mh-hf/exp-bucket

LeWM multi-domain robot corpus

Multi-domain robot data for latent world-model training and dataset-replay evaluation (reset an env to any stored frame, roll out a planner, score against the stored goal). Ten domain columns (eight simulators, OGBench locomotion, PushT), unified where it matters, documented where it can't be. Human and generated demonstrations both live here, tagged per episode, so a world model can be trained on the full behaviour distribution rather than on expert data alone. Since 2026-10-04 non-demonstration play data sits next to them: noisy replays, kept failures, and OGBench's goal-chaining play datasets; since 2026-10-08 also OCBench play, random subgoals carried out with human-like scripted motion; since 2026-10-13 PushT with six more block shapes (DINO-WM's PushObj).

Format: Lance. Each dataset is one <name>.lance table, one row per frame, episode_idx + step_idx addressing. Column-level detail for every task is in probe_manifest.json at the repo root (category / shape / dtype / frame / rot_encoding / entity / note per column, 0 unclassified; manifest_version 1.5). Task entries carry n_frames, n_episodes and, where a table mixes provenances, the per-tag episode counts.

Domains

domain	robot	scene	provenance	# tables	episodes	frames	hours	proprio	reset-state
robomimic	Panda 7-DoF (transport: 2× Panda)	fixed	ph + mh + mg + paired, one table per task (demo_source)	5	7,800	1,452,950	20.2	robosuite robot0_* (+robot1_*)	state, qpos
mimicgen	Panda 7-DoF; square / threading also IIWA, Sawyer, UR5e (6-DoF)	fixed	MimicGen-generated: core 26 + 14 robot / object variants + 3 keep-failed play tables (generation_success)	43	43,000	12,013,807	166.9	robosuite robot0_* + task-relative eef	state, qpos
libero	Panda 7-DoF	fixed	human 130 + 10 noisy-replay play tables (noise_std)	140	8,500	1,559,978	21.7	ee_pos / ee_ori / joint_states / gripper_states at the frame; *_next after the action	state, qpos (+ the demo's fixture poses)
robocasa	PandaOmron (mobile)	procedural	human atomic 65 + MimicGen-generated _mg 60 (300 ep each, separate tables)	125	25,356	5,368,125	74.6	robot0_base_* + robot0_base_to_eef_*	human tables: state, states (+ ep_meta, model_xml, dataset_meta); _mg tables: state only
dexmimicgen	2× Panda (bimanual)	fixed	DexMimicGen-generated	9	9,178	2,915,177	40.5	robot0_* / robot0_left_* / robot0_right_*	state, qpos
ogbench	6-DoF arm	semi	OGBench play release, first 1,000 episodes (10 tables; octuple: own collection, same recipe) + 9 noisy replays (data_type)	19	31,000	31,030,000	431.0	proprio/* namespace	qpos, qvel (+ privileged/button_*_state for scene / puzzle)
ogbench_loco	point mass / Ant / Humanoid	maze	navigate / stitch / explore (data_type)	29	118,000	57,118,000	1,252.8	observation (agent proprio)	qpos, qvel
ocbench	UR5e 6-DoF + Robotiq 2F-85	semi	OCBench human-like scripted motion: 13 play tables (random subgoals, data_type play) + 1 bowling scripted-push table (data_type demo), 1,000 episodes each	14	14,000	13,214,000	183.5	proprio/* namespace (OGBench names)	qpos, qvel (+ privileged/button_*_state) after rebuilding the episode's scene from its seed (play_segments.json)
pusht	2-D agent (pymunk)	fixed	T: LeWorldModel expert (HF quentinll/lewm-pusht) = DINO-WM's PushT, 185 Diffusion Policy demos replayed 101× with noise; L, Z, +, I, small T, square (PushObj, 6 tables): planner expert with closed-loop noise (noise_std)	7	24,685	2,802,233	77.8	proprio[4]	state[7]
molmo	Franka (DROID)	procedural	MolmoBot / ProcTHOR-Objaverse, 44 per-object tables (+ monolithic source table)	44	10,679	2,347,101	43.5	proprio[18] (native)	object / receptacle poses (no full sim state)
total				435	292,198	129,821,371	2,312.4

hours = frames / control rate / 3600 (see the per-task table below for the rates).

What changed on 2026-10-13

• PushObj tables (6, new) under pusht/: pushl_expert, pushz_expert, pushplus_expert, pushi_expert, pushsmallt_expert, pushsquare_expert. They are PushT with the block swapped for the other shapes of DINO-WM's PushObj generalization benchmark (AdaJEPA trains on T, L, Z, + and holds out I, small T and square), rendered and stored like pusht_expert: 224 px top-down JPEG pixels, action[2] (agent position delta in [-1, 1]², ×100 px, 10 Hz), proprio[4], state[7], the green target at the same pose for every shape.
• pusht_expert is DINO-WM's PushT data. Its 18,685 episodes are the 185 Diffusion Policy human demos (of 206) replayed open loop 101 times each from the demo's own start: once clean and 10 times at each of ten levels of Gaussian noise on the relative action (0.1, 0.3, 0.5, 1, 1.5, 2, 2.5, 3, 7, 15 px per step). Same episode and frame counts as DINO-WM's released pusht_noise/train.
• How the PushObj episodes were made. Human demos exist only for the T, so an oracle planner stands in: CEM over 20-step action sequences rolled out in the simulator from the true state, pushing the block onto the target (symmetry-aware pose error), smooth sampled sequences capped at 30 px per step (human-like speed), a small bias toward staying in contact with the block. Noise is injected in closed loop, so the planner corrects every drift: Gaussian noise on the executed action with a per-episode std (noise_std: 0, 2, 4, 7, 10 or 15 px, drawn uniformly, about twice DINO-WM's mix), paused while the agent has not touched the block for 5 steps. An episode ends at 95% coverage of the target, after 30 steps without progress, or at 300 steps; a start whose episode is shorter than 40 steps or touches the block in fewer than 20% of its steps is redrawn (up to 10 times). The point is that every stretch of an episode acts on the block: open-loop noisy replays of such demos left the agent without contact for a median 107-114 steps, against 27 in DINO-WM's human demos.
• Per row: noise_std[1] (per episode), data_type (demo), success[1] (95% coverage after the row's action), reward[1] (Diffusion Policy's PushT reward, coverage / 0.95 clipped to 1), robot_contact[1] (the agent touches the block after the action). Table metadata: task, shape, instruction, recipe.

table	shape	episodes	frames	length (median, range)	contact	longest no-contact run (median / p90)	reached 95% coverage	starts redrawn
pushl_expert	L	1,000	84,475	83 (40-209)	0.56	17 / 29	33%	65
pushz_expert	Z	1,000	77,387	74 (40-166)	0.45	17 / 29	27%	88
pushplus_expert	+	1,000	79,289	76 (40-189)	0.40	18 / 29	36%	204
pushi_expert	I	1,000	69,809	65 (40-178)	0.40	18 / 29	6%	141
pushsmallt_expert	small T	1,000	86,947	85 (40-205)	0.58	17 / 28	39%	53
pushsquare_expert	square	1,000	67,590	64 (40-167)	0.34	20 / 30	14%	327

What changed on 2026-10-12

• The 130 human LIBERO tables were replaced. Their rows, state, qpos, qvel, action and episode order are unchanged; their images and proprio columns are not.
• Images: 104 tables re-rendered in the scene the demos were recorded in. The earlier render built each scene from today's LIBERO bddl, which differs from the recording in five ways: the robot base of one task (LIVING_ROOM_SCENE2 butter, 14 cm), the salad-dressing model of four LIVING_ROOM_SCENE4 tasks (5 cm higher), the fixture poses LIBERO samples at every reset outside the state (88 tables, 5-51 mm; each demo's model_file holds the recorded ones), sRGB-tagged textures that MuJoCo 3.2 and later render darker (the floor plane, the salad dressing), and the stove burner LIBERO lights after each step. Against the official LIBERO HDF5 images (128 px) the median per-table mean absolute difference fell from 1.7 to 0.2 (agentview) and 2.2 to 0.4 (wrist), the worst table's from 38 / 42 to 3.1 / 2.0. The other 26 tables render identically and kept their images.
• Proprio is aligned with the frame. ee_pos, ee_ori, ee_states, gripper_states and joint_states are now read at the row's own state, frame 0 included. The previous values, which LIBERO's create_dataset.py read after the row's action, are kept as <col>_next, the convention of the noisy tables. joint_states_next[t] and gripper_states_next[t] equal the next row's values exactly; the end-effector columns differ from the next row's by one physics substep (0.06-0.43 mm median per table). Code that read ee_pos & co. from these tables as the post-action reading must switch to <col>_next.
• libero_90_LIVING_ROOM_SCENE2_pick_up_the_butter_and_put_it_in_the_basket also has a new wrist extrinsic (its robot sits 14 cm further back, as recorded).

