Known Defects

Genuine defects — sensor failures, missing scaling, frozen placeholders, swapped columns, or maintainer slips — that an estimator must detect and handle, and that should be corrected or excluded rather than learned from. Realistic in-flight difficulties a working sensor produces (dropouts, drift, fix transitions) are not defects; they live in Desirable Ugliness.

Data defects

  • RealSense IMU dead on mars_1, mars_3 (single row; gyro-x literally 0); degraded on mars_2 (~13.6 Hz), mars_7 (~87 Hz); absent on mars_6.
    • rs_odom is all-NaN on transition_2 and v100, single dead row on mars_1/mars_3; usable RealSense odometry on Mars exists effectively only for mars_2/mars_4.
    • mars_6 has neither RealSense IMU nor a nav camera.
  • LSM9DS1 accel scaling missing on indoor_3, outdoor_1, transition_1, transition_2, transition_3: gravity reads ≈ 6.67 m/s² (= g·16/24); multiply LSM accel by 1.5 there. The ×24/16 correction was applied inconsistently across the campaign.
  • LSM9DS1 gyro static bias ~1.1–2.0 rad/s on every Klagenfurt recording (indoor_*, outdoor_1, transition_*, and the 4 h static set; static per-axis [+1.04, +0.83, −0.94]); normal (~0.02–0.03 rad/s) on Mars. Plus isolated physically-impossible LSM gyro spikes (±240–245 rad/s, single samples) on outdoor_1/transition runs. The LSM gyro on Klagenfurt runs is unusable without removing a ~60–110 dps constant bias.
  • Nav-camera PNG tails missing on disk: mars_18 stems 1353..3715 absent (~63.6% of the run), mars_4 2032..2585 (554), mars_8 2876..3173 (298). mars_5 nav is header-only with an empty img/.
  • Referenced stereo pairs missing or corrupt on disk: the transition stereo timestamp files point to image pairs that are not usable from disk even though the rows are present. Missing-pair counts are transition_1: 8, transition_2: 6, transition_3: 4.
    • Additionally, transition_3/.../stereo_cam/img/7110_c2.png is 0 bytes, undecodable, and still referenced by the timestamps (its _c1 partner is intact, so that pair passes an existence check and fails only at decode).
  • Frozen/placeholder data: px4_gps no-fix sentinels on indoor/vibration (PX4 GPS); mars_18 mars_gps{1,2}_state timestamps frozen at 1633680000.0.
  • Build leakage: ~gvfKkcJ.tmp (1.2 MB partial-rename temp) in transition_3 nav img/; ~gvf0zbp.tmp (0 B) in vibration v100/; the reserve/ raw bag JSONL dumps for mars_12/14/15/18. None belong to navigation use.
  • Magnetometer spherical-column swap: the *_mag_data.csv / lsm9ds1_mag_data.csv family has spher_az/spher_el swapped on mars_4, mars_8, and all 18 insane_mag_calib_raw/klagenfurt files; the px4_mag/lsm_mag family is correct-order. Un-swap at load time.
  • LSM9DS1 magnetometer intrinsic block swap: the lsm_mag_cal_intr blocks — both the soft-iron transform and the hard-iron offset together — are swapped between the two site YAMLs (Calibration); the cross-site file is the correct one for LSM, the site-matched file for PX4. The offset injects an O(0.1) raw-stream bias differing 0.0623 cross-site.
  • Inconsistent calibration metadata: the two T_cn_cnm1 copies differ on klu1 (Calibration); 5 of 6 Mars runs with both timing files carry differing offsets (Time and Clocks).

Typos

Papers:

paper typo correction location reason this shouldn’t be corrected
INSANE (IJRR 2024) T256 T265 (no T256/T265 string exists in the download) uniform, e.g. Table 1, §3.2.3 N/A, it’s someone else’s article
INSANE (IJRR 2024) UST8310 IST8310 Table 1 N/A, it’s someone else’s article
INSANE (IJRR 2024) 483,298.5 nT ≈ 48,330 nT (misplaced decimal) Table 3 N/A, it’s someone else’s article
INSANE (IJRR 2024) Pixhawk v5 PixHawk 4 §3.2.1 N/A, it’s someone else’s article

Repos:

repo typo correction location reason this shouldn’t be corrected
insane_dataset_tools [1] tags 5,6 x = 48 cm x ≈ 84 cm (48↔︎84 digit swap) fiducial_marker/tag_calibration/gen_tag_board_calib.m upstream maintainer code; editing forks it from the published tool — correct at use time
insane_dataset_tools [1] data_sitching.m data_stitching.m README.md:308 upstream maintainer repo; editing forks it from the published source

Dataset:

typo correction location reason this shouldn’t be corrected
stereo_cam_tiemstamps.csv stereo_cam_timestamps.csv mars_10, mars_19 maintainer tooling branches on the misspelled name (post_scripts/csv2bag.py); renaming breaks it and the bit-for-bit provenance
transition_gt_in_moap.csv transition_gt_in_mocap.csv transition_1/2/3 maintainer tooling writes the misspelled name (data_extraction/data_stitching.m); same reasons

References

[1]
aau-cns, INSANE Dataset Tools. GitHub. Accessed: Jun. 09, 2026. [Online]. Available: https://github.com/aau-cns/insane_dataset_tools