DoomFly: A Fruit-Fly Connectome Controls Doom

What was actually connected to Doom

The first thing worth correcting in the viral framing is this: no living fruit fly was placed in front of Doom. Alex Wormuth’s open DoomFly repository describes a different, and no less unusual, experiment: the game is connected to a simulation of the MaleCNS v1.0 central nervous system connectome of a male fruit fly. By the time the post appeared, the source code was already available for inspection and reproduction.

The loop is closed. Doom frames are converted into stimulation for simulated sensory neurons, activity then propagates through the network, and selected output signals are translated into movement and shooting. The game changes after each command, and the next frame becomes input again, so the simulation is not merely processing prerecorded video.

The scale is genuinely striking: the project materials list 166,700 neurons and roughly 125 million synaptic contacts for MaleCNS v1.0. But a large number of nodes does not automatically imply intelligent gameplay. The crucial details are how the image is encoded, which motor outputs are selected, the temporal dynamics, and the learning mechanism.

The launch instructions specify Python 3.11 or newer, a C++ compiler, ViZDoom, and Node.js 22 for the observer interface. Separate stages prepare the connectome data and computational core, after which the server runs an experimental model with learning. This is not just a clip with an appealing caption; it is a technical artifact that can be examined layer by layer.

Public descriptions, however, mention only a limited arena and survival trials. There are no formal results on wins, training time, or transfer of behavior to other scenarios. It is therefore more accurate to call DoomFly a demonstration of a closed-loop architecture than proof that a fly has learned Doom.

Why the experiment is more useful than the viral headline

DoomFly’s main value is not gaming skill but a reproducible connection between sensory input, a biologically defined network, and motor output. For engineers, it is a compact reference for building a closed-loop neural interface where latency and conversion errors become immediately visible in the agent’s behavior.

• Researchers get an understandable environment with continuous feedback.
• Open code makes it possible to separate connectome properties from hand-designed decoder choices.
• Doom provides an immediately visible result without specialized laboratory equipment.

The first question to test is how much behavior comes from the neural dynamics themselves and how much comes from rules mapping signals to keys. Without comparisons to a simple controller and without ablations, the effect is easy to overstate. DoomFly is interesting precisely at this boundary: brain-like complexity is already present, while the real source of observed behavior still needs to be separated honestly from the engineering scaffolding.

We previously examined the Codex-on-Raspberry-Pi case and how striking embodied-AI demos differ from verifiable engineering implementations. The fruit-fly Doom experiment continues that discussion about the boundary between an impressive prototype and its actual substance.