dynamic_robot_bench_dr_scripted_14k

14,400 scripted-expert demonstrations across the 72 evaluated task families of dynamic-robot-bench — a conveyor-belt dynamic-manipulation benchmark (Franka Panda + wrist camera, ManiSkill 3 / SAPIEN GPU sim). One LeRobot v2.1 dataset: 200 episodes per family, success-filtered, every domain-randomization knob on, and belt speed uniform over 0.10–0.40 m/s.

1,118,617 frames · 209 distinct language instructions · 20 fps

The belt speed is uniform by construction, not by luck

The thing that makes this dataset different from a sweep is that the speed distribution is enforced. Collection stratifies the saved episodes over 10 equal-width bins of 0.030 m/s and holds an equal quota in each, discarding surplus successes rather than letting the success filter reshape the band — fast belts fail more often, so an unstratified success-filtered collection skews slow.

per-bin episode counts	1440, 1440, 1440, 1440, 1440, 1440, 1440, 1440, 1440, 1440
observed range	0.100 – 0.400 m/s
belt direction	+y: 7,244 / -y: 7,156

The equal counts are enforced, not observed. The collector retargets the draw at whichever bin is most deficient and discards surplus successes, so the counts are exact by construction and carry no information about how many attempts each bin cost. A goodness-of-fit statistic against an IID uniform null does not apply to them either — they have zero variance where an IID sample's would be multinomial — so none is quoted here. The table says the corpus is flat across the band; it does not say the collection was unbiased in any deeper sense.

Per-episode speeds are in meta/bench_episode_provenance.jsonl (14,400 rows), so this table can be recomputed from the files rather than believed.

Collection configuration

Identical for every family, from the recorded stamps:

knob	value
robot / control	panda_wristcam · pd_ee_delta_pose
scene	replicacad (index 0), shader default
randomization	--randomize — every boolean knob on the family's config, plus the scripted-expert velocity SDE
speed stratification	--speed-bins 10
scene-DR resampling	--reconfigure-every 8 → 25 independent scene draws per family
success filter	--filter-success (only successful episodes are saved)
parallel envs	4
episode cap	300 steps
seeds	one distinct --seed per family; the expansion is in each family's collection.main_seed in meta/ranges.json
code	874874b6

Scene-background randomization (workspace tint, lighting intensity/colour/direction, independent per-camera eye jitter, clutter props) is drawn once per reconfigure and shared by the batch, which is why the resampling count above is the number of distinct backgrounds a family carries — not 200.

Layout

Standard LeRobot v2.1, plus two bench files:

• meta/ranges.json — the family manifest: every family's episode range, frame count, and its own collection stamp. The stamp is per-family because bench_collection describes one collection run and this dataset is 72 of them; it is carried here rather than dropped.
• meta/bench_episode_provenance.jsonl — one row per episode: env_seed, belt_speed, belt_direction, object_id, the randomization snapshot, and the initial poses it was drawn into. gym_id names the owning family.

Features: exterior_image_1_left (224×224, three world-fixed eyes tiled), wrist_image_left (224×224), joint_position (7), gripper_position (1), actions (8 = 7 target joint angles + gripper, DROID convention).

