albagon/telearms_dr300_val

TeleArms DR300 validation (1 look)

LeRobot v3 dataset (albagon/telearms_dr300_val): bimanual YAM teleoperation demos from the TeleArms game, re-rendered in MuJoCo with visual domain randomization for sim-to-real transfer of a pi0.5 policy.

Episodes	1,271
Frames	668,013 (6.2 h at 30 Hz)
Tasks	47
Robot	bimanual_yam; 14-D state/action: left_joint1..6, left_gripper, right_joint1..6, right_gripper
Cameras	camera_front, camera_left (wrist), camera_right (wrist), 224x224, letterboxed from the real cameras' aspect ratios
Grippers	1.0 = open, 0.0 = closed
Action stats	relative (action - state for joints), pinned from helen9975/pi05_telearms_v4_fullft_lr7.07e-5@4a598baa0ecab581327eb77550820fdbc2b16341

• Validation set: up to 30 demos per task from the 10% holdout of ckpt0, so ckpt0 never trained on them; one DR look per demo. Pairs with albagon/telearms_dr300_1look and albagon/telearms_dr300_2looks.
• 2 episodes whose wrist view was near-black for the whole episode are excluded (provenance/exclude_dark_wrist_val.json).
• ckpt0's train/holdout split by demo UUID: provenance/ckpt0_split_by_uuid.csv. The DR shares in the table below are for the whole render; the dr_report summary covers all looks.

Domain randomization (observed shares)

Factor	Value	Configured	Observed
table source	procedural	0.20	0.202
table source	table_other	0.20	0.198
table source	table_real	0.40	0.410
table source	table_wood	0.20	0.190
lighting setup	lamp	0.10	0.103
lighting setup	original	0.15	0.159
lighting setup	panels	0.50	0.489
lighting setup	wide	0.15	0.151
lighting setup	window	0.10	0.098
floor replaced	False	0.10	0.090
floor replaced	True	0.90	0.910
skybox replaced	False	0.10	0.098
skybox replaced	True	0.90	0.902
procedural backdrop dark	False	0.40	0.400
procedural backdrop dark	True	0.60	0.600
JPEG applied (per camera)	False	0.50	0.503
JPEG applied (per camera)	True	0.50	0.497

Full config and per-factor ranges: provenance/dr_config_resolved.json; render report: provenance/report.html and provenance/summary.json.

Episodes per task

Task	Episodes
Adjacent Red Cube Placement	30
Align Blocks	30
Aligned Cube Row Placement	30
Aligned Letter Cube Row	17
Alphabet Cube Row Arrange	11
Alternate Green Yellow Row	29
Alternate Red Blue Row	30
Arrange Alphabetical Order	18
Arrange Shapes	30
Block Line Arrangement	23
Blue Base Red Pyramid	13
Blue Cube On Red Target	30
Blue On Red Stack To Green Target	30
Colored Cube Row Arrangement	30
Cube Center Placement	30
Cube Cluster Center	30
Cube Forward Right Placement	30
Cube Midpoint Placement	21
Cube Patch Placement	30
Cube Row Arrangement	28
Dog Cube Row Arrangement	22
Front Edge Cube Row	30
Front To Back Cube Line	29
Green Cube On Yellow Target	30
Hi Cube Placement	30
L Cube Arrangement	30
Letter Face Alignment	30
Microsoft Logo Cubes	30
Opposite Corner Cube Placement	23
Push Block to Target	30
Red Behind Blue Placement	30
Red Cube Center Placement	30
Red Cube In Front Of Blue Cube	30
Red Left Of Blue Placement	30
Red On Semantic Cube Stacking	30
Rotate Letter Block	30
Semantic Cube Blue B Placement	30
Semantic Cube Green Target Red Stack	30
Semantic Cube On Red Patch A Up	30
Semantic Cube Pyramid Build	27
Semantic Cubes 1x4 Line 1234	14
Sequential Push Block to Target	30
Spell Ok Cube Alignment	16
Three Cube Pyramid Stacking	30
Three Cube Wall Row	29
Top Letter Cube Stacking	21
Vertical Insert Two Blocks	30

Loading

from lerobot.datasets.lerobot_dataset import LeRobotDataset
ds = LeRobotDataset("albagon/telearms_dr300_val")