What thirty minutes of sitting still taught us about the IMU
A vehicle at rest is the only condition outside water where the true state is known exactly. Using that fact properly meant verifying the rest itself, not just trusting the clock.
Underwater, truth is never directly available — that's the whole reason the estimation stack exists. But there is exactly one condition where it is: a vehicle sitting still on a bench, where the true velocity, true angular rate and true position are all known by definition. That measurement had never been taken on this hardware, so it became the first one.
89,813 samples over 1,797.0 seconds at 49.98 Hz — and rest itself was verified rather than assumed: |accel| held at 9.7271 ± 0.0053 m/s² for the full duration, confirming nothing bumped the bench and that what follows are sensor statistics, not desk statistics.
The headline, and the one that isn't
Every gyro bias came back under 0.0004 °/s — three orders of magnitude below the per-sample noise floor of roughly 0.09 °/s. That number has a direct architectural consequence: the board is already estimating and removing its own gyro bias before the message ever reaches /mongla/imu, so the localization filter must not try to estimate that same bias a second time — double-correcting a bias that's already near zero would only add noise. Residual yaw drift, measured over the same window, came out to 0.1° per hour.
Allan deviation answers a different question than a standard deviation does, and it's the one that matters here.
A plain standard deviation says how much one sample scatters around the mean. Allan deviation says how long the sensor can be averaged before bias drift starts dominating over noise reduction — which is the actual number a process-noise model should be tuned against. These gyros kept improving out to 660–860 seconds of averaging, longer than any single mission leg this vehicle runs. Two numbers that looked like findings and weren't, logged specifically so they don't get "discovered" again later: the accel-x mean of −0.8718 m/s² is gravity on a tilted board mount, not a bias, and it's already accounted for elsewhere.