阿飛 AFEI

One fruit fly. A brain model you can watch.

AFEI is a small, slow observation game. One fruit fly lives in a glass box on a windowsill. You place fruit and tap the glass, then watch what the model and the physics body do. There are 20 tasks, most of them short; once they are done, there are no new goals.

What you do

• Place fruit. Up to three pieces: strawberry or banana, from fresh to mouldy, close to the fly or far away. Fruit ripens in real time (over hours).
  
• Tap the glass. Code sends a synthetic looming signal to the fly's looming-detector neurons. The model's take-off neurons trigger the take-off; code sets when a take-off can start and assists the jump and flight.
  
• Open the brain map. Show the whole central nervous system, the brain, or the smell path. Fibres are drawn from measured neuron shapes (simplified); the glow is the model's firing at that moment.
  
• Look closely. Orbit, zoom and follow the fly, and give it a name.
  
• Complete 20 tasks. From "Watch it eat" to "Weaker fruit reached first". Each task gives a card with one sentence on what the model does and what the code does. Finish all 20 for a full clear.

The science, plainly

Afei's brain is a leaky integrate-and-fire (LIF) model of all 165,122 neurons and about 25.6 million neuron-to-neuron connections in the Male CNS v1.0 connectome, the published wiring map of one male fruit fly's brain and nerve cord. It runs on your PC in 1 ms steps. The body is the NeuroMechFly physics model, running in MuJoCo. The model receives smell, taste, leg contact and, when you tap, a looming signal. It cannot see (the eyes are not wired in yet) and it does not learn. All neurons share one parameter set, and whether a connection excites or inhibits comes from the dataset's machine-learning predictions of each neuron's transmitter. All of Afei's movements are assisted by code, and the game marks the assists. The assists include: half of the walking drive, the flight itself, turning toward fruit, extending the proboscis, grooming, the walk/rest rhythm, wall-following, and a stronger smell input when it is hungry. Model time runs slower than real time: about half to nine-tenths of real time on the tested laptop.

No way to fail

The fly does not die and tasks never expire. The model pauses while the game is closed; fruit keeps ripening in real time (over hours). Rewards (keepsakes and scenes) are for you and never change the model.

AI assistance: code, images, the 3D model of the fly's head and body, text, music and sound effects were made with the help of AI tools and reviewed by the developer. The Japanese text is an AI-assisted translation, not yet checked by a native speaker. The Simplified Chinese text is converted from the Traditional Chinese, not yet checked by a mainland reader. Afei's brain is a simulation of a connectome, not generative AI.

Features

• A simplified brain model built from the published Male CNS connectome, computed on your PC in 1 ms steps
  
• A physics body with six legs, gravity and contact (NeuroMechFly in MuJoCo)
  
• A cartoon fly; leg shapes from the flybody model
  
• Brain map with simplified neuron fibres, the model's current firing, and readouts: smell input left/right, taste, foot contact, neural drive left/right
  
• Up to 3 fruits, 2 kinds x 4 ripeness stages; with one fruit out and the fly hungry, its tag can show the smell-path neurons' current firing rate
  
• 20 tasks (16 easy, 3 harder, 1 very hard), 20 cards and a full clear; keepsakes and two more scenes (a veranda and a night windowsill) as gifts
  
• Engineering assists marked with icons and notes; model time shown on the brain map
  
• Original music and sound effects
  
• Traditional Chinese, Simplified Chinese, English and Japanese, switchable in Settings (Simplified Chinese: converted from the Traditional Chinese; Japanese: AI-assisted translation)
  
• Autosave; optional background mode in windowed play

Real data, model and engineering assists

• Real data: neuron shapes and connections from the published Male CNS connectome; transmitter types are the data authors' machine-learning predictions.
  
• Model: Afei's brain is a simplified LIF model, not a real brain. Afei is not a complete fly either; its body runs on NeuroMechFly and MuJoCo physics. On-screen brain and body numbers are model output, not recordings; fullness and the meal line are engineering values.
  
• Engineering assists: the walking rhythm, steering toward fruit, the grooming and rest schedule, and the take-off and flight assists are program code, and the game labels them.

Data and tools

Neural wiring and neuron shapes: Male CNS v1.0 connectome (FlyEM at HHMI Janelia, University of Cambridge, MRC Laboratory of Molecular Biology, Google Research), CC BY 4.0. Simplified for this game (signed weights for a LIF model, simplified skeletons and region meshes); the data creators do not endorse this game. Body: NeuroMechFly v2 / FlyGym (Apache 2.0) and the MuJoCo physics engine (DeepMind Technologies Limited, Apache 2.0). Leg shape from flybody (TuragaLab, Apache 2.0). Fonts: Noto Sans TC, SC and JP (SIL Open Font License 1.1). Made with Unity.