README.md

---
license: cc-by-4.0
task_categories:
- object-detection
- audio-classification
tags:
- drone-detection
- counter-uas
- acoustic-array
- ptz-tracking
- multimodal
- gps-truth
- low-altitude
pretty_name: "nightjar field flight 2026-08-04 #2 — low-altitude, first flight of the cued zoom climb"
---

# nightjar flight — 2026-08-04, second sortie

The same rig and site as `nightjar-flight-20260804`, flown ~80 minutes later, but a
**deliberately different profile and the first flight of a control-loop change**. GPS truth is
again excellent: **7,817 valid UTC stamps out of 7,818**.

## Envelope — read this before comparing to the first sortie

| | this flight | first sortie |
|---|---|---|
| airborne | 784 s (13.1 min) | 1252 s (20.9 min) |
| **max height** | **16 m** | 276 m |
| max slant range | 374 m | 473 m |
| path flown | 2304 m | 3021 m |
| window (UTC) | 18:10:15 – 18:23:20 | 16:46:25 – 17:07:17 |

⚠ **This is a LOW flight.** 16 m of altitude at 374 m of range is an elevation angle of about
**2.4°** — the drone is essentially on the treeline, not against clean sky. That matters in
three independent ways, and it makes a naive flight-to-flight comparison invalid:

* the background is foliage and horizon clutter, not uniform sky, so the detection problem is
  categorically harder than the first sortie's high-altitude passes;
* a horizontal microphone array's **elevation sensitivity goes as sin(el)**, so at ~2° the
  array is at the null of its elevation response — near-zero information on that axis;
* conversely azimuth sensitivity goes as cos(el) and is near its maximum here.

Any per-flight difference in track time or detection rate between the two sorties is therefore
confounded by profile, not attributable to the code change below.

## What changed in the loop between the two sorties

This flight carries a fix to the **cued zoom climb**. The cue's containment focal was being
computed on every acoustic cue (~7.1 mm, from FOV ≥ 2·k·σ at the tilt-ladder spacing), stored,
and never commanded — the cued-search path had a goto that deliberately ignores zoom plus a
branch that only ever *widens*, so the climb half did not exist. The first sortie sat at a
median 4.31 mm — full wide — across the 69% of its time spent searching.

Measured effect here:

| | first sortie | this flight |
|---|---|---|
| SEARCH median zoom | 1.00× | **2.53×** |
| time above 2× | 4.4% | **61.6%** |

**And this flight exposed the follow-on bug**, which is why the data is worth keeping: the
climb drove at full lens speed into a 1 mm deadband and overshot badly — SEARCH focal ran
median 10.88 / p75 11.8 / **max 20.03 mm** against a 7.07 mm ceiling. Above the containment
ceiling the field of view no longer holds k·σ, and ~12 mm re-opens a ~10° vertical blind
stripe between 24° ladder rungs. The fix (proportional to the remaining gap) post-dates this
recording, so **this session shows the overshoot in the wild**.

Also live for the first time: a **state-dependent acoustic presence gate** (strict while
tracking, relaxed while searching). Within minutes it posted cues at 13 dB that the previous
fixed 14.6 dB threshold would have discarded. `audio/cue_log.jsonl` stamps `thr_db` and
`nj_state` on every candidate so each cue can be attributed to the threshold in force.

## 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        cue gate log, now carrying thr_db + nj_state per candidate
video_segments.tar         1080p native + 768x432 substream, H.264 -c copy, 60 s MPEG-TS
blackbox/state.jsonl       tracker FSM ~2.9 Hz: state, pose, focal, error, detection, track
blackbox/frames.tar        2,324 substream stills, named f<epoch_ms>.jpg
dji/*.csv, *.txt           decoded + original DJI FlightRecord (CUSTOM.dateTime = true UTC)
```

**Channel identity is not guessable.** UMA-16 0-based channel 9 is a dead electronic-floor
channel — exclude it and run on 15. SB-POLARIS has 3 live capsules of 8; **this session they
are `[0, 3, 5]`** (live ~−70 dBFS, dead ~−190). Detect them by RMS rather than hardcoding.

⚠ **The two arrays use different sample widths** — UMA S16_LE, SB S32_LE. Reading both as
int16 splits every SB sample in half, scrambles channel identity and destroys the spectrum.

## Calibration as flown

`az_offset_deg` **300.87**, `invert_az` true, re-solved the morning of this flight from the JBL
beacon (spread 0.10°, 41 dB). SB-POLARIS → camera was also solved for the first time that
morning — `az_offset_deg` **102.6**, `invert_az` **false**, rotation only — but **nothing
consumed SB bearings during this flight**; the cue chain is UMA-only. Cue producer was the
cheap daemon (`SUBHUNT=0`). Site notch 100–200 Hz. Start pose pan 45 / tilt 45.

## Caveat on GPS-referenced pixel work

Attempts to place the drone in the image from GPS reach a floor of about **5° of bearing
error** on this data. The cause is the DJI GPS itself: ±2–5 m of position at the 20–45 m ranges
where the optical correspondences live is 4–9° of bearing. Fitting a pose-vs-GPS time offset
does not help (best dt ≈ +0.05 s). A static-mount fit is impossible — of 2,281 blackbox frames,
123 carry a YOLO box and 686 have a still mount, and **zero have both**, because the mount
servos whenever it is tracking. Treat any in-frame or pixel-level claim derived from GPS here
as approximate.

## Elevation datum (fit_el_datum, 2026-08-16)

**Status: FEW_ON_DRONE** — this day's elevation datum is honestly UNSOLVABLE from
banked data (camera never/rarely locked on the drone).
Close-range elevation truth for this day remains datum-limited (~25 deg
floor). Tool: `sirch613/subhunt-v2` `v3/fit_el_datum.py`; summary:
`joshruby/acoustic-knowledge` -> `v3/assets/el_datum/SUMMARY.json`.