Step into the arena. Duel an experimental opponent
driven by the full network in your fly brain files.
Reward 0 · 0 feedback events · 0/51 learned weights
+1 land a hit · −1 take a hit · +5 win · −5 lose.
Training explores and learns. Evaluation freezes learning and removes exploration. Export before closing; import to continue later.
Your opponent can wait. Restarting clears placed blocks.
Building: press B or 2–4 for blocks. Aim at a surface within six blocks; E or right-click places, Q or left-click breaks your placed blocks. 1 selects the sword. Green outlines mark valid placement; red marks occupied or out-of-bounds space. You can jump onto blocks, and they obstruct vision and attacks. Maximum 512 blocks, eight blocks high. Builds reset each round and are not saved.
Every neuron and every connection row with known endpoints in your supplied FlyWire files is simulated in a background worker. The source connections are filtered; this is not every synapse in a living fly.
A 32 × 16 grayscale camera view stimulates annotated photoreceptor neurons. The worker receives only these pixels—not your position, distance, health, or a target identifier. The pixel-to-neuron mapping is artificial, because these files do not provide a validated retinal map.
Activity in descending neurons directly controls turning, forward/backward movement, and sword swings through an artificial readout with 51 trainable weights. There is no automatic pursuit, aiming, range-triggered attack, or obstacle-avoidance fallback. Collisions and sword hits still follow game physics. It may stand still, walk into walls, or swing at nothing.
This is a simplified recurrent rate model, not a verified digital copy of a fly. GABA and glutamate are assumed inhibitory, other transmitters excitatory; connection counts are normalized per receiving neuron. Neuron physiology, a reconstructed memory, a fly body, and biological validation are absent. A full wiring map alone cannot supply them.
Engineered reward learning: small random variations explore actions. A bounded Gaussian policy-gradient rule credits recent neural-to-action activity when the controller lands a hit (+1), takes a hit (−1), wins (+5), or loses (−5). It changes 51 added readout weights, not the millions of original fly connections. This is not a physiological dopamine model, and improvement is not guaranteed. Evaluation turns off exploration and learning. Rewards come from actual combat events; there is no distance-to-target reward.
Learning lasts for this page session. Export a small JSON learning file to keep it, and import it in a later session. It is not saved to your GKServer account. Restarting a round keeps learned weights but clears recent-action traces.
The small activity bars show sampled neuron rates; the picture below is the controller’s actual visual input. Neural simulation runs at up to 10 steps per second, slower on limited hardware. The first load downloads a compressed network; no extra downloads from FlyWire are needed. Your sound remains optional and plays on victory.
FlyWire: Dorkenwald et al., Nature (2024) ↗FlyWire data & annotations ↗