license: cc-by-4.0
task_categories:
- audio-classification
- object-detection
tags:
- drone-detection
- counter-uas
- acoustic-array
- cue-to-track
- autonomous-drill
pretty_name: 'nightjar 2026-08-11 — first autonomous hover drill: the loop closes in 18 s'
nightjar 2026-08-11 — the first autonomous cue drill (hover station)
A drone hovered at a fixed station above the WP beacon (~daylight, post-rain) while the rig ran the SELF-JUDGING look-away drill: point away → re-arm → the acoustic chain must re-find the drone → YOLO-vouched lock, scored against a self-bootstrapped truth pose. No operator in the loop: truth was acquired by ear (cue chain → slew → YOLO lock), the station gate froze at the WP bearing, and every trial was scored automatically.
★ Headline: the clean run closes the loop in 18–22 s at 0.2°
The final drill run (drill_20260811_151955, the only one with a SINGLE drill process — see
the incident below) is the honest measurement:
| trial | total time (incl. look-away + slew) | lock error | zoom at end |
|---|---|---|---|
| T1 | 22.2 s | 0.24° | 29.0 mm |
| T2 | 18.2 s | 0.15° | 25.5 mm |
Chain legs in the clean run: cue ≈ 0 s (drone continuously audible; a cue is always pending at re-arm), detection ≈ 0 s (the slew lands with the drone already in frame), vouch ≈ 0 s, then settle + zoom climb. Success bar was 5 consecutive locks <15 s / <2° / ≥30 mm — not reached only because the session ended after two trials; both trials beat the error bar by 10×, and zoom was still climbing when the run was stopped.
The acoustic side, 12 min of hover
172 cue candidates (14.4/min), az 139.6–144.8 (median 142.5, spread ±2.6°)
presence median 19.9 dB, max 25.0 · chirp_rejects 10 (guard off, no pinger)
Cue bias +5.8° — consistent, and NOT an error. The offset is anchored on the FAR beacon (JBL, parallax→0, matches drone-at-range geometry); the drone hovered above the NEAR WP beacon, so array↔camera parallax (~3 m baseline) shows up as a stable ~+5.8°. Containment absorbed it: the slew still put the drone in frame every time (det ≈ 0 s). At operational range this bias vanishes by construction.
⚠ THE INCIDENT: three concurrent drill processes (runs 151000 / 151031 / 151200)
Retried ssh launches all fired, stacking three drills onto one mount. They fought over
arm/disarm and pointing: interleaved trials, contradictory results for the same trial number,
"truth" yanked +14.4° while the drone hovered motionless, lock errors of 8–9.6°, vouch times
of 22–48 s. Do not use those three runs for latency or accuracy conclusions — every slow
number there is mount contention, not the tracker. (The clean run's vouch ≈ 0 s refutes the
"confirmation-confidence is the choke" hypothesis those runs suggested; a yolo_confirm_conf
change made on that basis was reverted the same day.)
Fixes shipped immediately: cue_drill.py now takes an exclusive flock (a second instance
refuses to start), and drill_start.sh / drill_stop.sh on-rig launchers keep dangerous
patterns off the ssh command line (a raw pkill in an ssh line self-matched and killed its own
session twice today). Double-start refusal verified live.
Contents
audio/uma16_*.wav UMA-16, 16 ch, 48 kHz, 5-min splits (12 min hover)
audio/cue_log.jsonl 172 candidates with az/presence/line_hz
video_segments.tar 1080p native + substream, 60 s MPEG-TS
blackbox/state.jsonl tracker FSM ~3 Hz
blackbox/frames.tar 8,753 substream stills
drill_20260811_151000/ trials.csv + events.jsonl ⚠ 3-process contention
drill_20260811_151031/ trials.csv + events.jsonl ⚠ 3-process contention
drill_20260811_151200/ trials.csv + events.jsonl ⚠ 3-process contention
drill_20260811_151955/ trials.csv + events.jsonl ✅ CLEAN single instance
Registration in force: UMA az_offset 323.30 / invert true (far-beacon anchored, verified by
pointing to 0.02° the previous evening). DJI FlightRecord not yet attached — add under dji/
when exported. UMA-16 ch9 (0-based) is dead; exclude it.