Building Proprio 🏋️‍♂️⌚

Part 0 of my first build-in-public series: turning a machine learning research project into a real smartwatch product.

Where this started

This project builds on a study by Dave Ebbelaar, written for his master's project at the Vrije Universiteit Amsterdam [1]. It asked a question I really liked: could a sensor on your wrist do part of a personal trainer's job? Using data from five people lifting in a real gym, he built a pipeline that works out two things:

1. which barbell exercise you're doing: bench press, squat, deadlift, overhead press, or barbell row, or whether you're just resting;
2. how many reps you did.

The first results looked great: a random forest scored 98.51% on sets it hadn't trained on. But the report didn't stop there. When the model was tested on a person it had never seen, accuracy dropped to 85.43%, and the author concluded that "the models have some problems generalizing to new data" [1].

That sentence is where Proprio starts. A product only matters for people the model has never seen. I wanted to actually wear this thing, on my own watch, in my own gym, and eventually ship it as a real product. This series is the story of that journey, from the original pipeline to the finished product.

I built the pipeline myself, following the YouTube series Ebbelaar later made from the project [2]. My best model scored 99.9% accuracy. That's higher than anything in the original report, on the same data, and it made me suspicious rather than happy. A number that good deserves a hard second look before I build anything on top of it. That gives me my first question:

Research Question 1: How accurate is the original pipeline on people it has never seen, when tested fairly?

Two more questions come up along the way.

In this part, I'll walk you through the original study and how its pipeline works, the gaps between a study and a product, who actually needs a product like this, and the development roadmap.

The original work

Before going any further, I want to give proper credit, because the original pipeline stands on a few layers of earlier work.

Ebbelaar collected the data himself, during real gym workouts, for his master's assignment at the Vrije Universiteit Amsterdam [1]:

• Participants: five people (himself included), wearing a MetaMotion wristband sensor from MbientLab [4].
• Exercises: the five lifts from the Starting Strength program.
• Sets: heavy sets of about 5 reps and medium sets of about 10.
• Weights: chosen relative to each person's estimated one-rep max.
• Rest: some resting periods were recorded too.

In its own words, the study set out to explore whether context-aware applications could "just like human personal trainers, track exercises, count repetitions, and detect improper form."

The original pipeline follows the course Machine Learning for the Quantified Self by Mark Hoogendoorn, and the book of the same name by Hoogendoorn and Funk [3]. A few of its helper modules come straight from the book's public code (ML4QS). Ebbelaar later walked through the whole pipeline in an eight-part YouTube series [2], which is how I built my own version.

The report tested its models three ways, and the harder the test, the lower the number:

• training on the first two sets of each exercise and testing on the rest: 98.51% with a random forest;
• training on heavy sets and testing on medium sets (and the other way round): about 80%;
• leaving one participant out at a time and testing on them: 85.43% on average.

So here's the honest starting point: a thoughtful study that was upfront about its limits, and two numbers that just don't fit together. The original report got 85% on people the model had never seen. My version of the same pipeline got 99.9%.

How the original pipeline works

If you're only here for the product story, feel free to skip ahead to "From study to product".

Here's the whole pipeline, from raw files to a predicted exercise:

Figure 1. The original pipeline, end to end. The rep counter works on the resampled data, separately from the models.

In plain words: the sensor files get merged and put on a common clock of five readings per second, and the obvious glitches are removed. Each moment is then described by a set of features, which are just numbers that summarize the last few seconds of motion, like the average acceleration or how rhythmically the signal repeats. Five classic models learn to map those features to an exercise. Separately, a rep counter (no learning involved at all) smooths the signal and counts the peaks.

Every step here is pretty standard for a course project. But keep two details in the back of your head, because they'll matter later. Neighbouring rows share most of their data, since each one summarizes nearly the same few seconds. And the "correct" rep counts come from the set type (heavy = 5, medium = 10), not from anyone actually counting. I'll come back to both in Part 1.

From study to product

A study is designed to answer a question. A product has to survive real people using it. And between those two, there's a long list of gaps:

• A different sensor. Nobody's buying a MetaMotion research band to go lift weights. People wear smartwatches, which come with different units, rates, axes and fit on the wrist. Part 1 gets into what that means for the data.
• New people. A product only matters for people the model has never seen before. The original report already found that results drop when the test person is new [1].
• New conditions. Different gyms, bars and grips. Walking between sets. Checking your phone mid-workout. All of it ends up in the data.
• An unchecked number. My 99.9% came out of one evaluation, run once. Before I trust it, I want to know exactly how the training and test data were split, and whether anything about the test made it easier than real life.

