Teach an ear (Wear tab, Ear gates) to see here whether what you're doing now would count as it.
Motor: –
While the wrist is measuring, the head's receiver waits for each of its calls. In bursts it switches on for ~32 µs at a time with a gap between, instead of staying on: the longer the gap, the less power while it waits (16 µs gaps: about 75% of the power; 32: about 60%; 64: about 45%; 96: about 40%; 128: about 36%; 160: about 33%; 200: about 30%), but too long a gap and it misses the calls. (Before 13e every gap failed: the bursts were set up with an on-time the chip doesn't accept.) Try a gap and watch the ranging success at the top of the page: if it drops, go back to a shorter gap or all the time. Needs firmware 13a or later on all three units.
The motion gate is trained with the model (Model page) and sent with it. While it is shut both radios sleep and the wrist only watches its motion sensor. It is never used while buzzing is off (strength 0), so drills and long recordings always measure.
Rest: when the wrist hasn't moved for this long it stops measuring and turns its LEDs off; any movement wakes it within a fraction of a second, before the hand can reach an ear. Off: after a long stillness (on a desk, overnight) it drops the Bluetooth links and powers down; moving it starts it again, but reconnecting takes 1–3 s, so the first touch after picking it up may be missed. Neither happens while this page is connected. A higher movement threshold lets it rest while worn but sitting quietly; raising the arm is always a big movement.
A head unit that hasn't moved isn't being worn, so it advertises less often (a wrist waking from off then takes a few seconds longer to find it); any movement switches it straight back. After a long stillness it switches off, and moving it starts it again. Only while no wrist is connected, so never while you're wearing both. The wrist passes these to the head when it connects; the head keeps them.
Besides your wrist's motion, the trees use what the radio says about where your hand is relative to the head unit: distance, how much of you is in the way, signal strength and path delay. This check holds each position for a few seconds at several power levels, switching between them every few seconds so a change of posture affects them all alike, and scores how well those radio readings tell your ears from everything else at each power. Results vary from run to run, so run each placement three times: the runs are combined. Then try other placements and compare them in the table.
For each position: rest with your arm down until the short buzz, then hold it (for an ear, play with it the way you normally do, moving around it) until the long buzz. The power switches back and forth while you hold. Keep the head unit where it is for the whole check. Your usual buzz is off meanwhile, and your power setting comes back at the end.
Not run yet for this person.
| Placement | Runs | Best power | Radio separates | Most mistaken for an ear | Ears read |
|---|
The head unit check above recommends this. Lower power reads more accurately up close but shortens the range; it must comfortably reach your furthest ear. The wrist passes it on to the head unit. It's kept as this person's power and set again whenever their trees are sent. The trees' signal input allows for the power, but distances shift a little, so record a few drills and train again after a big change.
This wrist still has a distance calibration table from earlier versions. It isn't needed any more: each ear's limits are learned from your own readings, so real centimetres don't matter. Keeping it does no harm. Removing it keeps a little more detail in very close readings, but the ears were taught with it, so re-teach both ears afterwards.
Set up also picks the transmit power (the range of the distance readings), which the gesture model uses too: do it once even if you use the model.
Do this once, wearing both units as you normally would. For each ear the page tries five transmit powers (30–86 on the 0–100 power scale), chooses the one that gives the clearest results, and teaches both ears from the readings at that power. The wrist learns each ear's arm pose and how far away it reads there for each kind of signal (through or round the head reads further than a clear path).
| Power | Left ear read | Right ear read | False buzzes | Notes |
|---|
What the wrist has learned. Teach again re-teaches one ear at the current power, e.g. after re-doing your scrunchie. A longer teach gives more readings: more blockage bands get a limit of their own, and the pose covers more of your natural range.
Every change this page makes to the ears (Setup, Teach, tightening, leaving a position out, adjusting by hand) is listed here with what the ears were just before it. Restore puts that version back on the wrist and saves it; the restore is listed too, so it can be undone in turn. Kept in this browser, for this wrist.
An ear touch is long and still. Require the hand to stay at the ear, with the distance and arm angle steady, before buzzing. This rejects a hand sweeping past the ear, at the cost of buzzing this much later. How much the distance may wobble is learned when you teach each ear.
