Desirable Ugliness

Using any of the particulars in this section inside the estimator is cheating.

The specific run names, indices, gap locations, and dropout timings listed here are an oracle. An estimator must discover and reject bad data online from the measurements and their reported quality, never from a hardcoded list of “where the bad data is.”

This section exists only as a sanity check:

• to confirm these phenomena are realistic degradations the estimator should be robust to, and that they are distinct from the genuine defects in Known Defects;
• to confirm we are not, through this list, smuggling in sensor/firmware knowledge that belongs in Platform but was never documented.

These are the realistic, desirable difficulties that make INSANE a fair test of robustness:

• Asynchronous, drifting rates. No hardware sync; per-stream rates wander (PX4 IMU 191–199 Hz, LSM IMU 634–928 Hz, mocap gross 298–359 Hz). The estimator must handle async arrival and estimate delays.
• Dropouts everywhere. Baro single-sample gaps (~18 Hz delivered vs 20 nominal); GT 8 Hz lane delivering ~4 Hz; UWB multi-second blackouts; mocap occlusion gaps; RealSense odometry NaN runs and long data gaps. Model irregular arrival; do not assume fixed-step.
• GNSS quality transitions. RTK float↔︎fixed transitions, no-fix (gps_fix_type 0/2) stretches near the building, per-receiver disagreement, and a genuine transition_2 GPS glitch — all flagged by the per-row covariance and status fields the estimator should weight on.
• LRF no-return spikes at the device floor (and rare spurious highs) — single-sample outliers to reject, with no validity flag.
• VIO drift and resets. RealSense odometry drifts to hundreds of metres with unflagged internal resets — relative, loosely-weighted, discontinuous.
• Vision degradation. Dust-occluded downwash imagery, lens flare (transition_3), changing light and magnetic field across the transition, low-texture planar ground (Mars Zone 3), and reduced visual sensors on cliff/wall traverses.
• Vibration coupling. The undamped LSM IMU and the high-lever-arm RealSense IMU carry resonance the damped PX4 IMU does not; the vibration bundles (v10–v100) characterize it at known motor commands.
• Magnetometer disturbance. Raw, uncompensated mag with field crew handling at experiment start and large indoor field swings — the reason GT downweights mag 50:1.