mliu48/telearms-yam-cube-v4

TeleArms Bimanual YAM — Cube Manipulation (v4, sim-time)

Teleoperated bimanual manipulation demonstrations collected in MuJoCo with a BimanualYAM robot, packaged in LeRobot v3.0 format for pi0.5 training.

Episodes	53,832
Frames	28,326,930
Tasks	48
Videos	161,496 (3 cameras per episode)
Control rate	30 Hz
Image size	224 x 224
Size	~180 GB

Observations and actions

Three RGB views per frame — camera_front, camera_left, camera_right (224x224, H.264). Robot state and action are both 14-dimensional: 6 arm joints plus 1 gripper per arm. A 128-dimensional non-robot state carries object/scene information for diagnostics.

Actions are stored as absolute joint targets. The training recipe converts the 12 arm dimensions to deltas relative to the current action-chunk anchor while the 2 gripper dimensions stay absolute. The normalization statistics in meta/stats.json are computed in that 50-step relative space — do not renormalize relative actions with absolute-action statistics.

For pi0.5, camera keys map as:

observation.images.camera_front  -> observation.images.base_0_rgb
observation.images.camera_left   -> observation.images.left_wrist_0_rgb
observation.images.camera_right  -> observation.images.right_wrist_0_rgb

Processing

• trajectory_resampling: simulation_time_pre_action_bounded_settled_idle
• camera_profile: workspace_v2 — a fixed workspace view, identical for every task, so framing leaks no ground-truth object positions
• Resampled on a uniform simulation-time grid (not wall clock); 30 Hz control is realized by phase-accumulating 16/17 sub-steps of the 500 Hz simulator
• State/action pairs use pre-action state; legacy recordings without pre-state pair post-state i with command i+1
• Settled-idle frames are trimmed only when the action is static, both arms are within 0.02 rad of the commanded target, and the whole scene is numerically still; ~1 s of context is retained across each trimmed span
• Every episode passes a recorded-state task-success replay plus a sampled one-step action-execution check

Recorded-state success is not the same as closed-loop policy success, and no per-demonstrator or skill weighting is applied. Task counts are heavily imbalanced — see below.

Tasks

Task	Episodes	Frames
Front To Back Cube Line	7,099	4,760,467
Letter Face Alignment	6,763	3,511,362
Align Blocks	4,529	1,428,963
Rotate Letter Block	3,001	1,584,212
Cube Cluster Center	2,917	2,390,362
Cube Patch Placement	2,599	1,038,440
Sequential Push Block to Target	2,146	1,090,311
Semantic Cube On Red Patch A Up	2,034	1,287,631
Vertical Insert Two Blocks	1,860	409,658
Hi Cube Placement	1,668	1,224,706
Front Edge Cube Row	1,506	923,475
L Cube Arrangement	1,501	1,207,114
Red Left Of Blue Placement	1,289	311,819
Blue On Red Stack To Green Target	1,034	398,264
Arrange Shapes	1,027	306,050
Red On Semantic Cube Stacking	1,005	282,141
Colored Cube Row Arrangement	827	732,908
Red Cube In Front Of Blue Cube	805	207,674
Semantic Cube Blue B Placement	800	353,780
Microsoft Logo Cubes	714	657,053
Adjacent Red Cube Placement	669	189,520
Red Behind Blue Placement	585	164,365
Aligned Cube Row Placement	577	269,119
Alternate Red Blue Row	527	333,498
Semantic Cube Green Target Red Stack	490	303,700
Green Cube On Yellow Target	458	71,488
Push Block to Target	426	276,953
Cube Center Placement	414	67,813
Blue Cube On Red Target	376	69,458
Red Cube Center Placement	365	62,940
Cube Forward Right Placement	355	61,173
Three Cube Pyramid Stacking	315	163,340
Cube Row Arrangement	288	198,751
Alternate Green Yellow Row	286	186,281
Three Cube Wall Row	285	125,818
Semantic Cube Pyramid Build	267	252,092
Block Line Arrangement	230	220,269
Opposite Corner Cube Placement	223	65,243
Dog Cube Row Arrangement	212	202,181
Cube Midpoint Placement	204	35,224
Top Letter Cube Stacking	204	169,531
Arrange Alphabetical Order	173	105,231
Aligned Letter Cube Row	161	132,116
Spell Ok Cube Alignment	159	120,869
Semantic Cubes 1x4 Line 1234	140	155,534
Blue Base Red Pyramid	123	118,981
Alphabet Cube Row Arrange	110	77,582
Dual Cube Side Placement	86	21,470

Splits

No train/eval split is baked in. The reference recipe holds out the last 10% of episodes per task by episode index (deterministic, not randomized, and not grouped by operator).

Limitations

• Task distribution is highly imbalanced; the largest task is ~13% of episodes
• 224x224 inputs limit recognition of small letter faces on cubes
• The per-task holdout is not isolated by operator or collection session