Touchstone
Robots learned to see from a billion photographs. Nothing, ever, recorded how hard a hand pressed.
Touchstone is an instrument, not a dataset vendor. A person works normally at a bench; the bench, the tools and the objects measure the contact; a band on the forearm reads the fingers from the muscles that move them. What comes out is not footage. It is force in newtons at a known point on a known object at a known instant, carrying a certificate that says how it was obtained and how far it can be trusted.
And because it stores the job the object experienced rather than the motion a particular hand made, one demonstration can be recompiled onto a body that looks nothing like yours.
- measured
- newtons at the contact, not inferred from pixels
- natural
- nothing on the hand, so the grip stays uncorrupted
- portable
- one demonstration compiles to many robot bodies
- certified
- every number traceable back to a reference standard
Force is invisible.
That is the entire problem.
A frame of someone holding a cup and a frame of someone crushing it are the same picture — right up until the moment it is too late. Every manipulation failure that embarrasses robots today is a force failure: the crushed egg, the dropped glass, the connector that never seats. None of it is visible in video, so none of it is in the training data.
Identical pixels. Seventeen times the force. Only one of these is a demonstration worth learning from, and nothing in the image says which.
One pass, end to end.
The whole system is one idea carried out carefully: do not instrument the human, and do not instrument the robot. Instrument the world. The table, the tools and the objects already touch everything that matters, and a band on the forearm covers the one thing they cannot see: what the fingers are doing while the object hides them.
- 01HUMANA person simply works. No glove, nothing on the hand.
- 02WORLD-SIDE CAPTUREThe bench, the tools and the objects do the measuring.load · contact onset · object pose
- 03FOREARM CAPTUREThe fingers are read from the muscles that actually move them.finger tension
- 04PHYSICS REFEREEEvery channel must satisfy one model of the hand. Disagreement is rejected, not averaged.newtons at the contact
- 05CONTACT RECORDWhere, how hard, in which direction, and exactly when.
- 06TOUCHSTONECalibration lineage, drift and per-field uncertainty sealed onto the record.certified
- 07COMPILE TO A BODYOne demonstration, recompiled per robot, with an honest verdict when it cannot be done.
- 02 · the benchThe work surface, the hand tools and a set of prepared objects each report what is happening to them. One channel is exact but slow, another is fast but relative; together they pin down both how hard and precisely when. Everything is timestamped against a common reference, because unsynchronised measurement is not measurement.
- 03 · the bandFingers contain almost no muscle. The muscles that drive them sit in the forearm, out of the way of whatever is being held, which is exactly why we read them there.
- 04 · the refereeNothing is averaged. Each channel is required to be consistent with a single physical model of the hand, and a reading that cannot be reconciled is flagged and set aside rather than blended into the result. Sensors that begin to drift are caught by this, not discovered months later.
- 06 · the sealA record only counts if someone else can trust it. Every one carries its calibration lineage, when it was last verified, how far it has drifted, and, field by field, whether a number was measured, distilled, inferred or simulated.
Cheap on purpose.
The industry collects contact data by buying a robot that costs as much as a car, bolting delicate sensors to it, and paying someone to drive it with a joystick. It is slow, it is fragile, and the data it produces fits exactly one robot. A Touchstone station is built from parts you could put on a credit card, and the person using it needs no training. They have been an expert at manipulation since they were one year old.
The table does the measuring.
When you lift a mug off the surface, the surface watches the weight leave it, which is precisely the moment your grip takes over. The tool in your hand knows how hard you are pushing while you are actually working. The object itself can report how hard it is being squeezed. Cameras give every measured force a precise address in three dimensions.
None of it touches the demonstrator. That is the point: the instant you put a sensor on a fingertip, you have changed the grip you were trying to record. Every glove-based approach in this market pays that price without saying so.
Read the fingers where nothing can hide them.
Your fingers are puppets. The muscles that pull them live in the forearm, connected by tendons, so the forearm knows what the hand is doing even when the hand is completely out of sight. Which matters, because the hand is most hidden exactly when force matters most: mid-grip, wrapped around the object.
Most wearables in this space chase the faint electrical signature of muscle activity. We deliberately did not. The band is built on the signals that predict finger force best and identify which finger best, and it is calibrated by the bench itself, in seconds, against forces the bench has already measured.
Don’t copy the dance.
Copy the job.
Your hand has twenty-odd joints and soft skin; a common gripper has two stiff fingers. Copying finger motion between them is hopeless. This is the embodiment gap, and it is why most human demonstration data dies with the body that recorded it.
