How it works

The whole method, including the parts that are guesses. If a number on your read looks wrong, this is where to find out why.

1. Twelve joints a frame, on your device

Your clip is stepped through thirty frames a second by MoveNet Lightning, a pose model that runs in the browser tab. It reports shoulders, elbows, wrists, hips, knees and ankles; the face is thrown away before anything else sees it. The video is never uploaded — not to us, not to anyone. What is saved, if you save anything, is the count and the faults.

Every length is measured in torso lengths (shoulder to hip), not pixels, so a clip filmed from far away reads the same as one filmed close up, and a tall athlete's numbers mean the same as a short one's.

2. Finding the jumps

A jump is a rise in your hips of at least 0.05 torso lengths (about two and a half centimetres) above where they stand over the surrounding second and a half. Hips, not feet, because the hips go where your body goes: a donkey kick lifts the feet a long way and the body hardly at all.

Each jump is then timed at the feet, which is where airtime actually happens. The feet are rarely caught at the exact frame they leave the floor, so the time they spend above a low threshold is corrected for the shape of a flight: a body in the air follows a parabola, and a parabola spends a known fraction of its flight above any given fraction of its height (above 25% of it, √(1 −0.25) of the time). On test clips with known timings this lands within about a hundredth of a second.

More than 0.6 seconds on the floor and the rope has stopped: whatever came before and after are different runs. "Best run" means unbroken.

3. Doubles, without seeing the rope

The camera cannot see the rope. It is a few millimetres across and moves faster than a phone records. So dulike estimates doubles from the one thing it can measure that a double needs: time in the air.

• If your jumps fall into two groups — short bouncy singles and longer doubles, as in single-single-double — the split is drawn between your own two groups. They have to differ by at least 30% with clear daylight between them, or it is one group with a wide spread, not two.
• If every jump looks the same, the whole clip is read against a fixed line of 0.28 seconds. A speed rope turns about five and a half times a second while doubling, so two passes need a bit over a third of a second — and some of the second pass happens on the way down.
• If you said "singles", every jump is a single. You were there.

When a clip sits close to either line, the read says "roughly". Your own count always wins: log it on Progress and delete the read.

4. The faults, and the one to fix

Each fault is judged on the median across your jumps, so one stumble does not become the headline. Some can only be seen from one angle — hand width from the front, knees coming up from the side — and a check the camera could not make is listed asnot checked, never as passed. Of everything found, the read puts one at the top: the most serious, and among equals the one that tends to cause the others.

These are coaching margins — roughly what a coach would let go without saying anything — reasoned from how the movement works. They are not fitted to anybody's data yet, and they will be tuned as real sets come in.

5. The ladder

Eleven rungs, each with a number you can count: from 30 singles to 100 unbroken doubles. Reaching a higher rung counts every rung below it, so nobody is sent back to prove they can do singles. Your rung sets today's session, which is built to fit the minutes you have — the cool-down and volume go first, the warm-up and your attempts never do.

Filming, briefly

Front-on, phone at hip height about three steps away, head to feet in frame, 10–30 seconds, camera still. Full tips on Train →