I Gave a Fruit Fly Connectome a Body—and Let It Learn in the Browser

A fruit fly wakes up on a kitchen floor, hungry and cold, surrounded by furniture. Its compound eyes perceive the scene, and its neural populations compete between food, warmth, light, rest, and exploration. As leg activity initiates a gait, the fly's wings accelerate, allowing it to leave the floor if the flight pathway is victorious.

FlyLab, an interactive browser experiment, is built around fruit fly connectome data. This project aims to transform static connectome maps into testable loops: brain → body → environment → sensors → brain. FlyLab allows observation of this loop, intervention within it, and measurement of changes, making it useful for experimentation, education, debugging, neuro-inspired control, and visualization.

Three examples illustrate FlyLab's potential applications. First, a developer may ask whether the brain truly controls the fly's wings. In FlyLab, the chain of events from neural activity to wing movement can be directly inspected. By silencing motor output, developers can conduct causal tests to determine if the wings stop producing thrust, confirming the causal relationship between neural activity and body movement.

Second, a student can use FlyLab to understand how a wiring diagram transforms into behavior. Starting with the FlyWire FAFB v783 dataset, the connectome contains 139,255 reconstructed neurons. After incorporating model-specific nodes, the runtime graph comprises 144,794 cells and 3,346,722 directed edges. FlyLab presents these data in a visually accessible format, enabling students to watch stimulus changes, observe neural population activation, and monitor the fly's response to the environment.

Lastly, an experimenter may inquire about which model works consistently across multiple trials. FlyLab allows users to train a second fly with its own neural graph while the primary fly observes the shared kitchen. The user can train or test various scenarios, such as food seeking, warmth and light seeking, obstacle escape, need-based choice, spectral cue discrimination, and hypotheses about neurotransmitters.

Each browser receives a unique random seed, enabling multiple experiments and contributing results to a shared archive. A strong model must successfully complete the task on held-out seeds before being accepted, ensuring reliable outcomes and preventing false-positive results.