# World Archive

> Robots have no internet-scale corpus to learn from. World Archive captures multi-modal egocentric data of skilled work across India, Uzbekistan, Türkiye and Romania — stereo RGB-D, dual wrist cameras, tactile gloves — then annotates, verifies and delivers it ready to train on, and helps teams measure whether it improved their policy.

World Archive builds the global data infrastructure for physical AI: the pipeline that turns real human work into robot training data. Not a footage marketplace — access, sync and calibration discipline, annotation throughput, and one delivery contract. Services: field capture, annotation & QA, and advisory (all available now); evaluation (in development); simulation (on the roadmap). Purpose-built capture rigs, sub-frame synchronisation, human-verified labels, consent-first collection, delivery in LeRobot / Hugging Face / MCAP.

## Start here

- [Homepage](https://worldarchive.co/): why policies plateau, what we do, the improvement loop, real footage, deliverables, coverage, process
- [Hardware & reference rig](https://worldarchive.co/hardware): stereo heads, wrist cameras, tactile gloves, sync protocol, rig specs
- [Data pack contract](https://worldarchive.co/data-pack): what ships in a delivery, L0–L3
- [Evaluation protocol](https://worldarchive.co/eval): how usefulness is measured
- [Privacy & governance](https://worldarchive.co/privacy): consent, PII handling, usage rights

## Sample data (inspect before commissioning)

- [Capability show](https://ggn-egocentric-data-sample.s3.ap-south-1.amazonaws.com/capability/index.html): Ego+Exo, Ego+IMU, Stereo, Stereo+Tactile capture demos
- [Density sample · Hugging Face](https://huggingface.co/datasets/WorldArchive/mono-india-workplace-sample): 9 clips, 8 annotation layers
- [LeRobot export](https://huggingface.co/datasets/WorldArchive/mono-india-workplace-lerobot): train-ready episodes
- [Interactive explorer](https://worldarchive.co/explorer): switch between modalities on real episodes
- [World Archive on Hugging Face](https://huggingface.co/WorldArchive): org hub

## Technical essays

- [Blog index](https://worldarchive.co/blog)
- [Annotation density advantage](https://worldarchive.co/blog/annotation-density): labels per minute vs raw video hours
- [Beyond the monocular plateau](https://worldarchive.co/blog/future-of-physical-ai-dataops): multi-modal DataOps for physical AI

## What we deliver

- L0 synced streams — time-aligned video from every camera, plus IMU and tactile
- L1 calibration & timing — intrinsics, extrinsics, per-frame timestamps
- L2 hands & objects — 21-point hand pose per frame, object boxes with persistent track IDs
- L3 actions & contact — verb–noun segments with timestamps, grasp/release, contact states
- Plus DATACARD, consent metadata, QA report, SHA256 manifests
- Formats: LeRobot, Hugging Face Datasets, MCAP, MP4, JSONL; RLDS/OXE on request

## Capture stack

- Stereo RGB-D head (ZED X Mini class), dual wrist RGB, IMU, optional tactile gloves
- Sub-frame synchronisation (clap board, dynamic QR, PTP); calibration re-checked hourly
- Purpose-built Z-kit field rigs, 50+ deployed
- Environments: factory floors, skilled trades, retail and service, craft, care, household
- Capacity: pilots typically 100–500 hrs/week, ramping toward ~2,000 hrs/week

## How engagements work

1. Scope — founder-led session on the behaviour, environment, modalities and success criteria
2. Pilot — a small scoped pack delivered against the contract, with QA report and eval hooks
3. Review — label quality, coverage and policy results assessed together
4. Scale — ramp with the protocol locked, exclusive or non-exclusive

## Contact

- Email: info@worldarchive.co
- Founder call: https://calendly.com/osman-worldarchive/30min

## Optional

- [llms-full.txt](https://worldarchive.co/llms-full.txt): extended notes for coding agents