[physcript]
TOUCHSTONEsheet 02 · the long sheet
← the atlascontact

TOUCHSTONE v3 · SHEET 02 · REV 0.2 · SPECIFIED, PARTLY VALIDATED, UNBUILT

Touchstone

the force layer for physical AImeasured · certified · portable

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; a sleeve on the forearm reads the muscles that move the hand; a handheld standard tells the sleeve what a newton is. What comes out is not footage. It is force in newtons on a known object at a known instant, with a stated interval around it — and because it records the job the object experienced rather than the motion one particular hand made, the same demonstration can be recompiled onto a body that looks nothing like yours.

This sheet is the long version. Everything freely shareable is here, including the parts that went wrong. Where a mechanism is deliberately held back, the sheet says so rather than glossing over it.

measured
newtons at the contact, not inferred from pixels
bare-handed
nothing on the palm or the fingertips
portable
one demonstration compiles onto many robot bodies
certified
every number traceable back to a physical reference
honest
and every number carries how far it can be trusted
01 — THE BLIND SPOT

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 it is too late. Every manipulation failure that embarrasses robots 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.

A tactile glove does not close the gap either. It gives a pressure pattern in relative units — rich spatial coverage, no newtons, and no stiffness channel at all. The missing thing is not resolution. It is the unit.

A correction we made to our own earlier position, and keep making in public: we used to say gloves corrupt the demonstration by construction. That was too absolute. Glove data can train force-sensitive policies, and at least one glove was engineered specifically to stay thin and preserve dexterity. Glove interference is a graded concern, not a disqualifier. Our real advantage over gloves is the stiffness channel and the calibration — not the bare fact of leaving the palm uncovered.

And underneath it, a second problem

A human hand has roughly twenty-three kinematic degrees of freedom and five deformable pads. A parallel jaw has two rigid jaws on one axis. Copying joint angles across that difference is close to meaningless, which is why retargeting research keeps fragmenting into one bespoke solution per robot.

The two problems have one answer. Stop recording the hand. Record what the hand did to the object, and how the arm was regulated while doing it. Force on the object does not care what applied it. Stiffness is a command any torque-controlled robot can render. Contact events are the subgoals of manipulation for any body at all. All three are measurable from the muscle side — and two of them only from the muscle side. That single fact is the entire strategic reason to build a wearable rather than anything else.

02 — THE INSTRUMENT

Three physical things, and nothing on the hand.

A forearm sleeve. A camera the customer already owns. A handheld calibration artifact. The palm, the fingers and the fingertips stay bare. Robots never wear any of it — the sleeve's signals are privileged training information, not a runtime dependency.

A compression sleeve rendered as a study model, with low-profile sensing modules seated along its length and a small electronics carrier on the dorsal side.
plate I · the sleeve, as a study modelconcept visualisation — not a photograph of finished hardware
The sleeve
Measurement equipment in the shape of performance apparel. Skin-contact sensing on the volar side, rigid mass on the dorsal side, one small status light and no display.
The camera
Customer-supplied and egocentric, exactly as the open capture rigs already do it. The Record has to be a strict superset of what a customer already collects, or nobody adopts it.
The artifact
A handheld object of known, stable, traceable properties. It is what makes every number on the sheet mean something in particular, and it is the reason this is metrology rather than a good estimate.
03 — FIVE BLOCKS

Fourteen ways to sense an arm, collapsed to five.

There are about fourteen credible ways to read what a forearm is doing. They are not fourteen separate choices — they are seven kinds of information, several of which more than one method can measure. Grouped that way, most of the fourteen either merge or fall out, and five blocks are left.

The selection rule mattered more than the list: kill a candidate only when another is a strict superset of its information. A naive keep-whatever-is-easiest rule would have killed ultrasound in favour of the cheap pressure ring — which would have been wrong, because pressure cannot see stiffness at all.

fig. 1 · what feeds whatcategories only · no layout, no componentsElectrode arraymyoelectric drive · co-contractionv0Ultrasound arraymuscle deformation · tissue stiffnessv1 · v3Tapper + accelerometerstendon tension · contact timingv0 · v2IMU clusterarm and wrist kinematicsv0FMG ringhow the sleeve was donnedv0Calibration artifactthe reference, in newtonsv0Egocentric cameracustomer-suppliedv0Kinematics200 HzWrench200 HzImpedance50 HzContact events±1 msUncertaintyper sampleprimary sourcecontributessupplies the referencecross-check only
Hover a row to isolate it. Note what the FMG ring is not wired to: it is capable of more, and it is deliberately kept out of the force path because its reading depends on how tightly the sleeve was put on. That dependence is a defect in most designs. Here it is the function.

