[physcript]
TOUCHSTONEatlas · v3

MVP · TOUCHSTONE v3 · ATLAS · sheet 01 · rev 0.2

A demonstration is not a trajectory. It is a trajectory, plus the force it put into the world — and nobody has ever written that part down.

Touchstone is an instrument, not a dataset vendor. A forearm sleeve reads the muscles that move the hand; a handheld standard tells the sleeve what a newton is; software compiles both into one record of what a person actually did to an object — with a stated interval on every force in it. This sheet is that whole system, at once, with the experiments that could end each part of it hanging underneath.

Measured
read from an instrument, in newtons
Certified
traceable to a physical reference
At risk
gated by an experiment not yet run
I

Two gaps, and they share one answer.

The first is that force is invisible. The second is that even if you could see it, a human hand and a robot gripper have so little in common that copying one onto the other means almost nothing. Both are usually treated as separate problems. They are not.

frame 0412 · rgbheld
peak normal force2.4 N
frame 0412 · rgbcrushed
peak normal force41.6 N

Identical pixels. Seventeen times the force. Only one of these is a demonstration worth learning from, and nothing in the image says which — which is why a tactile glove giving relative pressure in arbitrary units does not close the gap either. The missing thing is not resolution. It is the unit.

23 degrees of freedom against 2

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. There is no correspondence to copy, which is exactly why retargeting research keeps fragmenting into one bespoke solution per robot.

So 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 reason to build a wearable rather than anything else.

II

The vectors do not change. The body does.

Scroll, and the thing holding the object becomes something else each time. Watch what stays still.

the held object18.4 N12.0 N0.31 N·mHuman hand23 DOF · 5 deformable padsSourcethe wrench is fixed · only the body changes
  1. Human hand
  2. Five-digit hand
  3. Three-finger gripper
  4. Parallel jaw

Video sees five fingers and has to treat all five as meaningful. Measured force shows which two did the work.

Illustrative. No hardware exists yet — these are the shapes of the answers, not measured ones.

00the sheet is dormant

physcript · touchstone v3 · atlassheet 01 · rev 0.2 · not to scaleI · whyII · captureIII · compileIV · the recordV · consumeVI · deployThe force-data gapvideo has no force channelThe embodiment gapthere is nothing to copyThe thesisa demonstration isnot a trajectoryone answer to both gapsElectrode arraythe primary channelUltrasound arrayfails differentlyTapper + accelerometerstwo risks in one blockIMU clusterboring by designFMG ringdeliberately demotedCalibration artifactthe transfer standardEgocentric cameracustomer-suppliedL1Physics-anchored fusionnovelty 9 · risk highL2Conformal force intervalsnovelty 8 · risk mediumL3Hybrid physics-MLnovelty 4 · risk lowL4Missing-modality robustnessnovelty 3 · essentialTHE RECORDL5Kinematics200 HzWrench200 HzImpedance50 HzContact events±1 msUncertaintyper sampleintermediate representationL6Uncertainty weightingnovelty 7 · risk mediumL7Privileged distillationnovelty 3 · the sales argumentAny robot bodywrench is morphology-invariantForce-blind fleetsupgraded with no new hardwareA certified unit of forcetraceable and re-verifiedVII · what would kill whatfive pre-registered experiments · each with a written failure responseK1not runchannel valueK2not runfatigue robustnessK3not runwrist tensiometryK4not runmeasured stiffnessK5not runHD grid value

hover anything to isolate its lineage · click to read it · hover a gate to see what dies if it fails

record

The Record

R

This is the file format Touchstone actually sells. Every demonstration becomes one of these: where things moved, how hard they were pushed, how stiff the arm was, exactly when contact happened, and how confident we are in all of it.

Sleeve streams are the source language. Robot actions are the target. The Record is the intermediate representation between them, the way LLVM's IR sits between source code and machine code. Whoever defines the IR owns the ecosystem.

Each field is independently validated in the literature. The combination exists nowhere. Design decision: publish the format openly, keep the estimators — the value is not in the format, it is in being the only party who can fill it correctly.

III

What actually gets written down.

Five channels, one timebase. Four of them exist somewhere else in some form. The fifth — a stated interval around every force and stiffness value — exists in no dataset at any price, and it is the one that turns calibrated from an adjective into something a buyer can check.

one demonstration · 4.2 s · as it is written downtouchstone record v0.1touchloadliftslipreplacereleasekinematics200 Hzwrench200 Hzimpedance50 Hzcontact events±1 msquality1 Hzthe stated intervalwidens where the measurement is hardestcontact events are drawn across every channel because that is what they are — the joints of the task, not a sixth signal
a conventional dataset saysF = 4.2 N
the Touchstone Record saysF = 4.2 N, [3.6, 4.9] @ 90% coverage

Wrench is written in the object's frame, never the hand's. Hand-frame forces are morphology-specific; object-frame wrench is not. That one choice is what makes the record portable, and it is why the format will be published openly while the estimators that fill it will not.

IV

Where this actually stands.

Everything above is a specification and one validated corner of it. Said plainly, because a company whose entire pitch is calibrated numbers with stated error does not get to be vague about its own state.

  1. System specificationcomplete — 5 blocks, 7 layers, 5 gates, costedR
  2. Uncertainty validationrun on public data, externally reviewed, six of our own claims correctedR
  3. v0 hardwarenothing built · parts not ordered
  4. Gate K1untouched — no robot, no policy, no task anywhere in the programme yetH
  5. Gate K3untouched — highest risk and, since 2026-08-20, highest valueH
  6. Customer interviews3 of a target 15D
  7. Record schema, publishednot yet — cheap credibility, no dependencies

The measurement is the product. The chain that makes it believable is the position.

Next sheetThe long sheet — every block, the traceability chain, and the evidence in full →
Sheet
01 · atlas
Revision
0.2 · v3 architecture
Status
specified · partly validated · unbuilt
Enquiries
pranav@physcri.pt

physcriptthe force layer for physical AI