IMU Magnetometer Interference Indoors: Why Steel Gyms Cause Heading Drift

A strength coach we work with runs rotational power testing in the corner of a facility that used to be a loading dock — exposed steel columns on two sides, a rack of plates stacked against the wall, a wall-mounted speaker eighteen inches from the athlete's setup spot. For three weeks his medicine-ball throw heading numbers wandered session to session for no reason tied to the athletes: same kid, same throw, same stance, and the app's rotational bearing read 12 degrees off from one Tuesday to the next. He assumed the sensor was failing. It wasn't — it was doing exactly what a magnetometer does next to a wall of steel, reporting a magnetic north that had nothing to do with true north.

This gets confused with ordinary IMU drift constantly, but it isn't the same failure. Gyroscope drift accumulates over time regardless of environment; magnetometer interference is spatial, not temporal — it shows up the moment you're near a distorting object and vanishes the moment you step away, which is why a heading reading can be rock-solid in one corner of a gym and unusable six feet to the left. Below is the mechanism, two controlled studies that quantified it in exactly this kind of indoor setting, and the floor-mapping protocol that turns an unpredictable sensor into a predictable one.

A tilt-compensated compass heading — the number that tells a rotational asymmetry test or a change-of-direction metric which way the athlete's torso or hips are pointed — is built from three sensors, not one. The gyroscope tracks rotation rate, the accelerometer establishes down, and the magnetometer measures the local field vector to establish north. Fusion combines all three: gyroscope for smooth, fast, short-term rotation; magnetometer as a long-term reference that keeps the gyroscope's inevitable drift from wandering off over a session.

That last part is the trap. The magnetometer is only useful as a reference if the field it's measuring actually points at magnetic north. Earth's field at ground level runs roughly 25 to 65 microtesla depending on latitude, in a specific, locally predictable direction — and a magnetometer simply assumes it's reading that field. It has no way to tell Earth's field apart from a local field bent by a steel I-beam, a stack of cast-iron plates, or a running treadmill motor; it reports whatever vector it measures as if it were true north, and fusion trusts that report to correct the gyroscope. When the local field is distorted, the correction itself becomes the error source, and a persistent one rather than a random one, because fusion keeps pulling the heading back toward the wrong reference every cycle.

Outdoor open-field testing rarely runs into this because there's usually nothing ferrous close enough to matter. A commercial gym or a repurposed warehouse space is close to the worst-case environment a magnetometer can be asked to work in, for a specific reason: it combines structural steel with a dense cluster of ferrous training equipment and electronics, all at exactly the 0.5 to 3 meter range where distortion is strongest.

The distortions from these sources don't average out over a session — they're fixed to a location, so an athlete who does five reps of a rotational drill in the same spot gets the same wrong heading five times, which looks exactly like a reliable, repeatable measurement right up until someone moves the setup ten feet and the number changes.

Two studies, run about fifteen years apart with different goals, map onto this problem directly. Neither was done in a weight room, but both were done in the kind of indoor facility — motion-capture labs full of steel rigging and equipment — that behaves magnetically like a gym.

de Vries, Veeger, Baten, and van der Helm (2009, Gait & Posture) checked whether motion-analysis labs, which routinely use magnetic-inertial sensors to validate other motion-capture systems, were themselves magnetically clean enough to trust. Mapping the local field at a grid of points across two working labs, they found it was far from uniform: near steel structural elements, force plates, and lab equipment, the measured direction deviated from the expected geomagnetic direction by amounts large enough to produce clearly unusable heading estimates, while other zones in the same room stayed close to expected with negligible error. Their conclusion was blunt — a room isn't magnetically clean just because it looks like open floor, and every facility needs field-mapping zone by zone before trusting magnetic-inertial sensors for orientation work. The limitation: the numbers are tied to those two labs' construction, so a different building has a different footprint — the takeaway is the mapping method, not a universal safe distance.

Roetenberg, Luinge, Baten, and Veltink (2005, IEEE Transactions on Neural Systems and Rehabilitation Engineering) attacked the same problem from the correction side, building a Kalman-filter fusion algorithm that continuously checks the magnetometer's measured field magnitude and inclination against expected values and down-weights its influence whenever the reading falls outside tolerance — letting the gyroscope run uncorrected during a detected disturbance rather than letting a corrupted reading actively drag the heading wrong. Tested near a known steel source, standard fusion produced heading errors on the order of tens of degrees during the disturbance, while the disturbance-rejecting version held errors to roughly single digits over the same window. The trade-off: during rejection the algorithm runs on gyroscope integration alone, so a long enough disturbance eventually reintroduces ordinary drift, just later and slower.

Read together, the two papers say the same thing from opposite ends: the disturbance is real and location-specific (de Vries), and an algorithm that knows when to stop trusting the magnetometer recovers most of the accuracy (Roetenberg) — but neither one, and no algorithm since, makes the underlying physics go away. The fix is still spatial before it's computational.

You don't need a physics lab to find out whether your training floor is magnetically clean. You need a tape measure, a notebook, and about fifteen minutes the first time you set up a rotational testing station, then a re-check any time the equipment layout changes.

Start in what you think is the cleanest open area of the floor, at least 3 meters from any rack, plate stack, or speaker. Hold the sensor still and rotate it through a slow, level 360-degree turn over about 10 seconds, comparing the heading trace against the known rotation — it should climb smoothly and land back within a few degrees of the start. This is your reference reading for what clean looks like in your building, since even open floor in an older facility can carry some rebar signature.

Next, walk the sensor from that clean spot toward your intended testing zone in half-meter steps, holding orientation constant at each stop and logging the heading. In a clean zone the number won't move; near a distortion source it drifts steadily as you close the distance — that distance-correlated drift, rather than a sudden jump, is the signature of magnetic interference rather than a sensor fault. Repeat toward the squat rack, the speaker, and any wall-mounted electronics near your usual setup, and mark the point where drift first becomes noticeable.

The point of this exercise isn't a one-time certificate. Gyms rearrange equipment constantly — a new rack gets bolted where a bench used to be, a speaker gets remounted — and each change redraws the distortion map. Treat the floor map as something you check quarterly and re-verify any time furniture moves near a testing station, not something you set once and forget.

Once you know where the clean and distorted zones sit, the fix is mostly discipline rather than technology. First, run a standard figure-eight calibration — two slow sweeps of the sensor over about five seconds — in a zone you've already confirmed clean, never near the rack or speaker. Calibrating in a distorted spot bakes the distortion into the sensor's internal reference, making every later reading wrong in a consistent, hard-to-notice way — a more insidious failure than skipping calibration entirely.

Second, physically zone your testing floor. Mark a clean rotational-testing area at least 1.5 meters from any steel rack, plate stack, or speaker, based on your own walk-test results rather than a generic rule of thumb — construction varies enough that 1.5 meters is a starting point, not a guarantee. Run every heading-dependent drill — rotational power throws, change-of-direction bearing, hip-shoulder separation angle — inside that marked zone, and keep pure linear tests like vertical jump or sprint velocity, which don't depend on magnetometer heading, wherever is convenient.

Third, when a clean zone genuinely isn't available — a small facility where every corner sits within a meter of steel — switch heading calculation to gyroscope-and-accelerometer-only mode and reset the reference orientation manually at the start of every set by having the athlete square up to a fixed wall landmark. This sacrifices the long-term correction the magnetometer would normally provide, so expect roughly a few degrees of accumulated drift per minute of continuous tracking, based on uncorrected error rates from the fusion literature — small enough to ignore across a single 15-to-30-second set, worth resetting between sets rather than trusting across a full session.