Datasets:
audio audio |
|---|
nightjar flight — 2026-08-04
One 20.9-minute drone sortie recorded end-to-end by the nightjar rig: PTZ camera under closed-loop acoustic-cued optical tracking, two microphone arrays, full autopilot telemetry, and the tracker's own decision log. GPS truth is exceptionally clean here — 12,447 valid UTC stamps out of 12,448 rows (a prior session in this series had exactly 1, which made it unusable for anything truth-referenced).
Rig, site and method: see the companion sessions in this series and SENSIBELSPECS.md /
SUBHUNTBLACKPAPER.md in the project.
Envelope
| airborne | 1252 s (20.9 min), 12,448 GPS fixes at 10 Hz |
| max height | 276 m |
| max slant range | 473 m (furthest of the series to date) |
| max horizontal speed | 10.5 m/s (24 mph) |
| path flown | 3021 m |
| window (UTC) | 16:46:25 – 17:07:17 |
Contents
audio/uma16_<epoch>.wav UMA-16, 16 ch, 48 kHz, S16_LE, 5-min splits
audio/sb_<epoch>.wav SB-POLARIS, 8 ch, 48 kHz, S32_LE, 5-min splits
audio/cue_log.jsonl acoustic cue gate: start / cue_candidate / chirp_reject
video_segments.tar 1080p native + 768x432 substream, H.264 -c copy, 60 s MPEG-TS
blackbox/state.jsonl tracker FSM at ~2.9 Hz: state, pose, focal, error, detection, track
blackbox/*_frames.tar 7,135 substream stills
dji/*.csv decoded DJI FlightRecord (CUSTOM.dateTime = true UTC)
dji/*.txt original encrypted FlightRecord
Channel identity matters and is not guessable. UMA-16 0-based channel 9 is a dead
electronic-floor channel — exclude it, run on 15. SB-POLARIS has 3 live capsules of 8, and
this session they are [0, 3, 5] (measured: live ~−70 dBFS, dead ~−190 dBFS). The live
set has changed between sessions, so detect it by RMS rather than hardcoding. UMA channel
order is serpentine — load the geometry file, never hand-derive it.
⚠ The two arrays use DIFFERENT sample widths (UMA S16_LE, SB S32_LE). Reading both as int16 splits every SB sample into two halves, scrambles channel identity and destroys the spectrum. That bug produced a since-retracted long-range result in this series; see below.
Calibration as flown
| UMA-16 → camera | az_offset_deg 300.87, invert_az true — re-solved this morning from the JBL beacon (spread 0.10°, 41 dB); 1.1° drift from 08-03's 299.77 |
| SB-POLARIS → camera | az_offset_deg 102.6, invert_az false — solved for the first time this morning; two beacons 89.5° apart agreed to 0.9°, run-to-run 0.10°. Rotation only; mic positions are the ring model, not solved. Nothing consumed SB bearings during this flight — the cue chain is UMA-only. |
| cue producer | cheap daemon (SUBHUNT=0); ft16 starves under live tracker load |
| site notch | 100–200 Hz (widened from 140 Hz before this flight) |
| start pose | pan 45 / tilt 45 |
Results
Acoustic cue → optical lock
20 cues posted, 14 locked (70%), 11 chirp-rejects, 50 candidates not posted, 26 TRACK episodes. Rotor-band specifically: 10/15 locked. Time-to-lock ranged 1.7–41.3 s.
The site-emitter notch was widened to 200 Hz immediately before this flight, and it worked: no cue below 222 Hz was posted all session, where the previous session repeatedly fired false cues on a fixed 140–170 Hz machinery line and slewed the camera at nothing.
Where each array loses the drone
Per-second channel-agreement detection in the 240–460 Hz rotor band reads 100% at every range for both arrays — but that number is a trap, because a fixed site harmonic sits inside the band. The discriminator is the peak-frequency IQR: a real drone's rotor line wanders with throttle (wide IQR); a fixed emitter is pinned (tight IQR).
| range | UMA-16 peak Hz | SB-POLARIS peak Hz |
|---|---|---|
| 0–50 m | 342 IQR[319–357] | 312 IQR[258–342] |
| 50–100 m | 308 IQR[253–346] | 266 IQR[250–344] |
| 100–200 m | 326 IQR[242–348] | 260 IQR[245–324] |
| 200–400 m | 245 IQR[240–274] | 247 IQR[240–263] |
Both arrays converge on the same pinned ~245 Hz line beyond 200 m — that is the site emitter, not the drone. Neither array demonstrably hears the drone past ~200 m here, and the SensiBel shows no range advantage over the UMA.
Retraction. A previous session in this series reported the SensiBel still hearing the drone at ~330 m while the UMA had collapsed. That came from the int16/S32 misparse noted above. Recomputed with correct parsing, the SB's long-range IQR is tight and pinned, i.e. emitter, not drone. Treat any earlier SB-range claim in this series as withdrawn.
Optical: the zoom bootstrap trap
The camera held a median of 1.01× even while tracking, reaching 11.3× at most; 73% of
tracked time was still at full wide. Scoring the flight in pixels on target (0.35 m
drone, GPS slant range, deg_per_px from the blackbox) shows why that is decisive:
| range | median px on target | detection rate | px if held at 120 mm |
|---|---|---|---|
| 0–50 m | 27.1 | 46.6% | 454 |
| 50–100 m | 3.2 | 1.0% | 88 |
| 100–200 m | 1.6 | 0.0% | 44 |
| 200–300 m | 0.7 | 0.0% | 20 |
| 300–500 m | 0.7 | 0.2% | 20 |
Detection collapses exactly where apparent size falls below ~10 px, and the pixel model predicts the cliff. The counterfactual column is the finding: at the lens's usable tele the drone would have stayed above the detection floor out to ~300 m instead of dying at 50 m.
It did not, because zoom is gated on a track that can only exist where zoom is not needed — zoom climbs only under an established optical track, tracking requires detection, detection requires ~10 px, and inside 50 m the target is already 27 px. Outside 50 m there is no detection, so no track, so no zoom, so there will never be detection. The tracked median range was 10 m. This dataset is a clean instrumented record of that trap.
Known issues in this session
- Blackbox logs the zoom outcome (
focal_mm,zoom_speed) but not the decision — no containment ceiling, no σ, no binding axis. Zoom behaviour can be described from this session but not fully attributed. - SB-POLARIS elevation is not calibrated; azimuth registration exists but nothing consumed it.
OSD.heightin the decoded CSV is feet.- Blackbox sampling is ~2.9 Hz, well below video frame rate.
- Downloads last month
- 117