VR-VLA/EgoDex-Retarget-16robot

EgoDex Retarget — 16 robots

EgoDex egocentric human demonstrations retargeted onto 16 robot embodiments. For every episode, the human wrist and hand motion is solved into a joint trajectory for each robot, and the robot is rendered and composited into the original scene, over hand-erased frames and per-frame metric depth. Each (robot, task) pair is one LeRobot v3.0 dataset.

The repo holds two episode sets that share no episode. The first set covers all 16 robots and is described in Contents below. Set 2 adds 4,433 more episodes on the 14 single-arm robots, and is still being uploaded.

Inputs: VR-VLA/VR-egodex-inpaint-cache-full (hand-erased frames), VR-VLA/VR-egodex-depth-cache-full (depth), bi199797/VR-egodex-apple-lerobot (source episodes).

Contents

The first set has 6.5 hours of human demonstration (2,200 episodes, 702,843 frames at 30 fps) from 233 tasks (129 train, 104 test), each retargeted onto 16 robots: 102.3 hours of robot data in total.

robot	robot_type	state dims	datasets	episodes	frames	hours
xarm7_allegro	xarm7	14	233	2,199	702,372	6.5
panda	panda	14	233	2,200	702,843	6.5
fr3	fr3	14	233	2,200	702,843	6.5
ur5e	ur5e	12	233	2,200	702,843	6.5
ur10e	ur10e	12	233	2,200	702,843	6.5
sawyer	sawyer	14	233	2,200	702,843	6.5
gen3	gen3	14	233	2,200	702,843	6.5
iiwa	iiwa	14	233	2,200	702,843	6.5
jaco	jaco	12	233	2,200	702,843	6.5
arx_l5	arx_l5	12	233	2,175	680,286	6.3
viperx	viperx	12	232	2,171	686,044	6.4
widowx	widowx	12	232	2,114	650,533	6.0
piper	piper	12	231	2,144	657,112	6.1
yam	yam	12	232	2,156	665,456	6.2
g1_inspire	g1	26	232	2,140	681,864	6.3
h3_2_rh56e2	h3_2	38	233	2,200	702,843	6.5
total			3,722	34,899	11,049,254	102.3

Robot hours are below 6.5 where some episodes could not be placed (see below). The 14 single-arm robots are used as a pair, one per human arm; g1_inspire and h3_2_rh56e2 are humanoids. The table excludes the 5 older datasets described under Episode selection.

An episode is left out of a (robot, task) dataset when no base placement keeps that whole episode within the robot's reach, even after widening the placement search. 284 episodes are left out this way across all robots, mostly for widowx, g1_inspire, piper and yam; each dataset lists its own in meta/skipped_episodes.json. One pair, piper/test/sweep_dustpan, has no placeable episode and no dataset.

Episodes placed only by the widened search can track the human loosely; filter on meta/retargeting_errors.npz where that matters.

Layout

<robot>/<split>/<robot>__<task>_v0.1/
    data/chunk-000/file-000.parquet
    videos/observation.images.camera/chunk-000/file-000.mp4
    meta/info.json, stats.json, tasks.parquet, episodes/
    meta/embodiment.json             joint names per arm, canonical-vector conventions
    meta/retargeting_errors.npz      per-frame IK error
    meta/robot_solve_trace.parquet   per-frame named qpos, base pose, camera-to-robot transform
_human_skeleton/<split>/<task>.parquet   69-joint human skeleton, one copy per task

<robot>/<split>/<robot>__<task>_set2_v0.1/      Set 2, same layout as above
_human_skeleton/<split>/<task>_set2.parquet

meta/embodiment.json field human_skeleton holds the build server's absolute path; the file is _human_skeleton/<split>/<task>.parquet in this repo.

