Ember
Label the force the camera cannot see.
The wrist band that puts contact force and demonstrator state on the same clock as the video.
The graphite ribbon grows into a shallow dorsal sensing region and closes beneath the wrist. A mineral liner and subtle side squeeze complete the continuous future form.
Ember concept illustration. Design and engineering remain exploratory.
Ember is the one band Cr4dle builds, behind its dated gate. Bought bands stand in until then; no device is sold today.
Follow EmberSell the channel, never the hour.
Ember is a wrist band with 8-channel sEMG, a 9-axis IMU and PPG/EDA (BLE, 12 h) worn by people demonstrating tasks by hand. After a 15-minute per-user calibration it is to supply per-finger contact force, and a demonstrator-state column, to robot-training episodes on the same clock as the head-camera video (and the Lens scene video and gaze when the glasses are worn). It is the one piece of hardware that Cr4dle is to build.
- Intended sensing
- 8-ch sEMG · 9-axis IMU · PPG / EDA
- Channels
- Contact force (per finger, calibrated) · demonstrator state
- Build gate
- Q2 2027; needs three paid collections on bought bands and waits on the decision that replaces ADR-0136
- Egress
- Features only; raw sEMG never leaves the hub
One continuous gesture.
A narrow graphite ribbon follows the wrist, swelling gradually where sensing and electronics need space. A mineral inner liner and concealed overlap clasp keep the surface continuous.
Future ribbon section
17 mm wide / 3.65–6 mm thick
A variable-section target, with a 54 × 40 mm inner ellipse used only as a fit proxy. Real sizes, contact pressure, energy and haptics remain prototype work.
Contact
A ribbon, shaped from within.
The inner saddle places optical sensing between separate conductance contacts, with a thermal contact apart. The liner and several future sizes must establish fit without depending on excess pressure.
Anatomy
Distribute the working volume.
The dorsal substrate holds sensing and processing. Thin energy zones occupy one side, a haptic carrier the opposite shoulder, and a polymer span reserves radio space. This is a new package beyond the present bench components.
Touch & care
A cue you can answer.
A fingerable side squeeze explores deliberate marking and hold-to-pause. The inner liner detaches for care; a hidden palm-side closure opens the band onto a low charging saddle.
Bought bands now; the band to build behind a dated gate.
In an instrumented demonstration the band rides the recording path with the head camera and, when a buyer names gaze (a rule for a later phase, ADR-0136), the Lens glasses: sEMG and motion are calibrated against fingertip force sensors, the force estimate and the demonstrator-state column land in LeRobot-format episodes with sync error and consent lineage on the dataset card. Bought MindRove Armband 2 or OYMotion gForcePro+ bands stand in for Ember v1. Its build gate is Q2 2027, and the build needs three paid collections on bought bands, so it waits on the decision that replaces ADR-0136. Explore and Unwind remain a simulation-only background under ADR-0040.
Where we are.
What comes next.
Implemented reference: the family integration runs the band's four streams through the shared runtime on simulated or replayed data, records and replays the sEMG stream, and fills only the research arousal column; the force estimate stays empty until a calibrated EMG-to-force model is attached. The DRV2605L marker-cue bench is documented and unbuilt. The adapters for the bought bands, the 15-minute calibration and the force model exist as tested reference code in the capture kit, exercised on synthetic data only; nothing is measured on a person, and no band hardware is built before the gate.
A clear boundary
Contact force is a calibrated estimate reported with its held-out-user (LOSO) error, not a measurement of grip strength or effort. Demonstrator state never claims emotion, mood or stress; it names fatigue and safety-relevant physical states under a label-validity spec and never an inferred mental state. Arousal is a recorded research column, never a product.
Next steps and dated gates
- Designed forBought bands now; one band built later.
- The recording path drives the MindRove Armband 2 and the OYMotion gForcePro+ through their adapters; bought bands stand in for Ember, and none is bought yet: every unit enrolled today is the simulator's and both equipment rows read simulated. Ember v1 (8-ch sEMG, 9-axis IMU, PPG/EDA, BLE, 12 h, 25 units) is the planned build. Its build gate is Q2 2027, and the build needs three paid collections on bought bands, so it waits on the decision that replaces ADR-0136.
- Next proofPrint the force error before the pitch.
- The date of the 2026-10-05 gate (a cross-user EMG-to-force model at 1.5 N pooled or better after the 15-minute calibration, against a learned RGB vision baseline on at least five participants, with the sync error printed on the card) passed with no result recorded in the registry.
- Research horizonDoes force earn its place in the policy?
- The 2026-10-19 gate: a force-aware policy against a motion-only one on a deformable-object task, both policies trained on the same by-hand episodes, at least 40 robot trials per arm, a one-sided 95% Newcombe lower bound above zero and an observed gain of 15 points or more, a pre-registered minimum observed effect, not a certified margin; any other result is KILL; published either way.
Meet Lens.
The other half of the capture kit, on the same clock and the same recording path.