What changed on 2026-10-08

• OCBench tables (14, new) under ocbench/. OCBench (Park & Levine,
  2026. drives a UR5e + Robotiq 2F-85 with scripted controllers whose motion resembles human teleoperation (randomized spline keyframes, varied speed and grasps, occasional detours); it releases only fixed-task demonstrations. Here every subgoal of its controller is drawn at random and the episode never stops at task success, which turns the same motion into play: ocb_block_{single,double,triple,quadruple}_play (random block moves and stacks), ocb_chamber_{easy,medium,hard}_play (buttons, drawer, window and blocks; a block in a closed drawer is never picked), ocb_switch_{3x3,4x4,5x5}_play (random presses on the Lights Out board, random dropped buttons), ocb_hanoi_{single,double,triple}_play (random legal moves, random peg layout). Each holds 1,000 episodes of 1,001 rows (50 s); data_type = "play". ocb_bowling_push (1,000 × 201 rows, data_type = "demo") is OCBench's own randomized single push: bowling has no play variant (one push per episode, pins never reset).
• 20 Hz. OCBench runs at 50 Hz; these tables were collected at 20 Hz like the rest of the manipulation corpus (control_timestep 0.05, action_delta_scale 2.5, step-counted script constants rescaled). The rate change was validated on 500 seeds per task against 50 Hz: block-double 79.8 / 78.2 %, chamber-easy 98.0 / 97.6 %, switch-3x3 11.6 / 11.4 %, hanoi-single 99.2 / 99.4 % task success, bowling return 0.27 / 0.30.
• Action: action[7] is OCBench's normalized command in [-1, 1]: six arm joint-position deltas (× 0.45, 0.45, 0.45, 0.9, 0.9, 0.9 rad at this rate) and a gripper-opening delta (× 0.3; opening 0 = open, 1 = closed). The arm dimensions never saturate; the gripper dimension sits at ±1 while a grasp holds or the hand opens fully.
• Cameras: pixels = OCBench front, sideview = side, robot0_eye_in_hand = the real wrist camera (ur5e/wrist, fovy 75°), each with per-frame geometry; images and proprio/* are taken at the recorded qpos (verified: replaying a row's state reproduces its stored images to JPEG noise, and block centres project to their masks within 0.3 px median).
• Sidecars in each table folder: view_mapping.json (cameras, rate, seeds, retries) and play_segments.json (per episode: seed, task id, and every subgoal with its start, end and outcome).
• Quaternion order: privileged/*_quat in the OGBench and OCBench tables is (w, x, y, z), MuJoCo's order; earlier manifests labelled it quat_xyzw. Checked on stored rows: the yaw of the (w, x, y, z) reading matches privileged/*_yaw to 1e-7 rad, the (x, y, z, w) reading misses by up to 1.5 rad.

What changed on 2026-10-04

• OGBench play tables are OGBench's released play data. ogb_<env>_multiview for cube single / double / triple / quadruple, scene and puzzle 3x3 / 4x4 / 4x5 / 4x6 now hold the official *-play-v0 datasets, first 1,000 episodes each (1,001 steps per episode), replay-rendered in the three corpus views like the noisy tables (replay exact). They carry data_type = "play" and no target / success / control. OGBench releases no octuple set, so ogb_cube_octuple_multiview was collected with OGBench's play recipe (plan oracle, a new random goal after each one is reached, 1,000 episodes of 1,000 steps).
• The single-view names are gone. ogb_cube_single and ogb_scene_single would have held the same data as ogb_cube_single_multiview and ogb_scene_single_multiview, so they were removed. For one view, read pixels and its camera columns from the multiview table; Lance reads only the columns asked for, so this is as fast as a single-view table.
• Why: the tables these replace reached one cube goal per episode. stable-worldmodel's set_new_target() reuses the goal drawn at reset, so its collections never chain goals, and the removed ogb_cube_single (the LeWorldModel OGBench-Cube release) returned its cube to the same goal twice in 99.9% of episodes. OGBench's play recipe chains about 12 cube goals per 1,000-step episode (measured on a collection with the same recipe; the released files do not store goals).
• cube quadruple: release episode 431 cannot be replayed on MuJoCo 3.x (one of its states exceeds mj_maxContact), so episode 1000 takes its place at the end; the table's view_mapping.json has the map.
• Column names: every OGBench table uses proprio/<x> and privileged/<x>. The removed single-view tables used proprio_<x> / ep_idx; the quantities are the same, so code keyed on the old names needs the new ones.
• Play tables (13, new): libero/libero_10_<task>_noisy.lance replays each LIBERO-10 demo open loop with Gaussian noise of std 0, 0.1, 0.2 or 0.4 on the six arm dims (noise_std; action is the applied action, action_clean the recorded one, src_episode the demo); proprio is aligned with the frame, and <col>_next holds the reading after the action, which was then the convention of the human LIBERO tables (they carry both since 2026-10-12). mimicgen/mimicgen_<task>_d0_keepfail.lance (square, threading, coffee) are MimicGen generations that keep the failed attempts (generation_success per episode, success per frame).
• ogb_scene_noisy gained the camera columns of its sideview and robot0_eye_in_hand views.

What changed in the 2026-09 expansion

• robomimic is one table per task (robomimic/robomimic_<task>.lance) merging the proficient-human (ph), multi-human (mh), MimicGen-generated (mg) and paired (paired) releases. Every row carries demo_source and src_episode (its episode index in the source release); episode_idx is contiguous across sources (ph first). The v1.5 mh / mg / paired files have no end-effector velocities, so robot*_eef_vel_lin / robot*_eef_vel_ang are NaN on non-ph rows.
• MimicGen robot / object variants: mimicgen_square_d{0,1}_{iiwa,sawyer,ur5e}, mimicgen_threading_d{0,1}_{iiwa,sawyer,ur5e}, mimicgen_mug_cleanup_o{1,2} (1,000 demos each). Per-robot tables because gripper and joint dimensions differ (UR5e is 6-DoF; IIWA / Sawyer carry 6-d gripper qpos).
• RoboCasa MimicGen-generated data: robocasa/robocasa_<Task>_mg.lance, a seeded 300-episode sample of each atomic task's generated release (demo_source = "mg", src_episode = index in that release). Separate from the human tables: the generated release ships without the extras (states, ep_meta, model_xml, dataset_meta), so these tables have proprio + cameras but no reset columns.
• OGBench noisy manipulation: ogbench/ogb_<scene>_noisy.lance (cube single/double/triple/quadruple, scene, puzzle 3x3/4x4/4x5/4x6), replay-rendered from the official *-noisy-v0 datasets in the same three views, with the same verified camera geometry as the play tables. Replay is exact (joint positions match the source to 0.00). They carry data_type = "noisy" and no target / success / control (not derivable from the source).
• OGBench locomotion: ogbench/ogb_<pointmaze|antmaze|humanoidmaze|antsoccer>_<size>_<type>.lance (29 tables: navigate / stitch / explore, data_type = the type), one fixed top-down camera per maze (pixels + its geometry), observation = the agent's proprioceptive vector, qpos / qvel for reset.
• PushT: pusht/pusht_expert.lance, the LeWorldModel expert data (HF quentinll/lewm-pusht), 18,685 episodes at 10 Hz: pixels (single top-down render), action[2] (agent position delta), proprio[4] (agent xy + velocity), state[7] (agent xy, block xy, block angle, agent velocity). No pinhole camera model (orthographic 2-D render).

The unified interface (what every domain shares)

Same names, same conventions, across every table:

• RGB views, per-frame JPEG (224×224): pixels, and for the manipulation domains also sideview + robot0_eye_in_hand (OGBench locomotion and PushT have pixels only).
• Camera-geometry columns, one OpenCV cam2world convention, for each view: <view>_intrinsic [4] = fx,fy,cx,cy · <view>_extrinsic [16] = 4×4 world_T_cam · <view>_cam_raw [12]. Present for every view of every table except PushT.
• action, episode_idx, step_idx; reward / success / terminated / truncated where the source defines them.

The camera geometry is verified to match its image: projecting a known 3D world point (the end-effector, or the maze agent for locomotion) through each view's K + extrinsic lands on it in that view (reprojection-checked; K is for the stored 224² resolution).

Proprioception and the reset-state are intentionally kept in each domain's native names (the world model ignores proprio at train time; the eval reads the right column per domain via this card).

