The AI push-up counter, explained honestly
Every push-up app now says 'AI'. Almost none of them say what the model actually looks at, what it rejects, or where it fails. This page does, for the counter that ships in Workout Samurai on iOS and Android.
- Pose detection runs on-device: Vision on iOS, ML Kit on Android
- Four fused signals, not one fragile threshold
- Calibrated to your body, not to a generic elbow angle
- Structural gates reject reps that never actually happened
Sensei
“Almost at your word. Don’t stop. Show me more.”
The short answer
An AI push-up counter uses a pose-detection model to track your body's joints in each camera frame, then decides from how those joints move whether a rep happened. Workout Samurai fuses elbow flexion, shoulder descent, head descent and apparent scale into one rep-progress signal, normalises it to your own range, and counts a rep at a genuine descent-then-ascent. All of it runs on your phone.
What 'AI' means in a push-up counter
In a push-up counter, 'AI' means one specific thing: a pose-estimation model. It takes a camera frame and returns the estimated pixel position of your body's landmarks (nose, shoulders, elbows, wrists, hips, ankles), each with a confidence score. That is the whole of the machine learning. Everything after it, the part that decides whether a rep happened, is ordinary signal processing on those coordinates.
This matters because it tells you where these apps break. The model is good at finding a clearly visible body in decent light and unreliable when a joint is occluded, cropped, or lost in shadow. No amount of clever rep logic recovers a shoulder the model could not see. Camera setup is not a detail. It is the dominant variable in whether an AI push-up counter works for you.
Why one signal is not enough
The obvious way to count a push-up is to watch the elbow angle and count every time it goes below some threshold and back. It is also the way that fails first. Facing the camera, your forearms recede along the lens axis, so elbow angle is heavily foreshortened and the 'right' threshold differs per person and per phone position. Worse, at the bottom of a push-up your elbows are frequently occluded by your own torso.
Workout Samurai therefore fuses four signals into a single rep-progress value between 0 and 1, each weighted by how confident the model is in it at that instant:
| Signal | Role | Why it is included | Where it fails alone |
|---|---|---|---|
| Shoulder descent | Primary depth proxy | Survives elbow occlusion at the bottom of the rep | Drifts if you creep forward across a set |
| Elbow flexion | Supporting depth | Directly reflects the joint doing the work | Foreshortened facing the camera; often occluded |
| Head descent | Corroborating | Survives a single shoulder dropping out of frame | A head nod alone can fake it |
| Apparent scale | Weak supporting | You genuinely get closer to and further from the lens | Noisy; heavily down-weighted |
How the counter decides a rep happened
Turnarounds are detected as extrema in the fused rep-progress signal rather than by crossing a fixed number. You count a rep when the motion genuinely reverses, not when some angle passes 90 degrees. At the top of the return, the rep is checked against a short list of structural gates that an honest push-up always passes:
- Depth floor. Did you actually descend a meaningful fraction of your own calibrated range?
- Wrist plant. Are your hands on the ground where a push-up puts them?
- Rep structure. Was there a real descent followed by a real ascent, rather than a single jitter?
- Body present. Was enough of you in frame to make the above measurable?
If the gates pass, the rep counts and you feel a haptic pulse. If a supporting signal simply could not be evaluated, say you drifted half out of frame or the light dropped, the rep is still counted but recorded with lower confidence. That is the deliberate choice: an AI push-up counter that silently eats reps you really did is worse than one that occasionally counts a bad one, because the server checks the session afterwards anyway.
Does it count half reps?
Not as full reps. A rep that never crosses the depth floor does not pass the gate. But the honest answer is more nuanced than the marketing line every push-up app uses. The depth floor is calibrated to *your* range, which means a person with limited shoulder mobility and a person with a full chest-to-floor range are each judged against their own baseline. The counter is checking that you completed your push-up, not that you completed an idealised one.
If you want your depth to be objectively better rather than relatively consistent, that is a training problem, not a counting one. The push-up form guide covers what depth should look like and how to build the range to reach it.
Why the counting runs on your phone
Running pose detection on-device is slower to build than streaming frames to a server, and it is the right call for two reasons. The first is latency: a haptic that fires 300 milliseconds after your rep feels broken, and a round trip cannot beat local inference. The second is that it makes a privacy promise you can actually keep. There is no upload path to audit because there is no upload. See on-device rep counting for the full account of what is transmitted.
