FLYDINODEVELOPMENT LAB

A fly circuit. An endless runner.

FLYWIRE v783 · SAMPLED CIRCUIT · EXPERIMENTAL SIMULATION

CONNECTINGRUNNER / 01

SCORE00000
BEST00000
CHROME DINOCIRCUIT CONTROLLED
RUNNING · Chromium engine
ATTEMPT—SPEED—CLEARED—

RECORDED OUTPUT INPUT → CIRCUIT → ACTION

One server simulation shared by all connected viewers.

RECENT RUNS LATEST 20 · SAVED ON SERVER

Waiting for the first completed run.

Inside the experiment

The game runs the original Chromium 74.0.3719.3 Runner.update, Trex physics, obstacle generation, collision boxes, scoring and day/night logic (BSD license). A host adapter supplies a fixed 60 Hz clock, seeded randomness and shared canvas frames. The error-page intro and audio are omitted. Chromium license.

The controller simulates a sampled FlyWire FAFB v783 subgraph: 268 identified neurons and 1,429 directed connections with original synapse counts. These were extracted from the first 65,536-row batch of the official connectivity archive, not the complete connectome.

One simulated network receives a 120 × 30 luminance image reconstructed from the original game drawing commands. Four fixed image regions supply contrast and temporal-change signals to annotated visual projection cells LT1d, LT1a, LTe12 and LC31b. These receptive fields are engineered, not measured biological vision.

Activity propagates through the sampled connections. Four annotated descending neurons (DNp103, DNp45, DNpe050 and DNpe056) form two fixed output groups. Their relative firing determines jump, duck or no command. This button mapping is arbitrary, not a claim that these neurons biologically cause jumping or ducking. The controller does not receive obstacle labels, distance, height or game speed.

No learning or performance guarantee is implemented. Initial tests collided with the first obstacle after removing the scripted sensory routing. This is a partial connectome-based simulation, not a full fly brain.

The 3D brain displays one annotated location per neuron, not complete neuron skeletons. Lines show selected graph edges, not axonal trajectories. Orange points show computed firing in the active subgraph. The EPFL NeuroMechFly specimen is a 3D visual model, not a biomechanical simulation.

Live viewers share a server run. Replays preserve recorded frames. Blind-input and severed-connection controls are tested for loss of output activity; this establishes software dependence, not biological validation or superior playing ability.

Inspect circuit IDs, weights and sampling method ↗

FlyWire connectivity archive ↗
Neuron annotations v2.1.0 ↗
NeuroMechFly model · Apache 2.0 ↗

FlyWire data documentation ↗