A Game Boy Advance ROM featuring an articulated fruit fly, a small real FlyWire connectome, live neural and behavioral telemetry, and a persistent Tamagotchi mode.
Download the .gba from GitHub Releases and open it in mGBA or load it on a compatible GBA flash cartridge. A local build produces dist/fly.gba.
- An anatomical flybody model developed by Google DeepMind and HHMI Janelia: connected head and thorax, compound eyes, segmented abdomen, antennae, six articulated legs, wings with veins, and a proboscis.
- Offline rendering of the articulated 3D model into transparent, shaded sprites. The GBA selects the appropriate direction and animation frame; the arena uses software perspective projection. This keeps the model detailed without running MuJoCo or a large mesh renderer on the handheld.
- Idle, alternating tripod walk, flight with folded legs and moving wings, feeding, head grooming, and rest. Takeoff and landing follow the simulation's changing height.
- Tamagotchi care: feed, water, touch, play, clean, sleep, medicine, and hit. Health, food, water, energy, cleanliness, stress, bond, and age respond to care or neglect; hitting causes injury and a fleeing animation.
- Two-bank SRAM saves with checksums and recovery after an interrupted write.
- Autonomous sugar seeking, obstacle avoidance, hunger and energy, plus manual movement and turning.
- Standard, close, and overhead views; an arena minimap; separate energy and hunger indicators.
- A real reduced FlyWire FAFB v783 network: 128 neurons and 2,048 directed signed edges, including 20 surviving sugar sensory IDs and one MN9 from the lists published by Shiu et al.
- A neural activity map with spike flashes, selected-neuron connections, incoming/outgoing counts, root ID, and model voltage.
- Measured FPS and CPU utilization, simulated time, speed, distance, spikes per second, recently active neurons, contact, feeding and collision counters, spike history, and time spent in each behavior.
- Pause and session reset. A hardware timer advances simulation at 30 ticks per second independently of display refresh.
Use v0.2.1 or newer. Earlier builds force fast cartridge timings that are unsuitable for SuperCard RAM. The ROM now preserves the timings chosen by the firmware and resets the display after the launcher hands over control.
Copy the new .gba under a new filename and launch it as a fresh game, rather
than restoring an old savestate or Quick Load snapshot. Select No patching
for this homebrew ROM; it already uses native SRAM and preserves loader timings.
Disable the In-game
menu (IGM) for this ROM: it polls input and does not implement an IRQ handler
for firmware hooks. Use SRAM saves; SuperFW transfers battery-backed SRAM
to the SD card on reboot. Retain your existing .sav file for pet progress.
If corruption continues, disable Fast ROM loading in SuperFW and load the file again. The firmware documents slow-memory timing, IGM and loading issues in its troubleshooting guide. These changes are regression-tested in mGBA; confirmation on physical SuperCard hardware is still needed. Display speed depends on cartridge timings.
| GBA button | Action |
|---|---|
| L / R | Cycle arena, neural map, telemetry, help, and PET |
| Select | Toggle autonomous / manual control |
| Up / Down | Move forward / backward in manual arena mode |
| Left / Right | Turn in manual arena mode |
| Hold B | Fly; release to land |
| Hold A | Groom |
| Press A+B together | Standard → close → overhead view |
| Start | Pause / resume |
| D-pad on neural map | Select a neuron |
| A on neural map | Show / hide connections |
| B on telemetry page | General / body and behavior metrics |
| A on telemetry page | Reset lab telemetry while preserving the pet |
Press Select first for full manual control. In AUTO mode, the simulation continues on the other pages. Ground contact with sugar stimulates the sensory neurons; MN9 spikes gate food intake. Movement, energy, voltage, and distance use demonstration units.
Press L from LAB to enter PET. Use the D-pad to choose one of eight actions, B to open/close the care menu, and A to apply the selected action. Start pauses. Hold A+B for three seconds to create a new fly, replacing the previous pet.
Food is offered for four seconds; MN9 activity gates intake. Sleep restores energy, touch reduces stress and builds bond, and medicine restores health with a 30-second cooldown. Hitting removes 10% health, raises stress, reduces bond, and triggers fleeing. Neglect can end the pet's life; care does not resurrect a dead fly.
Pet time advances only while PET mode is active, including its neural, data, and help pages. Selecting LAB suspends pet time and restores the laboratory simulation. Age is shown in minutes; DATA shows feeding ticks and hits. Needs and bond use 0–100 game scales.
Care events and a 30-second checkpoint automatically save to 32 KiB SRAM. Enable save persistence in your emulator or flash cartridge and keep its .sav file. Reopening resumes the saved pet; elapsed time while the ROM is closed does not affect it. Up to 30 seconds of idle progress can be lost on closing. Health, stress, bond, and lifespan are game rules, not measured biological properties.