That last one is the first job, and it's exactly question 1. If the number doesn't hold up, everything I build on top of it inherits the problem.

But even if it does hold up, the first gap is still there. A model that only works on the sensor it was trained on can't become a product, which raises the second question:

Research Question 2: Can a model trained on research-sensor data work on a consumer smartwatch, once units, axes and sampling rates are converted?

Who actually needs this

Before putting months into this, I had to ask myself who would actually need it. I started out wanting an automatic exercise sets and reps logger for lifters like me. Turns out, that already exists in some forms. What I couldn't find was a watch that tells you what actually happened in a workout, beyond a rep count.

What a rep count doesn't tell you

Once I started looking past rep counts, four questions came up, each from a different kind of person.

1: Clinicians: did the prescribed exercise actually happen? Physiotherapists and coaches prescribe strength work that clients do at home, unsupervised. A 2021 paper sums up earlier research: "patient non-adherence with rehabilitation programs is 50%-70%" [5]. A 2023 survey of 149 clinicians found that the most common barrier to measuring adherence was simply not having a tool to do it [7]. Wrist accelerometers also show up as adherence measures in clinical trials [8]. But there's no product a clinician can just go and buy.

2: Remote trainers: what did the client actually do? Online coaching runs on what clients log themselves, and self-reports run optimistic. In a 2018 study, people with knee pain kept an exercise diary while an accelerometer hidden in their ankle weight counted what they actually did. The diaries recorded a median of 220 exercises; the hidden sensor counted 176 [6]. Nobody was lying badly, but the diaries were consistently optimistic.

Figure 2. Diaries overstate what people actually do. Data from Nicolson et al. (2018) [6]; the accelerometer was hidden in an ankle weight.

3: Lifters: how fast is the bar moving? Velocity-based training uses bar speed to guide how hard a set is, and it usually needs a sensor mounted on the bar. In a 2023 study, a wrist-worn Apple Watch estimated mean barbell squat velocity about as well as bar-mounted sensors [15]. There's early work on estimating effort too, though only as a preprint so far [17]. Reviews of commercial training sensors also warn that validation is often weak [16], which is a good reason to measure things carefully instead of trusting the marketing.

4: Anyone following a prescribed tempo: was each rep done at the right speed? Tempo prescriptions (for example, three seconds down and one second up) are common in tendon rehab and tempo training, but nobody checks them except the person counting in their head.

That's the opening. The research says a wrist can measure a lot more than reps, but none of it is in a product you can buy. That's the gap Proprio is built around, and it leads to the third question:

Research Question 3: Can a wrist-worn watch measure what actually happened in a workout (sets, reps, rest, bar speed, tempo) well enough for lifters, coaches and clinicians to rely on it?

Meet Proprio

Proprio (short for proprioception, your body's sense of its own position and movement) is my attempt at turning that study into something real. It's for lifters, and for the clinicians and coaches who prescribe their training. It has five parts:

1. A smartwatch records the raw motion of your whole workout.
2. A mobile app lets you label sets, fix mistakes, and see your history.
3. Backend services store the raw recordings, convert them, recognize exercises, count reps, and send the results back to your phone.
4. A web app for clinicians and trainers lets them prescribe exercises and see what their clients actually did.
5. An AI coach provides insights into your training in plain language, based on your own recorded history instead of generic advice.

I'm starting with a Galaxy Watch and an Android phone, simply because that's the watch I have to test on. An Apple Watch and iPhone app will come once the Android features are done. And since the raw recordings get converted on the server, supporting a new watch mostly means describing its sensor (units, axes, rates) and running a calibration session, not training a new model. That's the bet behind question 2.

The watch records exactly what its sensors measured, unconverted. Every conversion happens later, in one shared piece of code, so if it turns out to be wrong, it can be fixed and re-run.

The roadmap

Stage 1: prove it works (Phases 0–5)

Stage 2: Make It Useful

This is my answer to "who actually needs this?". Each phase takes one of the four questions from that section: clinicians, lifters, remote trainers, and tempo.

Stage 3: Grow It

Limitations

No results yet. Everything in this part is a plan. The numbers will come in later parts, including the ones that don't go my way.

