GGOSinon/robolab-bc27-chunked

RoboLab BC — chunked end-effector actions (19 tasks)

Behaviour-cloning data for training a vision-language model as a manipulation policy. Each sample is (two camera images + proprioceptive state + natural language instruction) → a JSON action, with no reasoning text.

Collected with a privileged scripted policy (cuRobo collision-free planning on ground-truth meshes) in RoboLab / Isaac Sim. Every attempt was kept -- successes and failures.

Action representation — and why it is deltas

The target is the next 10 end-effector displacements in millimetres relative to the current pose, plus the end orientation and jaw state:

{"d_mm": [[0.3,0.3,0.6], ..., [1.5,-4.0,0.8]],
  "quat": [0.0001, 0.9999, -0.0012, 0.0003],
  "jaw": "open"}

This matters more than it looks. An earlier version of this dataset used absolute TCP waypoints, and a 27B VLM trained on it learned a degenerate solution: the end-effector sits ~530 mm from the origin while a 10-step chunk moves only ~51 mm, so echoing the input state back scores ~90% of the token loss. The resulting policy produced perfectly-formed JSON and a healthy loss curve, yet its position error (49.1 mm) was exactly the error of predicting no motion at all (49.1 mm). Deltas make the signal the value itself.

If you build something similar: always score your policy against a hold-still baseline. A low absolute-coordinate error can mean the model learned nothing about motion.

Loading

from datasets import load_dataset
ds = load_dataset("GGOSinon/robolab-bc27-chunked")      # all rows
train = ds["train"].filter(lambda r: r["split"] == "train")

Stored as Parquet with the images embedded, so there are 28 files rather than 108k loose PNGs (the Hub caps a commit at 25k large files). image_overshoulder and image_wrist decode as datasets Image features; split is a column with values train / eval / failures.

Schema

Each line of train.jsonl / eval.jsonl / failures.jsonl:

column	meaning
image_overshoulder, image_wrist	the two camera views, 392 px long side
instruction	the task's natural-language goal
user	instruction + state + the request string (the full prompt)
assistant	the JSON action above
task, episode, step	provenance
split	train / eval / failures

The split is episode-disjoint (no episode contributes to both), because states within an episode are near-duplicates and a random split would leak.

Splits

file	samples	what it is
train.jsonl	26667	successful demonstrations, for behaviour cloning
eval.jsonl	3132	held-out episodes, same distribution
failures.jsonl	23907	failed attempts — NOT for cloning

failures.jsonl is the unusual part. Scripted-demonstrator datasets normally discard failures, but here every attempt was kept: 50 attempts per task, saved whether or not they succeeded. The failures are matched to the successes in task, scene and controller, which makes them usable for offline RL, failure analysis, or training a success classifier.

Three tasks have zero successes and appear only in failures.jsonl (NonHammerToolsInRightBin, JugsOnShelf, OneBottleOnShelf): the scripted policy attempts them and fails every time. Under a successes-only recipe these tasks are simply absent.

• 709 episodes across 19 tasks in the cloning splits
• 15 Hz control; a chunk spans ~3-4 cm of travel

task	train samples
AnimalsInBinTask	264
BlockStackingSpecifiedOrderTask	2717
BlocksInBinTask	4103
CannedFoodInBinTask	769
ClutterPlasticTask	2792
ClutterPumpkinTask	1393
ElectronicsInBinTask	1107
FoodPacking1BoxesTask	820
FruitsGreenLimesOnPlateTask	1925
FruitsMovingOrangeOrLimeTask	970
FruitsOnionTask	688
FruitsOnionToPlateTask	834
FruitsOrangesOnPlateTask	804
RedItemsInBinTask	966
RubiksCubeThenBananaTask	1872
SmallPumpkinInBinTask	830
SmartphoneInBinTask	820
SpoonsInPotTask	1965
WoodSpatulaToBowlTask	1028

Known limitations

• Five further tasks are absent entirely: the scripted grasp sampler finds no candidates on their meshes (pomegranate01, raisin_box, pink_spaghetti_spoon, the mugs), so no attempt is ever made — not even a failed one.
• Success rates vary enormously by task (50/50 down to 5/50), so the cloning splits are unbalanced: BlocksInBin contributes 4103 train samples, AnimalsInBin 264.
• Scenes are near-fixed. RoboLab benchmark tasks declare no randomisation events, so episodes of a task differ mainly by injected action noise (sigma 0.006, tau 12) and, for most tasks, a small pose jitter.

Provenance

Generated with policies/datagen on RoboLab, 224 px recording, two cameras, abs_joint_pos(7)+gripper(1) at 15 Hz; the chunked TCP action here is derived from the recorded obs/tcp_pose.

Reasoning annotations (new)

21.2% of rows (11,363 of 53,706, all in train) now carry natural-language reasoning about the action, generated by an agent shown the two camera images, the proprioceptive state and the reference chunk, and asked to produce the thinking that leads to that chunk -- at three depths.

column	meaning
has_reasoning	whether this row was annotated (false => the fields below are null)
reasoning_short	the immediate intention, under 12 words
reasoning_medium	two sentences: why this direction and this size of motion, now
reasoning_long	three to five sentences, with grounded spatial justification
grip_state	jaw state the annotator was given: open / shut / held / closing / opening
reasoning_model	model and effort used

The reference chunk is shown to the annotator but withheld from the output: the prompt forbids quoting coordinates, mm displacements or durations, so the text describes the motion spatially ("ease down onto the cup's rim") rather than restating numbers the row already contains.

Measured quality

check	result
direction words agree with the actual motion	96.8% (27,194/28,106)
leakage of mm / coordinates / durations / "expert"	0 / 0 / 0 / 0
distinct reasoning_medium / reasoning_long	100.0% / 100.0%
distinct reasoning_short	77.2%
fully-shut jaw narrated as empty	1 / 342 (0.29%)

Direction agreement is the check that matters: words like "lower", "left" and "lift" are compared against the sign of the recorded displacement on that axis. Chance is ~50%, so 96.8% confirms the text is about this state rather than generic manipulation prose. reasoning_short repeats across rows (77.2% distinct) because consecutive chunks genuinely are near-identical; the medium and long tiers are effectively unique.

The last row is a specific failure that was found and fixed: a fully-shut jaw reads the same whether it is closed on nothing or pinching a thin rim or stem, and an earlier prompt that called it "closed on nothing" made the annotator narrate whole carries as an empty gripper. The jaw description now states the ambiguity and defers to the wrist view.

Coverage is uneven

Episodes were sampled uniformly, so reasoned-row counts inherit the dataset's imbalance: BlocksInBinTask has 1,659, AnimalsInBinTask 141. eval and failures carry no reasoning at all. Filter on has_reasoning before training on these fields.