Families

family	episodes	frames	index range	mean belt speed (m/s)	distinct objects
ConveyorAxleSeat-v0	200	14235	0–199	0.251	—
ConveyorBeltCorner-v0	200	25201	200–399	0.250	16
ConveyorBeltLoad-v0	200	22203	400–599	0.250	6
ConveyorBeltSwap-v0	200	23612	600–799	0.250	8
ConveyorBoreGrip-v0	200	16111	800–999	0.250	—
ConveyorCapPress-v0	200	14211	1000–1199	0.250	—
ConveyorCapSwipe-v0	200	16810	1200–1399	0.249	—
ConveyorCartCouple-v0	200	15264	1400–1599	0.250	—
ConveyorClearArch-v0	200	13087	1600–1799	0.250	10
ConveyorClipCard-v0	200	10163	1800–1999	0.250	—
ConveyorColorPress-v0	200	12122	2000–2199	0.250	4
ConveyorCountPress-v0	200	15335	2200–2399	0.250	6
ConveyorCrushLimitPick-v0	200	15713	2400–2599	0.251	—
ConveyorCullFlawed-v0	200	24999	2600–2799	0.250	5
ConveyorDepthBand-v0	200	12021	2800–2999	0.249	—
ConveyorDivertPick-v0	200	17622	3000–3199	0.251	—
ConveyorDrawerOpen-v0	200	15024	3200–3399	0.250	—
ConveyorDualPin-v0	200	12681	3400–3599	0.250	—
ConveyorEdgeFlush-v0	200	15641	3600–3799	0.251	7
ConveyorEdgeRecenter-v0	200	20419	3800–3999	0.250	5
ConveyorEvenLoad-v0	200	15540	4000–4199	0.250	—
ConveyorFaceUp-v0	200	30191	4200–4399	0.250	—
ConveyorFlagHold-v0	200	16629	4400–4599	0.250	—
ConveyorFlapPost-v0	200	13087	4600–4799	0.250	8
ConveyorFlushFit-v0	200	13718	4800–4999	0.249	6
ConveyorFreeLane-v0	200	9442	5000–5199	0.250	8
ConveyorFruitDrop-v0	200	14046	5200–5399	0.250	6
ConveyorGapThread-v0	200	11514	5400–5599	0.250	4
ConveyorHandleGrasp-v0	200	17898	5600–5799	0.249	4
ConveyorHexSocket-v0	200	10784	5800–5999	0.250	—
ConveyorHookRail-v0	200	14030	6000–6199	0.250	—
ConveyorKeySlot-v0	200	13792	6200–6399	0.250	4
ConveyorKitFill-v0	200	15441	6400–6599	0.250	4
ConveyorMatchAngle-v0	200	11825	6600–6799	0.250	—
ConveyorMeterInsert-v0	200	8774	6800–6999	0.250	—
ConveyorNamedPick-v0	200	17835	7000–7199	0.250	24
ConveyorNutRunDown-v0	200	17637	7200–7399	0.250	—
ConveyorOrderPick-v0	200	18028	7400–7599	0.249	—
ConveyorOrientPlace-v0	200	16766	7600–7799	0.250	—
ConveyorPatchCover-v0	200	14351	7800–7999	0.249	6
ConveyorPegPull-v0	200	13836	8000–8199	0.251	—
ConveyorPourFill-v0	200	12796	8200–8399	0.251	—
ConveyorPressButton-v0	200	11828	8400–8599	0.250	—
ConveyorPullCord-v0	200	13650	8600–8799	0.250	—
ConveyorRailBalance-v0	200	14596	8800–8999	0.250	4
ConveyorRampRoll-v0	200	11658	9000–9199	0.250	2
ConveyorReach-v0	200	9753	9200–9399	0.251	—
ConveyorReadPresent-v0	200	29591	9400–9599	0.249	4
ConveyorRelPlace-v0	200	14160	9600–9799	0.250	12
ConveyorRingHang-v0	200	13814	9800–9999	0.250	—
ConveyorRingHangTilt-v0	200	11740	10000–10199	0.251	—
ConveyorRockerToggle-v0	200	12572	10200–10399	0.250	—
ConveyorScanAim-v0	200	11058	10400–10599	0.250	—
ConveyorScoopBall-v0	200	15772	10600–10799	0.251	3
ConveyorSelectorSet-v0	200	16567	10800–10999	0.251	3
ConveyorSheetEdgePick-v0	200	15100	11000–11199	0.251	—
ConveyorShoeMat-v0	200	14511	11200–11399	0.250	3
ConveyorShroudBore-v0	200	8153	11400–11599	0.249	—
ConveyorSideInsert-v0	200	10974	11600–11799	0.251	—
ConveyorSideLabel-v0	200	10487	11800–11999	0.250	—
ConveyorSoftSet-v0	200	13989	12000–12199	0.250	4
ConveyorSpringHold-v0	200	12946	12200–12399	0.249	—
ConveyorStampSeal-v0	200	10313	12400–12599	0.250	—
ConveyorTareLoad-v0	200	36305	12600–12799	0.250	3
ConveyorThreadLoop-v0	200	11341	12800–12999	0.251	—
ConveyorTopple-v0	200	18330	13000–13199	0.250	—
ConveyorToteUnload-v0	200	14163	13200–13399	0.250	12
ConveyorTrayLift-v0	200	18442	13400–13599	0.251	12
ConveyorTrayToBin-v0	200	15779	13600–13799	0.250	12
ConveyorTurntablePick-v0	200	13852	13800–13999	0.250	—
ConveyorUprightBottle-v0	200	32667	14000–14199	0.250	4
ConveyorWipeBoard-v0	200	14072	14200–14399	0.250	—

What this data is not

• Successes only. --filter-success keeps episodes the scripted expert solved, so the corpus carries no failure modes. Anything learning a recovery behaviour needs data collected without it.
• A scripted expert, not a human. Trajectories come from per-family analytic controllers with a bounded velocity SDE perturbing the per-step displacement budget; they are consistent in a way teleoperation is not.
• The evaluated families only. The benchmark registers 100 dynamic families; the 72 here are the ones that clear the suite's rate bar across the whole 0.10–0.40 m/s band under the shipped configuration. The rest are registered and importable but not maintained.

Asset attribution

The manipulated objects come from the sources below — mostly primitives built procedurally by this repository rather than scanned meshes. The dataset is released CC-BY-4.0, which is compatible with each source; attribution flows through to the original authors.

• Procedural (SAPIEN primitives built in this repository) — MIT, with the code (93 distinct objects, 6,190 episodes)
• YCB Object and Model Set — CC-BY 4.0 · https://www.ycbbenchmarks.com/ (6 distinct objects, 210 episodes)

The remaining 8,000 episodes are from families whose manipulated part is fixed, so they record no per-episode object identity. No Google Scanned Objects, RoboCasa or Poly Haven mesh appears in this dataset: the families that draw from those catalogs are the ones held out of the evaluated set.

Citation

The benchmark this was collected with is unpublished; cite the repository until a paper exists.