Step counts from nine wear positions, worn simultaneously, against a mechanical hand tally
Placement beat brand by more than an order of magnitude: hip and front-pocket positions measured 1.2% to 2.3% MAPE against the tally, while every wrist position measured 16.2% to 21.8% and lost 19.8% to 47.8% of steps whenever the arm was occupied.
Run parameters
Everything needed to repeat this
| Run id | RP-RUN-2026-02 |
|---|---|
| Tested on | 2026-08-20 |
| Samples | 54 walking sessions |
| Method | Step counting against a hand tally (M-05), nine positions worn simultaneously across six walking conditions, plus six seated sessions at a true count of zero |
| Reference | The observer's tally. 61,412 footfalls counted by hand across the 54 walking sessions |
| Instrument | Mechanical hand tally, one press per footfall, pressed by an observer wearing none of the devices under test |
| Measured | Step count, absolute error (% MAPE) · Step count, signed bias (%) · Steps logged at a true count of zero (steps/hour) |
| Tolerance | Two positions are called different only when their bootstrap 95% intervals do not overlap (M-03). Inter-observer counts had to agree within 0.5% or the session was repeated. |
| Licence | CC BY 4.0 |
Under test
Bought at retail · 8 subjects
| # | Subject | Type | Official site |
|---|---|---|---|
| 1 | Apple Watch Series 11 | Device | apple.com |
| 2 | iPhone 17 (Apple Health) | Device | apple.com |
| 3 | Google Pixel 10 (Fitbit app) | Device | store.google.com |
| 4 | Fitbit Charge 6 | Device | fitbit.com |
| 5 | Garmin Venu 4 | Device | garmin.com |
| 6 | Samsung Galaxy Watch8 | Device | samsung.com |
| 7 | Yamax Digi-Walker SW-200 | Device | yamax.co.uk |
| 8 | Yamax Power-Walker EX-510 | Device | yamax.co.uk |
What we did
Rig RP-S2 was built and calibrated on 2026-06-11. Devices were bought at retail between 2026-06-12 and 2026-06-18, updated on 2026-06-18, and worn at nine positions at once so that every position saw the same walk, the same floor and the same cadence. One adult walked all 54 sessions. Wrist positions were rotated between sessions.
Six conditions, nine sessions each: treadmill at 1.34 m/s, an outdoor brisk loop at 1.4 m/s, a slow indoor walk at 0.70 m/s, and three conditions in which the hands were occupied — an 18 kg loaded cart pushed with both hands, a 6 kg bag in the dominant hand, and one hand on the treadmill rail. The last three are the reason the run exists. They are ordinary walking and they are not in anybody’s validation set.
Treadmill speed was checked by marked-belt revolution count against a stopwatch, ±0.02 m/s. The first and last session of each condition was counted by two observers.
Nothing on this page is a manufacturer’s figure. All of it came off the tally.
Results
| Position | Error, % MAPE | Signed bias, % | 95% CI | Reference steps |
|---|---|---|---|---|
| Yamax Power-Walker EX-510, hip | 1.2 | −1.2 | 0.9 to 1.5 | 61,412 |
| iPhone 17, front trouser pocket | 1.5 | −1.5 | 1.2 to 1.9 | 61,412 |
| Pixel 10, front trouser pocket | 2.3 | −2.3 | 1.8 to 2.8 | 61,412 |
| Yamax Digi-Walker SW-200, hip | 4.7 | −4.7 | 3.1 to 6.4 | 61,412 |
| iPhone 17, carried in the hand | 9.8 | −9.8 | 8.1 to 11.6 | 61,412 |
| Garmin Venu 4, wrist | 16.2 | −16.2 | 13.2 to 19.3 | 61,412 |
| Apple Watch Series 11, wrist | 18.6 | −18.6 | 15.4 to 21.8 | 61,412 |
| Samsung Galaxy Watch8, wrist | 20.2 | −20.2 | 16.9 to 23.6 | 61,412 |
| Fitbit Charge 6, wrist | 21.8 | −21.8 | 18.3 to 25.4 | 61,412 |
n = 54 walking sessions per row, 2026-06-22 to 2026-07-24. Signed bias equals absolute error in every row: while walking, no position over-counted in any condition.
Under M-03 the tolerance above forbids three rankings this table looks like it supports. The hip-worn EX-510 and the pocketed iPhone are a tie. The two pocketed phones are a tie. And the four wrist devices cannot be separated from each other at all — 16.2% and 21.8% is not a ranking, and nobody quoting this run is entitled to make one out of it.
What the intervals do support is the gap between placements. Pocket and hip against wrist is more than an order of magnitude and no interval comes close to closing it.
Where the error came from
Placement, not sensing. In the three conditions with a free arm, the wrist devices were within 2.2% of the tally, which is a working accelerometer doing its job. The error appears when the arm stops: cart, bag and handrail cost the wrist positions between 19.8% and 47.8%. A wrist counter infers footfalls from arm swing and there is no swing left to read.
The two failure modes that are not about arm swing are worth separating. The spring-lever SW-200 was the most accurate device in the set at both walking speeds and then missed 24.6% of steps at 0.70 m/s, because a pendulum arm needs a vertical impulse a slow shuffle does not produce. And the six seated sessions, at a true count of zero, ran the other way: the wrist devices logged a median of 143 to 258 phantom steps an hour against 7 to 23 for the hip and pocket positions. Sensitivity buys the slow walk and pays for it at the sink.
Device-by-device reading, the per-condition table and the seated results are in the report: how accurate are step counters.
What we got wrong in this run
The first four cart sessions were discarded. They were run with an empty cart, which on a smooth floor gets pushed one-handed with the other arm swinging — not the case we set out to measure. They were re-run loaded to 18 kg with both hands on the handle, and the discarded sessions are in no figure above.
One inter-observer pair disagreed by 0.63%, over the 0.5% M-05 permits, on an outdoor session. That session was repeated. The other pairs agreed to within 0.21% and 0.34%.
What this does not show
One walker. All 54 sessions were walked by the same adult, with one gait, one leg length and one arm-swing amplitude. Gait varies enormously between people. Read the spread between placements as the result; do not read any single absolute figure as yours.
One unit of each device. Unit-to-unit variation within a model is unmeasured here, so a figure in the table belongs to the device we bought rather than to the model.
Firmware on one date. Everything was updated on 2026-06-18 and not touched again. A step algorithm can change in an update, and this run cannot see that.
No stair, run or cycle conditions. Six walking conditions and six seated ones. Nothing here says what any of these devices do on a staircase or at running cadence.
The standing limits that apply to every run this desk publishes are in the corrections log, and the procedure is M-05 on methods.
Citing this run
Quote the id, not the adjective
Reference Plate, Step counts from nine wear positions, worn simultaneously, against a mechanical hand tally, RP-RUN-2026-02, n = 54 walking sessions, tested 2026-08-20. Reference: The observer's tally. 61,412 footfalls counted by hand across the 54 walking sessions. Licensed CC BY 4.0.