Physical AI

Physical AI data,captured where the task happens.

Robots and embodied models learn from people doing real work. Saolabs captures that work in kitchens, factories, warehouses, roads and simulations, then labels it with the same vetted experts behind our model data.

01

Egocentric video

First-person video of people doing real tasks

02

Depth and 3D

RGB-D, stereo and LiDAR capture of real scenes

03

Teleoperation

Human-guided robot demonstrations, episode by episode

04

Video and sensor annotation

Segmentation, captions, masks and success labels

05

Bring your own rig

Your hardware and protocol, our crews and QA

Who does the capture

Capture is done by people who know the task. A line cook records cooking, a warehouse picker records picking, a technician records repairs. People who do the work move the way the work is really done, and that is what a policy needs to learn.

Operators are briefed on a written protocol for each task, and every one of them signs consent before recording. People who appear on camera without being part of the task are blurred or kept out of frame.

Before an episode counts

Every episode is checked before it is delivered. An episode that fails a check is re-shot or rejected, and you are not billed for it.

  • Framing: the task and hands stay in view
  • Exposure, focus and motion blur within agreed limits
  • Sync across cameras and sensors
  • Task compliance: the episode follows the protocol from start to end
  • Metadata complete: setting, device, operator and timing
  • Privacy: faces and personal details handled as agreed
  • Labels agree with a gold set before a batch ships

How it is delivered

Episodes are delivered to your cloud storage in the region you choose, in your file layout and schema. Each episode carries its metadata, consent reference and review history.

Questions, answered.

How does a physical AI pilot start?

Tell us the task, the setting and the hardware. Within 48 hours we send a written spec and quote. The pilot is a small paid batch with a QA report, so you can judge the data before committing to volume.

Can you capture in a specific setting or on our hardware?

Yes. Tell us the setting, the task and the rig. We can run capture on your hardware and protocol, or source the devices and settings ourselves.

Do you capture simulation and gameplay data?

Yes. Operators can play missions and tasks in a simulator or game while their screen and inputs are recorded, for agents that learn from human play.

Who owns the data?

Ownership and licence terms are set in the order form for each programme. Commissioned data is not resold to other customers.

Every safe model
has an expert behind it.