Variable matrix

variable	robomimic	mimicgen	libero	robocasa	dexmimicgen	ogbench	ogbench_loco	ocbench	pusht	molmo
Observation
RGB frames (per-frame JPEG)	✅	✅	✅	✅	✅	✅	✅	✅	✅	✅
flat observation vector	·	·	·	·	·	✅	✅	✅	·	·
Camera geometry
intrinsics K	✅	✅	✅	✅	✅	✅	✅	✅	·	✅
extrinsics world_T_cam	✅	✅	✅	✅	✅	✅	✅	✅	·	✅
raw camera pose	✅	✅	✅	✅	✅	✅	✅	✅	·	✅
Proprioception
eef position	✅	✅	✅	✅	✅	✅	·	✅	·	✅
eef orientation	✅	✅	✅	✅	✅	✅	·	✅	·	✅
eef velocity	✅	✅	·	·	·	·	·	·	·	·
joint position	✅	✅	✅	·	✅	✅	·	✅	·	·
joint velocity	✅	✅	·	·	✅	✅	·	✅	·	·
gripper	✅	✅	✅	✅	✅	✅	·	✅	·	·
mobile base pose	·	·	·	✅	✅	·	·	✅	·	✅
flat proprio vector	·	·	·	·	·	·	·	·	✅	✅
Env reconstruction
flat state vector	✅	✅	✅	✅	✅	·	·	·	✅	·
full MuJoCo state (reset)	·	·	·	✅	·	·	·	·	·	·
qpos / qvel	✅	✅	✅	·	✅	✅	✅	✅	·	·
ep_meta scene recipe	·	·	·	✅	·	·	·	·	·	·
model_xml exact scene	·	·	·	✅	·	·	·	·	·	·
dataset_meta env_args	·	·	·	✅	·	·	·	·	·	·
privileged sim state	·	·	·	·	·	✅	·	✅	·	·
World & object
object state / poses	✅	✅	·	·	✅	✅	·	✅	·	✅
articulated / button state	·	·	·	·	·	✅	·	✅	·	·
Task & meta
action	✅	✅	✅	✅	✅	✅	✅	✅	✅	✅
reward / success	✅	✅	✅	✅	✅	✅	✅	✅	✅	✅
goal / target state	·	·	·	·	·	✅	·	·	·	·
scene id	·	·	·	·	·	·	·	·	·	✅
object identity (JSON)	·	·	·	·	·	·	·	·	·	✅
grasp pose	·	·	·	·	·	·	·	·	·	✅
Provenance
demonstration source (per episode)	✅	·	·	✅	·	·	·	·	·	·
source episode index	✅	·	✅	✅	·	·	·	·	·	·
dataset split tag	·	·	·	·	·	✅	✅	✅	✅	·

Per-task hours

hours = stored frames / stored rate / 3600 — stored frames counted once per table (not per camera) over the domain's recording rate: robomimic / mimicgen / libero / robocasa / dexmimicgen 20 Hz; ogbench manipulation 20 Hz (control_timestep 0.05); ocbench 20 Hz (control_timestep 0.05, collected at this rate instead of OCBench's 50 Hz); ogbench locomotion 10 Hz for point / ant (timestep 0.02 × frame_skip 5) and 40 Hz for humanoid (0.005 × 5); molmo 15 Hz; PushT 10 Hz (control_hz of swm/PushT-v1). This is not wall-clock. molmo is represented by its 44 per-object tables; the monolithic molmo table is the source they are carved from and is not counted here.

robomimic (one table per task: ph + mh + mg + paired) — 5 tables

task	episodes	frames	hz	hours
robomimic_can	4,600	690,758	20	9.59
robomimic_lift	2,000	265,793	20	3.69
robomimic_square	500	110,885	20	1.54
robomimic_tool_hang	200	95,962	20	1.33
robomimic_transport	500	289,552	20	4.02
subtotal	7,800	1,452,950		20.2

mimicgen (core 26 + robot/object variants 14 + keep-failed play 3) — 43 tables

task	episodes	frames	hz	hours
mimicgen_coffee_d0	1,000	223,130	20	3.10
mimicgen_coffee_d0_keepfail	1,000	218,752	20	3.04
mimicgen_coffee_d1	1,000	224,403	20	3.12
mimicgen_coffee_d2	1,000	224,204	20	3.11
mimicgen_coffee_preparation_d0	1,000	689,273	20	9.57
mimicgen_coffee_preparation_d1	1,000	687,674	20	9.55
mimicgen_hammer_cleanup_d0	1,000	285,359	20	3.96
mimicgen_hammer_cleanup_d1	1,000	286,847	20	3.98
mimicgen_kitchen_d0	1,000	616,751	20	8.57
mimicgen_kitchen_d1	1,000	619,273	20	8.60
mimicgen_mug_cleanup_d0	1,000	338,136	20	4.70
mimicgen_mug_cleanup_d1	1,000	338,034	20	4.70
mimicgen_mug_cleanup_o1	1,000	333,018	20	4.62
mimicgen_mug_cleanup_o2	1,000	336,548	20	4.67
mimicgen_nut_assembly_d0	1,000	358,907	20	4.99
mimicgen_pick_place_d0	1,000	677,340	20	9.41
mimicgen_square_d0	1,000	153,477	20	2.13
mimicgen_square_d0_iiwa	1,000	150,404	20	2.09
mimicgen_square_d0_keepfail	1,000	155,158	20	2.15
mimicgen_square_d0_sawyer	1,000	153,765	20	2.14
mimicgen_square_d0_ur5e	1,000	152,649	20	2.12
mimicgen_square_d1	1,000	152,400	20	2.12
mimicgen_square_d1_iiwa	1,000	151,959	20	2.11
mimicgen_square_d1_sawyer	1,000	153,470	20	2.13
mimicgen_square_d1_ur5e	1,000	152,827	20	2.12
mimicgen_square_d2	1,000	153,112	20	2.13
mimicgen_stack_d0	1,000	107,590	20	1.49
mimicgen_stack_d1	1,000	108,233	20	1.50
mimicgen_stack_three_d0	1,000	254,810	20	3.54
mimicgen_stack_three_d1	1,000	255,096	20	3.54
mimicgen_threading_d0	1,000	224,508	20	3.12
mimicgen_threading_d0_iiwa	1,000	235,741	20	3.27
mimicgen_threading_d0_keepfail	1,000	220,644	20	3.06
mimicgen_threading_d0_sawyer	1,000	229,992	20	3.19
mimicgen_threading_d0_ur5e	1,000	236,079	20	3.28
mimicgen_threading_d1	1,000	223,115	20	3.10
mimicgen_threading_d1_iiwa	1,000	234,203	20	3.25
mimicgen_threading_d1_sawyer	1,000	229,401	20	3.19
mimicgen_threading_d1_ur5e	1,000	232,928	20	3.23
mimicgen_threading_d2	1,000	227,084	20	3.15
mimicgen_three_piece_assembly_d0	1,000	336,695	20	4.68
mimicgen_three_piece_assembly_d1	1,000	334,869	20	4.65
mimicgen_three_piece_assembly_d2	1,000	335,949	20	4.67
subtotal	43,000	12,013,807		166.9

