Robot learning data
Every skill starts as a human frame.
Egocentric demonstrations of skilled professional work, captured in real workplaces with consent built in, and delivered as training-ready datasets.
The problem
Robots need structure, not more raw video
Robot foundation models need human data from real environments, with labels and task structure. Buyers report that much of the data on the market fails quality checks: missing calibration, unusable hand tracking, no task structure, unclear consent.
Most available capture comes from private homes. Professional service work, such as hotel housekeeping, laundries and commercial kitchens, is largely unrecorded, and that is where service robots will work first.
What we are building
Skilled work, captured where it happens
We capture egocentric video of skilled professional work in real workplaces, with employer and worker consent, and turn it into training-ready datasets.
Our first focus is hospitality (hotel housekeeping, laundries, kitchens) in Serbia and the Western Balkans. Other environments, such as manufacturing, retail, logistics and facility cleaning, on demand.
Humanframe is built by Orange Hill, a software and AI engineering company: we adapt schema, labels and delivery to your model and embodiment.
At a glance
- Real workplaceshotels and hospitality first, any industry on demand
- Consent by designvoluntary, paid, never used to evaluate workers
- Spec to sampleabout 4 weeks from agreed spec, in LeRobot, RLDS or your schema
How it works
From a shift in a hotel to a dataset
1 · Capture
Head-mounted wide-angle camera as standard; wrist cameras and motion sensors on request. Natural work during regular shifts, plus optional short structured sessions.
2 · Processing
Synchronized streams, camera trajectory and 3D hand pose with per-joint confidence. Faces, guest belongings, screens and documents blurred before data leaves the site.
3 · Annotation
Task and sub-task segments checked by people, EPIC-KITCHENS-compatible action labels, success flags and a dataset card.
4 · Delivery
Standard open formats used by robot-learning teams (LeRobot, RLDS), or the schema your pipeline expects.
Working together
We don't sell a catalogue. We capture to your spec.
From agreed spec to first sample in about 4 weeks. Hotel and hospitality capture sites lined up in Serbia and the region.
- Week 0Spec call
Your robot, tasks, sensors and acceptance criteria.
- Weeks 1–3Capture
Recorded on a live hotel site, processed and annotated.
- Week 4Sample
Delivered in LeRobot format with a QA report.
- ThenProduction
Regular batches against the same spec.
| Spec & Sample | Production | Exclusive | |
|---|---|---|---|
| Purpose | Your tasks, sensors and acceptance criteria, captured on a real hotel site as a sample | Recurring delivery of processed hours | Your environments (any industry), tasks and sensors |
| Scope | A focused set of housekeeping and hospitality tasks | Volume by agreement | Scoped with you |
| Timeline | First sample in about 4 weeks from agreed spec | Regular batches | Phased plan |
| License | Evaluation | Non-exclusive | Exclusive, full IP assignment |
Spec & Sample
- Purpose
- Your tasks, sensors and acceptance criteria, captured on a real hotel site as a sample
- Scope
- A focused set of housekeeping and hospitality tasks
- Timeline
- First sample in about 4 weeks from agreed spec
- License
- Evaluation
Production
- Purpose
- Recurring delivery of processed hours
- Scope
- Volume by agreement
- Timeline
- Regular batches
- License
- Non-exclusive
Exclusive
- Purpose
- Your environments (any industry), tasks and sensors
- Scope
- Scoped with you
- Timeline
- Phased plan
- License
- Exclusive, full IP assignment
You set the acceptance criteria; we deliver against them. Every delivery includes a QA report: hand and tracking coverage, and agreement on a double-labelled sample.
Built for European compliance
Privacy is part of the product
- Recording is voluntary and paid for each worker and never used to evaluate their performance.
- Signed consent per worker and site, provenance per clip, impact assessment per location.
- Anonymization at the source; raw footage kept short-term, only anonymized datasets retained.
- GDPR-aligned processing; no recording in guest-occupied rooms, restrooms in use, changing or break rooms.
Why Serbia and the region
Close to you, real variety on site
- A large hospitality sector with real variety, plus access to manufacturing and service sites across the region.
- Same time zone as Germany and Austria, a short flight from your team.
- Competitive cost for processed, well-documented data.
- Operated by Orange Hill GmbH in Vienna, engineering team in Belgrade.