At your ear, your hand and head weaken the signal. When you teach an ear, its usual signal strength is learned; a reading much stronger than that (the wrist near the head with a clear path, e.g. arm raised beside it) doesn't count as at that ear. Learned at the current transmit power: teach the ears again (or run Setup) after changing the power.
After a buzz, the wrist re-arms once the distance goes this far past the ear's trigger distance, or the arm leaves the pose.
Connect and teach an ear to see each live reading checked against it.
| Test | Now | Taught ear | Margin |
|---|
Hold the position (or keep doing the movement) and press the button: the page records 5 s, keeps it in Your data (Collect tab) as "Checked now" (not a touch; change the type there if it was one), and, if it would buzz, shows ways to tighten the ear to leave it out, with what that does to your other recordings, before changing anything.
Each circle is a taught ear pose (solid: inside counts; dashed: 5° slack that keeps a touch going). The dot is your arm now, with its last 3 seconds as a trail.
Many short, labelled examples, quickly. Pick a set and press Start: rest with your arm down (or carry on with whatever the set says) until a short buzz, then do the movement until the long buzz. The gaps are random, so you can't anticipate them. Each one is saved under Your data below, labelled, with its 6-axis motion. Mix the ear-habit sets across positions, and do the ones that shouldn't buzz too: the model learns as much from those.
Not running.
| Set | Type | Done |
|---|
Wear the units and go about your day while this page records everything (distance and signal, the 6-axis motion, tilt, buzzes) at full rate. When you notice yourself touching your ear, press SW2 on the wrist (the button that isn't reset: the motor ticks to confirm), or the button below. You don't have to catch every touch, or press at the start of one: the model allows for that. Everything that isn't marked counts as "not a touch", which is how it learns what ordinary life looks like.
Keep this page open and the phone or computer nearby and awake (the page asks the screen to stay on where it can). While recording, the wrist measures all the time rather than only in an ear pose, which uses more battery. Recordings are saved in this browser as you go; export them to keep them or to send them.
The guess comes from the trees you last trained (Model tab); before you've trained any, it can't ask. It asks at most every 30 s.
Not recording.
Everything the models learn from, for this person. With the model is what the model you're looking at on the Model tab does with each one, tested without having learned from it (train first). View shows it moment by moment on the Model tab, where you can also change a long recording's labels.
No drills yet.
| Drill | Label | With the model | Recorded |
|---|
| Recording | Labels: touches marked · yes / no / no answer | With the model | Started · length |
|---|
Answers lists every question and SW2 press in a recording, so you can fix one you got wrong (pressed twice instead of once, or pressed by mistake).
What the wrist keeps while you wear it with the buzz on, without this page: your SW2 verdicts on buzzes (once = right, twice = false), touches it missed (one press with no buzz) and an ordinary moment every few minutes (taken as not a touch). Copied when the page connects; while it stays connected, new ones appear about 15 s after a press (ordinary moments within a few minutes). Change a label if a press was wrong; that changes this page's copy for training, not the wrist.
Nothing copied from the wrist yet.
| When | What | Trees' score then | Label for training |
|---|
Record your ear habit and the things that shouldn't buzz. Each is labelled and kept above, with what the ear gates would do with it. Drills do this faster; this is for one particular movement. Record each a few times (Guided does three, with your arm down in between).
Guided: three recordings with your arm down in between. The wrist buzzes briefly to start and longer to stop, so you don't need to watch the screen. Start each one with your arm down and bring your hand up as you normally would: the 6-axis sensor records the approach too.
While this page is connected, every touch the wrist detects is kept here with the readings from the 4 seconds before it and 1.5 seconds after. Mark each one real or false as it happens, and give false ones a name (what you were doing) so they can be grouped. Add marked buzzes to Your recordings then adds them to Your recordings (real ones as ear habits, false ones as not a touch), so they're checked and scored with the rest.
No buzzes logged yet. Connect, wear the units and go about your day; this page has to stay connected.
Nothing collected yet: start with some drills (Collect tab).
Import loads an Export everything file back, as a new person or in place of the current person's data (Person, at the top). One file with all you've collected for this person in this browser: drills and recordings (with their 6-axis motion), long recordings (with your answers, presses and corrections), your SW2 verdicts, the touch history, the buzz log, model versions and the wrist's settings. Compressed (.json.gz); send it as it is.
