Learn from data. Test what transfers.
Illustration of one clock: every stream logs the same flash instant. Not data.
Record. Sync.
Export under consent.
Future human recordings could address gaps in public data. The recording design aligns sensor clocks, redacts on the device and checks consent at export. Its current end-to-end demo uses synthetic data.
- 01 / Record
Record on two clocks.
Bought sensors on a person demonstrating by hand. Every frame keeps its own device-monotonic time beside the hub time.
Raw sEMG and gaze stay on device - 02 / Sync
Measure the sync error.
An LED flash the camera sees and the band logs. Offsets are the median over events; the residual is the error, floored at what the detectors resolve.
Measured, never asserted - 03 / Export
Export under consent.
Faces and bystanders redacted on device; consent checked again at export, with revocation; a card printed with every episode.
No consent → nothing leaves
Streams syncedEvery stream saw the flash. The offsets are measured, the residual is the sync error, and the episode can be exported.
Knowing what
not to claim.
A number has to be measured.
The sync error comes from an LED flash every stream sees or logs, floored at what the detectors resolve and re-derived from the raw area. A stream with no event is marked unverified rather than given a number.
Consent is checked twice.
At capture, on every episode, and again at the instant of export, with revocation. An episode whose consent is missing, mismatched or revoked is not exported, and the discard is counted on the card.
What has not been run says so.
A card prints the held-out-user force error and the ablation result as numbers or as not run with the dated gate by which they are due. Nothing is left blank and nothing is asserted.
Different sensors.
One clock, one card.
The sEMG/IMU band supplies force and a demonstrator-state column; the head camera is the default video source from which hand pose is estimated, and head pose is planned, because the reference estimator gives no measured head pose; the Lens glasses add the gaze target, and their own egocentric video, when a buyer names gaze (a rule for a later phase, ADR-0136). The same runtime, redaction stage and exporter carry every stream, and the same card prints what each one measured.
Today the path runs end to end on synthetic data and the simulator; the vendor adapters, the calibration and the redaction stage exist as tested reference code. No collection on real sensors has been sold, and no sensing performance has been validated.
Force at the fingertip.
Pads left bare.
A glove puts a sensor between the finger and the object. We plan to measure fingertip force from the fingernail instead. The module is in design; it is not built, and it has no measured performance.
It reads from the nail.
When a fingertip presses, the nail bed changes colour and the nail strains. The module is to sense one or both from the back of the finger, after the fingernail sensors of Mascaro and Asada. The finger pad stays bare, so nothing covers the grasp and the video shows bare hands.
It calibrates against a rig.
A fingertip-force rig, checked against a force gauge at the start and end of each calibration block, is to be the reference. Each label is to ship with its error on people held out of training, as a number or as not run with a dated gate.
It is a module, not a family.
If its bet passes, the module is to join a later revision of Ember, the wrist band. Raw signals are to stay on the device, and only features are to leave it. No module is built, and no sensing performance is measured.
Score the labels.
From any dataset.
We plan an open harness that scores the force labels of any dataset: ours against the fingertip-force rig, and an external corpus against the reference sensor that it carries. It reports three numbers, or says which of them it could not measure.
Held-out-user error.
Compare a dataset's force labels with its reference sensor (the rig for Cr4dle data) on people that its model did not see in training. A label without that number is reported as unverified.
Sync error.
Measure the time offset between the force stream and the video from a shared event, such as an LED flash, rather than accept a stated sync.
Coverage.
Count the fingers, grasps and episodes that carry a usable label, and the discards with their reasons, so a dataset user sees what the labels cover.
The kit is to calibrate.
The model is to scale.
Contact Lift is a research model that is to estimate per-finger contact force and contact events from bare-hand egocentric video alone. Paired kit episodes train it: the video, and the calibrated force from the band, and from the nail module if its bet passes. A gloved kit is a poor source for it, because gloved-hand pixels do not match bare-hand video.
The bet (id video-contact-lift) is proposed, and its kill criterion is registered with a dated gate, and its margin waits for founder attestation. On people and objects held out of training, the video-only estimate must have a lower per-finger force error than a baseline that uses only hand pose, by a stated margin. If it fails, contact-force labels come only from kit-instrumented episodes, and the result is published either way. No paired episode is recorded yet, and a lifted label is an estimate, not a measured channel.
One question.
A reproducible test.
Build a small model and our own evaluation before depending on the robot of a research partner. An offline result, a physical result and demand in a later phase (ADR-0136) are separate things to establish.
Start with a small question.
Choose a physical interaction to predict, a permitted dataset and simple alternatives to compare against. Begin with free, open tools and a bounded experiment; a robot or a paid model service is not needed to start.
Test on what was held apart.
Keep evaluation examples separate from training and tuning. Measure prediction errors, test changed conditions and publish failures alongside gains. A simulator result does not establish useful transfer to a real robot.
Learn what would be useful.
Ask potential users which failures matter and what improvement would help. Cr4dle is a non-commercial research programme now (ADR-0136); the plan for a later phase is a data offer to robot-learning teams: training data, tasks and rubrics. Feedback and measured model limits guide the next milestone; new recordings or robot trials need a specific question to answer.
Contact first.
Then models, then agents.
Each step is a dated bet with a kill criterion. Agents in bought bodies start after the Human Model gate, not after the world model. Nothing on this path is proven, and nothing is offered for sale.
Contact truth.
Calibrated contact-force labels from bare hands, each with its held-out-user error and measured sync error. This is rung one, and it comes first.
The Human Model.
What a person reaches for, when and how hard, to be learned from paired episodes. Cross-user calibration is the hard problem; useful transfer is unproven until a real-robot test measures it.
A force-conditioned world model.
Human- and force-conditioned models, to be used first as evaluators and data engines, built on open weights under research bets with kill dates.
Agents in bought bodies.
Agents built from the Human Model are to run in bodies that we buy. They start only after the Human Model gate, and no such agent exists today.