Teaching robots to feel.

First collection begins in Tamil Nadu, September 2026. Leave an address and we will send the first release when it is published.

Read the thesis

The thesis

India never finished automating. That is the asset.

The work still done by hand here survived because machines could not do it. Cloth, leather, wire, sheet brass. Deformable, contact rich, two handed, irregular. Not an accident of development. A filter, and it selected for exactly the problems robotics has left.

Our approach

A camera records where a hand went. It does not record how hard it pressed, when contact began, or whether the material slipped. We intend to measure that at the fingertip, a hundred times a second, and release it openly.

See the spec

Placeholder footage, illustrative only. Not our capture data.

Force and video, on one clock.

Load at grip0.00 N
Planned rate100 Hz
Threads taut0

Physics simulation, not sensor data. Drag the weave.

What is settled

Human video scales. Published work shows manipulation performance rising monotonically with hours of egocentric footage, and a million hours are now free to download.

What is not

Whether contact must be measured or can be estimated from video. Recent work argues both ways. We are running the cheap version of that experiment before building at scale, and we will publish the answer either way.

Fingertip normal forcerare10 ch · 100 HzCalibrated to newtons against known masses, shipped with a per-unit curve rather than raw ADC counts.
Egocentric RGB1080p · 30 fpsHead mounted. Fisheye intrinsics solved per device, Kannala-Brandt, four coefficients.
Exocentric RGB2 × 1080p · 30 fpsFixed at the bench, extrinsics solved per workstation, so ego and exo share a frame.
Hand orientation9-axis · 100 HzOne IMU per hand, onboard fusion, quaternion output.
Skill graderareper workerYears in trade, recorded on the same task at the same bench. Apprentice failures and recoveries kept, not filtered out.
Shear forcenot in v0Arguably what governs whether cloth slips. We do not have it yet and will say so plainly at release.

The rig

Reference design · nothing built yet

128° 1 2 3 4 5 6 Open frame, side view
  • 1Camera moduleMonocular, 1080p at 30 fps, 128 degree fisheye. Intrinsics calibrated per device and shipped with the footage.
  • 2Brow padLoad spread across the forehead. It has to be wearable for an eight hour shift or the data never gets collected.
  • 3Head strapRigid open frame with an elastic crown strap. No nose bridge, no prescription fitting, works over existing eyewear.
  • 4Recorder and batteryLocal microSD at roughly 1.6 GB per hour. Uploads overnight on factory wifi, never mid shift.
  • 5Port, status, kill switchUSB-C charge, one status light, one press to stop recording. That last part is why factory owners say yes.
  • 6Field of viewAngled down to hold both hands and the working surface of a bench in frame.

Fingertips stay open. A tailor who cannot feel fabric takes the glove off in ten minutes, and then there is no data. Shear is not measured in this version.

Target sites

Force-dominant trades · in approach

01

Garment construction

Mass Knit Garments

Velampalayam, Tiruppur, Tamil Nadu

Collar attach, hemming, gathering. The canonical deformable problem, done by people who do it four thousand times a week.

Prospective site · no agreement yet
02

Hand chasing, sheet brass

A. J. Brass Emporium

Rehmat Nagar, Moradabad, Uttar Pradesh

Sheet brass moves only under correct pressure. Video of hand chasing is close to useless on its own. Force is what makes it legible.

Prospective site · no agreement yet

Placeholder footage, illustrative only. Not our capture data.

Access

The first release will be open.

Apache 2.0, WebDataset format, RLDS on request. If you train manipulation policies and there is a material, task or force profile you need, tell us now and we will collect for it.

We reply to every lab within two working days.