libero (human 130 + noisy-replay play 10) — 140 tables

task	episodes	frames	hz	hours
libero_10_KITCHEN_SCENE3_turn_on_the_stove_and_put_the_moka_pot_on_it	50	13,298	20	0.18
libero_10_KITCHEN_SCENE3_turn_on_the_stove_and_put_the_moka_pot_on_it_noisy	200	53,192	20	0.74
libero_10_KITCHEN_SCENE4_put_the_black_bowl_in_the_bottom_drawer_of_the_cabinet_and_close_it	50	12,434	20	0.17
libero_10_KITCHEN_SCENE4_put_the_black_bowl_in_the_bottom_drawer_of_the_cabinet_and_close_it_noisy	200	49,736	20	0.69
libero_10_KITCHEN_SCENE6_put_the_yellow_and_white_mug_in_the_microwave_and_close_it	50	15,232	20	0.21
libero_10_KITCHEN_SCENE6_put_the_yellow_and_white_mug_in_the_microwave_and_close_it_noisy	200	60,928	20	0.85
libero_10_KITCHEN_SCENE8_put_both_moka_pots_on_the_stove	50	20,794	20	0.29
libero_10_KITCHEN_SCENE8_put_both_moka_pots_on_the_stove_noisy	200	83,176	20	1.16
libero_10_LIVING_ROOM_SCENE1_put_both_the_alphabet_soup_and_the_cream_cheese_box_in_the_basket	50	13,476	20	0.19
libero_10_LIVING_ROOM_SCENE1_put_both_the_alphabet_soup_and_the_cream_cheese_box_in_the_basket_noisy	200	53,904	20	0.75
libero_10_LIVING_ROOM_SCENE2_put_both_the_alphabet_soup_and_the_tomato_sauce_in_the_basket	50	14,700	20	0.20
libero_10_LIVING_ROOM_SCENE2_put_both_the_alphabet_soup_and_the_tomato_sauce_in_the_basket_noisy	200	58,800	20	0.82
libero_10_LIVING_ROOM_SCENE2_put_both_the_cream_cheese_box_and_the_butter_in_the_basket	50	13,021	20	0.18
libero_10_LIVING_ROOM_SCENE2_put_both_the_cream_cheese_box_and_the_butter_in_the_basket_noisy	200	52,084	20	0.72
libero_10_LIVING_ROOM_SCENE5_put_the_white_mug_on_the_left_plate_and_put_the_yellow_and_white_mug_on_the_right_plate	50	12,909	20	0.18
libero_10_LIVING_ROOM_SCENE5_put_the_white_mug_on_the_left_plate_and_put_the_yellow_and_white_mug_on_the_right_plate_noisy	200	51,636	20	0.72
libero_10_LIVING_ROOM_SCENE6_put_the_white_mug_on_the_plate_and_put_the_chocolate_pudding_to_the_right_of_the_plate	50	12,756	20	0.18
libero_10_LIVING_ROOM_SCENE6_put_the_white_mug_on_the_plate_and_put_the_chocolate_pudding_to_the_right_of_the_plate_noisy	200	51,024	20	0.71
libero_10_STUDY_SCENE1_pick_up_the_book_and_place_it_in_the_back_compartment_of_the_caddy	50	9,470	20	0.13
libero_10_STUDY_SCENE1_pick_up_the_book_and_place_it_in_the_back_compartment_of_the_caddy_noisy	200	37,880	20	0.53
libero_90_KITCHEN_SCENE10_close_the_top_drawer_of_the_cabinet	50	3,828	20	0.05
libero_90_KITCHEN_SCENE10_close_the_top_drawer_of_the_cabinet_and_put_the_black_bowl_on_top_of_it	50	10,662	20	0.15
libero_90_KITCHEN_SCENE10_put_the_black_bowl_in_the_top_drawer_of_the_cabinet	50	6,208	20	0.09
libero_90_KITCHEN_SCENE10_put_the_butter_at_the_back_in_the_top_drawer_of_the_cabinet_and_close_it	50	9,440	20	0.13
libero_90_KITCHEN_SCENE10_put_the_butter_at_the_front_in_the_top_drawer_of_the_cabinet_and_close_it	50	9,070	20	0.13
libero_90_KITCHEN_SCENE10_put_the_chocolate_pudding_in_the_top_drawer_of_the_cabinet_and_close_it	50	8,969	20	0.12
libero_90_KITCHEN_SCENE1_open_the_bottom_drawer_of_the_cabinet	50	7,157	20	0.10
libero_90_KITCHEN_SCENE1_open_the_top_drawer_of_the_cabinet	50	4,736	20	0.07
libero_90_KITCHEN_SCENE1_open_the_top_drawer_of_the_cabinet_and_put_the_bowl_in_it	50	10,010	20	0.14
libero_90_KITCHEN_SCENE1_put_the_black_bowl_on_the_plate	50	6,628	20	0.09
libero_90_KITCHEN_SCENE1_put_the_black_bowl_on_top_of_the_cabinet	50	6,811	20	0.10
libero_90_KITCHEN_SCENE2_open_the_top_drawer_of_the_cabinet	50	3,765	20	0.05
libero_90_KITCHEN_SCENE2_put_the_black_bowl_at_the_back_on_the_plate	50	5,538	20	0.08
libero_90_KITCHEN_SCENE2_put_the_black_bowl_at_the_front_on_the_plate	50	6,910	20	0.10
libero_90_KITCHEN_SCENE2_put_the_middle_black_bowl_on_the_plate	50	5,393	20	0.07
libero_90_KITCHEN_SCENE2_put_the_middle_black_bowl_on_top_of_the_cabinet	50	6,821	20	0.10
libero_90_KITCHEN_SCENE2_stack_the_black_bowl_at_the_front_on_the_black_bowl_in_the_middle	50	6,415	20	0.09
libero_90_KITCHEN_SCENE2_stack_the_middle_black_bowl_on_the_back_black_bowl	50	5,730	20	0.08
libero_90_KITCHEN_SCENE3_put_the_frying_pan_on_the_stove	50	9,979	20	0.14
libero_90_KITCHEN_SCENE3_put_the_moka_pot_on_the_stove	50	7,563	20	0.10
libero_90_KITCHEN_SCENE3_turn_on_the_stove	50	4,645	20	0.07
libero_90_KITCHEN_SCENE3_turn_on_the_stove_and_put_the_frying_pan_on_it	50	14,013	20	0.20
libero_90_KITCHEN_SCENE4_close_the_bottom_drawer_of_the_cabinet	50	5,832	20	0.08
libero_90_KITCHEN_SCENE4_close_the_bottom_drawer_of_the_cabinet_and_open_the_top_drawer	50	11,773	20	0.16
libero_90_KITCHEN_SCENE4_put_the_black_bowl_in_the_bottom_drawer_of_the_cabinet	50	6,643	20	0.09
libero_90_KITCHEN_SCENE4_put_the_black_bowl_on_top_of_the_cabinet	50	7,910	20	0.11
libero_90_KITCHEN_SCENE4_put_the_wine_bottle_in_the_bottom_drawer_of_the_cabinet	50	6,376	20	0.09
libero_90_KITCHEN_SCENE4_put_the_wine_bottle_on_the_wine_rack	50	12,093	20	0.17
libero_90_KITCHEN_SCENE5_close_the_top_drawer_of_the_cabinet	50	3,762	20	0.05
libero_90_KITCHEN_SCENE5_put_the_black_bowl_in_the_top_drawer_of_the_cabinet	50	6,107	20	0.09
libero_90_KITCHEN_SCENE5_put_the_black_bowl_on_the_plate	50	6,665	20	0.09
libero_90_KITCHEN_SCENE5_put_the_black_bowl_on_top_of_the_cabinet	50	7,576	20	0.10
libero_90_KITCHEN_SCENE5_put_the_ketchup_in_the_top_drawer_of_the_cabinet	50	10,539	20	0.15
libero_90_KITCHEN_SCENE6_close_the_microwave	50	10,160	20	0.14
libero_90_KITCHEN_SCENE6_put_the_yellow_and_white_mug_to_the_front_of_the_white_mug	50	6,282	20	0.09
libero_90_KITCHEN_SCENE7_open_the_microwave	50	7,414	20	0.10
libero_90_KITCHEN_SCENE7_put_the_white_bowl_on_the_plate	50	9,702	20	0.14
libero_90_KITCHEN_SCENE7_put_the_white_bowl_to_the_right_of_the_plate	50	6,946	20	0.10
libero_90_KITCHEN_SCENE8_put_the_right_moka_pot_on_the_stove	50	10,588	20	0.15
libero_90_KITCHEN_SCENE8_turn_off_the_stove	50	8,795	20	0.12
libero_90_KITCHEN_SCENE9_put_the_frying_pan_on_the_cabinet_shelf	50	9,457	20	0.13
libero_90_KITCHEN_SCENE9_put_the_frying_pan_on_top_of_the_cabinet	50	9,272	20	0.13
libero_90_KITCHEN_SCENE9_put_the_frying_pan_under_the_cabinet_shelf	50	8,612	20	0.12
libero_90_KITCHEN_SCENE9_put_the_white_bowl_on_top_of_the_cabinet	50	7,519	20	0.10
libero_90_KITCHEN_SCENE9_turn_on_the_stove	50	5,871	20	0.08
libero_90_KITCHEN_SCENE9_turn_on_the_stove_and_put_the_frying_pan_on_it	50	13,047	20	0.18
libero_90_LIVING_ROOM_SCENE1_pick_up_the_alphabet_soup_and_put_it_in_the_basket	50	6,939	20	0.10
libero_90_LIVING_ROOM_SCENE1_pick_up_the_cream_cheese_box_and_put_it_in_the_basket	50	7,426	20	0.10
libero_90_LIVING_ROOM_SCENE1_pick_up_the_ketchup_and_put_it_in_the_basket	50	8,665	20	0.12
