mh-hf/exp-bucket

robosuite + LeRobot demonstrations for world-model training (Lance, 3-view)

Preview bucket for the world-forge data pipeline. It holds 235 manipulation datasets converted to the Lance format used by the stable-worldmodel platform, across five sources:

• 5 robomimic proficient-human tasks
• 26 MimicGen task/level variants
• 130 LIBERO tasks (the LIBERO-130 split: 90 + 10 + object + spatial + goal)
• 65 RoboCasa atomic kitchen tasks
• 9 DexMimicGen bimanual dexterous tasks

Total ≈ 14.11M frames. Every dataset is a single-version <name>.lance/ directory. Each per-task suite is nested under its own subdir — robomimic/, mimicgen/, libero/, robocasa/, dexmimicgen/, plus the mirrored ogbench/. Single multi-task corpora (droid, robotwin) sit at the top level.

These are source demonstrations, not policy rollouts. robomimic, MimicGen, LIBERO and DexMimicGen were replayed in robosuite from recorded MuJoCo states and re-rendered to 224×224 from three camera views. RoboCasa ships as pre-rendered LeRobot videos (v0.2), so its three views are decoded directly from those videos (no replay) and resized to 224×224 for consistency.

Cameras (3 views)

Every step carries three JPEG-encoded 224×224 RGB views, following the DROID two-exterior-plus-wrist layout:

column	robomimic / MimicGen / LIBERO	RoboCasa	role
pixels	agentview	robot0_agentview_left	primary exterior
sideview	sideview	robot0_agentview_right	second exterior
robot0_eye_in_hand	wrist	robot0_eye_in_hand	in-hand / eye-in-hand

The PickPlace-family envs (mimicgen_pick_place_d0, robomimic_can) do not define a sideview camera natively, so the standard robosuite world-level sideview is injected at render time. All replay-rendered datasets therefore share the same three views; RoboCasa's three come from its native left/right/wrist LeRobot videos.

Schema

All datasets share the framework columns below, named to match the stable-worldmodel / ogbench convention:

column	type	meaning
episode_idx	int32	episode index
step_idx	int32	step within episode
pixels / sideview / robot0_eye_in_hand	binary (JPEG)	the three 224×224 RGB views
action	float[]	controller action (robosuite 7-dim single-arm / 14-dim bimanual; RoboCasa 12-dim)
state	float[]	per-step state vector
reward	float[1]	task reward (sparse)
terminated / truncated / success	float[1]	episode-end / success flags

Replay-rendered sources (robomimic, MimicGen, LIBERO) additionally carry qpos / qvel (MuJoCo generalized position / velocity), render_time, id, and each source's own named low-dim observation keys, kept verbatim — e.g. robosuite's robot0_eef_pos, object, …; LIBERO's ee_pos, ee_ori, gripper_states, joint_states. There state is the flattened MuJoCo state used to drive replay.

RoboCasa differs: it is derived from LeRobot recordings, not a MuJoCo replay, so it has no qpos/qvel/sim-state. Its state is the LeRobot 16-dim observation.state proprioception and action is the 12-dim LeRobot action. This per-source variation follows the same convention as other datasets in the bucket (e.g. DROID keeps its own obs_*/act_* names).

Datasets

robomimic — proficient-human (ph, low_dim source), re-rendered

dataset	frames
robomimic_can	23,207
robomimic_lift	9,666
robomimic_square	30,154
robomimic_tool_hang	95,962
robomimic_transport	93,752

MimicGen — reset-distribution levels

dataset	frames		dataset	frames
mimicgen_coffee_d0	223,130		mimicgen_square_d0	153,477
mimicgen_coffee_d1	224,403		mimicgen_square_d1	152,400
mimicgen_coffee_d2	224,204		mimicgen_square_d2	153,112
mimicgen_coffee_preparation_d0	689,273		mimicgen_stack_d0	107,590
mimicgen_coffee_preparation_d1	687,674		mimicgen_stack_d1	108,233
mimicgen_hammer_cleanup_d0	285,359		mimicgen_stack_three_d0	254,810
mimicgen_hammer_cleanup_d1	286,847		mimicgen_stack_three_d1	255,096
mimicgen_kitchen_d0	616,751		mimicgen_threading_d0	224,508
mimicgen_kitchen_d1	619,273		mimicgen_threading_d1	223,115
mimicgen_mug_cleanup_d0	338,136		mimicgen_threading_d2	227,084
mimicgen_mug_cleanup_d1	338,034		mimicgen_three_piece_assembly_d0	336,695
mimicgen_nut_assembly_d0	358,907		mimicgen_three_piece_assembly_d1	334,869
mimicgen_pick_place_d0	677,340		mimicgen_three_piece_assembly_d2	335,949

The D0/D1/D2 suffix is the environment reset-distribution level, not a data-size level. D0 initializes objects over a region resembling the source demonstrations; D1 and D2 broaden the initial-pose distribution. See the MimicGen dataset docs.

LIBERO — libero/ (130 tasks, 1,007,618 frames)

The LIBERO-130 lifelong-learning suite, replayed from its recorded states:

suite	tasks	frames
libero_10	10	138,090
libero_90	90	669,043
libero_goal	10	63,728
libero_object	10	74,507
libero_spatial	10	62,250

Each task is libero/<suite>_<scene>_<language-goal>.lance.

