An October 14 PX4 pull request provides an unusually concrete example. Its author measured an accelerometer response to commanded heater steps on an ARK FPV board and proposed limiting heater-power changes during flight. The patch was still open at the research cutoff on October 15, so its results are contributor bench measurements, not independent validation or a released fix. The broader engineering problem extends beyond this board: temperature compensation corrects measured sensor behavior over a calibration range; it does not automatically cancel every short-lived mechanical or electrical effect of a nearby heater.
What the ARK FPV measurements actually show
The PX4 contributor describes a 1 W heater placed near an IIM-42653 IMU. Switching it on reportedly moved the accelerometer's Z output by approximately -0.6 m/s² within 0.3 seconds. After about ten seconds, the offset settled to roughly -0.16 m/s² per watt of heater power. Switching the heater off produced a reported +0.55 m/s² transient.
The author says the Z-axis response followed heating rather than the 10 ms power-control PWM. That observation is consistent with a thermal or thermo-mechanical mechanism, but the PR does not independently establish which component changes: the sensor die, PCB strain, package stress, mounting, or some combination. It would be premature to generalize these amplitudes to different airframes or IMUs.
The proposed HEATERn_SLEW parameter limits duty-cycle change per second while armed. On the tested board the author used 0.01/s, equivalent to one percentage point of duty cycle per second. Disarmed startup heating remains unrestricted. In a bench sequence with temperature targets of 10, 45, 42, 45 and 10 °C, the contributor reported accelerometer Z variation over 1-30 seconds falling from 0.041 to 0.0048 m/s² RMS, and the largest excursion from 0.556 to 0.019 m/s².
Those figures describe a single reported test setup. They do not establish how much an aircraft's navigation error would change in flight, or whether a slower heater is acceptable in every ambient temperature.
Temperature calibration solves a different part of the problem
PX4's thermal calibration guidance describes corrections for accelerometer, gyroscope, magnetometer and barometer bias as a function of sensor temperature. The software fits measurements acquired during a temperature sweep to polynomial coefficients. For accelerometers and gyroscopes, the documented offset model uses a third-order polynomial; the relevant temperature is the sensor's own measurement.
That model is valuable when a sensor's bias changes reproducibly as it warms. It becomes less complete when heater power changes much faster than the internal sensor-temperature reading, or when two points on the same PCB are temporarily at different temperatures. Two samples recorded at the same die temperature can have different offset behavior if nearby material is still expanding or cooling.
Calibration also has limits of its own. PX4 warns that active heating elements can affect magnetometer readings and recommends keeping them inactive or sufficiently distant during calibration. Its documented process calls for a stationary board, controlled temperature change, raw logging and a final accelerometer calibration within the covered temperature range. Thermal-compensation coefficients are only as useful as the conditions represented by the calibration data.
ArduPilot's IMU temperature-calibration instructions make another relevant distinction: multiple IMUs inside one autopilot need not reach the same temperature. On certain Cube controllers, a third IMU is farther from the heater and may require a lower calibration ceiling than the other two. One thermostat setpoint is therefore not evidence that every inertial measurement is operating at that temperature.
Why an estimator can care about a short transient
An inertial navigation system integrates acceleration after accounting for attitude and gravity. A biased accelerometer can look like a genuine change in vehicle motion until other measurements constrain the estimate. Whether it actually shifts the vehicle's estimated position depends on the error direction, estimator design, aiding sensors, filtering and duration.
A simple illustration shows the scale without pretending to predict flight behavior. If a constant 0.5 m/s² horizontal acceleration error went uncorrected for one second, straightforward integration would imply 0.5 m/s of velocity error and 0.25 m of position error. Those are uncompensated mathematical values, not measured drift on the ARK FPV board. A vertical-axis heater transient cannot simply be substituted into that calculation as a horizontal navigation error; attitude and gravity projection matter.
For multirotors, heater-load changes may coincide with airflow changes as the vehicle moves or climbs. A controller that responds aggressively to a temperature setpoint can then introduce a sensor disturbance at the same moment the navigation estimator is processing genuine vehicle acceleration. An intermittent bias is harder to characterize than a stable warm offset.
Three interventions and their trade-offs
Thermal design first. Physical separation, less concentrated heat flux and mechanically stable mounting may reduce coupling between heater power and the IMU. These are engineering options, not proven fixes for the particular ARK FPV measurements. They should be evaluated against board volume, warmup time, vibration isolation and power consumption.
Calibrate across the intended conditions. Both PX4 and ArduPilot document stationary cold-to-warm calibration. Coverage should include the temperatures the aircraft will actually reach, while acknowledging that a slow static sweep may not reproduce rapid heater-power transitions. Calibration is not a substitute for investigating transients.
Constrain heater changes during flight. The PX4 proposal favors full-power warmup while disarmed and slower changes while armed. That separates the period when quick heating is useful from the period when a sudden offset can contaminate navigation. The trade-off is weaker immediate temperature control in changing airflow. A designer needs to show that the resulting temperature excursions stay within the calibrated range.
These approaches are complementary, not interchangeable. Software rate limiting can reduce an observed transient without proving that the underlying hardware coupling is gone.
A bench test that can distinguish the effects
A practical investigation can be performed with the aircraft restrained, propellers removed and the flight controller securely stationary. The objective is to isolate the sensor response from any real vehicle motion.
- Log raw accelerometer samples, sensor temperature, heater output, heater target, supply voltage and estimator innovations where the firmware exposes them. Use consistent timestamps.
- Hold the board at several temperatures and record a baseline after it settles. Note the warmup time and whether other sensors on the board reach different temperatures.
- Introduce controlled heater-demand steps while the board remains stationary. Compare the first few hundred milliseconds with the following 10-30 seconds; do not rely only on the final temperature or a long averaging window.
- Repeat under different cooling conditions if the setup permits. Airflow over the flight controller changes heater demand, and the purpose of this test is to find whether that demand correlates with accelerometer errors.
- Compare the original heater-control behavior with a rate-limited version, while checking that temperature regulation remains adequate. Recheck calibration and power behavior before considering a flight test.
The measurements to preserve are peak error, time to recovery, RMS error over declared windows, and the difference between the heater command and the temperature sensor's response. A benchmark that reports only steady-state bias may miss the problem.
What would establish a general fix?
The PX4 bench report gives a mechanism worth examining and a candidate mitigation with quantified results. It does not answer whether the proposed setting is appropriate for cold-weather takeoff, all board revisions, or other IMU packages. Tests on multiple controllers, varying supply voltages and realistic cooling profiles would help distinguish a board-specific coupling from a broadly reproducible fault.
The engineering lesson is narrow but useful: a stable temperature target does not guarantee a stable inertial signal during the transition toward that target. Both the slow temperature-bias curve and the fast response to heater power deserve measurement.