A small dataset. The original data comes from five people, all doing the same five lifts. That's enough to learn from, but not enough to prove a product works for everyone.

One platform and one wrist to start. Proprio starts on Galaxy Watch and Android, worn on the left wrist. Other watches, and the other wrist, need their own testing.

What's next

The first job is the one I started with. The original report got 85% on people the model had never seen. My version of the same pipeline got 99.9% from the same data. They can't both be describing the same reality.

So in Part 1, I'm taking my implementation apart, making it reproducible, and finding out which number to believe.

I'll be writing one part per phase as I build it, and that includes the things that fail. If you're working on something similar, or just curious how a research project actually turns into a product, follow along!

References

1. Ebbelaar, D. Exploring the Possibilities of Context-Aware Applications for Strength Training. Mini Master Project, Vrije Universiteit Amsterdam. (Linked from the tutorial's source material.)
2. Ebbelaar, D. Full Machine Learning Project — Coding a Fitness Tracker with Python (video series, 8 parts). https://youtube.com/playlist?list=PL-Y17yukoyy0sT2hoSQxn1TdV0J7-MX4K
3. Hoogendoorn, M. & Funk, B. (2018). Machine Learning for the Quantified Self: On the Art of Learning from Sensory Data. Springer. https://link.springer.com/book/10.1007/978-3-319-66308-1 · Code: https://github.com/mhoogen/ML4QS/
4. MbientLab. MetaMotion sensor. https://mbientlab.com/metamotions/
5. Picha, K. et al. (2021). Physical therapists' assessment of patient self-efficacy for home exercise programs. International Journal of Sports Physical Therapy, 16(1), 184–194. https://doi.org/10.26603/001c.18957 (the 50–70% figure is from the paper's introduction, summarizing earlier studies)
6. Nicolson, P. J. A. et al. (2018). Self-reported home exercise adherence: a validity and reliability study using concealed accelerometers. Journal of Orthopaedic & Sports Physical Therapy, 48(12), 943–950. https://doi.org/10.2519/jospt.2018.8275
7. McArthur, C. et al. (2023). Older adults, clinicians, and researchers' preferences for measuring adherence to resistance and balance exercises. BMC Geriatrics, 23, 530. https://doi.org/10.1186/s12877-023-04237-x
8. ClinicalTrials.gov. Adherence in Upper Extremity Home-based Rehabilitation (NCT05032638). https://clinicaltrials.gov/study/NCT05032638
9. Samsung Community. "Galaxy Watch 4 Classic rep tracking is not very accurate" (user forum thread). https://eu.community.samsung.com/t5/wearables/galaxy-watch-4-classic-rep-tracking-is-not-a-very-accurate/td-p/4955247
10. TechRadar. "Why Garmin's strength training mode needs to be improved – or scrapped." https://techradar.com/features/why-garmins-strength-training-mode-needs-to-be-improved-or-scrapped
11. the5krunner (2026). "New WHOOP Strength Trainer update: Passive MSK explained." https://the5krunner.com/2026/02/28/new-whoop-strength-trainer-update/
12. Motra (formerly Train Fitness), App Store listing. https://apps.apple.com/us/app/motra-formerly-train-fitness/id1548577496
13. Burns, D. M. et al. (2018). Shoulder physiotherapy exercise recognition: machine learning the inertial signals from a smartwatch. arXiv:1802.01489. https://arxiv.org/abs/1802.01489
14. Burns, D. M. et al. (2020). Adherence tracking with smart watches for shoulder physiotherapy in rotator cuff pathology: protocol for a longitudinal cohort study. JMIR Research Protocols, 9(7), e17841. https://doi.org/10.2196/17841
15. Achermann, B. et al. (2023). Velocity-based strength training: the validity and personal monitoring of barbell velocity with the Apple Watch. Sports, 11(7), 125. https://www.mdpi.com/2075-4663/11/7/125
16. Weakley, J. et al. (2021). The validity and reliability of commercially available resistance training monitoring devices: a systematic review. Sports Medicine, 51, 443–502. https://doi.org/10.1007/s40279-020-01382-w
17. Thomas, J. & Wahlström, J. (2025). Estimation of resistance training RPE using inertial sensors and electromyography. arXiv:2510.03197 (preprint). https://arxiv.org/abs/2510.03197