We film real work so robots can learn it.

First-person task data for robotics, captured by our own workforce in Bali, Mexico, China, Brazil and Denmark.

Robots learn physical work by watching people do it. We record that work from the worker's own point of view, in real hotels, kitchens, workshops and warehouses, with consent and rights documented for every clip. Our aim is the largest consented library of real work in the world, filmed by the people who do it for a living.

What people do
What robots learn
Built for humanoids

Real work, filmed where it happens.

Two hands smoothing a white bed sheet, seen from above

Hospitality

Rooms, linen, turndown

Commercial kitchens

Prep, cooking, dishwashing

Close-up of hands soldering a green circuit board

Light manufacturing

Assembly, soldering, sewing

Warehouse and packing

Picking, packing, labelling

Gloved hands wiping a desk with a cloth and spray bottle, seen from above

Homes

Cleaning, laundry, tidying

A hand stacking tomatoes in a shop display

Retail

Restocking, shelving, unpacking

Maintenance and trades

Repairs, tools, installation

Your task list

Send the tasks, objects and edge cases your model is missing.

Tell us

Footage a model can read.

We frame every task so hands, objects and contact stay in view. That is what pose and action models need.

318
labelled steps from one 46 minute housekeeping session
Keypoints per hand
21, 2D and 3D
Step names
Checked by a person
Format
LeRobot v3 or yours
Consent
Signed, AI training
Recorded to one client's brief. We record to yours.

Real ROCKnWOOL footage. Hand keypoints are drawn from the dataset labels.

What ships

Clips, metadata, consent and a QA report in every batch. Labels, pose and sensors when you need them.

1080p–4KResolution
30–60 fpsFrame rate
Head or chestCamera mount
Wide lensHands stay in frame
IMU + depthSynced, on request
Your rigOr ours
MP4 + JSONLeRobotHDF5RLDSMCAPWebDatasetYour schema

Data spec

Our defaults. Send your own spec and we record to it.

Four checks on every clip

Rejected clips are not billed. The acceptance rate ships with every batch.

Sample pack

20 labelled housekeeping steps in LeRobot v3. Load them before you sign anything.

Request a sample pack

More than footage.

Five ways to work with us today. Four more we scope on request.

01

One task, five countries

The same protocol recorded in every location. Different tools, rooms and habits in one dataset.

02

Real sites for robot teams

Working hotels, villas and kitchens for capture, pilots and evaluation. Owner approval in writing.

03

Failure and recovery

Scripted slips, drops and retries, each one time-stamped. The cases natural footage rarely contains.

First person
Third person
04

Two views of the work

The worker's view and the room's view, for teams training world models.

05

Non-exclusive datasets

Common tasks recorded once and licensed to several buyers, at a lower price per hour.

On request

Closer to the robot

Scoped and priced per project
  • Depth, motion and hand poseDepth and motion sensors synced to the video, with checked 3D hand pose.
  • Handheld gripper crewsOur operators work with your handheld grippers, in real rooms.
  • Teleoperation operatorsTrained people who pilot your robots remotely, in shifts.
  • Whole-body and touchBody tracking and tactile gloves for humanoid programmes.
Ask about a capability

Sample first. Then a pilot. Then volume.

  1. Sample

    Real clips in your format, to load before you commit.

  2. Protocol

    Tasks, device, metadata and acceptance criteria, agreed in writing.

  3. Pilot

    A small trained crew records a first batch. You judge real footage.

  4. Production

    More operators and more sites, supervised on location. Delivered in batches.

Danish-run, with a workforce on four continents.

We recruit through hotel, villa and restaurant networks in every location. A project starts with a small trained crew and can grow to thousands of operators, all supervised in person.

Work with us
Who owns the footage?

You choose. Footage can be exclusive to you, or non-exclusive at a lower price. It is agreed per project and written into the contract.

Can we see a sample first?

Yes. Ask for a sample pack: real clips with metadata, labels and hand pose in your format, to load into your pipeline before you commit.

What do you film with?

Head-mounted phones and action cameras are the usual setup, at 1080p to 4K and 30 to 60 frames per second. If your model needs specific hardware, operators use the device you specify or ship to us.

What do we receive?

Video with metadata for every clip: task, location, operator ID and consent record, plus a QA report for every batch. MP4 and JSON by default, or LeRobot, HDF5, RLDS, MCAP or your own schema. On request we add task labels, action timestamps and hand pose data.

How do you check quality?

Every clip passes four checks: on the device, on upload, by a human reviewer and against earlier footage for duplicates. Rejected clips are not billed, and the acceptance rate is reported with every batch.

Is this egocentric video data?

Yes. Every clip is egocentric video: first-person footage of a real person doing real work. It is the training data robotics and physical AI teams use for humanoids, manipulation and embodied AI models.

Which kinds of work do you cover?

Hospitality, homes, commercial kitchens, light manufacturing, warehouse and packing, retail, and maintenance and trades. For anything else, tell us the task and the industry. We find people who do that work for a living and record them, on your equipment or ours.

How fast can you scale?

The pilot crew is small on purpose. Once the data is right, we add operators through our hospitality networks, from tens to thousands, with team leads on site.

What does it cost?

We price per accepted hour. The rate depends on the task, the sensors and the rights. Tell us your volume and budget and we scope the project to fit.

How do we start?

Ask for a sample pack, or send the tasks and the spec. We reply with a pilot proposal covering scope, timeline and price.

What is your model missing?

Ask for a sample pack or a pilot proposal.

hello@rocknwool.com
WhatsApp +45 53 57 72 17