Initialization and Practical Estimator Guidance

• Static start.
  • Every run opens with the vehicle on the ground: PX4 IMU sees ~9.8 m/s² with near-zero rotation.
  • Use the leading samples — gated by low |gyro| and |a|≈g, not a fixed count or an external takeoff flag — for gravity-vector attitude initialization and accel/gyro bias seeding.
  • The 4 h static set gives Allan-variance noise parameters and per-IMU residual scale.
• Choose the IMU deliberately. PX4 (ICM20689, damped, near vehicle centre) is the cleanest accelerometer (≈ g). LSM (undamped, high-rate) needs the Known Defects corrections on Klagenfurt runs. RealSense (highest lever arm) is degraded/dead on several Mars runs.
• Respect the clocks. Fuse px4_gps, ground_truth_*, *_data_revised directly on the system clock; convert raw rtk_gps{1,2} by the per-run t_pximu_imugt first, or use the _data_revised lanes instead.
• Score on the real overlap window. ground_truth_8hz/80hz start after px4_imu, and ground_truth_80hz can end much earlier than the IMU span on the long and transition runs (Ground-Truth Products).
• Weight on reported quality.
  • Use rtk_status/gps_fix_type and the per-row covariances.
  • RTK σ from √cov, floored 1.4 cm/1.0 cm; PX4 GPS σ 0.8–4 m.
  • Reject cov_p > 1e12 PX4 rows outright.
• Replay in file order; treat t as approximate. Do not dedupe or reorder by timestamp. For images, key stereo pairs by _c1/_c2 + equal t[ns], verify file existence, and verify PNG decode: some referenced transition stereo pairs are absent on disk and one referenced transition_3 frame is zero-byte.
• Apply the corrections of Schema and Parsing and Known Defects at load time:
  • LSM accel ×1.5 on the five runs
  • LSM gyro bias removal on Klagenfurt
  • mag column un-swap on the _data family
  • headerless reads for _revised/px4_imu_data
  • duplicate-key YAML parse
• LRF. No camera↔︎LRF or LRF↔︎platform transform ships in any YAML (Camera-LRF Extrinsic). The boresight is essentially nadir / image-center; see Camera-LRF Extrinsic for the derived value and its uncertainty.
  • MaRS defines no LRF/range update — the LRF is recorded but not consumed by the estimator (it feeds the flight stack’s landing detector only).
• Reported usability benchmarks (paper).
  • MaRS EKF tracks transition_1 well with indoor vision-drift divergence
  • OpenVINS on mars_18 reaches ATE position 7.5 m / rotation 7° (Umeyama-aligned)