Physical AI needs a billion hours of human hands.

Handtrace will pay people everywhere to film everyday tasks first-person, and our engine turns every clip into data that's ready to train. We open to filmers and labs soon.

▶Watch the 2:31 demothe real product and engine ✋Try the tracker on your handslive, in your browser

Every skill a robot needs already lives in a human hand.

Humans are the most capable machines on earth. Handtrace captures how they do it, first-person. Every tile below is a first-person view clip run through our engine, both hands tracked joint by joint. This is the kind of footage that becomes robot training data.

Wash a glassPOV

Type on a keyboardPOV

Wrap a giftPOV

Fold paperPOV

Chop vegetablesPOV

Free-licence stock clips, traced by the Handtrace engine: 21 joints per hand, every frame.

Physical AI is learning from almost nothing.

Robots learn from human demonstrations, and almost none of them are on video. Robot companies already spend $100M+ a year buying real-world data.

MIT Technology Review, Apr 2026.

First-person view of hands wrapping a notebook, both hands tracked joint by joint by the Handtrace engine.

Labs post the jobs. Everyone else films them.

People anywhere pick up jobs with a phone and the Handtrace Band, and get paid for every clip that passes. The engine checks and labels each one, and labs receive data that's ready to train.

Launching first · drag to turn · a goal map, nothing on it is live yet

Where we need hands.

Robots that will work in homes everywhere have to learn from hands everywhere. We launch everywhere at once: anyone with a phone, in any country, can film. In places like India a Handtrace job pays about double the local rate for this work.

Sample jobs, the kind labs will post. No lab has posted yet.

Every clip, passed or failed, is checked.

Real results from our engine on first-person view stock clips.

Labs post a job and a price per accepted clip

Filmers claim a batch of 10 and film it first-person

The engine checks every clip, then a vision model confirms the task

Accepted clips pay the filmer. Rejected ones cost the lab nothing

Labs download LeRobot episodes, ready to train

We're building the largest library of human skill on earth, filmed by everyone, for the machines that will learn it.

Mecka

Sells real-world robot-training data to labs

~$100M/yr (Sacra)

Handtrace

Open to anyone with a phone
Paid for every clip that passes our checks
Every clip checked by software

Film for an hour. We keep the parts that matter.

Filmers don't have to stop and start. Record a whole session and our cutter keeps only the stretches where your hands are working, then checks each one like any other clip. It's built the way our Picture Lock engine cuts raw footage down to the good parts.

0:00A real test recording, 3:103:10

190.7 s

recorded, with idle time between four tasks

65.8 s

kept, as four clips, one per task

3.6×

faster than real time on one laptop CPU

idle stretches left in

One clip in. A ready-to-train episode out.

HP-01

21 joints × 2 handstracked every frame

ST-02

Step 1/3 · lift

PASS

Every check passedhands 100% · sharp 94% · framed 79%

Episode fh_7f4eb4e2b753

LeRobot · first-person

Wash a glass

203 frames · 24 fps · 8.46 s

Pass

liftgraspplace

0.002.794.338.46 s

Hands2 × 21 joints

Two-hand180 / 203

Steps3, labelled

TrackerMediaPipe

AudioDropped

FormatLeRobot

Duration8.5 s

Hands visible100%

Sharp94%

Framed79%

Filmer paidevery check passed

"t": 2.50, "side": "Right", "score": 0.972,
"landmarks": [[0.1933, 0.7102, 0.0],
  [0.2577, 0.7079, -0.0737],
  [0.3248, 0.6115, -0.0922], ...]

Real engine output on a first-person view clip

The same clip, as a ready-to-train episode

Every clip, measured by software.

Wrap a giftboth hands · 181 frames · real tracked joints

Episode: fh_f25dc97b351b
Tracker: MediaPipe hand landmarker, 21 joints per hand
Frames: 181 at 25 fps · 7.24 s
Checks: Length7.2 srule 5 to 180 s

Hands in view100%rule 60% or more

Sharpness84%rule 70% or more

Framing79%rule 70% or more

Resolution506 pxrule 480 px or more
Steps: wrap 0.0 to 0.9 s · hold 0.9 to 4.0 s · release 4.0 to 7.2 s
Privacy: audio dropped · faces blurred automatically
Verdict: PassThe filmer is paid. The lab gets a LeRobot episode.

joints tracked per hand, every frame

first-person view clips run through the engine so far

4 of 17

passed the engine's quality checks. The rest failed for real reasons: blur, cropped fingers, low resolution.

0.84–0.99

hand-tracking confidence, on every clip

Two ways in. One engine checks both.

First-person view clip of wrapping a notebook, with the engine's hand trace.

Anyone, anywhere, films.

A phone and the Handtrace Band are all it takes. Every clip goes through the same checks before a lab sees it.

First-person view clip of washing a glass, with the engine's hand trace.

Tested on stock footage first.

Free-licence stock clips run through the same engine. It is how we test the checks before any filmer joins.

Every clip, passed or failed, tunes the checker.

Every clip, passed or failed, is stored as a labelled example we use to tune the checker. The more clips we run, the better it gets at telling a good clip from a bad one.

1
Clips come in
Passed and failed.
2
Every verdict is stored
As a labelled example.
3
We tune the checker
On those examples.

Every hand leaves a trace.

The engine runs today. We open to filmers and labs soon. Join the list to be first in.

You're on the list.We'll email you when early access opens.

Robot lab? Talk to us