And they arrive in a deliberate order

The cheapest, most-proven configuration ships first and answers the most commercially important question. Every later stage adds exactly one genuinely new kind of information, and each one sits behind an experiment that can stop it.

  1. v0
    electrodes · accelerometers · IMU · FMG · artifact · cameraforce, contact events, kinematics, uncertainty
    → K1
  2. v1
    A-mode ultrasound, dry silicone couplingfatigue-robust muscle deformation
    → K2
  3. v2
    the micro-tapper → tendon tensiondirect-physics tension; enables L1
    → K3
  4. v3
    ultrasound → imaging + elastographymeasured, not inferred, stiffness
    → K4
  5. v4
    electrodes → high-density gridmotor-unit-level neural drive
    → K5

Where each block sits on the arm, how the emitting ones are time-multiplexed so they do not step on each other, and what any of it costs per line are not published. Naming what kind of sensing exists is fine; handing over the layout is not.

04 — THE STANDARD

Why a squeezable object is a metrology instrument.

In metrology this is called a transfer standard: a physical object of known, stable, traceable properties, used to tie an instrument's readings to real units. Every serious measuring instrument has one. No robot-data-capture product ships one. That absence is the opening.

It does a second thing that matters more. The laboratory method for measuring limb impedance is perturbation: displace the limb suddenly and measure the restoring force. The instrument that does this costs tens of thousands of dollars. An object that can change its own stiffness while being gripped runs the same experiment in the user's hand — which is what makes the stiffness channel possible at all outside a lab.

What the object is made of, how it changes stiffness, and the range it can cover are held back. The purpose, the metrology framing and the fact that it changes stiffness at all are said freely — those are the parts worth being generous about.

fig. 2 · the chain of custody for a newton01Certified referenceexternal standard

a force reference with a certificate we did not write

02The artifactserialised · re-verified

a handheld transfer standard that ships with every sleeve

03Per-session check≈4 min per wearing

8–10 minutes once, then a short check each time the sleeve goes back on

04The Recordan interval per sample

the number, and how far it can be trusted, written together

05The robotnumbers with a provenance

and able to refuse a grasp its body cannot actually make

Cut any link and the sheet below it is still full of numbers — they simply stop meaning anything in particular. What you have then is relative muscle activity in arbitrary units, which is what everyone already has. Calibration does not remove error. It converts an unknown error into a stated one, and that is the whole of the promise.
05 — THE RECORD

The thing that is actually being sold.

Everything upstream exists to produce this. Everything downstream exists to consume it. If this is right and the hardware is mediocre, the product still works. If this is wrong, no amount of sensing rescues it.

01

Kinematics

200 Hz

hand + object pose, arm dynamics

commodity — everyone has this

02

Wrench

200 Hz

forces + torques induced ON the object, in the object's frame

object-frame wrench validated as an embodiment-agnostic space by CHORD (NVIDIA, 2026)

03

Impedance

50 Hz

task-space stiffness ellipsoid over time

tele-impedance and variable-impedance control, a mature literature

04

Contact events

±1 ms

touch · load · lift · hold · slip · replace · release

Johansson & Flanagan — contact events are the subgoals of manipulation

05

Uncertainty

per sample

a stated interval on every force and stiffness sample

conformal prediction — proven elsewhere, unclaimed here

Each field is independently validated in the literature. The combination exists nowhere. One well-known dataset has wrench but no impedance, no events and no uncertainty. Another has dense tactile patterns but no SI units and no stiffness channel. The forearm band that came closest has force and none of the rest.

demonstration:
  meta:
    subject_id, session_id, sleeve_serial, artifact_serial
    enrollment_ref, verification_ref, config, clock

  stream kinematics    @ 200 Hz  → SE(3) hand/object pose + arm state
  stream wrench        @ 200 Hz  → R^3 force, R^3 torque, object-fixed frame
  stream impedance     @  50 Hz  → R^3x3 stiffness (SPD), principal ellipsoid
  events  contacts     async     → touch | load | lift | hold | slip | replace | release
  stream quality       @   1 Hz  → donning_index, channel_health, drift_estimate

One design choice worth remembering

Wrench is recorded in the object's frame, not the hand's. Hand-frame forces are morphology-specific — a five-fingered grasp and a two-jaw grasp produce completely different hand-frame numbers for the same outcome. Object-frame wrench does not care who or what applied it. That is what embodiment-agnostic means concretely, and it is what makes the Record directly consumable by cross-embodiment methods.

