A fly wakes up on a kitchen floor.
It is hungry, slightly cold, and surrounded by furniture. Its compound eyes receive the scene. Competing neural populations choose between food, warmth, light, rest, and exploration. Leg activity starts a gait. If the flight pathway wins, the wings accelerate until the body actually leaves the floor.
This is FlyLab, an interactive browser experiment built around fruit fly connectome data.
Open FlyLab: flylab.ankhit.com
FlyLab running in the browser: the shared kitchen, neural activity, compound-eye view, needs, and movement are visible at the same time.
Why build this?
A connectome is a map of connections. By itself, it cannot show whether a visual signal can guide a body around a table, whether a fly can switch from walking to flight, or what happens when one neural pathway is removed.
FlyLab turns that static map into a testable loop:
brain → body → environment → sensors → brain
You can observe the loop, intervene in it, and measure what changes. That makes the project useful for experimentation, education, debugging, neuro-inspired control, and visualization.
Here are three examples of what that means in practice.
Story 1: A developer asks, “Does the brain really control the wings?”
Imagine you are building a neural simulation. The fly moves, but you do not know whether the movement comes from its neural state or from a hidden steering rule.
In FlyLab, you can inspect the chain directly. Activity from descending neurons contributes to walking, steering, and flight permission. Wing power produces modeled forces. The body takes off only after those forces overcome support and gravity.
Now silence the relevant motor output. If the wings stop producing thrust, you have a causal test—not just a moving 3D model.
The same approach works for other questions:
- Does removing vision change the route?
- Does body feedback help the fly escape an obstacle?
- Can it walk with silent wings?
- Can it land when the neural goal changes to rest?
Every intervention connects a visible behavior to a measurable part of the model.
Story 2: A student asks, “How does a wiring diagram become behavior?”
FlyLab starts with the FlyWire FAFB v783 dataset. The source connectome contains 139,255 reconstructed neurons. After model-specific nodes are added, the runtime graph contains 144,794 cells and 3,346,722 directed edges.
Those numbers are difficult to reason about as a table. FlyLab makes them observable.
One screen shows the fly’s compound-eye panorama, active neural populations, internal needs, wing output, leg contact, current goal, physical movement, and trail through the kitchen. A student can watch a stimulus change photoreceptor activity, see a goal population win, and then watch the body respond.
The interface also exposes the difference between intention and action. A fly may choose food while physically blocked by the kitchen island. The brain goal remains visible while contact feedback triggers a turn or climb response.
That distinction is one of the most useful lessons in the project: deciding what to do and successfully doing it are separate problems.
Story 3: An experimenter asks, “Which model works across repeated trials?”
The Train button creates a second experimental fly with its own full neural graph. The main fly remains in the shared kitchen while the experimental fly enters a controlled training environment.
The user can train or test:
- food seeking;
- warmth and light seeking;
- obstacle escape;
- need-based choice;
- spectral cue discrimination;
- hypotheses about uncertain neurotransmitters.
Each browser receives a reproducible random seed. Different browsers can test different candidates and contribute results to a shared archive.
The current live-lab-v4 protocol does more than compare reward scores. A candidate must complete the task on two held-out seeds before FlyLab can accept it.
In the spectral task, the fly must reach reinforced food without touching an equal-brightness decoy. The current model completed a five-seed probe with five correct food contacts and zero decoy contacts.
FlyLab then tested a new candidate. Both models completed the held-out task, but the current model kept the higher mean score: 3.020 versus 2.690. The system preserved the stronger model and archived the rejected candidate, its parameters, and its trajectory.
That is useful because optimization systems often celebrate a better-looking number even when the agent never completes the task. FlyLab requires an observable outcome.
One browser, one fly
Every open browser computes its own neural model locally in a dedicated Web Worker. WebGL renders the kitchen and brain activity. A small Python aiohttp server connects participants through WebSocket.
When two people enter the same room, their flies share the kitchen. They receive one another’s position, posture, leg pose, and wingbeat state. More importantly, they can see one another through the compound-eye model. Furniture and nearer flies can occlude those behind them.
This creates a simple path from individual experiments to multi-agent ones. A future task might test following, avoidance, competition for limited food, or communication through modeled signals.
The brain runs locally in the browser. Once the data is prepared, neural computation does not depend on an external AI service or paid API.
What FlyLab can be used for today
- Causal testing: disable a pathway and measure which behavior disappears.
- Model comparison: evaluate candidates with fixed seeds, controls, and held-out trials.
- Distributed experiments: let multiple browsers explore different parameter hypotheses.
- Teaching: demonstrate the path from sensory input to neural activity and movement.
- Simulation debugging: separate goal selection, motor output, physics, and feedback.
- Neuro-inspired robotics: study competition between needs and switching between locomotion modes.
The model layers remain explicit: neural dynamics, sensory conversion, motor readouts, body physics, and added decision circuits can each be inspected or replaced. This traceability is what turns a visually interesting fly into an experimental tool.
What comes next
The next steps are to benchmark one neural graph against two, run more independent trials, improve receptor and neuromodulator models, add richer leg and joint mechanics, and design social tasks for multiple flies.
The long-term goal is to make every behavior easier to explain, challenge, and reproduce.
What would you test first?
If you could intervene in the fly’s vision, needs, synaptic hypotheses, or motor output, what experiment would you run first?
What result would convince you that the behavior came from the connectome rather than from the surrounding model?
Share a test in the comments. The best suggestions have a clear control, a measurable outcome, and a result that could prove the hypothesis wrong.
Try the live project: flylab.ankhit.com
Top comments (0)