Training on your own comes with a familiar problem: you can't see yourself. Is your back rounding on the deadlift? Are your knees caving in on the squat? Without a mirror, a training partner or a coach, you're flying blind β and in the worst case you grind in mistakes you only notice once they start to hurt. A South Korean research team has now tested whether a smartphone app with AI-powered real-time movement recognition can help here.
Smartphone spots form errors β and beginners stick with it
Christopher Klenk7 min readAt a glance
146 young adults without gym access trained for 16 weeks with an app that analyses their exercise form in real time via the smartphone camera. The gains in strength and body composition match what you would expect from beginners on a structured programme. More interesting: the movement recognition agrees 95.8 % with physiotherapists' assessments β and participants stuck with an unsupervised home programme remarkably consistently.
The underlying problem: training alone means training blind
Most people who train at home or without a coach have no reliable feedback on their exercise form. A mirror only gives one perspective and distracts during the movement. Filming yourself and reviewing the footage afterwards is cumbersome and doesn't help in the moment the mistake happens. The result: many train with suboptimal technique without noticing β and build in movement patterns that are at best less effective and at worst lead to overuse problems.
This is exactly where the technology examined in this study comes in. The app uses Google's MediaPipe β an open-source framework for body recognition β to capture 33 body landmarks in real time via the smartphone camera. From these it calculates joint angles and compares them with reference values for correct exercise form. If it detects deviations, you get immediate feedback: visually as a colour-coded skeleton overlay (green for clean form, red for errors), along with text prompts and voice cues. Everything runs directly on the smartphone β no internet connection needed, no data gets uploaded.
WHAT IS MEDIAPIPE?
MediaPipe is an open-source framework from Google. Its pose-estimation component (BlazePose) detects 33 three-dimensional body landmarks in real time on mobile devices β without cloud connectivity. The technology is already used in various fitness and AR applications and is freely available.
What the study looked at
The team led by Seoyoon Heo (Kyungbok University, South Korea) had 216 young adults (mean age 23.8 years, 69 % male) train with the app over 16 weeks. Participants had not done regular strength training in the previous six months and had limited access to gyms. The programme consisted of three 30-minute sessions per week with seven basic exercises: squat, bench press, deadlift, barbell row, military press, pull-ups and dips. Participants provided their own equipment β barbell or dumbbells. 146 people (67.6 %) completed the programme in full.
The study used a pre-post design without a control group. That means: all participants used the app, and there was no comparison group with conventional training or no app at all. The changes therefore can't be attributed unambiguously to the AI component β they could just as well be the result of the structured training itself.
The results
The training outcomes were in line with what you'd expect from previously untrained participants on a progressive strength programme:
Metric | Before β after | Effect size |
|---|---|---|
1RM squat | 34.67 β 39.06 kg (+12.7 %) | d = 1.148 (large) |
Body fat | 23.3 β 20.0 % (β3.3 pp) | d = β1.373 (large) |
Skeletal muscle mass | 37.8 β 40.5 % (+2.19 kg) | d = 1.433 (large) |
FMS score | +0.29 points (median at 14) | d = 0.285 (small) |
VO2max | 39.2 β 40.8 mL/kg/min (+4.8 %) | d = 0.917 (large) |
The numbers are solid but not surprising. Meta-analyses show that beginners on progressive resistance training typically gain 10β15 % in strength over comparable time frames β regardless of whether they train with an app, a coach or a fixed paper plan. What this study does show: app-based training doesn't produce worse results. Whether it delivers better results than training without movement analysis can't be answered here, because the comparison is missing.
Where it actually gets interesting
Two aspects stand out, beyond the expected strength gains.
First: people stuck with it. Among participants who completed the programme, average adherence was 88 % β that is, 42 out of 48 scheduled sessions. 76.7 % completed more than 85 % of all sessions. For a 16-week home-training programme without a personal trainer, that's a solid figure. Whether the real-time form feedback, the app's gamification elements (achievements, streak counter, progress dashboard) or simply taking part in a study is responsible can't be separated out. It's plausible that the combination is what works: immediate feedback on movement quality builds confidence, while gamification provides structure and small sense-of-progress wins. The study can't isolate the effect of the individual components, though.
Second: the movement recognition works. The app's pose classification agreed 95.8 % with physiotherapists' assessments (ICC = 0.94). Key-point detection reached 97.2 % accuracy at 28.6 milliseconds of processing time β fast enough for smooth real-time feedback. For a pure smartphone-camera solution without additional sensors, that's impressive. In plain terms: the technology can now tell whether you're doing a squat cleanly or whether your knees are clearly caving in β and flag it at the moment it happens. That said, the detection works with defined angle thresholds: gross deviations are reliably picked up, while subtle faults can slip through.
Especially for beginners, who don't yet have a fine-tuned sense of movement patterns, that's practically relevant. If you're deadlifting for the first time and have no one watching your spinal position, you at least get a basic safety net. It's not a substitute for an experienced coach, who also catches things a 2D camera can't see β but it's a clear improvement over completely unsupervised training. That camera-based movement analysis works beyond pure form feedback is also shown by the BackTracker study on back-pain classification: there, a neural network identifies the pain type from a simple video with 92 % accuracy β based on the same underlying technology.
LIMITS OF THE TECHNOLOGY
A smartphone camera works in two dimensions. When body parts occlude each other β for example when the torso hides the hip during a deadlift β accuracy in the study dropped by 4.3 %. Poor lighting or unfavourable camera angles are likely to make this worse in practice. No 2D system can reliably pick up subtle compensatory patterns or individual limitations. On top of that there's a mundane practical issue: the smartphone has to be set up stably, at the right distance and angle to the exercise. The study doesn't describe how participants solved that β the measured accuracy of 97.2 % comes from controlled conditions, not from a phone propped up against a water bottle.
What we don't know
Several things remain open. The study has no control group β we don't know whether adherence with the app is higher than without. The 32.4 % drop-out rate (of which 10.6 % due to Korean military service) puts things in perspective. The sample was young, predominantly male and limited to a Korean university context. How the technology performs with older trainees, with more advanced lifters or under less controlled conditions is unclear.
The app itself isn't publicly available β it's a research application. Whether and when a commercial product will emerge from it, the authors don't say. There's also no follow-up: we don't know whether participants continued training after the study ended. The authors themselves call for a randomised controlled trial with a comparison group and a more diverse population as the next step.
What you can take away
The study doesn't prove that AI-powered apps deliver better training results than conventional training. What it does show: real-time movement-recognition technology via smartphone is now mature enough to analyse strength exercises at a level approaching professional assessment. And it appears β at least in this setting β to help beginners train regularly over a meaningful period. Movement analysis is therefore one of the areas where AI in fitness already delivers concrete practical value today.
MediaPipe is freely available as open-source technology and runs on off-the-shelf smartphones. It's entirely plausible that we'll see this kind of movement analysis in a lot more fitness apps over the coming years. If you train at home and don't have a training partner, apps like this could help you catch at least the obvious form errors early β before they get ingrained. For advanced lifters, it doesn't replace individual coaching, but as an extra set of eyes on your form for standard exercises, the technology is genuinely worth a look.
Source
Heo, S.; Choi, T.; Choi, W. (2026): Clinical Validation of an On-Device AI-Driven Real-Time Human Pose Estimation and Exercise Prescription Program; Prospective Single-Arm Quasi-Experimental Study. Healthcare, 14(4), 482. doi.org/10.3390/healthcare14040482 (Open Access)
