31f35c6f2f
Measured on the cartridge from the rung-10 checkpoint: `wXCoord` and `wYCoord` change at the *end* of a sixteen-frame step while the background scrolls throughout it, so for fifteen frames of every sixteen the screen buffer is centred one tile ahead of the coordinates. The cross-check compared the decode of the fly's own tile with the screen's reading of the tile ahead and refused. Pewter City decoded on 118 of 120 standing frames and on none of the moving ones, so every walk the fly actually took was re-planned over the ten-by-nine window, which is the oscillation of row 23. Nothing in the pinned symbol table says "a step is in progress" and a new address cannot be pinned without the disassembly `gen_symbols.py` reads, so the anchor is measured rather than named: the screen is centred on the fly's tile or on one of its four neighbours, and the one it is centred on is the one whose whole neighbourhood agrees with the decode. A decode with a wrong stride, a wrong quadrant or a half-loaded map agrees with none of the five, and neither does the mid-warp tear the per-serve check was added for, so both refusals stand. `state::step_destination` names the tile the step is landing on, for the ledger the next commit writes. The town fixture gains two landmarks beside the fly, because a neighbourhood that is the same tile id in every direction cannot tell one anchor from another. |
||
|---|---|---|
| .forgejo/workflows | ||
| .github/workflows | ||
| apps/stage | ||
| data/fafb-v783 | ||
| docs | ||
| infra | ||
| packages | ||
| services | ||
| tools | ||
| .gitignore | ||
| CLAUDE.md | ||
| CONTRIBUTING.md | ||
| LICENSE | ||
| LICENSES.md | ||
| Makefile | ||
| NOTICE | ||
| package-lock.json | ||
| package.json | ||
| README.md | ||
| ROM-POLICY.md | ||
flybrain
A simulated fruit-fly brain that plays video games. A connectome-constrained spiking network reads the screen, its population rates become controller inputs, and a scalar reward nudges a bounded set of Kenyon-cell to MBON gains.
The library is @flybrain/brain in packages/brain. It holds the connectome dataset format, the
LIF kernel, the plasticity rule, the population-rate readout and the activity-map geometry. It
holds no game, no emulator and no reward rules: those belong to whatever embeds it.
Workspace layout
| Path | Contents |
|---|---|
packages/brain |
the library (@flybrain/brain) |
data/fafb-v783 |
FlyWire-derived browser artifacts (CC BY-NC 4.0) |
tools/ |
the Python builder that regenerates data/ from official Codex exports |
docs/ |
overview, dataset format, model, plasticity, readout, integration, limitations, verification |
services/flysim |
the Rust service: brain, emulator, snapshot feed, control API, checkpoints |
apps/ |
planned: one directory per game demo |
infra/ |
planned: deployment for the 24/7 stream (see docs/streaming-plan.md) |
Quick start
npm ci
npm test
npm run typecheck
76 tests, about 7 seconds. There is no build step. npx tsx packages/brain/examples/node-random-frames.ts 60
runs the full 139,255-neuron brain with the Game Boy readout on noise frames in plain Node
(about 0.9x Game Boy real time single-threaded on a WSL laptop).
Usage
import { NeuralAgent, gameboyDecoderConfig, toButtonMask } from '@flybrain/brain';
import { loadBrainDatasetFromDir } from '@flybrain/brain/node';
const dataset = await loadBrainDatasetFromDir('data/fafb-v783');
const agent = new NeuralAgent(dataset, { decoder: gameboyDecoderConfig() });
agent.warmup(firstFrame); // 2,500 ms with plasticity off, then calibrate
// every emulator frame:
const { active } = agent.tick(framebuffer, { // RGBA 160x144 by default; any size via config
rewards: [{ value: 0.5 }], // scalar rewards your game adapter detected
boot: !inGame, // relaxes Start/Select throttling on title screens
});
emulator.setButtons(toButtonMask(active));
const checkpoint = agent.exportState(); // bit-exact resume, validated on import
The lower layers (LifNetwork, RewardModulatedStdp, PopulationDecoder) are exported too for
hosts that want to run the loop themselves.
integration.md has the full per-frame loop, the fractional frame timing and the checkpoint contract.
Documentation
- Overview: the pipeline, the layer map and the design principles.
- Dataset format: artifacts, CSR layout, weight encoding, every role and its count, fingerprinting, regeneration, license.
- Model: the 1-ms LIF kernel step by step, every default constant, the retina projection, the RNG, state export and version strings.
- Plasticity: edge selection, the eligibility and reinforcement equations, statistics, topology hash and the explicit non-claims.
- Readout: scores, exclusive groups, pulse channels, the blocked-direction cooldown, the Game Boy preset table and checkpoint versions.
- Integration: what a game must provide, the Pokemon Red integration as a worked example, and a sketch of a platformer adapter.
- Limitations: what is not claimed, what is unproven, measured throughput.
- Verification: the oracle-test strategy and what each test file covers.
- Streaming plan: headless capture, Twitch, VM design and the phased plan for a 24/7 stream.
- Artifact builder: how to regenerate and verify
data/fafb-v783. - Data attribution: source, license and citations.
Provenance
The network, plasticity rule, readout, FlyWire pipeline and activity viewer were extracted from
the fly-plays-pokemon prototype so several game demos can share one core. The default
configuration reproduces that prototype's kernel bit for bit, and verbatim copies of its modules
live in packages/brain/tests/legacy/ as oracles. Built with Astra.
Licensing
The data/fafb-v783 artifacts are derived from the FlyWire FAFB public Codex v783 exports and are
licensed CC BY-NC 4.0. That is a
non-commercial license, so a commercial demo needs a different data source or separate permission.
Citations and the list of modifications are in
data/fafb-v783/ATTRIBUTION.md.
No license has been chosen for the code in this repository yet.
ROMs and save states never enter this repository.