Platform and Fixed Hardware Invariants

These are the constants a manufacturer/firmware datasheet would carry: the same regardless of which flight.

Device identities are paper-attributed — only the strings LSM9DS1 and Decawave actually appear in the download (in README.txt/YAML); every other model name is paper-only.

Vehicle / compute.

Sensor suite (per-run measured rates from the README.txt files; nominal device specs from the papers):

Stream (file) Device Rate (per-run range / nominal) Units / notes
px4_imu TDK ICM20689 (autopilot, elastomer-damped) ~191–199 Hz body-frame m/s², rad/s
rs_imu BMI055 (inside RealSense unit) ~200 Hz (degraded/dead on some Mars runs) body-frame; unit mounted rotated
lsm_imu ST LSM9DS1 (standalone, undamped) ~634–928 Hz, per-run captures vibration; two site defects (Known Defects)
Bosch BMI055 (2nd autopilot IMU) deactivated never appears in the data
px4_mag PixHawk magnetometer (IST8310) ~85–88 Hz Tesla; raw/uncorrected
lsm_mag LSM9DS1 magnetometer ~20 Hz (14–18 Hz on mars_2/mars_7) different units; raw/uncorrected
px4_baro MS5611 20 Hz nominal (≈18 Hz delivered) static pressure, Pa
px4_gps PixHawk GNSS ~5 Hz non-RTK, meters-level
rtk_gps1/2 2× u-blox C94-M8P 8.00 Hz dual RTK; UHF link replaced by ZigBee
lrf_range Garmin Lidar-Lite v3 ~30 Hz downward, raw slant range, m
uwb_range Decawave TREK1000 ~1.7–6.2 Hz delivered (7 Hz nominal) 3 anchors (4th dropped)
rs_odom RealSense T265 (paper prints “T256”) V-SLAM ~200 Hz local VIO pose, reset-prone
nav cam IDS UI-3270LE-M-GL, lens BM4018S118C 20 Hz 2056×1542 mono global shutter, downward, co-aligned with LRF
stereo cam RealSense T265 fisheye pair 30 Hz 848×800 mono, 64 mm baseline, forward, optical axis ~30° below horizontal (60° from nadir)
px4_rpm motor command (PX4 mixer) ~99 Hz normalized 0–1999, not tachometer RPM
mocap OptiTrack (VRPN), 37-camera ~334 Hz median indoor/transition only

RTK geometry (fixed): two antennas on aluminium rods, baseline 1.16 m (sensor_calibration.yaml norm_rtk1_rtk2; the paper rounds to 1.2 m; measured median baseline on dual-fixed rows 1.15–1.19 m), rotated 45° from the body x-axis, virtual-GNSS baseline direction rtk2 − rtk1.

Worst-case heading error from per-antenna position noise scales as ε ≈ η/B, inversely with baseline. The paper’s 0.95° (Figure 8) is η = 1 cm over its rounded B = 1.2 m; at the measured 1.16 m baseline the same 1 cm noise gives ε ≈ 0.98°.

Per-site ENU reference coordinates and magnetic constants (fixed per site, used by the extraction):

Site Runs ENU ref (lat, long, alt) Magnetic mag_var [inc, dec] rad
Klagenfurt (KLU) indoor/transition indoor_*, transition_* 46.613559, 14.262388, 489.494 inc −1.089745, dec +0.059505 (mag_calibration_klu.yaml)
KLU model airfield outdoor_1 46.606867, 14.279121, 484.017 (KLU mag calibration)
Negev / Mars analog mars_1..19 30.599929, 34.867308, 526.594 inc −0.808604, dec −0.084160 (mag_calibration_mars.yaml)

Transition runs additionally carry a building-entry coordinate 46.613331, 14.262166, 489.474.

The magnetic inc/dec are stored as literal signed radians; use them as-is.

Camera-calibration set per run (fixed):

sensor_calibration.yaml (RTK geometry, IMU/baro transforms) is single and run-invariant.

(The run→set mapping is the conventional reading; the download carries no field that states it, but the per-run ENU reference coordinates corroborate it.)

References

[1]
M. Scheiber et al., “CNS flight stack for reproducible, customizable, and fully autonomous applications,” IEEE Robotics and Automation Letters, vol. 7, no. 4, pp. 11283–11290, 2022, doi: 10.1109/LRA.2022.3196117.