The wrist decides with decision trees (40 small ones) learned from your labelled data: distance, blockage, signal strength and path delay, the 6-axis motion (how hard the wrist turns about each axis, hand shake, how quick the movement is) over the last second, the wrist's tilt (which way gravity points), turning and shake over the last 2 s, and whether the distance is shrinking. A moment counts as a touch when the trees' score is high enough and the hand is within ear range (a fixed distance limit the trees can't override, so movements they've never seen can't buzz far from your head). It buzzes when that holds for most of a second.
Testing: it always tests on data it didn't learn from. With data from two or more days, it's tested on each day after learning from the others (the fair test). With one day, even one or two short sessions, it still works: each recording, and each stretch of a long recording (10 minutes, or shorter for a short one), is left out in turn and tested on. That flatters it a little (same day, same scrunchie position), so the result says so. Ear-habit recordings count from 2–2.5 s after the start (the hand arriving isn't judged); long recordings count 5 s before to 1 s after each marked touch as a touch, leave the 12 s around it out (you may have marked late or early), and leave out the time around any buzz you didn't mark.
Collect some data, then train. Nothing on the wrist changes until you send it (On the wrist, below).
| Data | Touches caught | False buzzes | Time to buzz |
|---|
Why the trees score a moment the way they do: each bar is how much one measure pushed the score towards a touch (right) or away from one (left) on the way down the trees, compared with an average moment. Uses the trees you trained above (or, before you train here, the last ones saved).
Connect, and train a model, to see the live score explained.
Send the trained trees, with the score and ear range set above, to the wrist. They're the model trained on all your data (the test models above each leave a part out, only to measure it). The wrist decides with them with or without this page. Your buzz settings, cool-down and hold still apply. Switched off, the wrist detects nothing (it still records for this page); the trees stay stored on it.
Connect to see what the wrist is using.
Each training is kept here (the last 8, in this browser). The two bold numbers matter most: how many real touches it catches, and how often it would buzz when you're not touching your ear. Send an earlier one back if a newer one turns out worse. A new one's numbers follow the sliders until you train again or send it.
Every recording (Collect tab, Your recordings) is replayed through several versions of the rules. Missed = ear-habit recordings that wouldn't buzz; false = other recordings that would. The ears' limits come from the taught recordings (not scored). Each ear's gesture (a fingerprint of how your wrist moves while you rub that ear) is learned from its taught recording and, with the box below ticked, its ear-habit recordings too; a recording is never compared with a gesture that includes itself, so the numbers aren't flattered. Nothing on the wrist changes.
Teach the ears (Wear tab, Ear gates), record your habit and some things that shouldn't buzz (Collect tab), then press Work it out.
Move a gate and the table updates. Looser lets more through (fewer missed touches, more false buzzes); tighter the opposite. These only change this table, not the wrist.
| Rules | Missed touches | False buzzes | Compared with now | Readings counting: ear habit / other | Time to buzz | Scored on |
|---|
Each column is one setting of the chosen gate (the others stay where the sliders are): missed touches · false buzzes, out of the recordings each version can score. The best trade-off is the column with the fewest of both.
Wearing it without the page: after a buzz, press SW2 on the wrist within 10 s (or 5 s after your hand leaves): once = it was right, twice quickly = false. If you catch yourself touching your ear and it didn't buzz, press SW2 once: that's kept as a touch it missed. About every 4 minutes it also keeps an ordinary moment (taken to be "not a touch"). The wrist keeps each with the trees' 15 inputs (the last 384 moments), copies them here when it connects, counts them above, and the Model tab trains on them. No press is fine.
The wrist keeps about the last 1,000 touches; this page keeps everything it has copied, in this browser on this device. Export a CSV now and then as a backup: clearing the browser's site data also clears the history.
Both units keep their recent log (about 11 KB of text, the same lines as the serial log) in flash, through restarts, off mode and a flat battery. Download it here to see what happened while they were on battery.
For the head log, the wrist lets go of the head for a moment, then Chrome asks you to pick UWB-Head1; after the download the wrist reconnects by itself (within 2 minutes even if something goes wrong).
Every change is saved on the wrist as soon as you make it.