And one commercial one

The schema will be published openly and permissively licensed. The value is not in the format — it is in being the only party who can fill it correctly. An open format recruits the ecosystem that makes the data valuable; a closed one guarantees competing alone against better-funded labs.

06 — THE COMPILER

Seven layers, scored on themselves.

Most of these use well-known techniques from other fields, and say so. Three are, as far as we have been able to find, unclaimed — and all three only work if you also own the hardware, which is what makes them hard to copy rather than merely hard to think of.

sourcesleeve streamsfront endL1 – L4the IRL5 · the Recordback endL6 – L7targetany robot bodywhoever defines the intermediate representation owns the ecosystem
  1. L1

    Physics-anchored fusion

    flag planted

    Our channels are not peers. One of them is grounded in direct physics; the others are statistical estimators. So the grounded one continuously corrects the rest, and calibration does not go stale during a shift.

    The loop itself, and the window it exploits, are held back pre-term-sheet.

    novelty
    9/10
    risk
    high
    position
    flag planted
  2. L2

    Conformal force intervals

    flag planted

    Turns a promise of quality into a number you can audit: an interval with a coverage level, from a method with distribution-free guarantees and no priors to argue about.

    Validated against public data before any of our own hardware existed — see the evidence below.

    novelty
    8/10
    risk
    medium
    position
    flag planted
  3. L3

    Hybrid physics-ML estimator

    Constrain the estimator with musculoskeletal forward dynamics so its outputs are physically realisable rather than merely plausible.

    Good engineering, mature literature, not a breakthrough. This layer is not where the moat lives.

    novelty
    4/10
    risk
    low
    position
    adopt
  4. L4

    Missing-modality robustness

    One software stack serves every hardware tier, because customers will own different configurations and nobody wants a model per SKU.

    Commercially essential rather than novel. Without it the staged roadmap multiplies engineering cost by five.

    novelty
    3/10
    risk
    low
    position
    adopt
  5. L5

    The Record — the IR

    flag planted

    The format itself is a layer. Sleeve streams are the source language, robot actions the target, and this sits in between.

    To be published openly. The value is not the format; it is being the only party who can fill it correctly.

    novelty
    8/10
    risk
    medium
    position
    standards play
  6. L6

    Uncertainty-weighted learning

    Error bars only earn their keep if they improve training. Downweight the uncertain samples, or hand the interval width to the model as a feature.

    Every existing method weights by demonstrator suboptimality. None weights by sensor-derived measurement uncertainty, because no dataset has ever carried it.

    novelty
    7/10
    risk
    medium
    position
    new — because the data is
  7. L7

    Privileged distillation

    Train with the sleeve's rich channels as privileged information; deploy on a robot that has a camera and nothing else.

    One sleeve-collected dataset upgrades an entire fleet of force-blind robots, permanently, with no hardware change on the robots.

    novelty
    3/10
    risk
    low
    position
    adopt — and it is the sales argument
07 — THE EVIDENCE

What we actually know, including where we were wrong.

Before any hardware existed, the uncertainty claim was tested on a public dataset — twenty subjects, 256-channel forearm recordings, per-finger force ground truth, seventy-four thousand windows. Every prediction was written down and frozen before a model was fitted. Then the write-up was handed to an external reviewer, who was right on every substantive point, and six claims were corrected or withdrawn.

This is the only part of this sheet with real evidence behind it. It is also the part most companies would not publish.

fig. 4 · predicted before the data was touched · observed afterbracket = frozen prediction · dot = what happenedWithin-subject coveragenominal was 90. It landed on 90.50100 %90.3Cross-subject coveragewe predicted our own method would fail here. It did not. The claim is retired.50100 %90.4Cross-subject accuracycoverage held by widening intervals 1.77×. Accuracy did not.00.8 R²0.233Unlabelled adaptation recoveryno ground truth needed — the user just has to wear the thing and move.0100 %54Reverse-session coveragerun backwards, the collapse is the same. It is re-donning, not the second day.50100 %60.8Recalibration gapwhat a labelled recalibration buys back, in coverage points.050 pts26.5
Two of these are our own predictions being wrong in public, and one of them is the prediction the earlier version of this company was built on. It was retired the day the data disagreed with it. Anyone still telling you cross-subject intervals silently under-cover is quoting a claim we withdrew.
the claim we retired

