Pegaso
Send a spec

Your robot has never worked a shift at a factory.

We've recorded in factories, kitchens and workshops: 197 workplaces across South America so far. Pegaso turns that kind of footage into multimodal datasets for embodied AI, captured to the spec your team writes.

Send us a spec

Someone does their job. We record what their hands see.

A baker filling empanadas. A mechanic backing out a seized bolt. Our operators wear the rig you pick and work the way they always do, with consent signed before the first frame.

Every sensor runs on one clock.

Video, IMU and audio share a single monotonic timebase, stamped per frame and per sample. When your model asks what the wrist was doing at frame 4,812, the answer is already in the file.

People check it. Then people label it.

Reviewers throw out sessions with dropped frames or a camera pointed at the ceiling. What survives gets timestamped action segments, English narrations and hand keypoints, in your taxonomy if you have one.

The rest is your robot's job.

You get data your training stack reads on day one: your schema, your splits, a dataset card, and the consent trail behind every clip.

Everything between a person at work and a file your model can read.

Hire us for one piece of it, or hand us the whole program.

  1. 01

    Custom datasets

    Multimodal data built to your spec. You choose tasks, environments, rigs, hours and label schema. We run the operation and deliver in your format.

    Video · IMU · Audio · Labels
  2. 02

    Capture rigs

    Mono, multicam, stereo or wrist, alone or combined.

    4 rigs
  3. 03

    QA and annotation staff

    A trained bench of reviewers you can bring into your own pipeline, on your tools and your guidelines.

    Staffing
  4. 04

    Capture app

    The iOS app our operators record with, licensed to teams that collect their own data.

    iOS
  5. 05

    Field operations

    Recruiting, training, devices, consent and logistics. Based in Argentina, expanding across South America project by project.

    AR · SA

Mono, multicam, stereo, wrist.

Four ways to put cameras on a person. Tell us what your policy needs to see and we'll fit the rig to it.

Mono

One head-mounted camera at eye line. The most hours per dollar, and the view most egocentric models are trained on.

Streams
1 RGB + IMU
Best for
Volume, task diversity

Multicam

A head camera plus two or three fixed cameras around the workspace, for the context the wearer can't see.

Streams
3 to 4 RGB
Best for
Full-body context, occlusion

Stereo

A calibrated pair at eye line for depth. The calibration travels with the data, session by session.

Streams
2 RGB, synced
Best for
Depth, 3D hand pose

Wristcam

One camera on each wrist, close to the grasp, plus one on the head. The wrists see what the head loses behind the fingers.

Streams
3 RGB + IMU
Best for
Contact, fine manipulation

197

Workplaces recorded so far.

Real workplaces in Argentina, recorded between November 2025 and August 2026.

What the operator sees.

Clips from real sessions, compressed for the web. Full-resolution samples come with their sensor files, on request.

Filling empanadasPOV · Food
Polishing a car doorPOV · Auto
Screen printing bagsPOV · Textile
Assembling a framePOV · Craft
Measuring steel stockPOV · Metal
Bagging dried fruit at the counterPOV · Retail

Borrow our reviewers.

We trained a large bench of QA reviewers for egocentric data. They check sync, framing, task compliance and label quality against written guidelines, and they're starting on our own sessions now. Your operation can use them too, with a lead on our side who answers for the numbers.

Every frame starts in our app.

Our operators record with an iOS app we built for this work. It stamps each frame and each IMU sample against the same clock, keeps the camera's capabilities with the session, writes a receipt when the upload lands, and even notes why a recording stopped. If your team needs to collect data in-house, we license it and set it up for your protocol.

Ask about licensing →

A year in the field.

Pegaso is run by the people who spent the last year running egocentric capture for physical AI programs. We recruited and trained hundreds of operators, wrote the recording protocols, shipped phones to other provinces and answered for every hour a client rejected.

Consent comes first. Every operator signs before recording, and every clip traces back to a person, a place and a date.

Start with a paid pilot

  1. Send a spec. A paragraph also works: we'll write the spec with you.
  2. We record a pilot batch in environments that fit the task.
  3. You get it with its QA report and run it through your evals.
  4. We scale whatever moved your numbers.

Send us a spec.

A full document or three lines both work. We reply with samples that match and a proposal.

contact@pegaso.ar
What do you need?

By sending this form you agree that we use these details to reply to you, as set out in our privacy policy.