The cheap air rower that learned to coach.

Gamifying a dumb machine

An air rower is honest but mute: a fan, a cord, a seat. It never tells you whether the stroke was good — and without feedback, training decays into pulling hard. This project is the story of giving that machine a voice, in five chapters.

Three signals, one stroke

Force alone is blind to coordination — it only sees the sum of what legs, trunk and arms do. The sensor choice follows directly from that: the handle position is the sum of the segments, the seat position is the legs alone, so their difference isolates the upper body. That decomposition powers all coordination metrics.

Ghost curves are computed, not drawn

The target curve you chase is not a designer's sketch. It is the force that physics predicts when an ideal body rows this machine at your chosen rhythm.

The quadratic air-drag model and the cube law it implies are the standard physics of ergometers[4]. The shape comes from kinematics: each body segment follows a minimum‑jerk profile — the velocity profile experimentally confirmed for skilled human movement[5] — sequenced the way rowing technique demands: legs 0–60 % of the drive, trunk 30–85 %, arms 55–100 %[1].

You prescribe the rhythm (stroke rate and drive:recovery ratio), never the power — power falls out of the model with the cube law P ∝ R³/d². And because force is blind to coordination, the firmware generates three coupled ghosts: force (how much, what shape), seat (when the legs drive), arms (when the upper body closes). Same clock, three answers.

Shape comes from kinematics; amplitude from the fan; duration from rhythm; watts are an output, not an input.

One cue per stroke, straight from the literature

Every metric is judged against reference bands published in rowing biomechanics research: the force‑curve shape bands and catch‑factor/RSF optima come from Kleshnev's Biorow work[1], the link between stroke technique and performance from Holt et al.[2], and the phase‑normalised curve analysis from Warmenhoven's functional data studies[3]. And like a real coach, the system never shouts five corrections at once: faults are ranked — sequence first, then loading, then shape — and only the highest‑priority one speaks.

The target shape itself is the Beta family φ(u; p,q) on normalised drive fraction — preset Technique (peak 33 %, Φ 0.56) for low rates, Race (peak 40 %, Φ 0.61) for 30+ spm. A Python analyzer replays every session offline with the same model, plus functional PCA: your stroke is compared to your own best modes, not to a population ideal.

The workouts are written by the same equations

Most training libraries are folklore: someone's favourite intervals, copied forever. Here the pipeline runs the other way — every model this page describes was turned into a generator, and the generator wrote the training library: 148 workouts and four multi‑week plans, in EXR and Zwift formats, loaded in the Web‑app below and downloadable from the repo.

And the loop closes on the machine: the coach that grades your strokes uses the same models that wrote the workout you are doing. Browse the library →

This is the actual onboard Web-app

Not a mockup: the frame below runs the very web/index.html the ESP32 serves on the rower, connected to a simulated machine instead of the real sensors. Try every mode — Dashboard with live force, cable and seat traces, the Coach tab with per‑stroke cues and catch factor, History with saved sessions and ghost replay, Setup with thresholds, ghost parameters and athlete profile. Every control speaks the same command protocol as the firmware.

This is what your phone shows when you sit on the machine — here fed by a simulated rower. Prefer a full window? Open the Web-app full-screen → (a small pill takes you back here). Sessions you save land in your browser's local storage, exactly as on the rower.

Racing yourself is the oldest game in rowing

White papers & deep dives

Every design choice was written down before it became code. These documents live in the repo and read as small white papers — key extracts below.

Scientific references

1. Kleshnev, V. — The Biomechanics of Rowing, Crowood Press, 2016; and the Biorow technical newsletters (force‑curve shape, peak position 32–40 %, catch factor, Rowing Style Factor). biorow.com
2. Holt, A. C., Aughey, R. J., Ball, K., Hopkins, W. G., Siegel, R. — “Technical Determinants of On‑Water Rowing Performance”, Frontiers in Sports and Active Living, 2020.
3. Warmenhoven, J., et al. — functional data analysis of rowing force–angle profiles: Journal of Science and Medicine in Sport (2018) and Sports Biomechanics (2017‑19); basis of the phase‑normalised ensemble curves used here.
4. Dudhia, A. — The Physics of Ergometers, Oxford University (online monograph): quadratic air drag F = k·v², drag‑factor independence, the power–pace cube law.
5. Flash, T., Hogan, N. — “The Coordination of Arm Movements: An Experimentally Confirmed Mathematical Model”, Journal of Neuroscience, 5(7), 1985 (minimum‑jerk velocity profiles).
6. Skiba, P. F., Chidnok, W., Vanhatalo, A., Jones, A. M. — “Modeling the Expenditure and Reconstitution of Work Capacity above Critical Power”, Medicine & Science in Sports & Exercise, 44(8), 2012 (W′‑balance).
7. Monod, H., Scherrer, J. — “The Work Capacity of a Synergic Muscular Group”, Ergonomics, 8(3), 1965 (the critical‑power / W′ framework).
8. Banister, E. W. — “Modeling Elite Athletic Performance”, in Physiological Testing of the High‑Performance Athlete, Human Kinetics, 1991 (TRIMP training impulse).
9. Keytel, L. R., et al. — “Prediction of Energy Expenditure from Heart Rate Monitoring during Submaximal Exercise”, Journal of Sports Sciences, 23(3), 2005.
10. Concept2 — “Watts Calculator / How Calories Are Calculated” (technical documentation): the pace–power convention W = 2.80 / pace³ used by all ergometer displays.