Dataset Viewer
Auto-converted to Parquet Duplicate
The dataset viewer is not available for this split.
Parquet error: Scan size limit exceeded: attempted to read 511395999 bytes, limit is 300000000 bytes Make sure that 1. the Parquet files contain a page index to enable random access without loading entire row groups2. otherwise use smaller row-group sizes when serializing the Parquet files
Error code:   TooBigContentError

Need help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.

RoboLab det10 — deterministic scripted manipulation, successes and failures (87 tasks)

Behaviour-cloning data for training a vision-language model as a manipulation policy. Each row is (two camera images + proprioceptive state + natural language instruction) → a JSON action. 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.

What makes this set different: determinism

The scripted policy ran with --action-noise 0. Episodes of the same task therefore differ only by their initial state, not by injected action noise. Object placements were randomised per episode (xy offset and yaw), with a fallback ladder — 2 cm / 0.1 rad, then 1 cm / 0.05, then 0.5 cm / 0.02, then none — because the most constrained tasks reject aggressive randomisation outright ("no grasp candidates").

This is deliberate. An earlier collection replayed only 2–6 distinct initial states per task 4–41 times each, with action noise as the only variation, and 93% of its "held-out" evaluation scenes also appeared in training. Here each task was attempted 10 times from 10 distinct initial states, and the split is episode-disjoint.

These magnitudes are ours, not RoboLab's. The perturbation function is RoboLab's own reset_pose_uniform, but the ranges are not a RoboLab protocol: RoboLab ships tasks/randomize_initial_pose/*_uniform_10cm at ±10 cm and its runtime default is ±5 cm, both without yaw. ±2 cm was chosen here because the scripted planner reports "no grasp candidates" on some tasks at wider ranges. This is therefore a narrower, easier layout distribution than RoboLab's randomised benchmark variants, and results on it should not be quoted as if they were on *_uniform_10cm. 104 of 108 tasks kept the full ±2 cm; four needed 0.5–1 cm; none fell to zero.

Splits

split rows what it is
train 178029 successful episodes
eval 22163 successful episodes, held out (episode-disjoint)
failures 208585 failed attempts — not for cloning

failures is slightly the larger half, and that is not a mistake: a failed episode runs to the full step budget while a success terminates as soon as the task predicate is satisfied, so failures are about 1.7× longer each (median 399 rows per episode against 235). 446 failed episodes therefore produce marginally more rows than 627 successful ones.

Do not train a policy on failures by cloning it. These are drops, missed grasps and collisions; imitating them reproduces them. They are published because they are matched to the successes in task, scene and controller, which makes them usable for offline RL, failure prediction, reward modelling and value learning. Twenty of the 108 attempted tasks produced no successful episode and appear only here.

Action representation — and why it is deltas

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

{"d_mm": [[5.0,0.6,-0.8], ..., [60.7,11.9,-15.6]],
  "quat": [0.0299, 0.9731, 0.2252, -0.0384],
  "jaw": "open"}

An earlier version used absolute TCP waypoints, and a VLM trained on it learned a degenerate solution: the end-effector sits ~530 mm from the origin while a chunk moves only ~50 mm, so echoing the input state back captures most of the token loss. The policy produced well-formed JSON and a healthy loss curve while its position error exactly equalled the error of predicting no motion. Deltas make the signal the value itself.

If you build something similar: always score against a hold-still baseline. On this data that baseline is 53.2 mm of position error, and several representations that looked healthy by loss or token accuracy scored exactly there.

Schema

column meaning
image_overshoulder, image_wrist PNG bytes, 392 px long side
user instruction + state + the request string
assistant the JSON target, with the fields below
instruction the task's natural-language goal
task, episode, step provenance

The assistant JSON carries three target families on the same frames, so one download serves several policy formulations:

field what it is
d_mm, quat, jaw the next 10 TCP displacements (mm, relative to the current pose), end orientation, jaw state
joint_pos, gripper_close absolute joint command — the env's native action space (abs_joint_pos(7)+gripper_close(1)), exactly what cuRobo emitted
goal_mm, goal_quat, phase, goal_steps displacement to the phase-segment goal, its orientation, the motion phase, and steps remaining

Why joint_pos is worth using. d_mm is derived from obs/tcp_pose, the achieved pose. The arm is position controlled and never reaches its setpoint within a step: median joint tracking error 0.033 rad, and the command leads the response by about 4 steps (17 mm). Training on achieved poses and then feeding the prediction back as a command is not self-consistent, and realising a TCP delta needs an IK step whose minimum-norm solution discards the elbow configuration cuRobo chose for collision avoidance. joint_pos avoids both: it is what the environment consumes, with no IK in the loop.

Why goal_mm is worth using. It is a subgoal — median 22 steps (1.5 s) ahead, max 61 — recovered from the TCP pose at each phase boundary. The executor settles before a phase change, so the achieved pose there is the pose cuRobo was aiming at: 0.0019 rad tracking error at boundaries against 0.0329 overall. It is piecewise-constant across a phase, and immune to the dead zone described under limitations.