The mug does not care that you used five fingers. It cares that it was pressed at these points, this hard, in these directions, over this time. Store that, and you have written the job description rather than the choreography. And a job description can be handed to any hand, claw or jaw. Select one below.
compile targetanthropomorphic robot hand
- Force closure
- satisfied
- Required twist
- within capacity
- Margin
- 1.8×
ReadingNear-direct transfer. The job description barely has to be re-solved.
- workers vs restersA camera-only system sees five fingers on the cup and must assume all five matter. Measured force shows that two carried the load and three were along for the ride. Knowing which is which is the difference between a transferable plan and a guess.
- the tripod factA camera tripod stands on three legs. With friction, roughly three well-chosen contacts hold most objects steady, which is exactly why a three-finger robot can do almost anything a hand can, if it presses the right places by the right amount.
- an honest noWhen no arrangement of a given robot’s fingers can reproduce the measured job, we say so and say why, then propose the workaround a person would use, whether that is fixturing it, re-gripping it or picking up a tool, then re-checks that plan before offering it. A trustworthy refusal is worth more than a confident failure.
Why the name.
You trust your kitchen scale because its calibration traces, link by link, back to a reference standard. That discipline is called metrology, and it is centuries old. It has simply never been applied to robot training data. A hundred labs each holding a hundred hours in their own format, with their own sensors and no way to prove whose numbers are right, is a hundred small piles. Every leap in modern AI came from one large pile.
So every record leaves the station with a passport, and stations are re-verified against a reference on a schedule. Drift gets caught, not discovered. This is the part that makes data from somebody else’s bench safe to train on, and it is the layer nobody in this market has built.
TOUCHSTONE RECORD
TS-0114-0296 · specimen
- Station
- ST-004, field class certified
- Lineage
- ST-004 → REF-01 → national standard traceable
- Last verification
- 6 h ago, pass logged
- Drift since
- within band logged
- Contact points
- 5 on the scanned surface measured
- Normal force
- 00000000 measured
- Tangential force
- 0000000000 measured
- Contact onset
- sub-millisecond, synchronised measured
- Finger tension
- 0000000000 distilled
- Friction coefficient
- from observed slip at known load inferred
- Object pose
- 6-DoF against scanned mesh measured
- the referenceOne station is the reference kilogram of the network. Every other station’s numbers are only as good as their chain back to it, and that chain is written down.
- the ritualEach station periodically performs a short scripted routine on a standard object. If its answers stop matching what physics says they must be, the station is quarantined until it is re-certified.
- honest labelsMeasured, distilled, inferred, simulated. A user always knows which they are looking at. Honest uncertainty beats confident nonsense, and it is the only basis on which two labs will ever pool anything.
- it improves backwardsEvery accidental slip at a known load teaches the network how slippery that material really is. Old recordings get more accurate as the library grows.
What everyone else is missing.
Human demonstration data became a procurement category the moment the frontier labs proved it transfers to robots at scale. Every source they can buy today is force-blind, force-corrupted, or locked to the robot that recorded it.
| human video | teleoperation | glove / mocap | touchstone | |
|---|---|---|---|---|
| uncorrupted natural grip | ● | ○ | ○ | ● |
| force in newtons at the contact | ○ | ◐ | ◐ | ● |
| contact onset and texture | ○ | ◐ | ◐ | ● |
| portable to other robot bodies | ◐ | ○ | ◐ | ● |
| traceable to a reference | ○ | ◐ | ○ | ● |
| cost per demonstration hour | ◐ | ○ | ◐ | ● |
The published anchor for contact-rich robot data, and it cannot be offshored, because latency destroys precision exactly where precision matters.
Natural-hand capture has no latency budget, so it can be collected anywhere. The force channel is included, not extra.
Because we store the job rather than the motion, an hour collected today still compiles onto the hand that ships next year.
This is the shape of the instrument, not its internals.
Sensor topology, the estimator that turns the forearm signal into per-finger newtons, the certification mathematics, the contact-space format and the compiler are documented — elsewhere. If you are a lab that needs force-rich manipulation data, a builder who wants to certify a rig, or an investor who would like to see the falsification plan and the numbers behind it, the technical dossier is available under NDA.
request the dossier- sheet
- 01 / system overview
- revision
- 0.1
- stage
- MVP-0 · in progress
- drawn by
- Physcript
- status
- concept