The premise this company was previously built on was that a population-calibrated interval would be too narrow for a new user and would silently under-cover. It does not. Coverage held at nominal; the method absorbed the shift by widening intervals 1.77× instead. Conformal prediction behaved correctly under a shift it was not guaranteed to survive. Any deck still asserting otherwise — including ours, for a while — is wrong.

what replaced it

Something narrower and better evidenced. Per-user calibration is worthless on average — it moves mean coverage by −0.5 points — and it removes the tail entirely: under population calibration 20% of subjects fall below 85% coverage; under per-user calibration none do. One user in five gets materially wrong error bars without it.

fig. 5 · take the sleeve off, put it back on tomorrowwhat was true — coverage against a nominal 90%nominal 90%90.3% · same day59.5% · next day, electrodes re-appliedwhat the model said — stated interval width, %MVC11.6911.69identical to two decimal placesthe errors exploded; the stated uncertainty did not move
One subject received intervals labelled 90% that covered 17.2% of the time. This is the exact failure this company exists to sell against, and we found it in our own pipeline — which is the only reason we can claim to know it is there.

We tried to kill our own product

If a smarter model could fix re-donning with zero burden on the user, the per-session ritual would not be justified and the calibration artifact would lose most of its rationale. So we pre-registered the falsification: if the field's standard fix — randomised channel masking during training — lifted cross-session coverage to 83.3% or better, the ritual was unjustified.

It reached 57.1%. It moved coverage the wrong way by 2.4 points, while buying an accuracy gain that does not approach usable. The artifact survived the hardest test available on public data — which is worth more precisely because the test was set up to go the other way.

Read honestly: channel masking is documented to help cross-day classification. It did not help cross-day regression here. That is a real, pre-registered negative result on the method the field would reach for first — on one dataset, with one model class.

And then we caught ourselves doing it

Coverage was being computed from hundreds of overlapping windows treated as independent. Windows are not exchangeable units. Trials are. Recomputed properly:

calibrationwindow-levelcertifiable at 90%
4 trials · 100 s86.6%0 of 20 subjects
9 trials · 225 s89.0%20 of 20 · intervals 2.15× wider

At four calibration trials the honest radius is infinite for every single subject — a 90% guarantee mathematically does not exist at that budget — while the comfortable-looking window-level number reads 86.6%. That gap, a figure that looks acceptable against a guarantee that is not there, is the exact failure mode Touchstone exists to sell against. We shipped it in our own write-up for a day. The corrected spec is longer and more expensive, and it is the one we quote.

The label on every claim

Every factual claim this company makes, internally and publicly, carries a label for how solid it actually is. The rule that gives it teeth: no hypothesis appears in customer-facing material until there is an experiment behind it and the label has changed. A promise to sell calibrated numbers with stated error is worthless from a company that will not hold itself to the same standard.

DDemonstrated
independently demonstrated in peer-reviewed work
CCompany-reported
claimed by a company, not independently verified
RResearch prototype
shown once in a lab, not productised
IInference
our own reasoning from evidence, not itself measured
HHypothesis
speculative — must be gated by an experiment
08 — THE GATES

Five ways to find out we are wrong, cheaply.

Each is a specific experiment with a written pass/fail line decided before it runs, and a stated plan for what we do if it fails. A failed gate is not a setback. It is a cheap way of learning not to spend money on something, before the money is spent.

  1. K1

    the channel-value experiment

    not run

    the one that can end the premium tier

    Do the impedance and contact-event channels measurably improve force-sensitive policy performance beyond wrench alone?

    requires
    v0 hardware only
    if it fails
    Ship v0 only. Abandon the premium tier. Do not reframe the claim.

    Third-party evidence that force helps at all is strong and multiplying — around 22–23% success-rate improvements from force-grounded policies in 2026 work. All of that validates wrench beating kinematics-only. None of it touches K1's question, which is whether contact schedule, impedance and uncertainty weighting beat wrench alone. No third party has run that comparison.

  2. K2

    fatigue robustness

    not run

    Does ultrasound fusion hold force accuracy across a full four-hour capture shift, where the electrode channel alone degrades?

    requires
    v1 hardware
    if it fails
    Drop the ultrasound block. Rely on anchoring and more frequent re-verification.
  3. K3

    wrist tensiometry

    not run

    highest risk in the programme — and, now, highest value

    Does direct-physics tendon tension sensing work at the wrist, with usable signal-to-noise, across people?

    requires
    v2 hardware — cheap relative to its information value
    if it fails
    L1 degrades to conventional fusion. The highest-novelty hardware claim is withdrawn.

