Physical Intelligence

Embodied data, collected by people who do the task.

Egocentric video, depth, teleoperation and handheld-gripper demonstrations — recruited, consented and QA-reported on the same bench that has shipped four thousand hours of speech. The people on camera are doing work they actually do.

Start a pilot
Early access · first programs being scoped

Perit has shipped 4,000+ hours of speech through this bench. Physical-AI capture runs on the same recruiting, consent and QA machinery, and the first collections are being scoped with early-access partners now. Nothing on these pages is a delivered volume — the first hours will be listed here the way our speech hours are, after they ship.

Four kinds of capture

Pick the signal the policy needs.

Human video for breadth, depth for geometry, teleoperation when the embodiment has to match, handheld grippers for grasps in rooms a robot has never seen. Hover a card to see the capture.

Egocentric

First-person RGB video of real people doing real tasks.

You receive: raw video per episode, original device capture kept

Depth

RGB-D, stereo and LiDAR capture for 3D scene understanding.

You receive: rgb video and aligned depth per episode

Teleoperation

Human-guided robot demonstrations, one episode at a time.

You receive: synchronised camera streams per episode

UMI Gripper

In-the-wild wrist-gripper demos for grasp generalisation.

You receive: fisheye video per episode
Environments

Rooms where the task actually happens.

Tasks are recorded where they actually happen — not in a studio dressed to look like a kitchen. Hover a room to see the kind of setting the protocol calls for.

Kitchens
Warehouses and light industrial
Offices
Retail floors and small shops
KitchensBedrooms and living roomsBathroomsOfficesRetail floors and small shopsWarehouses and light industrialWorkshops and labsCustom or proprietary sites, sourced to spec
Same bench, same machinery

The record is in speech. The process is not.

Recruiting to a locale, consent on the file, a calibration gate before paid work, QA sampling and a report per batch — none of it cares whether the file is audio or a first-person video.

4,000+
Hours of audio delivered
650
Transcribers and aligners on the bench
50 hrs
Delivered every working day
95%+
QA-sample accuracy, checked against WER

These are speech figures. Physical-AI hours will be listed separately once the first collections ship.

How a pilot runs

Small first, on purpose.

Ten to thirty episodes before anything scales — because the protocol always breaks somewhere, and it should break cheaply.

  1. 01
    Brief

    The task family, the environments, the rig, the sensors, the episode count. One page, agreed before anything is bought or anyone is recruited.

    day 0
  2. 02
    SOP and kit

    A written capture protocol — framing, lighting, where a task starts and ends, what counts as a failed take — and the kit list. Operators are trained and tested on it before a single episode is paid for.

    week 1
  3. 03
    Pilot episodes

    A small batch, ten to thirty episodes, delivered with QC status and a quality report. This is where the protocol breaks — on purpose, and cheaply.

    week 2
  4. 04
    Review and scale

    You review the pilot, we fix the protocol, then the run scales on a weekly delivery cadence with the same report attached to every batch.

    ongoing
With every episode

What ships alongside the capture.

  • Task and environment metadata per episode
  • QC status and reviewer chain per episode
  • Consent reference on every file
  • Quality report per batch: accepted, rejected, and why
  • Weekly drops to your bucket, region pinned at kickoff

Tell us the task, the room and the rig.

One page is enough to scope a pilot. If the protocol is wrong, the pilot is where we find out — before the volume.

Start a pilot