CYBERORIGIN — Ground Truth for Embodied Intelligence

Data Annotation Capture Research Careers Request access

First-person capture of real human work

123,000+ verified hours

339,000+ clips · 2.42M+ annotations

aggregated & de-identified · capture-quality snapshot · 2026-06

Try CyberCode→ Data viewer↗ Get in touch

The library

Indexed by task, not by folder.

Every clip sits in an ontology — occupation × scene × procedure. Pick an industry and pull exactly the slice you need, down to the bench it was shot at.

30s

11s

30s

CNet ontology · 2,660 occupations · 299 scenes · 19,147 processes

Open the data viewer →

The annotation

One clip, taken apart.

Every object the hands touch, tracked and named. Every motion placed in 3D. Every segment written out — and the reason it followed.

objects · tracked · namedDragon fruit, foam-net bag, picking tongs — boxed per frame, not per clip.

3D pose · object topologyBody, head and hands in 3D — plus how the objects relate to each other.

action · environment · execution · inferenceFour causality layers, so a model learns why a motion followed.

plain-language segmentsWritten out on a shared timeline. Delivered LeRobot v3 / RLDS.

Capture

Hands and body, recovered in 3D.

The hand mesh placed back into a real point cloud, and full-body pose across two scenes — all recovered off ordinary first-person capture.

Hand tracking · 2D keypoints + mesh in point cloud · 1:11

Full-body pose · outdoor · input · mesh · skeleton · 0:33

Full-body pose · workshop · input · mesh · skeleton · 1:33

Egocentric capture · synchronized view · 0:30

Research

Research and open tools.

37 ★ on GitHub OpenCyberGlove SDK Capture-glove SDK, open-sourced. Python API, ROS2 wrappers, Inspire-hand retargeting. Open the repo ↗ arXiv:2609.34817 · 2026 ESTHER: Egocentric Stereo Hand Estimation and Reconstruction in the Wild Metric 3D hand reconstruction from wearable stereo, designed for real-world egocentric capture. By Hongyu Ma, Hairong Qu, Shiqi Zhao, Yongsong Yang, and Peng Yin. Read the paper ↗

Tell us what you're training.

What tasks, what environments, what keeps failing. We'll tell you whether we already have it.

The physical world was never written down.
We are writing it.

The origin point of machine reality

We're hiring Shenzhen · Singapore · remote considered · [email protected]

Building the capture layer nobody else owns. These are open.

01Data engineeringPipeline · storage · delivery

02Capture standardsProtocol · fleet ops · field

03Quality systemsAuto QC · human review

…3D perception, and moreWrite to the address above