Where other aircraft ask satellites, this one would have to remember. The open question: can a neural network small enough to fit in 4 MiB memorize what its flight area looks like from above — by day, by dusk, by night — well enough to turn one glance of a low-light camera into (lat, lon, confidence)? No satellites to jam or lose, no internet — if it can be done at all. Nobody knows yet; finding out is the project. And the research is not done by me: an autonomous loop of coding agents designs, trains, and scores one pre-registered experiment at a time, keeping only what measurably helps. This site is its live lab notebook.
The model family is scene-coordinate regression: one compact network per flight area that encodes "what does this place look like from above, under which lighting" directly into its weights — the pipeline works for any bounding box on Earth, but each trained model knows exactly one patch of it by heart. No reference imagery on the aircraft, no retrieval, no matching. Training data comes from a frozen pipeline that fetches open-licensed aerial orthophotos for any bounding box and re-renders them under six lighting conditions, from morning to night, as seen by a simulated starlight-class low-light sensor (Sony STARVIS2 / IMX585 class — the aircraft's chosen camera). That sensor choice is the premise of the whole project: at high gain it keeps daylight-like scene structure deep into the night — trading it for noise and washed-out color, which is exactly what the night renders below show — so a single compact model can localize from morning to midnight.
The research loop is Karpathy-style autoresearch: each iteration, a headless coding agent reads the full experiment history, pre-registers ONE focused change — hypothesis, method, expected outcome, and a hand-drawn architecture figure — then the harness trains and scores it against a single frozen ruler: the worst median position error across 6 lighting conditions × 4 German test areas (dense Berlin, rural Prignitz, Munich, Frankfurt), on held-out map crops. Improvements are kept as git commits; everything else is reverted but stays in the record. Hamburg is never touched by the loop — it exists only as a blind check that the method generalizes.
What the loop is actually allowed to search over. The gray endpoints are fixed by the harness — one low-light camera crop in, one (lat, lon, confidence) answer out — and the dashed box is the entire search space: every experiment in the model designs gallery is one way of filling it. The ochre lane underneath holds the training-only signals — losses, targets, samplers — that shape the weights but are torn down before anything flies.
GPS-denied visual localization is not an unsolved problem. The established, field-tested family matches live camera frames against georeferenced reference imagery carried on the aircraft — e.g. WildNav (Vision-based GNSS-Free Localization for UAVs in the Wild), which matches drone photographs against satellite tiles with deep feature matching and demonstrated GNSS-comparable accuracy in real flights. If you need working GPS-denied navigation today, start there, not here.
This project deliberately walks a different road, for two reasons. Licensing: the strongest matchers in that stack — Magic Leap's SuperGlue / SuperPoint pretrained networks — are licensed for noncommercial research only, without the right to sublicense, which is incompatible with a fully open-sourceable, commercially usable system (permissive alternatives like LightGlue (Apache-2.0) exist, but the whole approach still means shipping reference imagery on the airframe — this project's hardest constraint rules that out). Curiosity: the actual motivation is ground-level research — can an autonomous loop of coding agents discover a genuinely different approach, with no reference imagery on board and the map living entirely in the network's weights, under a low-light premise? It may or may not end up competitive with the field-tested systems above. Saying so out loud is part of the experiment.
The vision, if it works: a novel algorithm that lets anyone draw a bounding box around a place they care about and generate their own tiny, punchy, self-contained geo-boxed model — reliable, personalized, and free. Open data in, open weights out; no reference imagery to license, no vendor to call, no cloud to depend on. The map is yours, and it lives in a few megabytes you own.
updated 2026-07-22 21:02 UTC · experiments run around the clock on a RunPod Secure Cloud pod — one RTX 4090 (24 GB) at $0.69/hr; the loop commits every result to git as it goes