Features

feature	shape	content
observation.state	robot joints	right arm, then left arm
action	robot joints	observation.state of the next frame (absolute joint targets)
observation.ee_pose.{right,left}_wrist	7	IK target from the human wrist: x y z qw qx qy qz
observation.gripper_pose.{right,left}_gripper	8	gripper pose and width from the human hand keypoints: x y z qw qx qy qz width
observation.canonical_state, canonical_action	80	embodiment-agnostic vector; layout and mask in meta/embodiment.json
observation.images.camera	1080×1920×3	robot composited into the scene, AV1

• Paired single-arm robots use the same joint names for both arms; the first half of the state is the right arm.
• Arm state and action carry no gripper dimension. Gripper width is observation.gripper_pose.*[7], in metres.
• Humanoid state includes the hand joints (g1_inspire 6 per hand, h3_2_rh56e2 12 per hand) and no waist or legs.
• observation.ee_pose.* is the target the IK aimed at, not forward kinematics of observation.state. The gap per frame is in meta/retargeting_errors.npz.

meta/retargeting_errors.npz holds columns and data with one row per frame: episode, frame, pos_err_m, ori_err_deg, base_feasible_frac, degenerate_frac.

Measured trajectory quality

Measured on 153–194 datasets per robot (630–810 episodes each), part-way through the build:

robot	frames with wrist error > 5 cm	episodes with a joint step > 0.5 rad between frames
ur10e	1.1%	72%
ur5e	2.5%	72%
sawyer	3.5%	28%
xarm7_allegro	5.9%	23%
iiwa	6.9%	41%
gen3	9.1%	43%
panda	9.7%	39%
fr3	12.9%	43%
viperx	13.4%	39%
g1_inspire	13.6%	91%
widowx	21.3%	44%
jaco	46.4%	58%
h3_2_rh56e2	47.1%	99%
piper	57.4%	52%
yam	58.0%	70%
arx_l5	58.6%	68%

• Joint steps above 0.5 rad affect 0.3–3.9% of frames depending on the robot. On the humanoids they are mostly in the hand joints.
• All joint positions are within the robots' URDF limits; there are no NaN values.
• Joint velocities at 30 fps exceed URDF velocity limits on 0–75% of frames depending on the robot.

Episode selection

Episodes are those listed as batch2 and batch3 in batches.json of the depth repo, minus 8 whose hand-erased frames were partial or missing. batch3 is the single task train/add_remove_lid_v2.1 (1,156 episodes).

The train/*__fold_unfold_paper_origami_v0.1 datasets of g1_inspire, piper, viperx, widowx and yam were uploaded earlier from a separate build. They hold different episodes, use an older layout (human skeleton inside the dataset, no observation.gripper_pose), and were left unchanged.

Set 2

Set 2 is a second, disjoint set of EgoDex episodes for the same tasks, retargeted the same way onto the 14 single-arm robots. The two humanoids are not included. Its datasets carry _set2 before the version in their name, for example ur5e/test/ur5e__add_remove_lid_set2_v0.1, and use the same features and metadata files as the first set.

	episodes	frames	human hours	tasks
batch 1 of the depth repo, not retargeted before	659	282,240	2.6	239
episodes not listed in batches.json of the depth repo	3,774	1,057,472	9.8	224
total	4,433	1,339,712	12.4	240

• Set 2 covers 240 tasks (129 train, 111 test). Seven test tasks are new and not in the first set: flip_coin, fry_bread, fry_egg, set_up_clean_up_chessboard, wash_fruit, wash_kitchen_dishes and wash_put_away_dishes.
• 23 episodes were left out because their hand-erased frames were partial or missing.
• As in the first set, an episode no base placement can reach is left out of that robot's dataset and listed in meta/skipped_episodes.json.
• Depth for Set 2 is in the set2/ folder of VR-VLA/VR-egodex-depth-cache-full.

Status (2026-10-09): the build is paused. 2,724 of the 3,360 (robot, task) datasets are on the Hub. The 45 tasks with the longest episodes, 630 datasets, are not built yet. Six (robot, task) pairs have no episode any base placement can reach, so they have no dataset: test/fry_bread for arx_l5, piper, widowx and yam, and test/wash_fruit for piper and yam. The per-robot table of episodes and hours will be added here when the build is complete.