02Compare the approaches
Video, gloves, bare hands.
Video shows the motion of a hand, not its contact force. A glove adds force but covers the hand. We plan a bare-hand kit whose labels are to carry a measured error. The table compares the three by category; each Cr4dle cell names its state, and nothing below is validated product performance.
01Three approaches
Who measures the contact?
The table names categories, not companies. A planned or research cell is a dated plan, not a result. Each label is to carry its held-out-user error before we state a force number.
Scroll horizontally to compare every column
| Question | Video-only egocentric hoursHead-camera video of hands at work | Glove and EMG collection serviceSensor glove, forearm EMG and a head camera, run as a service | Bare-hand calibrated kitCr4dle, in research |
|---|---|---|---|
| Is the hand covered? | No. The hands are bare. | Yes. The gloves in the offers we read cover the finger pads. | No. An sEMG/IMU band is to keep the finger pads bare. A fingernail module can join a later revision of Ember if its own bet passes.in design |
| Is the force ground truth published? | No force channel. | Not published in the offers we read. | A fingertip-force rig, checked against a force gauge at the start and end of each calibration block, is to be the reference for each label.planned |
| Is the held-out-user error on the card? | Not applicable: no force label. | Not published in the offers we read. | The card prints it as a number, or as not run with a dated gate. The synthetic demo card shows the field.planned |
| Is the sync error measured? | Not applicable: one stream. | Streams are said to be in sync; no number is published in the offers we read. | An LED flash that each stream sees is to give a measured sync error for each episode. The synthetic demo prints the field from simulated streams, not from a device.planned |
| Consent lineage and revocation for each episode? | Varies by vendor. | Not described in the offers we read. | Consent is checked at capture and again at export against the consent store. A revoked episode is not exported, and the card counts it as a missing-consent discard. The synthetic demo uses a simulation-only consent mode.reference code |
| Does the raw biosignal leave the device? | No biosignal is recorded. | Not stated in the offers we read. | No. Raw sEMG stays on the device. Only features cross, through the egress contracts.design rule |
| Does it scale to commodity video? | It is the commodity video, without force labels. | Gloved-hand pixels do not match bare-hand video. | Contact Lift is to estimate contact from bare-hand video, trained on paired kit episodes. It is a bet with a dated kill criterion.research |
| Delivery formats | Varies by vendor. | Not named in the offers we read. | LeRobot v2.1, LeRobot v3.0 is written in the synthetic demo. is planned. The MCAP source of record of each episode stays on the capture hub and is never delivered.reference code |
02Pair explorer
Choose two families to compare.
Choose two distinct families. Each carries the posture its registry row gives it; no family is offered for sale today.
Comparing Ember and Lens.
- Form
- Ember: Capture band
- Lens: Eye-tracking glasses
- Next proof
- Ember: Per-finger contact force and demonstrator state on the video clockDated gate or kill date; published either way
- Lens: Gaze target on the video clock, beside the hand pose of the head-camera adapter; head pose is plannedDated gate or kill date; published either way
- Reads the registry row declares
- Ember: Contact force · demonstrator stateLabelled channels, each under a label-validity spec and a dated gate; hardware unvalidated
- Lens: Gaze target · hand pose · head poseLabelled channels, each under a label-validity spec and a dated gate; hardware unvalidated
- Posture
- Ember: Capture kit · the one band Cr4dle builds
- Lens: Capture kit · bought device
- Registry status
- Ember: Cr4dle build (gate Q2 2027)
- Lens: Bought device
03Every navigable family
A view across the registry.
The kit families and the research tiers have different evidence paths. Tier describes sensing or invasiveness, not product quality; status is the build/buy posture from the registry row.
Scroll horizontally to compare every column
| Family / tier | Form | Next proof | Posture | Registry status |
|---|---|---|---|---|
| EmberL0 | Capture band | Per-finger contact force and demonstrator state on the video clockDated gate or kill date; published either way | Capture kit · the one band Cr4dle builds | Cr4dle build (gate Q2 2027) |
| LensL0 | Eye-tracking glasses | Gaze target on the video clock, beside the hand pose of the head-camera adapter; head pose is plannedDated gate or kill date; published either way | Capture kit · bought device | Bought device |
| LumenL2 | Forehead patch | Research-only physiological columns beside the kit, in bounded sessionsDated gate or kill date; published either way | Research tier · bounded sessions | Research tier |
| CrownL3 | Headband | Whether a wider EEG and fNIRS montage recovers anything the kit does notDated gate or kill date; published either way | Research tier · bounded sessions | Research tier |
| WhisperL1 | Earbud + sEMG | None planned; the family is deferredDated gate or kill date; published either way | Deferred · no work planned | Deferred |
Start with the registry.
Follow the evidence.
Read how we separate a registry row, a reference implementation and a claim supported by a held-out-user number.