RoboCasa — robocasa/ (65 atomic tasks, 1,495,313 frames)

RoboCasa v0.2 atomic kitchen tasks (PandaOmron mobile manipulator, procedurally generated kitchens), decoded from the released LeRobot videos:

dataset	frames
robocasa_AdjustToasterOvenTemperature	21,328
robocasa_AdjustWaterTemperature	20,953
robocasa_CheesyBread	31,141
robocasa_CloseBlenderLid	36,933
robocasa_CloseCabinet	27,754
robocasa_CloseDishwasher	15,781
robocasa_CloseDrawer	15,670
robocasa_CloseElectricKettleLid	7,530
robocasa_CloseFridge	26,888
robocasa_CloseFridgeDrawer	14,946
robocasa_CloseMicrowave	20,075
robocasa_CloseOven	18,230
robocasa_CloseStandMixerHead	11,593
robocasa_CloseToasterOvenDoor	19,815
robocasa_CoffeeServeMug	16,921
robocasa_CoffeeSetupMug	23,636
robocasa_LowerHeat	31,174
robocasa_MakeIcedCoffee	29,048
robocasa_NavigateKitchen	79,550
robocasa_OpenBlenderLid	20,124
robocasa_OpenCabinet	37,492
robocasa_OpenDishwasher	18,086
robocasa_OpenDrawer	20,488
robocasa_OpenElectricKettleLid	10,928
robocasa_OpenFridge	33,138
robocasa_OpenFridgeDrawer	18,517
robocasa_OpenMicrowave	26,017
robocasa_OpenOven	15,555
robocasa_OpenStandMixerHead	13,411
robocasa_OpenToasterOvenDoor	15,469
robocasa_PackDessert	27,994
robocasa_PickPlaceCabinetToCounter	20,201
robocasa_PickPlaceCounterToBlender	38,892
robocasa_PickPlaceCounterToCabinet	24,225
robocasa_PickPlaceCounterToDrawer	28,225
robocasa_PickPlaceCounterToMicrowave	42,012
robocasa_PickPlaceCounterToOven	32,014
robocasa_PickPlaceCounterToSink	22,410
robocasa_PickPlaceCounterToStandMixer	25,467
robocasa_PickPlaceCounterToStove	24,039
robocasa_PickPlaceCounterToToasterOven	24,313
robocasa_PickPlaceDrawerToCounter	31,819
robocasa_PickPlaceFridgeDrawerToShelf	26,396
robocasa_PickPlaceFridgeShelfToDrawer	27,047
robocasa_PickPlaceMicrowaveToCounter	38,729
robocasa_PickPlaceSinkToCounter	26,397
robocasa_PickPlaceStoveToCounter	23,003
robocasa_PickPlaceToasterOvenToCounter	19,323
robocasa_PickPlaceToasterToCounter	26,907
robocasa_PreheatOven	21,102
robocasa_SlideDishwasherRack	19,052
robocasa_SlideOvenRack	23,958
robocasa_SlideToasterOvenRack	11,496
robocasa_StartCoffeeMachine	13,722
robocasa_TurnOffMicrowave	15,233
robocasa_TurnOffSinkFaucet	12,309
robocasa_TurnOffStove	32,741
robocasa_TurnOnBlender	11,698
robocasa_TurnOnElectricKettle	12,460
robocasa_TurnOnMicrowave	14,010
robocasa_TurnOnSinkFaucet	23,795
robocasa_TurnOnStove	17,910
robocasa_TurnOnToaster	10,042
robocasa_TurnOnToasterOven	17,051
robocasa_TurnSinkSpout	11,130

DexMimicGen — dexmimicgen/ (9 bimanual tasks, 2,915,177 frames)

DexMimicGen bimanual dexterous tasks (NVIDIA, ICRA 2025) across three robot configs — bimanual Panda, Panda + dexterous hands (PandaDexRH/LH), and the GR1 humanoid — replayed from recorded states. Native cameras are agentview + two wrist views, so sideview is the injected world camera. action is 14-dim (bimanual Panda) or 24-dim (dexterous / GR1); proprio obs keys span robot0/robot1 (and GR1 left/right).

dataset	frames
dexmimicgen_two_arm_box_cleanup	234,398
dexmimicgen_two_arm_can_sort_random	322,073
dexmimicgen_two_arm_coffee	326,707
dexmimicgen_two_arm_drawer_cleanup	298,235
dexmimicgen_two_arm_lift_tray	516,848
dexmimicgen_two_arm_pouring	338,519
dexmimicgen_two_arm_threading	218,858
dexmimicgen_two_arm_three_piece_assembly	239,827
dexmimicgen_two_arm_transport	419,712

Loading

from huggingface_hub import snapshot_download
import lance

local = snapshot_download(
    "mh-hf/exp-bucket", repo_type="dataset",
    allow_patterns="robomimic/robomimic_lift.lance/*",  # or "libero/*"
)
ds = lance.dataset(f"{local}/robomimic/robomimic_lift.lance")
print(ds.schema.names)               # note: pixels, sideview, robot0_eye_in_hand
print(ds.to_table(limit=2).to_pandas())

Provenance and licensing

Rendering and Lance conversion were done with the stable-worldmodel platform (three 224×224 cameras). The underlying demonstrations come from:

• robomimic — Mandlekar et al., What Matters in Learning from Offline Human Demonstrations for Robot Manipulation, CoRL 2021.
• MimicGen — Mandlekar et al., MimicGen: A Data Generation System for Scalable Robot Learning using Human Demonstrations, CoRL 2023.
• LIBERO — Liu et al., LIBERO: Benchmarking Knowledge Transfer for Lifelong Robot Learning, NeurIPS 2023.
• RoboCasa — Nasiriany et al., RoboCasa: Large-Scale Simulation of Everyday Tasks for Generalist Robots, RSS 2024 (v0.2 LeRobot release).
• DexMimicGen — Jiang et al., DexMimicGen: Automated Data Generation for Bimanual Dexterous Manipulation via Imitation Learning, ICRA 2025.

Upstream sources are MIT-licensed; this bucket redistributes derived renders for research use. It is a throwaway preview account for collaboration review.