Skip to content

announced a strategic investment and partnership with Arabic.ai.

Read the announcement
Talk to our team
First-person view of gloved hands slicing tomatoes at a restaurant prep station
First-person view of a suture simulation on a training pad
First-person view of gloved hands tightening screws in an electrical control panel
First-person view of hands with a wrench on an engine mount bolt

Physical AI

Teach machines the physical world.

First-person demonstration data for physical AI: video, hand motion and task labels on one clock, recorded by specialists, delivered rights-clean.

Sample episode · 001

See one hour of Arabic.AI data.

One real task, recorded first-person: the video, both hands and every step on a single clock. Drag the timeline to explore it. Ask our team about the full hour.

Video
Left hand
Right hand
Steps
00:00.00
Video1080p · 30 fps · 110° FOV, plus a chest camera
MotionWrist IMU on both hands · 100 Hz
SyncOne episode clock · frame-aligned (≤ 33 ms)
LabelsSteps · verbs · objects · tools · hands; boxes on request
FormatsLeRobot · RLDS · MCAP
RightsSigned release · consent ID per clip · faces blurred

Every label, box and trace under this clip is timed to its own frames. The hand traces are measured from the video. These specs are an example: final specs are agreed for each project, recorded to your task script.

Filmed where the work happens.

Real sites with releases in place: factories, warehouses, restaurant kitchens, stores, workshops and simulation labs. Your task script decides what is performed; our capture teams decide nothing but the camera.

Factory
Warehouse
Restaurant kitchen
Retail
Auto repair
Healthcare sim
First-person view, factoryFirst-person view, warehouseFirst-person view, restaurant kitchenFirst-person view, retailFirst-person view, auto repairFirst-person view, healthcare sim
FactoryAssembly, fit and fastening at the bench: alignment, torque, inspection.01 / 06
  1. Factory Assembly, fit and fastening at the bench: alignment, torque, inspection.
  2. Warehouse Picking, scanning, packing and putting away across racking and docks.
  3. Restaurant kitchen Prep, line cooking and plating at service speed.
  4. Retail Shelf stocking, facing and price checks on the shop floor.
  5. Auto repair Diagnosis and repair in the engine bay, with hand tools and torque.
  6. Healthcare sim Suturing, dressing and instrument handling on training models.

Every frame is data.

What the model sees

Outside the glass, a mechanic tightening a bolt. Inside it, what we deliver: the hands, the tool, the part and the step, labelled on every frame.

Move the glass · click it to change its shape · lens

From task script to dataset.

A 300-hour pilot across three environments, delivered in your format with a datasheet and consent records.

A first-person view of gloved hands in an electrical control panel
01

Task script

Read the step
A certified electrician wearing smart glasses at a distribution panel
02

Specialists and consent

Read the step
A technician wearing smart glasses, a chest camera and wrist sensors
03

Capture

Read the step
First-person view of hands with a wrench on an engine
04

Review and annotation

Read the step
First-person view of gloved hands slicing tomatoes in a kitchen
05

Delivery

Read the step

Why Arabic.AI

What makes the data usable.

Select a slice to open it.

Specialists, not a crowd.

Anyone can wear a camera. Few can wire a panel or close a wound. We pick people who do the task for a living and brief them on your script.

  1. Specialists, not a crowd. Anyone can wear a camera. Few can wire a panel or close a wound. We pick people who do the task for a living and brief them on your script.
  2. Filmed where the work happens. Real sites with releases in place: factories, kitchens, warehouses and workshops, not a studio set.
  3. Motion and intent, not just pixels. The wrist trace and the step labels ride on the video’s clock, so a model sees what the hand did and why.
  4. Labelled, frame by frame. Steps, hands, tools and objects on one clock, with boxes and keypoints on annotated projects.
  5. Rights-clean before it ships. Signed releases, blurred faces and a consent ID on every clip. Ready for your legal review on day one.
  6. Delivered in your format. LeRobot, RLDS or MCAP, with checksums, a datasheet and the consent records in every drop.

Three tiers, one pipeline.

Consent and rights.

Physical AI is part of Arabic.AI Data Services: collection, labeling and curation under one consent record.

With every clip

  1. A signed contributor releaseIn the contributor’s own language, naming the use: training machine-learning models, with the right to sublicense to your organisation.
  2. A site releaseFor every workshop, kitchen, store or lab we film in.
  3. A consent IDIn the metadata sidecar, so any clip traces to its paperwork and can be withdrawn if a contributor asks.
  4. A datasheetWhat was collected, where, under which script, with known gaps stated.

What we never do

  • Film anyone under 18, or anyone who has not signed.
  • Scrape, buy or resell footage we did not record ourselves.
  • Film in public spaces without a site release.
  • Reuse a dataset collected under an exclusive agreement for another customer.

Questions labs ask.

Who owns the data?

You do, under the licence in your agreement. Exclusive collection is available: a dataset recorded under an exclusive agreement is never reused for another customer.

Can we send our own task script and format?

Yes, and we prefer it. Send the tasks, the camera placement, the label taxonomy and the file format you train on; we return a task script for approval and deliver in your format.

How is hand motion captured?

A wrist IMU on each hand at 100 Hz, timestamped on the same clock as the glasses and the chest camera, delivered raw as CSV. Finger-level pose comes on annotated projects as estimated keypoints from the video, marked as estimated.

How long does a pilot take?

It depends on your task script, the environments and the hours you need. Whether it's a 300-hour pilot across three environments or a different scope, we give you a timeline once the scope is agreed with our team.

Do you offer teleoperation data?

Not yet at scale. Teleoperation rigs for robot-arm demonstrations are planned next, after pilots on the three tiers above. If your lab needs it, tell us and we will plan the rig around your platform.

How is a project quoted?

By tier, on your task script. Talk to our team for a quote, including a 50-hour starter dataset.

With every delivery

Three promises in one folder.

00 / 03Open the folder.

Each file inside is something that ships with the data. Click one to read it.

  1. One clock Video, motion and labels timestamped together, so a frame, a wrist reading and a step name always agree.
  2. Your spec, our capture Task script, label taxonomy and file format are yours; the kit, the calibration and the people are ours.
  3. Rights, registered Every clip carries a consent ID that traces to a signed release. Nothing ships that cannot be traced.

Start with a 50-hour dataset.

Recorded to your task script and scoped with our team.

Figures are confirmed per task script.

Smart glasses with a camera at the hinge beside a compact drive, lit violet from behind

This page does not set marketing or analytics cookies, so there is nothing to change here.