Training data for humanoid manipulation

Human hands, machine-readable.

Digitus Labs builds position and pressure mapping gloves, every joint angle and every newton of grip is recorded straight off your hand. The firmware exports episodes in the format your training stack can read, avoiding occlusion, increasing accuracy, and overall making training data more clean, accurate, and precise.

Force
0.1–0 N
calibrated fingertip pressure, normal + shear
Position
0 DoF @ 0 Hz
every finger joint, measured on the hand
Sync
<0 ms
force and position on one clock, verified per episode
Calibration
0 s
put the glove on and start recording

Video shows the task. Force and position finish it.

Videos get robots most of the way there, but they fail at contact. The failure modes are clustered at grasp, where fingers are hidden behind the object and where the camera has no way to know how hard the hand is pressing.

“Video alone cannot capture the rich contact signals critical for mastering manipulation.” — Meta FAIR
washing up, egocentric · fingers behind dishes for most of the task · footage: EPIC-KITCHENS · accuracy figures:

What each kind of data actually captures

Every way robots learn manipulation today, side by side. Only one of them measures force.

Data source Finger position Grip force During contact
(fingers hidden)
Cost to scale
Internet & egocentric videoGuessed from pixelsNoneAccuracy halvesCheap, scrapable
Vision hand-trackingEstimatedNoneFails at the graspCheap
Robot teleoperationRobot jointsRare, uncalibratedYesA robot per operator
SimulationPerfect but syntheticNo validated contact physicsYesNear-free
Digitus gloveMeasured · 26 DoF @ 120 HzCalibrated · 0.1–100 NUnaffectedHuman-speed, no robot

● measured / holds up  ·  ◐ partial or estimated  ·  ○ missing or breaks

fragile-object handling·sub-5 mm insertion·in-hand reorientation·cable & connector routing·fabric manipulation·tool use·bimanual assembly·food & produce·

Robots can see okay, but not well. And they can't sense at all.

Humanoids learn by watching video, that's why one can pour coffee in a demo and still crush a strawberry in actual practice. It has no idea how hard it's gripping. Touch is the other half of manipulation, and there is no internet to scrape that knowledge from. Someone has to record it from real human hands, and that's the gap we aim to address.

Language models had the entire internet to train off of, but robots have near nothing. We're building the training data for the physical world, and every glove shipped adds to it.

For consumers. Labs, use the form below.

The first production run goes to research partners; a consumer edition with the same sensing comes after, for creators, VR, and telepresence. This waitlist is for consumers only — if you're a lab, use the partner form in the next section.

You're on the list. We'll email you when pre-orders open.

Tell us where your policies drop the egg.

We work directly with humanoid and foundation-model labs and with academic research groups. Tell us the hand you're training and the format you need. You'll have a draft LOI and a quote within two business days.

Industry LOI: reserve gloves from the first production run at locked pricing.

Academic partnership: research pricing, free pipeline software, co-authorship.

Consumer edition: not yet. Join the waitlist above and we'll email you first.

Your email draft is ready. Send it and we'll reply within two business days.