    The whole published literature for this measurement is lower-limb; the wrist is unclaimed territory. Since 2026-08-20 its payoff is measured rather than asserted: without it, re-donning the sleeve costs the user minutes of labelled calibration every session. K3 decides between two different products — one where putting the sleeve back on is invisible, one where it costs four minutes every time.

  4. K4

    measured versus inferred stiffness

    not run

    Does direct elastography beat the cheaper inferred estimate, judged against the artifact's own perturbation ground truth?

    requires
    v3 hardware
    if it fails
    Keep the cheaper inferred impedance. Do not build v3.
  5. K5

    high-density grid value

    not run

    Do motor-unit features improve per-finger resolution enough to justify a tenfold increase in electrode channel count?

    requires
    v4 hardware
    if it fails
    Stay at low-density electrodes permanently.

The numeric threshold behind each of these is written down and decided in advance. It is not published here, because in the wrong room it is a shortcut to somebody else's validation design.

09 — WHERE THIS SITS

An empty band, and two curves moving the right way.

Below about $130 you can buy relative, uncalibrated signal. Above about $4,900 you can buy professional mocap and haptic hardware that was never designed for force capture. Between the two, essentially nothing carries a public list price. Every number below is checkable by anyone.

fig. 6 · every price here is a public list priceessentially nothing is sold here$10$100$1,000$10,000$100,000STAG tactile gloveuncalibrated pressure mapcheap smart glovesrelative tactileUMIvision + pose, no forceUMI-3Dvision + 3D, no forceSenseGlove R1haptic output, not captureHaptX G1haptic feedbackManus Metagloves Proprofessional mocapSleeve v0force + events + kinematics, calibratedSleeve, full+ stiffness + tendon tensionbelow $130 the signal is relative and uncalibrated · above $4,900 the hardware is mocap and haptics, never designed for force captureone dense-tactile capture station is estimated inside the middle band, but its price is not published
fig. 7 · the commodity channel deflates · the missing one does notfully-loaded teleoperation data, $ / hour$340early 2024$118march 2026−65% in two years. Observed, not forecast.force-labelled share of that spend8%202715%202922%2031Modelled at 47%/yr — the broader physical-AI rate, not the200%+ this category actually grew. Deliberately understated.

The size of it, bottom-up

Roughly $277.5M a year is already being spent on manipulation data — twenty frontier labs at about $6M each, seventy mid-stage robotics companies at about $1.8M, and several hundred academic and industrial teams at about $70k. Cross-checked a second way: at the March 2026 benchmark that implies about 2.35M data-hours a year industry-wide, or around twenty large capture operations across the whole industry — which is roughly what exists.

20292031
Data$132M$419M
Hardware$22M$75M
Recurring calibration$11M$45M
Total$165M$540M

The reframe that matters

The companies raising serious money to build wearable capture networks are building the distribution infrastructure for exactly this — and none of them captures force. We are not competing with them for the same buyer. We are the missing physical channel inside collection networks that are already being deployed by the thousand.

And the honest weak point

The open capture rigs are $371 and open-source. The closest forearm band is likely under $150 of parts. Our premium cannot be defended by having more sensors — only by channels that do not exist at any price: calibrated newtons, stiffness, and a stated interval. If gate K1 does not show those improve policy performance, the premium is indefensible, and the honest response is to ship v0 alone rather than to reframe the claim.

There is a clock on this, too. The labs that proved the physics still exist, all now see the same newly-paying market, and several have working prototypes we do not. “Nobody built the product” can mean there were structural reasons — and there were — or it can mean somebody has not got round to it yet. No framing changes which one it turns out to be.

10 — STATE

Said plainly.

  1. System specificationcomplete — five blocks, seven layers, five gates, costed at three volumesR
  2. Uncertainty validationseven arms run on public data, externally reviewed, six of our own claims correctedR
  3. v0 hardwarenothing built. Parts not ordered. This is the binding constraint on everything below—
  4. Gate K1untouched — there is no robot, no policy and no task anywhere in the programme yetH
  5. Gate K3untouchedH
  6. Customer interviews3 of a target 15 — the cheapest high-value work available, and the most consistently avoidedD
  7. The Record schema, publishednot started—

Calibration does not remove error. It converts an unknown error into a stated one. That is the whole promise, and it is the one no competitor currently makes.

Back toThe atlas — the whole system on one sheet →
Sheet
02 · the long sheet
Revision
0.2 · v3 architecture
Status
specified · partly validated · unbuilt
Enquiries
pranav@physcri.pt

physcriptthe force layer for physical AI