How Reachy Mini turns 2.7 million camera numbers into “follow my hand” — about 30 times every second.
BIG robot first · live demo · then we ride the pipeline, station by station

Figure’s smartest computer brain yet — it watches the camera and moves the whole body. It walks into homes it never saw before.
living rooms
fold + basket
make the bed
sees a picture
finds your hand
follow you
Big question: HOW does a picture become a head move? Let’s ride the pipeline.
each number goes from 0 (none) to 255 (max)
1280 × 720 = 921,600 dots
× 3 numbers = 2,764,800 numbers
about 30 times every second
Once it finds you, it just follows the box picture to picture. It only re-searches the whole picture if it loses your hand.
Looks only inside the box. Same 21 jobs every time: wrist + each finger’s 3 knuckles and tip. Nobody wrote a rule — it learned.
0 wrist · 1–4 thumb · 5–8 pointer · 9–12 middle · 13–16 ring · 17–20 pinky
x (how far right) · y (how far down) · z (how close)
21 dots × 3 = 63 numbers · it also says LEFT or RIGHT hand
Picking the clues (grown-ups say feature extraction) = a recipe a person wrote: a few dots in → one clue out. Yellow math — not a learned brain.
Recipe: robot-x = (0.5 − x) × 2 · robot-y = (y − 0.5) × 2 → (0.3, 0.6) becomes (+0.4, +0.2). Middle of the picture = (0, 0) = look straight ahead.
hold up the head
turns it · joins in on big turns
one motor each
✓ = the two moves this app uses · no single motor owns a move — all six work together
no tilt · no slide · the app never asks for the body — but the robot turns it itself when the neck runs out (~65°)

Looks at the whole picture → draws a box around the person (not just a hand)
Works inside the box → marks 33 dots · about 18 pictures a second
Six moves. Each one is the same recipe: a few dots in → one clue → one move out. Yellow math — a person chose every row.
9 dots do all of it · nose · 2 ears · 2 shoulders · 2 wrists · 2 hips