libero_90_LIVING_ROOM_SCENE1_pick_up_the_tomato_sauce_and_put_it_in_the_basket	50	8,088	20	0.11
libero_90_LIVING_ROOM_SCENE2_pick_up_the_alphabet_soup_and_put_it_in_the_basket	50	7,034	20	0.10
libero_90_LIVING_ROOM_SCENE2_pick_up_the_butter_and_put_it_in_the_basket	50	6,609	20	0.09
libero_90_LIVING_ROOM_SCENE2_pick_up_the_milk_and_put_it_in_the_basket	50	5,875	20	0.08
libero_90_LIVING_ROOM_SCENE2_pick_up_the_orange_juice_and_put_it_in_the_basket	50	8,383	20	0.12
libero_90_LIVING_ROOM_SCENE2_pick_up_the_tomato_sauce_and_put_it_in_the_basket	50	5,306	20	0.07
libero_90_LIVING_ROOM_SCENE3_pick_up_the_alphabet_soup_and_put_it_in_the_tray	50	6,004	20	0.08
libero_90_LIVING_ROOM_SCENE3_pick_up_the_butter_and_put_it_in_the_tray	50	5,069	20	0.07
libero_90_LIVING_ROOM_SCENE3_pick_up_the_cream_cheese_and_put_it_in_the_tray	50	7,697	20	0.11
libero_90_LIVING_ROOM_SCENE3_pick_up_the_ketchup_and_put_it_in_the_tray	50	7,986	20	0.11
libero_90_LIVING_ROOM_SCENE3_pick_up_the_tomato_sauce_and_put_it_in_the_tray	50	5,571	20	0.08
libero_90_LIVING_ROOM_SCENE4_pick_up_the_black_bowl_on_the_left_and_put_it_in_the_tray	50	6,176	20	0.09
libero_90_LIVING_ROOM_SCENE4_pick_up_the_chocolate_pudding_and_put_it_in_the_tray	50	7,660	20	0.11
libero_90_LIVING_ROOM_SCENE4_pick_up_the_salad_dressing_and_put_it_in_the_tray	50	5,970	20	0.08
libero_90_LIVING_ROOM_SCENE4_stack_the_left_bowl_on_the_right_bowl_and_place_them_in_the_tray	50	10,771	20	0.15
libero_90_LIVING_ROOM_SCENE4_stack_the_right_bowl_on_the_left_bowl_and_place_them_in_the_tray	50	11,734	20	0.16
libero_90_LIVING_ROOM_SCENE5_put_the_red_mug_on_the_left_plate	50	7,239	20	0.10
libero_90_LIVING_ROOM_SCENE5_put_the_red_mug_on_the_right_plate	50	7,032	20	0.10
libero_90_LIVING_ROOM_SCENE5_put_the_white_mug_on_the_left_plate	50	5,327	20	0.07
libero_90_LIVING_ROOM_SCENE5_put_the_yellow_and_white_mug_on_the_right_plate	50	5,377	20	0.07
libero_90_LIVING_ROOM_SCENE6_put_the_chocolate_pudding_to_the_left_of_the_plate	50	6,195	20	0.09
libero_90_LIVING_ROOM_SCENE6_put_the_chocolate_pudding_to_the_right_of_the_plate	50	4,473	20	0.06
libero_90_LIVING_ROOM_SCENE6_put_the_red_mug_on_the_plate	50	6,668	20	0.09
libero_90_LIVING_ROOM_SCENE6_put_the_white_mug_on_the_plate	50	7,907	20	0.11
libero_90_STUDY_SCENE1_pick_up_the_book_and_place_it_in_the_front_compartment_of_the_caddy	50	8,742	20	0.12
libero_90_STUDY_SCENE1_pick_up_the_book_and_place_it_in_the_left_compartment_of_the_caddy	50	7,938	20	0.11
libero_90_STUDY_SCENE1_pick_up_the_book_and_place_it_in_the_right_compartment_of_the_caddy	50	7,831	20	0.11
libero_90_STUDY_SCENE1_pick_up_the_yellow_and_white_mug_and_place_it_to_the_right_of_the_caddy	50	6,405	20	0.09
libero_90_STUDY_SCENE2_pick_up_the_book_and_place_it_in_the_back_compartment_of_the_caddy	50	7,336	20	0.10
libero_90_STUDY_SCENE2_pick_up_the_book_and_place_it_in_the_front_compartment_of_the_caddy	50	7,582	20	0.10
libero_90_STUDY_SCENE2_pick_up_the_book_and_place_it_in_the_left_compartment_of_the_caddy	50	6,104	20	0.09
libero_90_STUDY_SCENE2_pick_up_the_book_and_place_it_in_the_right_compartment_of_the_caddy	50	7,115	20	0.10
libero_90_STUDY_SCENE3_pick_up_the_book_and_place_it_in_the_front_compartment_of_the_caddy	50	8,738	20	0.12
libero_90_STUDY_SCENE3_pick_up_the_book_and_place_it_in_the_left_compartment_of_the_caddy	50	7,310	20	0.10
libero_90_STUDY_SCENE3_pick_up_the_book_and_place_it_in_the_right_compartment_of_the_caddy	50	6,774	20	0.09
libero_90_STUDY_SCENE3_pick_up_the_red_mug_and_place_it_to_the_right_of_the_caddy	50	5,896	20	0.08
libero_90_STUDY_SCENE3_pick_up_the_white_mug_and_place_it_to_the_right_of_the_caddy	50	5,158	20	0.07
libero_90_STUDY_SCENE4_pick_up_the_book_in_the_middle_and_place_it_on_the_cabinet_shelf	50	8,055	20	0.11
libero_90_STUDY_SCENE4_pick_up_the_book_on_the_left_and_place_it_on_top_of_the_shelf	50	8,547	20	0.12
libero_90_STUDY_SCENE4_pick_up_the_book_on_the_right_and_place_it_on_the_cabinet_shelf	50	5,111	20	0.07
libero_90_STUDY_SCENE4_pick_up_the_book_on_the_right_and_place_it_under_the_cabinet_shelf	50	5,988	20	0.08
libero_goal_open_the_middle_drawer_of_the_cabinet	50	7,027	20	0.10
libero_goal_open_the_top_drawer_and_put_the_bowl_inside	50	10,208	20	0.14
libero_goal_push_the_plate_to_the_front_of_the_stove	50	7,638	20	0.11
libero_goal_put_the_bowl_on_the_plate	50	4,669	20	0.07
libero_goal_put_the_bowl_on_the_stove	50	5,081	20	0.07
libero_goal_put_the_bowl_on_top_of_the_cabinet	50	5,094	20	0.07
libero_goal_put_the_cream_cheese_in_the_bowl	50	5,349	20	0.07
libero_goal_put_the_wine_bottle_on_the_rack	50	8,808	20	0.12
libero_goal_put_the_wine_bottle_on_top_of_the_cabinet	50	5,394	20	0.07
libero_goal_turn_on_the_stove	50	4,460	20	0.06
libero_object_pick_up_the_alphabet_soup_and_place_it_in_the_basket	50	7,808	20	0.11
libero_object_pick_up_the_bbq_sauce_and_place_it_in_the_basket	50	7,348	20	0.10
libero_object_pick_up_the_butter_and_place_it_in_the_basket	50	7,865	20	0.11
libero_object_pick_up_the_chocolate_pudding_and_place_it_in_the_basket	50	7,982	20	0.11
libero_object_pick_up_the_cream_cheese_and_place_it_in_the_basket	50	7,194	20	0.10
libero_object_pick_up_the_ketchup_and_place_it_in_the_basket	50	8,056	20	0.11
libero_object_pick_up_the_milk_and_place_it_in_the_basket	50	7,281	20	0.10
libero_object_pick_up_the_orange_juice_and_place_it_in_the_basket	50	6,968	20	0.10
libero_object_pick_up_the_salad_dressing_and_place_it_in_the_basket	50	6,641	20	0.09
libero_object_pick_up_the_tomato_sauce_and_place_it_in_the_basket	50	7,364	20	0.10
libero_spatial_pick_up_the_black_bowl_between_the_plate_and_the_ramekin_and_place_it_on_the_plate	50	5,068	20	0.07
libero_spatial_pick_up_the_black_bowl_from_table_center_and_place_it_on_the_plate	50	5,882	20	0.08
libero_spatial_pick_up_the_black_bowl_in_the_top_drawer_of_the_wooden_cabinet_and_place_it_on_the_plate	50	7,479	20	0.10
libero_spatial_pick_up_the_black_bowl_next_to_the_cookie_box_and_place_it_on_the_plate	50	6,312	20	0.09
libero_spatial_pick_up_the_black_bowl_next_to_the_plate_and_place_it_on_the_plate	50	5,963	20	0.08
libero_spatial_pick_up_the_black_bowl_next_to_the_ramekin_and_place_it_on_the_plate	50	6,707	20	0.09
libero_spatial_pick_up_the_black_bowl_on_the_cookie_box_and_place_it_on_the_plate	50	5,052	20	0.07
libero_spatial_pick_up_the_black_bowl_on_the_ramekin_and_place_it_on_the_plate	50	5,796	20	0.08
libero_spatial_pick_up_the_black_bowl_on_the_stove_and_place_it_on_the_plate	50	7,111	20	0.10
libero_spatial_pick_up_the_black_bowl_on_the_wooden_cabinet_and_place_it_on_the_plate	50	6,880	20	0.10
subtotal	8,500	1,559,978		21.7

