doomfly-rl: a network wired like a fruit fly's brain, playing Doom

Three checkpoints from nonatofabio/doomfly-rl, all step 60,000 of the same distillation recipe on the five standard ViZDoom scenarios, plus the connectome file they need. The project was built end to end by the Strands harness; the story is here.

Read a score as "connectome-constrained RL agent", never as "a fly". Two thirds of the parameters are an ordinary conv stem and MLP readout, and the controls below show the wiring adds nothing measurable on these scenarios.

The malecns49k_v2_final checkpoint on the five ViZDoom scenarios, one clip each, median episode of ten.

Files

Path	What it is
malecns49k_v2_final/	The real connectome: 49,393 MaleCNS neurons, 9,050,172 edges, frozen wiring and signs, one learned gain per synapse. 48.2 M params. The footage in the blog post.
malecns49k_shuffled_s0/	Control. Same graph with edges shuffled degree-preserving (seed 0): same sparsity, same fan-in/out, same signs, no biology. 48.2 M params.
malecns49k_noconn/	Control. No neuron layer at all, stem wired straight to the decoder. 27.7 M params.
connectome_malecns49k.npz	The connectome the first two checkpoints run on. Derived from fernandofernandes/fly-connectome-49k, CC BY 4.0. connectome_malecns49k.json has the source revision and manifest hash.

Each checkpoint directory holds model.safetensors and config.json (FlyNetConfig plus the training step).

Results (10 episodes, mean ± std of return, sampled policy)

checkpoint	basic	defend_the_center	health_gathering	deadly_corridor	defend_the_line
malecns49k_v2_final (connectome)	78.2 ± 7.5	18.6 ± 1.8	1728 ± 616	2280.5 ± 2.4	21.0 ± 5.9
malecns49k_shuffled_s0	78.2 ± 7.5	19.0 ± 1.7	1602 ± 597	2280.6 ± 2.2	21.7 ± 6.1
malecns49k_noconn	77.5 ± 8.5	18.3 ± 1.6	1373 ± 555	2279.9 ± 2.4	23.0 ± 5.0

Every gap is inside one eval standard deviation. Full protocol and training curves: docs/controls.md in the repo.

Use

git clone https://github.com/nonatofabio/doomfly-rl && cd doomfly-rl
uv venv && uv pip install -e . huggingface_hub
hf download fabiononato/doomfly-rl --local-dir hf

# play 10 episodes of every scenario and write the median one per scenario as MP4
python -m doomfly.evaluate --ckpt hf/malecns49k_v2_final --connectome hf/connectome_malecns49k.npz \
    --episodes 10 --gif-dir /tmp/clips --tics --pick median --fmt mp4 --device cpu

# the no-connectome control needs no connectome file but the loader still takes the path
python -m doomfly.evaluate --ckpt hf/malecns49k_noconn --connectome hf/connectome_malecns49k.npz --episodes 10

In Python:

from doomfly.surgery import load_flynet
model, cfg = load_flynet("hf/malecns49k_v2_final", "hf/connectome_malecns49k.npz", "cpu")
model.eval()

load_flynet widens the stored 22-action, 5-scenario head to 23/6 in memory so the same checkpoint runs on the five scenarios and in Freedoom II free play. On the five scenarios the output is bit-identical to the original head.

Model

FlyNet follows mlabonne/chessfly. Input: a 4-frame stack of 72×96 grayscale Doom frames through a conv stem onto the visual sensory neurons. Core: the connectome as a sparse recurrent layer unrolled 5 steps, with BatchNorm and homeostatic scale/shift, sign and topology frozen, one log-gain per edge learned. Readout: central, descending and motor neurons → 512 → 22 actions with a per-scenario legality mask, plus a 64-bin value head.

Training: PPO teachers (Stable-Baselines3, one per scenario) → ε-greedy rollouts → distillation for 60k steps, batch 128, lr 3e-4 (stem/readout) and 3e-3 (connectome gains), on one L40S. Details and every launch script are in the repo.

Limitations

• Five saturated ViZDoom scenarios only. The controls show no connectome effect there; whether a wiring prior matters when the policy has to learn rather than imitate is the open question (free play on Freedoom II MAP01 is next in the repo).
• No FlyWire-783 checkpoint here yet; the 138k-neuron backbone trains 22× slower and its final run is in S3, not yet published.
• Nothing here is a claim about biology.