# FlyBrain Drone Lab

First prototype: seeded forest flight, Chase (default) and FPV cameras, bilateral visual samples, identified LPLC2/LC4 stimulation, connectome readout, direct vision baseline, circuit silencing, 0–200 ms added delay and CSV export. Expanding-object and uniform-dimming benches hold the drone still.

## Run

```sh
python3 -m http.server 8000
```

Open http://localhost:8000/experiments/drone/. No npm install or data build is needed to run the site: MaleCNS v1.0 and its index are committed under data/male-cns-v1.0/. Both apps select their data through data/dataset.json. See [dataset provenance and rebuilding](../../docs/DATASETS.md).

## GitHub Pages

In Settings → Pages, choose **Deploy from a branch**, **master**, and **/(root)**. Every ordinary push to master publishes the committed static files. The neuron index is included in the branch, and .nojekyll disables Jekyll processing. No custom Actions deployment or build step is required.

Open https://bquast.github.io/flybrain/experiments/drone/ . The Drone Lab checks workflow performs validation only; GitHub's own Pages workflow handles publishing. The upstream custom domain is not used.

## Methods and limits

- Each eye renders 32 × 25 grayscale perspective samples. The hexagonal display is an approximation, not reconstructed ommatidial optics.
- Gain/offset-normalized images supply dark connected components. Positive angular-area growth becomes engineered looming features injected into identified LPLC2/LC4 cells. This is not a complete visual circuit model.
- The MaleCNS v1.0 index uses ascending body IDs, stored as strings. It retains 166,700 annotated neurons and all 25,582,938 edges between them. Side uses somaSide with rootSide fallback for sensory neurons. This is a routing assumption, not a measured receptive-field map.
- A bounded rate model uses the full graph. Incoming absolute weights are normalized, recurrent gain is 0.9, time constant 25 ms, and integration steps are at most 10 ms. Signs use MaleCNS consensus transmitter labels with the prototype convention (GABA negative, all others positive); these dynamics are not fitted physiology.
- The hand-tuned, untrained readout uses bilateral LPLC2/LC4 rates for steering and speed. Those cells are directly stimulated, so successful flight alone does not establish a benefit from recurrent wiring.
- LC11, GF/DNp01, descending and motor cells are observation probes. VNC cells and their edges are simulated, but drone steering still reads LPLC2/LC4. Conventional attitude and altitude control handles simplified drone physics; obstacle coordinates enter rendering/collision detection only.
- Each sensory frame advances 1/30 second and four physics steps. Worker calls are sequential. Slow hardware lowers simulation speed, displayed in the interface. Added delay is in simulation time; compute and round-trip wall times are separately logged. Network jitter/loss is not modeled.
- Changing settings resets the seeded trial. Hiding the tab pauses it. CSV logging is capped at 18,000 frames (10 minutes).

## Checks

```sh
node --test experiments/drone/tests/core.test.cjs
node --test experiments/drone/tests/real-data.test.cjs
```

CI also checks the actual browser at a /flybrain/ subpath, cell silencing, flight, pause, CSV export and mobile layout, and saves screenshot evidence.

## Attribution

FlyBrain by snedea, MIT (../../license.md). Data and citations:

- MaleCNS by HHMI Janelia/FlyEM, Cambridge, MRC LMB and Google Research: https://male-cns.janelia.org/ (CC BY 4.0; see ../../docs/DATASETS.md for transformations)
- Earlier FlyWire baseline: https://doi.org/10.1038/s41586-024-07558-y
- https://github.com/flyconnectome/flywire_annotations
- https://pmc.ncbi.nlm.nih.gov/articles/PMC5340600/
- Inspiration: https://x.com/c10ned/status/2099183080065941919 (independent implementation; the author's source was unavailable).

## Future gym

The shared dataset loader is independent of drone sensing and control. A launcher for interchangeable experiments is planned later; see [the gym direction](../../docs/GYM.md).