Sampling density

Every simulation step is a row (stride 1). An earlier revision sampled every 10th step; this one does not skip frames, so consecutive rows overlap heavily and differ by one control step. Filter on step % k == 0 for a sparser view.

Images are embedded in the parquet shards, so the dataset is self-contained.

Known limitations

  • Per-task depth is uneven. 32 tasks have the full 10 successful episodes; 56 have fewer, and 6 have exactly one. Task breadth is good, per-task generalisation evidence is thin.
  • A grasp needs millimetre precision this data does not obviously teach. A 0.8B and a 4B VLM trained on the successes both reach ~35 mm held-out position error — well above the hold-still baseline of 53.2 mm, so they have learned direction and grasp timing (96%+ jaw accuracy), but far short of what closing on an object requires. Closed-loop success on a four-task subset was 0/192 episodes across every configuration tried.
  • The first ~10 steps of every episode are stationary. cuRobo's trajectories start at zero velocity, so motion is exactly 0.000 mm through roughly step 9, and the step-0 d_mm target is a median 0.36 mm (84.8% under 1 mm). A policy trained on d_mm emits a sub-millimetre first command, and if your controller rejects small commands as no-ops that is an absorbing state — measured here as 225 consecutive rejections and a wholly wasted episode. goal_mm does not have this problem: at the same step it reads ~256 mm.

Per-task sample counts

task train samples
AnimalsInBinTask 1341
AppleAndYogurtInBowlTask 2844
BBQSauceInBinTask 466
BagelsOnPlateTask 1027
BananaInBowlTask 1490
BananaOnPlateTask 1496
BananaThenRubiksCubeTask 3252
BananasInBinOneMoreTask 2713
BananasInBinThreeTotalTask 1802
BananasInCrateTask 1785
BananasOutOfBinTask 3345
BigPumpkinInBinTask 1691
BlockStackingOrderAgnosticTask 3028
BlockStackingSpecifiedOrderTask 4354
BlocksInBinTask 6691
BowlInBinTask 2392
BowlStackingLeftOnRightTask 1259
BowlStackingRightOnLeftTask 996
ButterAboveRaisinTask 2185
CannedFoodInBinTask 1373
ClampInRightBinTask 283
ClutterPlasticTask 5031
ClutterPumpkinTask 3688
CookingClearPlateTask 4152
CubesAndBlocksInBinTask 13281
DishesInBinTask 3271
ElectronicsInBinTask 1856
FoodPacking1BoxesTask 1736
FoodPacking1CansTask 524
FoodPacking2BoxesTask 356
FruitsGreenLimesOnPlateTask 3380
FruitsMovingOrangeOrLimeTask 1653
FruitsMovingTask 2369
FruitsOnPlate3Task 3129
FruitsOnionTask 674
FruitsOnionToPlateTask 1106
FruitsOrangesOnPlateTask 1327
GrabABagelTask 1154
GrabAFruitTask 1271
GreenSpoonsInPotTask 2861
MarkerInMugTask 2359
MouseOnKeyboardTask 2616
MoveBananaToBagelPlateTask 3014
MustardAboveRaisinTask 1180
MustardInLeftBinTask 1281
MustardInRightBinTask 1129
OneBottleInSquarePailTask 881
PhoneOrRemoteInBinTask 2139
PickGlassesTask 1071
PickOrangeObjectTask 722
PickUpBluePitcherTask 176
PickUpGreenObjectTask 91
PinkSpoonInPotTask 1975
PlasticBottlesInSquarePailTask 1432
PutBowlOnShelfTopTask 1103
PutTwoMugsOnShelfTask 2106
RecycleCartonTask 634
RedDishesInBinTask 1540
RedItemsInBinTask 391
RubiksCubeAndBananaTask 2553
RubiksCubeBehindBowlTask 1297
RubiksCubeInFrontOfBowlTask 1167
RubiksCubeLeftOfBowlTask 1229
RubiksCubeOrBananaTask 1155
RubiksCubeRightOfBowlTask 1443
RubiksCubeTask 1356
RubiksCubeThenBananaTask 2750
RubiksCubesInBinTask 4832
SauceBottlesCrateTask 1950
SmallPumpkinInBinTask 1932
SmallerObjectButterInBinTask 1051
SmartphoneInBinTask 1119
SpoonsInPotTask 4062
Stack3RubiksCubeTask 3212
StackYellowOnRedTask 1378
TakeMeasuringSpoonOutTask 1643
TakeSpatulaOffShelfTask 190
ThrowAwayAppleTask 1451
ThrowAwaySnacksTask 3577
ToolsPickingHammerTask 200
ToyInBinTask 766
UnstackRubiksCubeTask 1702
UtensilsInMugTask 2392
WhiteMugsInBinTask 2382
WoodSpatulaToBowlTask 1624
YellowAndWhiteObjectsInBinTask 3819
YogurtInBowlTask 1325
Downloads last month
202