1 episodes · 30 fps · 3 cameras · 1600×1300
BaseMatrix EGO Desktop Demo
Human hand manipulation demonstrations captured with binocular egocentric cameras + exocentric view, processed through a 6-stage pipeline into structured observations for robot learning.
This is a preview dataset. Full-scale datasets with hundreds of episodes are available upon request — contact info@basematrix.ai.
Quick Start
from lerobot.common.datasets.lerobot_dataset import LeRobotDataset
dataset = LeRobotDataset("basematrix/ego-desktop-demo")
print(dataset)
print(dataset[0].keys())
Dataset Summary
| Property | Value |
|---|---|
| Format | LeRobot v3.0 |
| Episodes | 1 |
| Frames | 2,114 |
| Duration | ~70 seconds |
| FPS | 30 |
| Scene | Desktop object manipulation |
| Capture device | Orbbec EGO binocular (stereo fisheye, 1600x1300) + iPhone EXO (1920x1080) |
| Time sync | Hardware PTS (< 40 us L-R stereo) + audio beep sync (< 3.3 ms EXO) |
Features (24 channels)
Observation — Vision (3 video streams)
| Feature | Shape | Description |
|---|---|---|
observation.images.ego_left |
(3, 1300, 1600) | Left stereo fisheye, 30fps H.264 |
observation.images.ego_right |
(3, 1300, 1600) | Right stereo fisheye, 30fps H.264 |
observation.images.exo_iphone |
(3, 1080, 1920) | Exocentric iPhone view, 30fps |
Observation — State (11 channels)
| Feature | Shape | Description |
|---|---|---|
observation.state.camera_pose |
(7,) | ORB-SLAM3 stereo-inertial 6DoF: tx, ty, tz, qx, qy, qz, qw |
observation.state.left_wrist |
(3,) | Left wrist 3D position in world frame (meters) |
observation.state.right_wrist |
(3,) | Right wrist 3D position in world frame (meters) |
observation.state.left_keypoints_3d |
(63,) | Left hand 21 joints x 3 coords (MANO topology, world frame) |
observation.state.right_keypoints_3d |
(63,) | Right hand 21 joints x 3 coords (MANO topology, world frame) |
observation.state.left_contact |
(1,) | Left hand contact state (0/1) |
observation.state.right_contact |
(1,) | Right hand contact state (0/1) |
observation.state.left_force |
(3,) | Reserved — force/torque (zeros, sensor not connected) |
observation.state.right_force |
(3,) | Reserved — force/torque (zeros, sensor not connected) |
observation.state.left_tactile |
(48,) | Reserved — tactile array (zeros, sensor not connected) |
observation.state.right_tactile |
(48,) | Reserved — tactile array (zeros, sensor not connected) |
observation.state.gripper_force |
(2,) | Reserved — gripper force (zeros, sensor not connected) |
Action (4 channels)
| Feature | Shape | Description |
|---|---|---|
action.left_wrist_delta |
(3,) | Left wrist displacement (dx, dy, dz in meters) |
action.right_wrist_delta |
(3,) | Right wrist displacement (dx, dy, dz in meters) |
action.left_grasp |
(1,) | Left hand grasp signal (0 = open, 1 = grasping) |
action.right_grasp |
(1,) | Right hand grasp signal (0 = open, 1 = grasping) |
Processing Pipeline
Each recording goes through 6 automated stages:
| Stage | Tool | Output |
|---|---|---|
| L0 | Audio beep detection (1kHz) | Multi-camera time alignment |
| L1 | Nearest-neighbor PTS matching | Stereo frame pairs + EXO sync |
| L2 | ORB-SLAM3 (stereo-inertial) | 6DoF camera trajectory |
| L3 | Hands23 + WiLoR + HaWoR | Contact detection, 3D hand skeleton, world-frame wrist trajectory |
| L4 | GPT-4.1 Vision | Semantic annotation (objects, phases, sub-activities) |
| L5 | Cross-layer verification (24 checks) | Quality score + dataset card |
| L6 | LeRobot v3.0 packaging | This dataset |
Hand Tracking Stack
Three complementary models provide different levels of hand understanding:
- Hands23 (HOI-DETR): Hand-object interaction detection — bounding boxes, contact state, grasp classification
- WiLoR-mini: Single-frame 3D hand pose — 21 MANO keypoints per hand
- HaWoR: Temporal hand-in-world reconstruction — globally consistent wrist trajectories in ORB-SLAM3's world coordinate frame
Retargeting
This dataset provides human-frame observations (hand skeleton + wrist trajectory). To use with a specific robot embodiment:
- For parallel-jaw grippers (Franka, ALOHA, UR5): Map
left/right_wrist+left/right_graspto end-effector pose + gripper open/close via inverse kinematics - For dexterous hands (LEAP, Allegro, Shadow): Map
left/right_keypoints_3d(21 MANO joints) to robot finger joint angles via fingertip-position optimization
The action.{left,right}_wrist_delta features can serve as a starting point for action-space definition, though most VLA pipelines will want to define actions in robot joint space after retargeting.
Coordinate System
- World frame: ORB-SLAM3 stereo-inertial SLAM coordinate system. Origin at first tracked frame. Right-handed, Y-up.
- Hand keypoints: Relative to wrist (wrist = origin), following MANO joint topology (0=wrist, 1-4=thumb, 5-8=index, 9-12=middle, 13-16=ring, 17-20=pinky).
- Camera pose: Position (tx, ty, tz) in meters + orientation as unit quaternion (qx, qy, qz, qw).
Capture Hardware
- Head-mounted: Orbbec Femto Mega (EGO configuration) — 140g, dual global-shutter fisheye, 1000Hz IMU, 120mm stereo baseline, H.264 encoding
- Exocentric: iPhone — standard wide lens, 1080p 30fps, audio beep sync
Citation
@misc{basematrix2026ego,
title={BaseMatrix EGO Desktop Demo: Binocular Egocentric Hand Manipulation Dataset},
author={BaseMatrix},
year={2026},
url={https://huggingface.co/datasets/basematrix/ego-desktop-demo}
}
License
CC BY-NC 4.0 — free for research and evaluation. Contact info@basematrix.ai for commercial licensing.
Contact
- Website: basematrix.ai
- Email: info@basematrix.ai
- Pipeline demo: basematrix.ai/pipeline-showcase
- Downloads last month
- 81