Where AI rep counting still loses
- Poor light. The pose model degrades in the dark before anything else does.
- Bad framing. A phone lying flat on the floor sees a foreshortened blur. It needs to stand up.
- Baggy or low-contrast clothing. Joint estimation gets worse when the model cannot find your limb edges against the background.
- Very fast partial reps. Bouncing at speed produces a signal that looks less like a push-up than you would expect, and gates may reject it.
- Multiple people in frame. The counter calibrates to one body per session.
None of these are fixable by claiming better AI. They are fixable by standing your phone up properly and turning a light on.
Where the counted reps go
The counter is a component, not the product. Every verified rep clears ground in the dojo campaign, pays XP toward a samurai title, and adds to your season total on the ranked leaderboard, which is what turns an AI rep counter into a push-up game instead of a novelty. The push-up counter app page covers the counting end to end, and clans and seasons covers what happens to the number afterwards.
Workout Samurai: Push-Ups is out now on iOS and Android, free to download and free to play. If you want to try the counter against your own framing rather than read about it, that is the shortest route: get the app and run one set.
Frequently asked questions
How does an ai push up counter work?
An AI push-up counter runs a pose-estimation model on each camera frame to locate your body's joints, then analyses how those joint positions move over time. When the motion shows a genuine descent followed by a genuine ascent that clears a depth threshold, it counts a rep. The 'AI' is the joint detection; the counting itself is signal processing.
Does an AI push-up counter count half reps?
A well-built one does not. Workout Samurai requires each rep to clear a depth floor calibrated to your own range before it counts, so a shallow bounce is rejected. The depth floor is relative to your calibrated range rather than an absolute chest-to-floor standard, which is what makes it fair across different body proportions and mobility.
Is AI rep counting accurate?
Accuracy is dominated by camera setup rather than by the model. Well-framed, evenly lit, phone standing upright two to three feet away, whole upper body visible. Under those conditions counting is reliable. Cropped, dark or side-on framing degrades it sharply, because the pose model cannot estimate joints it cannot see.
Can you cheat an AI push-up counter?
On the device, deliberately so: the counter is built to be generous rather than to drop honest reps. That is why scoring is separate. Every ranked session is validated on the server, physically impossible input is rejected, and suspicious sessions are flagged for human review. Cheating the count does not translate into rank.
Does an AI push-up counter need internet?
Not to count. The pose model and all the rep logic run locally on your phone, so counting works with no connection at all. A ranked session does need a live connection, because the server issues and verifies it. That is what keeps the leaderboard honest.
Which camera should I use, front or rear?
The front camera, facing you, is what Workout Samurai's counting is tuned for. The phone stands in front of you and you face it. The rear camera generally means you cannot see the screen or confirm your framing before starting, which makes setup mistakes much more likely.
Can an AI push-up counter tell diamond push-ups from regular ones?
Workout Samurai does not. It counts standard push-ups, and it does not classify variations as separate movements. Hand-position classification is a different and considerably harder problem than rep counting, and claiming it without doing it would be dishonest.
Why did the app miss some of my push-ups?
Almost always framing or light. If a shoulder, your head or your hands leave the frame, or if the room is dim enough that joint estimation degrades, the counter loses the signals it needs. Reposition the phone so your whole upper body and both hands are visible, and add light before you blame the model.
Does the AI push-up counter work with a phone case or on a tripod?
Both are fine, and a small tripod is genuinely the easiest way to get consistent framing. Anything that holds the phone upright and steady at roughly floor level, facing you, will do. A wall, a water bottle and a book all work.
Is on-device AI worse than cloud AI for rep counting?
Not for this task. Pose estimation is exactly the kind of workload modern phones run well, and doing it locally removes the round-trip latency that would make the per-rep haptic feel broken. The trade-off is engineering effort, not accuracy, and it is what makes it possible to promise that no frame leaves the device.
Which push-up app has the best AI rep counter?
Judge one by what it will tell you about itself. An app that names the pose framework it uses, says which signals it fuses, states where it fails, and admits it counts only standard push-ups is describing something real. Workout Samurai is on iOS and Android, free to download, so the cheapest test is one set.
Do I have to pay for the AI push-up counter?
No. The camera counter is part of the free game, alongside the dojo campaign, samurai titles, season divisions and clans. Dojo Premium is an optional subscription and it changes nothing about how reps are counted or scored. Leaderboard rank is earned the same way whether you subscribe or not.