robocasa (human atomic 65 + MimicGen-generated _mg 60) — 125 tables

task	episodes	frames	hz	hours
robocasa_AdjustToasterOvenTemperature	107	21,328	20	0.30
robocasa_AdjustToasterOvenTemperature_mg	300	63,770	20	0.89
robocasa_AdjustWaterTemperature	106	20,953	20	0.29
robocasa_AdjustWaterTemperature_mg	300	64,492	20	0.90
robocasa_CheesyBread	101	31,141	20	0.43
robocasa_CloseBlenderLid	106	36,933	20	0.51
robocasa_CloseBlenderLid_mg	300	108,173	20	1.50
robocasa_CloseCabinet	105	27,754	20	0.39
robocasa_CloseCabinet_mg	300	77,352	20	1.07
robocasa_CloseDishwasher	111	15,781	20	0.22
robocasa_CloseDishwasher_mg	300	48,949	20	0.68
robocasa_CloseDrawer	110	15,670	20	0.22
robocasa_CloseDrawer_mg	300	46,168	20	0.64
robocasa_CloseElectricKettleLid	102	7,530	20	0.10
robocasa_CloseElectricKettleLid_mg	300	25,100	20	0.35
robocasa_CloseFridge	106	26,888	20	0.37
robocasa_CloseFridgeDrawer	111	14,946	20	0.21
robocasa_CloseFridgeDrawer_mg	300	44,945	20	0.62
robocasa_CloseFridge_mg	300	68,902	20	0.96
robocasa_CloseMicrowave	105	20,075	20	0.28
robocasa_CloseMicrowave_mg	300	64,032	20	0.89
robocasa_CloseOven	107	18,230	20	0.25
robocasa_CloseOven_mg	300	55,206	20	0.77
robocasa_CloseStandMixerHead	110	11,593	20	0.16
robocasa_CloseStandMixerHead_mg	300	31,417	20	0.44
robocasa_CloseToasterOvenDoor	110	19,815	20	0.28
robocasa_CloseToasterOvenDoor_mg	300	59,622	20	0.83
robocasa_CoffeeServeMug	108	16,921	20	0.23
robocasa_CoffeeServeMug_mg	300	51,885	20	0.72
robocasa_CoffeeSetupMug	105	23,636	20	0.33
robocasa_CoffeeSetupMug_mg	300	69,414	20	0.96
robocasa_LowerHeat	110	31,174	20	0.43
robocasa_MakeIcedCoffee	102	29,048	20	0.40
robocasa_NavigateKitchen	503	79,550	20	1.10
robocasa_OpenBlenderLid	106	20,124	20	0.28
robocasa_OpenBlenderLid_mg	300	60,374	20	0.84
robocasa_OpenCabinet	107	37,492	20	0.52
robocasa_OpenCabinet_mg	300	128,939	20	1.79
robocasa_OpenDishwasher	111	18,086	20	0.25
robocasa_OpenDishwasher_mg	300	52,061	20	0.72
robocasa_OpenDrawer	102	20,488	20	0.28
robocasa_OpenDrawer_mg	300	64,829	20	0.90
robocasa_OpenElectricKettleLid	110	10,928	20	0.15
robocasa_OpenElectricKettleLid_mg	300	33,338	20	0.46
robocasa_OpenFridge	105	33,138	20	0.46
robocasa_OpenFridgeDrawer	107	18,517	20	0.26
robocasa_OpenFridgeDrawer_mg	300	51,848	20	0.72
robocasa_OpenFridge_mg	300	144,395	20	2.00
robocasa_OpenMicrowave	105	26,017	20	0.36
robocasa_OpenMicrowave_mg	300	79,914	20	1.11
robocasa_OpenOven	108	15,555	20	0.22
robocasa_OpenOven_mg	300	47,886	20	0.67
robocasa_OpenStandMixerHead	109	13,411	20	0.19
robocasa_OpenStandMixerHead_mg	300	41,079	20	0.57
robocasa_OpenToasterOvenDoor	105	15,469	20	0.21
robocasa_OpenToasterOvenDoor_mg	300	88,700	20	1.23
robocasa_PackDessert	108	27,994	20	0.39
robocasa_PickPlaceCabinetToCounter	106	20,201	20	0.28
robocasa_PickPlaceCabinetToCounter_mg	300	61,614	20	0.86
robocasa_PickPlaceCounterToBlender	106	38,892	20	0.54
robocasa_PickPlaceCounterToBlender_mg	300	111,899	20	1.55
robocasa_PickPlaceCounterToCabinet	108	24,225	20	0.34
robocasa_PickPlaceCounterToCabinet_mg	300	70,173	20	0.97
robocasa_PickPlaceCounterToDrawer	107	28,225	20	0.39
robocasa_PickPlaceCounterToDrawer_mg	300	76,527	20	1.06
robocasa_PickPlaceCounterToMicrowave	110	42,012	20	0.58
robocasa_PickPlaceCounterToMicrowave_mg	300	97,118	20	1.35
robocasa_PickPlaceCounterToOven	111	32,014	20	0.45
robocasa_PickPlaceCounterToOven_mg	300	82,040	20	1.14
robocasa_PickPlaceCounterToSink	108	22,410	20	0.31
robocasa_PickPlaceCounterToSink_mg	300	70,429	20	0.98
robocasa_PickPlaceCounterToStandMixer	107	25,467	20	0.35
robocasa_PickPlaceCounterToStandMixer_mg	300	69,911	20	0.97
robocasa_PickPlaceCounterToStove	108	24,039	20	0.33
robocasa_PickPlaceCounterToStove_mg	300	70,555	20	0.98
robocasa_PickPlaceCounterToToasterOven	108	24,313	20	0.34
robocasa_PickPlaceCounterToToasterOven_mg	300	70,632	20	0.98
robocasa_PickPlaceDrawerToCounter	103	31,819	20	0.44
robocasa_PickPlaceDrawerToCounter_mg	300	90,344	20	1.25
robocasa_PickPlaceFridgeDrawerToShelf	106	26,396	20	0.37
robocasa_PickPlaceFridgeDrawerToShelf_mg	300	76,293	20	1.06
robocasa_PickPlaceFridgeShelfToDrawer	111	27,047	20	0.38
robocasa_PickPlaceFridgeShelfToDrawer_mg	300	75,776	20	1.05
robocasa_PickPlaceMicrowaveToCounter	112	38,729	20	0.54
robocasa_PickPlaceMicrowaveToCounter_mg	300	91,956	20	1.28
robocasa_PickPlaceSinkToCounter	108	26,397	20	0.37
robocasa_PickPlaceSinkToCounter_mg	300	75,998	20	1.06
robocasa_PickPlaceStoveToCounter	109	23,003	20	0.32
robocasa_PickPlaceStoveToCounter_mg	300	65,715	20	0.91
robocasa_PickPlaceToasterOvenToCounter	106	19,323	20	0.27
robocasa_PickPlaceToasterOvenToCounter_mg	300	60,469	20	0.84
robocasa_PickPlaceToasterToCounter	105	26,907	20	0.37
robocasa_PickPlaceToasterToCounter_mg	300	80,917	20	1.12
robocasa_PreheatOven	107	21,102	20	0.29
robocasa_PreheatOven_mg	300	60,555	20	0.84
robocasa_SlideDishwasherRack	100	19,052	20	0.27
robocasa_SlideDishwasherRack_mg	300	62,600	20	0.87
robocasa_SlideOvenRack	113	23,958	20	0.33
robocasa_SlideOvenRack_mg	300	67,541	20	0.94
robocasa_SlideToasterOvenRack	105	11,496	20	0.16
robocasa_SlideToasterOvenRack_mg	300	36,091	20	0.50
robocasa_StartCoffeeMachine	108	13,722	20	0.19
robocasa_StartCoffeeMachine_mg	300	40,055	20	0.56
robocasa_TurnOffMicrowave	108	15,233	20	0.21
robocasa_TurnOffMicrowave_mg	300	45,082	20	0.63
robocasa_TurnOffSinkFaucet	106	12,309	20	0.17
robocasa_TurnOffSinkFaucet_mg	300	39,462	20	0.55
robocasa_TurnOffStove	109	32,741	20	0.46
robocasa_TurnOffStove_mg	300	86,321	20	1.20
robocasa_TurnOnBlender	107	11,698	20	0.16
robocasa_TurnOnBlender_mg	300	34,219	20	0.47
robocasa_TurnOnElectricKettle	108	12,460	20	0.17
robocasa_TurnOnElectricKettle_mg	300	39,391	20	0.55
robocasa_TurnOnMicrowave	107	14,010	20	0.20
robocasa_TurnOnMicrowave_mg	300	41,267	20	0.57
robocasa_TurnOnSinkFaucet	107	23,795	20	0.33
robocasa_TurnOnSinkFaucet_mg	300	72,828	20	1.01
robocasa_TurnOnStove	104	17,910	20	0.25
robocasa_TurnOnStove_mg	300	49,255	20	0.68
robocasa_TurnOnToaster	108	10,042	20	0.14
robocasa_TurnOnToasterOven	103	17,051	20	0.24
robocasa_TurnOnToasterOven_mg	300	60,329	20	0.84
robocasa_TurnOnToaster_mg	300	31,808	20	0.44
robocasa_TurnSinkSpout	109	11,130	20	0.15
robocasa_TurnSinkSpout_mg	300	34,852	20	0.48
subtotal	25,356	5,368,125		74.6