The anatomical edges and root IDs come from real FlyWire data. The ROM contains a small extracted subgraph, not the complete fly brain. The neural map is a schematic layout without anatomical soma coordinates.
The 2,048 strongest internal connections among the selected 128 nodes are retained. Published signs are preserved; aggregated weights are quantized. The integer LIF demonstration uses 33.3 ms steps, threshold 100, leakage of 1/8, and a one-tick refractory period. It does not reproduce the calibrated temporal parameters of the Shiu model.
Body shape and joint attachments derive from flybody. Meshes are simplified and shaded offline; the gait, wing cycle, grooming, and feeding poses are illustrative animations. The ROM does not include flybody's physics, learned controllers, or biomechanical validation. Wing animation depicts flight and does not reproduce the fly's biological wingbeat frequency.
A compact heuristic controller handles locomotion, flight, grooming, and energy. The extracted neural circuit controls the feeding gate through MN9. The ROM identifies these limitations in its help page.
The GBA has 256 KiB EWRAM and 32 KiB IWRAM. Sprite data stays in cartridge ROM; a small subgraph fits in memory in place of the full connectome. See the GBA memory map.
make # ARM-capable Clang, LLD, llvm-objcopy, and Python 3
make test # Native ASan/UBSan, behavior, provenance, sprites, and ROM header
make preview # Host-rendered PPM frames using the same rendererOn Ubuntu, install clang lld llvm. tools/build.py searches PATH and Homebrew's LLVM directory; it can also use LLD bundled with Rust. Override paths with CLANG, LD_LLD, and LLVM_OBJCOPY. The ROM build needs neither devkitARM nor MuJoCo, Python scientific packages, or downloaded models: generated assets are committed.
The ROM, ELF, and link map are written to build/; a ROM copy is written to dist/. make clean removes build/.
The real cartridge test uses the mGBA 0.10.5 core and reads emulated RAM and framebuffers:
python3 tools/run_emulator_check.py --source /path/mgba --build /path/mgba-build
python3 tools/export_media.py # Pillow; uses actual emulator capturesIt covers boot, movement, flight/landing, grooming, feeding, all three views, neural and telemetry pages, pause, pet care, neglect, reset, SRAM reboot persistence, and interrupted-save recovery. See validation for results and limits. Physical GBA hardware has not been tested.
Connectome extraction needs NumPy, PyArrow, Connectivity_783.parquet, and figures.ipynb from the authors' repository:
python3 tools/import_connectome.py /path/Connectivity_783.parquet /path/figures.ipynb
makeSelection uses three-step forward/backward anatomical relevance between sugar receptors and MN9, followed by induced-subgraph extraction and absolute-weight ranking. Quantization is sign(w) * round(sqrt(abs(w)) * 7), clamped to ±96. Original weights and IDs are in connectome_edges.csv; source hashes and extraction settings are in connectome.json. One of the 21 sugar IDs and the second MN9 were absent from this v783 table and were excluded.
Optional body asset authoring instructions and exhaustive pose checks are described in body-model.md. The pinned source, original joint anchors, reduced mesh, and generation scripts are included for reproducibility.
Every branch push and pull request builds the ROM, runs native ASan/UBSan checks, validates data and cartridge assets, and runs the actual cartridge in mGBA. Tested ROMs, release archives, checksums, and screenshots are uploaded as workflow artifacts.
A vX.Y.Z tag triggers the release workflow. After validation, it publishes a .gba, a ZIP containing documentation and licenses, SHA256SUMS, and a manifest. The tag must match VERSION. Tags ending in -alpha.N, -beta.N, or -rc.N publish prereleases. Publishing uses the repository's standard GITHUB_TOKEN.
make all test
python3 tools/package_release.py --tag v0.2.1See CONTRIBUTING.md for the release procedure. Actions and downloaded test sources are pinned by SHA/hash.
- FlyWire, data guidelines, and connectome publication.
- Shiu et al., 2024: A Drosophila computational brain model reveals sensorimotor processing and published data and neuron lists.
- flybody, Google DeepMind / HHMI Janelia; anatomical model and derivative body assets under Apache-2.0.
- mGBA and devkitPro gbafix for the cartridge header specification.
Original project code is MIT. FlyWire-derived data is CC BY-NC 4.0; its attribution and noncommercial conditions also apply to the included derivative data in the ROM. The flybody-derived graphics retain Apache-2.0 terms. See LICENSE, THIRD_PARTY_NOTICES.md, and bundled third-party licenses. This project is not an official FlyWire or flybody product.