albagon/telearms_dr300_2looks

TeleArms DR300 train, 2 looks per demo

LeRobot v3 dataset (albagon/telearms_dr300_2looks): 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	24,652
Frames	12,800,291 (118.5 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

• Training demos: up to 300 per task from the training split of checkpoint helen9975/pi05_telearms_v4_fullft_lr7.07e-5 (ckpt0); none come from ckpt0's holdout (checked against provenance/ckpt0_holdout_uuids.json).
• Two independent DR looks per demo (__dr00, __dr01). The __dr00 episodes are albagon/telearms_dr300_1look.
• 18 episodes whose wrist view was near-black for the whole episode are excluded (provenance/exclude_dark_wrist_train.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	0.200
table source	table_other	0.20	0.196
table source	table_real	0.40	0.404
table source	table_wood	0.20	0.200
lighting setup	lamp	0.10	0.102
lighting setup	original	0.15	0.145
lighting setup	panels	0.50	0.505
lighting setup	wide	0.15	0.149
lighting setup	window	0.10	0.099
floor replaced	False	0.10	0.100
floor replaced	True	0.90	0.900
skybox replaced	False	0.10	0.104
skybox replaced	True	0.90	0.896
procedural backdrop dark	False	0.40	0.395
procedural backdrop dark	True	0.60	0.605
JPEG applied (per camera)	False	0.50	0.501
JPEG applied (per camera)	True	0.50	0.499

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	600
Align Blocks	599
Aligned Cube Row Placement	599
Aligned Letter Cube Row	288
Alphabet Cube Row Arrange	198
Alternate Green Yellow Row	513
Alternate Red Blue Row	599
Arrange Alphabetical Order	310
Arrange Shapes	600
Block Line Arrangement	414
Blue Base Red Pyramid	220
Blue Cube On Red Target	599
Blue On Red Stack To Green Target	600
Colored Cube Row Arrangement	599
Cube Center Placement	600
Cube Cluster Center	600
Cube Forward Right Placement	597
Cube Midpoint Placement	364
Cube Patch Placement	600
Cube Row Arrangement	518
Dog Cube Row Arrangement	380
Front Edge Cube Row	599
Front To Back Cube Line	600
Green Cube On Yellow Target	600
Hi Cube Placement	600
L Cube Arrangement	600
Letter Face Alignment	600
Microsoft Logo Cubes	599
Opposite Corner Cube Placement	400
Push Block to Target	600
Red Behind Blue Placement	599
Red Cube Center Placement	600
Red Cube In Front Of Blue Cube	600
Red Left Of Blue Placement	599
Red On Semantic Cube Stacking	599
Rotate Letter Block	600
Semantic Cube Blue B Placement	600
Semantic Cube Green Target Red Stack	600
Semantic Cube On Red Patch A Up	600
Semantic Cube Pyramid Build	480
Semantic Cubes 1x4 Line 1234	252
Sequential Push Block to Target	600
Spell Ok Cube Alignment	286
Three Cube Pyramid Stacking	566
Three Cube Wall Row	512
Top Letter Cube Stacking	366
Vertical Insert Two Blocks	598

Loading

from lerobot.datasets.lerobot_dataset import LeRobotDataset
ds = LeRobotDataset("albagon/telearms_dr300_2looks")