dexmimicgen — 9 tables

task	episodes	frames	hz	hours
dexmimicgen_two_arm_box_cleanup	1,016	234,398	20	3.26
dexmimicgen_two_arm_can_sort_random	1,020	322,073	20	4.47
dexmimicgen_two_arm_coffee	1,014	326,707	20	4.54
dexmimicgen_two_arm_drawer_cleanup	1,026	298,235	20	4.14
dexmimicgen_two_arm_lift_tray	1,033	516,848	20	7.18
dexmimicgen_two_arm_pouring	1,009	338,519	20	4.70
dexmimicgen_two_arm_threading	1,025	218,858	20	3.04
dexmimicgen_two_arm_three_piece_assembly	1,006	239,827	20	3.33
dexmimicgen_two_arm_transport	1,029	419,712	20	5.83
subtotal	9,178	2,915,177		40.5

ogbench (manipulation: OGBench play release + noisy) — 19 tables

task	episodes	frames	hz	hours
ogb_cube_double_multiview	1,000	1,001,000	20	13.90
ogb_cube_double_noisy	1,000	1,001,000	20	13.90
ogb_cube_octuple_multiview	1,000	1,000,000	20	13.89
ogb_cube_quadruple_multiview	1,000	1,001,000	20	13.90
ogb_cube_quadruple_noisy	5,000	5,005,000	20	69.51
ogb_cube_single_multiview	1,000	1,001,000	20	13.90
ogb_cube_single_noisy	1,000	1,001,000	20	13.90
ogb_cube_triple_multiview	1,000	1,001,000	20	13.90
ogb_cube_triple_noisy	3,000	3,003,000	20	41.71
ogb_puzzle_3x3_multiview	1,000	1,001,000	20	13.90
ogb_puzzle_3x3_noisy	1,000	1,001,000	20	13.90
ogb_puzzle_4x4_multiview	1,000	1,001,000	20	13.90
ogb_puzzle_4x4_noisy	1,000	1,001,000	20	13.90
ogb_puzzle_4x5_multiview	1,000	1,001,000	20	13.90
ogb_puzzle_4x5_noisy	3,000	3,003,000	20	41.71
ogb_puzzle_4x6_multiview	1,000	1,001,000	20	13.90
ogb_puzzle_4x6_noisy	5,000	5,005,000	20	69.51
ogb_scene_noisy	1,000	1,001,000	20	13.90
ogb_scene_single_multiview	1,000	1,001,000	20	13.90
subtotal	31,000	31,030,000		431.0

ogbench_loco (locomotion) — 29 tables

task	episodes	frames	hz	hours
ogb_antmaze_giant_navigate	500	1,000,500	10	27.79
ogb_antmaze_giant_stitch	5,000	1,005,000	10	27.92
ogb_antmaze_large_explore	10,000	5,010,000	10	139.17
ogb_antmaze_large_navigate	1,000	1,001,000	10	27.81
ogb_antmaze_large_stitch	5,000	1,005,000	10	27.92
ogb_antmaze_medium_explore	10,000	5,010,000	10	139.17
ogb_antmaze_medium_navigate	1,000	1,001,000	10	27.81
ogb_antmaze_medium_stitch	5,000	1,005,000	10	27.92
ogb_antmaze_teleport_explore	10,000	5,010,000	10	139.17
ogb_antmaze_teleport_navigate	1,000	1,001,000	10	27.81
ogb_antmaze_teleport_stitch	5,000	1,005,000	10	27.92
ogb_antsoccer_arena_navigate	1,000	1,001,000	10	27.81
ogb_antsoccer_arena_stitch	5,000	1,005,000	10	27.92
ogb_antsoccer_medium_navigate	4,000	4,004,000	10	111.22
ogb_antsoccer_medium_stitch	8,000	4,008,000	10	111.33
ogb_humanoidmaze_giant_navigate	1,000	4,001,000	40	27.79
ogb_humanoidmaze_giant_stitch	10,000	4,010,000	40	27.85
ogb_humanoidmaze_large_navigate	1,000	2,001,000	40	13.90
ogb_humanoidmaze_large_stitch	5,000	2,005,000	40	13.92
ogb_humanoidmaze_medium_navigate	1,000	2,001,000	40	13.90
ogb_humanoidmaze_medium_stitch	5,000	2,005,000	40	13.92
ogb_pointmaze_giant_navigate	500	1,000,500	10	27.79
ogb_pointmaze_giant_stitch	5,000	1,005,000	10	27.92
ogb_pointmaze_large_navigate	1,000	1,001,000	10	27.81
ogb_pointmaze_large_stitch	5,000	1,005,000	10	27.92
ogb_pointmaze_medium_navigate	1,000	1,001,000	10	27.81
ogb_pointmaze_medium_stitch	5,000	1,005,000	10	27.92
ogb_pointmaze_teleport_navigate	1,000	1,001,000	10	27.81
ogb_pointmaze_teleport_stitch	5,000	1,005,000	10	27.92
subtotal	118,000	57,118,000		1,252.8

ocbench (OCBench play 13 + bowling scripted push 1) — 14 tables

task	episodes	frames	hz	hours
ocb_block_double_play	1,000	1,001,000	20	13.90
ocb_block_quadruple_play	1,000	1,001,000	20	13.90
ocb_block_single_play	1,000	1,001,000	20	13.90
ocb_block_triple_play	1,000	1,001,000	20	13.90
ocb_bowling_push	1,000	201,000	20	2.79
ocb_chamber_easy_play	1,000	1,001,000	20	13.90
ocb_chamber_hard_play	1,000	1,001,000	20	13.90
ocb_chamber_medium_play	1,000	1,001,000	20	13.90
ocb_hanoi_double_play	1,000	1,001,000	20	13.90
ocb_hanoi_single_play	1,000	1,001,000	20	13.90
ocb_hanoi_triple_play	1,000	1,001,000	20	13.90
ocb_switch_3x3_play	1,000	1,001,000	20	13.90
ocb_switch_4x4_play	1,000	1,001,000	20	13.90
ocb_switch_5x5_play	1,000	1,001,000	20	13.90
subtotal	14,000	13,214,000		183.5

