Contact data for physical AI
Manipulation models train on pixels and joint angles. Neither records the signal that decides whether the glass lifts or shatters: how hard the hand squeezed, and how fast it corrected when the object began to slip. We build the instrument that records it.
The whole argument, in 2.4 seconds
Somebody picks up a glass and it starts to slide. Watch what each sensor has to say about the moment it happens. Drag the trace to scrub.
The glass begins to slide at the marked instant. Grip force collapses, then the corrective squeeze lands about seventy milliseconds later. The pose track moves under a millimetre, which is inside the noise floor of every hand tracker there is. A camera can watch this happen and record nothing about it.
The bottleneck
Everyone is recording video. Nobody records force.
Contact is the least accurate part of every physics engine. Friction, deformation and slip are approximated, so teleoperating a simulated arm yields a clean trajectory through an interaction that never had to obey a real surface. That is the sim to real gap, and it is a contact gap.
You can estimate a hand's pose from footage to within a few millimetres. You cannot estimate the newtons it applied, or the correction it made forty milliseconds after the object shifted in its grip. That correction is the entire skill, and it is invisible.
Force sensing gloves exist, bolted to a bench in a handful of research groups, producing hours. A model needs years. Nobody has put contact sensing on a hand that walks out of the building and into a kitchen.
The instrument
A wrist unit worn while you do ordinary work. Four printed parts, a removable bay cap, and a camera that comes off with a quarter turn.
An open C that springs over the wrist, so there is no buckle to thread one handed. The bore is an ellipse because a wrist is: model it round and it rocks, digs in at the styloid, and reads its sensors through a moving air gap.
A lip drops into the bay mouth on a 0.25 mm printed snap fit, and a thumb notch on the proximal edge gives you somewhere to lever it off without taking a glove off first. Vented, because the bay holds a radio and a battery.
Three lugs, drop in and twist 60 degrees to a hard stop. A bayonet rather than a thread because it is one handed, tool free, and cannot back itself out under vibration, which matters when the entire point is that somebody wears it while actually working. It looks down the hand, tilted 20 degrees palmar, so the fingers and whatever they hold are in frame.
Reaches round to the mound at the base of the thumb on a deliberately thin arm. The arm has to flex, because the mound changes shape as the thumb opposes, and a rigid one would lift the pad off at exactly the moment the grip it is measuring begins.
Vision, inertial and contact are sampled against a shared timebase in the device, not aligned in software afterwards. A correction that lands seventy milliseconds after a slip is only legible if the two streams agree on when now was.
Geometry is watertight and audited: every bayonet check is proven to fail on a build that is not twisted home, or it does not count as a check. The Band is in design and not yet a product you can buy.
The proof
A wearable sensor is a hardware problem long before it is a data problem. Hotaru is the evidence we take a device from parameters to a shipped object, in public.
A desk robot with four servos, a microphone and a speaker. It leans toward whoever is talking and folds down when the room goes quiet. Twenty two printed parts, every one on a 180 mm bed, CAD open in pure Python with public interface audits.
The engine
Every failure a buyer reports becomes a capture target we pay for. The network gets more valuable the more it is used, which is the only kind of data business worth building.
Our own device, so we control what is sensed and how it is synchronised. Vision, inertial and contact on one clock.
Contributors record real tasks in real kitchens and workshops. Consent is captured with the data, not asserted about it later.
Each accepted contribution is written on chain with its terms and its author. Provenance and payment are the same record.
Buyers take slices by task, object and material. Their failures come back to us as the next thing we pay to record.
| Approach | What it captures | What it misses | Scale |
|---|---|---|---|
| Browser teleoperation of simulated arms | Trajectories, object poses, task metadata | Real friction, deformation, slip | Very high |
| Scraped and egocentric video | Hand pose, scene, intent | All force, all correction | Very high |
| Lab force rigs | Contact, richly | Everyday tasks, variety, volume | Hours |
| Thenar | Vision, inertial and contact force, one clock | Whole body motion, for now | Network |
Why the name. The thenar eminence is the mound of muscle at the base of your thumb. Its job is opposition, swinging the thumb across to meet each fingertip, and modulating how hard it presses once it arrives. It is the anatomy that separates a hand from a claw, and it is the exact capability robots have not learned. The mark is that arc.
Honest status
Hotaru is real hardware with open CAD, printable STLs, and interface audits anyone can run.
The Thenar Band is a specification and a sensor selection, not a product you can buy. We are building it around the first dataset we sell rather than guessing at a spec first.
The contributor network is the model we are building toward, not something you can join this week. We would rather say that here than have you find it out after signing something.