Present on indoor_3, outdoor_1, transition_1..3 (+ static).
Decawave TREK1000; only 3 of 4 anchors published; no anchor-position table ships — but the three ground anchors are physical modules mounted at known positions on the door fiducial boards (boards 1/2/7; paper §3.2.6–3.2.7, Figures 10 and 12), which are a proven-fixed rig (Motion Capture). The anchor geometry is recoverable, not absent. Figures 10 and 12 place U0 on board 2 (tag ids 15–29), U1 on board 7 (tag ids 86–96, its 0.5 m anchor tag id 91), and U2 on board 1 (tag ids 0–14) — see the fiducial-board structure on What Is Not in the Download.
The underlying Decawave ROS message shape carries four distance[] / valid[] slots, but the public CSV export preserves only m1..m3; the fourth lane is dropped at extraction time and later reconstructed as (distance=0, valid=false) by the maintainer csv2bag.py path.
valid_mk is perfectly redundant with dist_mk > 0: only two states occur — missing (dist == 0 & valid == 0) and valid (dist > 0 & valid == 1). Gate on either.
Delivered rate 1.7–6.2 Hz on flights (7 Hz nominal) with multi-second blackout gaps (max dt 18–71 s).
Time-gate updates; do not interpolate across gaps.
Valid ranges reach ~40–65 m outdoors; reliable below ~15 m. Two distinct NLOS regimes, not one:
distance tail — ranges beyond ~15 m carry a growing negative (non-line-of-sight) bias as the door-mounted anchor is occluded by the building edge far outdoors; gate the long-range tail rather than treating it as a defect.
in-gap doorway occlusion — once the vehicle crosses the Dronehall doorway the door anchors are geometrically occluded, so in-gap ranges run 0.2–0.7 m long on transition_1/transition_3 (worst on the deep 11 s transition_3 gap), while only transition_2’s shallow gap stays line-of-sight (in-gap bias 1–12 cm). A <15 m range gate therefore guarantees line-of-sight on the approach but not in the gap.
The anchor geometry is observable by trilateration from the moving vehicle. With the RTK+IMU-fused vehicle trajectory and ranges gated to the sub-15 m line-of-sight regime, a robust trilateration of each fixed anchor reaches GDOP 0.5–2.1 (the outdoor approach sweeps each anchor from >5 m down to ~2 m, a well-conditioned wide baseline) at p95 range residual ≤42 cm. The recovered anchors sit ~3.3–3.5 m from the trajectory’s door end at z ≈ 0.7–2.5 m. Cross-run agreement is only coarse, however: Procrustes-aligned cross-run RMS 0.503 m, max pairwise spread 0.669 m, and raw inter-anchor edge spread reaching ~1.04 m — so the recovery is a per-run ENU ranging witness, not a stable shared board-frame door triangle.
The vehicle main-anchor lever arm is absent from sensor_calibration.yaml but is recoverable. A joint fit over the three runs — one shared PX4-IMU-body offset, observable because the on-approach attitude varies (yaw ~360°, body-z spread 8–17°) — returns [−0.42, −0.44, 0.09] m, ‖0.61‖ m; modeling it shifts the per-run anchors 0.4–0.8 m and drops the worst-anchor range-residual median from 16 cm to 5 cm. Measured by the UWB anchor recovery.