molmo_obj — 44 tables

task	episodes	frames	hz	hours
molmo_obj_Irishpotato	479	114,701	15	2.12
molmo_obj_apple	189	42,991	15	0.80
molmo_obj_atomizer	464	111,981	15	2.07
molmo_obj_boiler	67	17,652	15	0.33
molmo_obj_bottle	71	18,938	15	0.35
molmo_obj_bowl	291	71,402	15	1.32
molmo_obj_butterknife	228	50,985	15	0.94
molmo_obj_candle	31	1,486	15	0.03
molmo_obj_cellulartelephone	198	47,385	15	0.88
molmo_obj_cookingpan	28	6,332	15	0.12
molmo_obj_creditcard	200	24,939	15	0.46
molmo_obj_cup	1,307	319,150	15	5.91
molmo_obj_egg	3	73	15	0.00
molmo_obj_fork	127	28,033	15	0.52
molmo_obj_knife	291	57,248	15	1.06
molmo_obj_ladle	103	22,593	15	0.42
molmo_obj_objabellpepper	729	174,364	15	3.23
molmo_obj_objabowl	7	901	15	0.02
molmo_obj_objaclock	42	11,540	15	0.21
molmo_obj_objacorkscrew	33	8,174	15	0.15
molmo_obj_objacrownmolding	24	2,203	15	0.04
molmo_obj_objadigitalcamera	256	61,426	15	1.14
molmo_obj_objaegg	68	14,505	15	0.27
molmo_obj_objaeggcarton	54	13,224	15	0.24
molmo_obj_objafantasybook	174	30,633	15	0.57
molmo_obj_objafossilreplica	470	52,057	15	0.96
molmo_obj_objaglove	57	11,352	15	0.21
molmo_obj_objahandplane	565	136,712	15	2.53
molmo_obj_objamechanicalgauntlet	15	2,090	15	0.04
molmo_obj_objastoneartifact	872	135,263	15	2.50
molmo_obj_papertowel	249	63,193	15	1.17
molmo_obj_pen	55	10,572	15	0.20
molmo_obj_peppershaker	304	64,680	15	1.20
molmo_obj_plate	35	3,775	15	0.07
molmo_obj_pot	21	4,573	15	0.09
molmo_obj_remotecontrol	877	191,792	15	3.55
molmo_obj_saltshaker	770	190,509	15	3.53
molmo_obj_soapdispenser	238	61,449	15	1.14
molmo_obj_spatula	272	67,741	15	1.25
molmo_obj_spoon	257	64,070	15	1.19
molmo_obj_statue	13	2,920	15	0.05
molmo_obj_tissuepaper	69	17,623	15	0.33
molmo_obj_vase	48	9,447	15	0.17
molmo_obj_watch	28	4,424	15	0.08
subtotal	10,679	2,347,101		43.5

pusht (LeWM expert) — 7 tables

task	episodes	frames	hz	hours
pushi_expert	1,000	69,809	10	1.94
pushl_expert	1,000	84,475	10	2.35
pushplus_expert	1,000	79,289	10	2.20
pushsmallt_expert	1,000	86,947	10	2.42
pushsquare_expert	1,000	67,590	10	1.88
pusht_expert	18,685	2,336,736	10	64.91
pushz_expert	1,000	77,387	10	2.15
subtotal	24,685	2,802,233		77.8

Grand total: 435 tables · 292,198 episodes · 129,821,371 frames · 2,312.4 hours

Storage format

Images are per-frame JPEG blobs (binary columns) in the main table in every domain, random-access per frame, which is what the world-model loader samples (history + frameskip). molmo was decoded from its original per-episode mp4 into this layout.

Camera views: same name, different physical camera

The three view names are shared, but the physical camera differs by domain. robot0_eye_in_hand is the wrist camera in every domain except OGBench manipulation, which has no wrist camera: there it is a second fixed front camera (OGBench's front). pixels is a main / front-ish exterior view but the pose varies (robocasa's rides the mobile base; ogbench manipulation is more top-down; ocbench's front looks back at the robot from beyond the table; ogbench locomotion is a fixed top-down maze camera). sideview is a true ~90° side for the robosuite tabletop domains, but for robocasa it is a second exterior view (agentview left/right) close to pixels. The per-view extrinsic / intrinsic columns record the real pose, so downstream code reads the actual geometry regardless of the shared name.

Environment reconstruction (dataset-replay reset)

Fixed-scene domains (robomimic / mimicgen / libero / dexmimicgen) restore a frame from a state vector + the fixed model: env.reset_to({"states": state[t]}). LIBERO adds two things the state does not hold: the poses of its fixtures (cabinets, stoves, microwaves...), which a reset samples anew and the demo's model_file in the LIBERO HDF5 records (write them into the fixture bodies' body_pos / body_quat after the reset), and the object-state visuals such as a lit stove burner, which LIBERO sets in _post_process() after each step (call it after setting a state).

robocasa is procedural, a new kitchen per episode, so its human tables also carry the scene recipe: states (per-frame flattened MuJoCo state, dim varies per episode), ep_meta (layout / style / object_cfgs / cam_configs, on step 0), model_xml (exact gzipped scene XML, on step 0, required), dataset_meta (env_args to instantiate the env, on step 0). Reset: env.reset_to({"states": states[0], "model": model_xml, "ep_meta": ep_meta}), then per-frame env.reset_to({"states": states[t]}). Reproduction was pixel-verified. Requires a complete robocasa asset install on the render machine + robosuite 1.5.2 / env 0.5.1. The _mg tables cannot be reset this way (no extras in the generated release).

ogbench (manipulation and locomotion) resets with env.set_state(qpos, qvel), plus the button states (privileged/button_*_state) for scene and puzzle; PushT with env.reset(options={"state": state[t]}) in swm/PushT-v1, and the PushObj tables with env.reset(options={"state": state[t], "variation_values": {"block.shape": env.unwrapped.shapes.index(shape)}}), shape from the table's metadata (L, Z, +, I, small_tee, square).

ocbench (pip install ocbench==1.0.1) randomizes the scene layout at reset (chamber drawer / window / button positions, hanoi peg layout, switch dropped buttons), so a frame is restored by first rebuilding its episode: env = ocbench.make(<env id>, control_timestep=0.05, action_delta_scale=2.5), then env.reset(seed=seed, options={"task_id": task}) with the episode's seed and task from the table's play_segments.json (env ids in view_mapping.json), then qpos / qvel and, for chamber and switch, the privileged/button_*_state values (env.unwrapped._cur_button_states, then _apply_button_states()). privileged/dynamics_info (chamber, hanoi) records the rebuilt layout for checking.

Gotchas

• robomimic robot*_eef_vel_* are NaN on mh / mg / paired rows (see above); filter by demo_source if you consume them.
• libero proprio changed meaning on 2026-10-12: ee_pos, ee_ori, ee_states, gripper_states and joint_states of the human tables are now the reading at the frame; the earlier post-action reading is <col>_next (see above).
• robocasa _mg tables have no reset columns and no success semantics beyond the source's reward / success flags; the human tables are unchanged.
• ogbench play and noisy tables have no target / success / control (the released datasets do not store them; ogb_cube_octuple_multiview does have them). data_type tells play from noisy.
• ogbench single view: there are no single-view manipulation tables; read pixels from the _multiview table (it is the camera the removed ogb_cube_single / ogb_scene_single used).
• ogbench locomotion tables are single-view (pixels only) but do carry that view's camera geometry.
• ogbench proprio names are proprio/<x> in every table since 2026-10-04 (the removed single-view tables used proprio_<x>).
• molmo (MolmoBot / ProcTHOR-Objaverse) has 2 RGB views (pixels = exo_camera_1, robot0_eye_in_hand = wrist; no sideview). Its cameras are parametric (robot-mounted), so intrinsic is from FOV and extrinsic is best-effort (mount spec + robot pose, not reprojection-verified); <view>_cam_raw holds the exact raw mount spec. Pose quaternion order is wxyz. molmo carries no full sim reset state, so it cannot be reset-replayed like robocasa.
• ocbench rows hold the state before the row's action and that action; the episode's last action was applied but the state after it is not stored. The first subgoal of every episode starts from a random arm pose. One pick-and-place takes 10-25 s of human-like motion, so a 50 s play episode holds about 3-4 block moves (chamber about 6 subgoals, switch about 9 presses). ocb_bowling_push episodes are 10 s with one push each; about the last 40% of every bowling episode is the arm holding still.
• ocbench gripper action is close to binary (±1 while holding or fully open); the six arm dimensions are smooth and never saturate.
• PushObj episodes vary in length (they end at success or on stalling) and their action is the executed, noisy one; pusht_expert has none of the extra columns (noise_std, success, reward, robot_contact).
• proprio / reset-state names are domain-native by design; use this card + probe_manifest.json to select the right column per domain.

Files

• <domain>/<task>.lance — the datasets (robomimic/, mimicgen/, libero/, robocasa/, dexmimicgen/, ogbench/, ocbench/, pusht/ (T and the six PushObj shapes), molmo_obj_<category>/, molmo/); the ocbench and ogbench play folders also carry view_mapping.json, and the ocbench ones play_segments.json.
• robotwin/robotwin.lance — RoboTwin (27,500 episodes), on the bucket but not yet in the manifest or the tables above.
• probe_manifest.json — full per-column schema for every table (0 unclassified columns).