

episode_index, num_frames, fps (always 15), instructions (DROID's three annotator wordings), caption, caption_qualitative, and points — an ordered list of atomic facts, each with kind (move/grasp/release/hold), frame range, and for moves direction, distance_cm, rotation_deg, carrying. Grasps carry aperture; grasp/release may carry a target with target_source recording whether the name was parsed from the instruction or observed by the teacher.

Measured quality

Scored by parsing each caption back into motion points and matching against the log, over all 74,591 episodes:

metric	value	mismatched-episode control
recall	0.9978	0.631
precision	0.9768	0.617
gripper event recall	0.9970	—
direction accuracy	0.9807	0.236
distance MAE	0.3038 cm	14.50 cm
rotation hit rate	0.9988	—
carrying accuracy	0.9976	—

The control column scores every caption against a different episode's truth. It is the floor any real number has to clear, and without it a direction accuracy of 0.6 would look like a result rather than label imbalance.

Two caveats worth stating plainly:

Recall is 0.9978, not 1.000, because 172 episodes (0.23%) exceed the 40-clause cap and have their middles elided into a [N further moves] stub. That is the only mechanism that lowers it.
Precision is understated. The scorer reads the put X on Y inside a sub-task lead as an extra release. Single-sentence captions score precision 1.0000; multi-sentence ones 0.9275. The model emits no actions at all, so the gap is entirely the round-trip parser, not invented content.

Object naming, the teacher's actual contribution, cannot be scored against the log. The available proxy: on the slots where the instruction had already named the object, the teacher's independent visual naming agrees 0.8797 of the time (57,699/65,589).

Provenance

Derived from allenai/MolmoAct2-DROID-Dataset (Apache-2.0). episode_index indexes into that dataset's episode list — it is not interchangeable with cadene/droid_1.0.1_v30 or other DROID repackagings, which number their episodes differently. Joining against the wrong variant pairs every caption with another episode's video and produces no error of any kind, so check it: episode 0 is 237 frames and its instruction is "Pick up the blue ring from the table and put it in the wooden tray".

Videos are not included; they are reconstructible from the source dataset.

Downloads last month: 26

Total file size:

111 MB