A LeRobot v3.0 dataset of egocentric bimanual manipulation from human demonstrations, featuring world-frame 3D hand poses, full MANO hand meshes, camera trajectories, and dense depth maps.

Designed for training Vision-Language-Action (VLA) models, world models, and imitation learning policies from human hand demonstrations.

Key Features

Feature	Description
World-frame hand poses	6 keypoints per hand (wrist + 5 fingertips) in metric 3D world coordinates
Full MANO mesh	778 vertices per hand per frame for detailed hand shape
Bimanual actions	40D world-frame delta actions (both hands + gripper openness)
Camera tracking	6-DOF camera-to-world SE(3) poses
Dense depth	Per-frame metric depth maps + colorized depth video
Gripper proxy	Thumb-index distance as grasp openness signal

Dataset Summary

Property	Value
Episodes	2
Total Frames	514
FPS	10 Hz
State Dim	40 (world-frame bimanual hand)
Action Dim	40 (world-frame hand delta)
Hand Model	MANO (778 vertices, 1538 faces)
Camera	Egocentric (head-mounted)
RGB Resolution	480 x 640
Depth Resolution	384 x 512

Episodes

#	Task	Frames	Duration
0	Tidy the desk by organizing and rearranging items	220	22s
1	Fold and tidy clothes on the table	294	29.4s

State and Action Representation

observation.state / action — 40D float32, world frame
┌──────────── Left Hand (20D) ──────────────┐
│  [0:3]   wrist xyz (meters)               │
│  [3:6]   thumb tip xyz                    │
│  [6:9]   index finger tip xyz             │
│  [9:12]  middle finger tip xyz            │
│  [12:15] ring finger tip xyz              │
│  [15:18] pinky finger tip xyz             │
│  [18]    gripper openness (thumb-index m) │
│  [19]    detection confidence [0,1]       │
├──────────── Right Hand (20D) ─────────────┤
│  [20:38] (same layout as left)            │
│  [38]    gripper openness                 │
│  [39]    detection confidence             │
└───────────────────────────────────────────┘
action[t] = state[t+1] - state[t]

Quick Start

import pyarrow.parquet as pq
import numpy as np

data = pq.read_table("data/chunk-000/file-000.parquet")

# World-frame hand state and action
state = np.array(data["observation.state"][50].as_py())   # (40,)
action = np.array(data["action"][50].as_py())             # (40,)

# World-frame hand keypoints (6 points x 3D per hand)
left_kps = np.array(data["observation.hand_left_keypoints_world"][50].as_py()).reshape(6, 3)
right_kps = np.array(data["observation.hand_right_keypoints_world"][50].as_py()).reshape(6, 3)

# Camera-to-world pose
c2w = np.array(data["observation.camera_pose"][50].as_py()).reshape(4, 4)

# Full MANO mesh vertices (778 x 3)
mesh = np.array(data["observation.hand_left_vertices_camera"][50].as_py()).reshape(778, 3)

All Features (28 columns)

Column	Shape	Frame	Description
`observation.state`	(40,)	World	Bimanual hand pose
`action`	(40,)	World	Hand pose delta
`observation.camera_pose`	(16,)	World	Flattened c2w 4x4 SE(3)
`observation.camera_intrinsics`	(4,)	Pixels	[fx, fy, cx, cy]
`observation.camera_translation`	(3,)	World	Camera position (m)
`observation.camera_rotation_quat`	(4,)	World	Quaternion (w,x,y,z)
`observation.hand_{L/R}_keypoints_world`	(18,)	World	6 keypoints x 3D
`observation.hand_{L/R}_keypoints_camera`	(18,)	Camera	6 keypoints local
`observation.hand_{L/R}_vertices_camera`	(2334,)	Camera	Full MANO mesh
`observation.hand_{L/R}_cam_t`	(3,)	Camera	Hand 3D translation
`observation.hand_{L/R}_bbox`	(4,)	Pixels	[x1,y1,x2,y2]
`observation.hand_{L/R}_gripper`	(1,)	World	Thumb-index distance
`observation.hand_{L/R}_confidence`	(1,)	—	Detection score
`observation.hand_{L/R}_depth`	(1,)	Camera	Depth at hand (m)
`episode_index`	int64	—	Episode ID
`frame_index`	int64	—	Frame within episode
`timestamp`	float32	—	Time (seconds)
`index`	int64	—	Global frame index
`task_index`	int64	—	Task ID
`next.done`	bool	—	Episode end flag

Coordinate Frames

All hand keypoints in observation.state and hand_{L/R}_keypoints_world are in a fixed world frame (Y-down OpenCV convention). Hand mesh vertices in hand_{L/R}_vertices_camera are in the hand detector's local camera frame. To reconstruct the full mesh in 3D camera coordinates: mesh_3d = vertices + cam_t.

World-frame projection: 2D hand keypoints are projected onto the depth map, backprojected to 3D using camera intrinsics, then transformed to world frame via the camera-to-world pose.

Data Quality

Metric	Value
Left hand detection rate	98.6%
Right hand detection rate	100%
Wrist-to-camera distance	0.3 - 1.0 m
Gripper openness range	0.00 - 0.31 m
Action magnitude (mean)	0.035 m/frame

Citation

@misc{egoworld2026,
  title={EgoWorld: Egocentric Bimanual Manipulation with 3D World-Frame Hand Poses},
  author={Haoyang Li},
  year={2026},
  note={LeRobot v3.0 format}
}

Downloads last month: 49

Total file size:

35.1 MB