Race FLYNN

Code on GitHub · Paper (arXiv)

FLYNN is a recurrent neural network wired like a fruit fly's brain: every unit is a neuron and every connection a synapse from the FlyWire connectome. It was trained to drive a two-wheeled robot to a goal around obstacles. Here you drive the same robot through the same arena against it, and you only get what FLYNN gets.

Keyboard: arrow keys and Space. On a phone or tablet, on-screen buttons appear; hold the phone upright.

Hold the phone upright for the biggest view.

How to play

• ↑ / ↓ drive forward / backward at FLYNN's cruising speed; ← / → turn (on the spot if you aren't driving). On a touch screen, hold the arrow buttons: turn with your left thumb, drive with your right.
• Reach the white ball first. You have 30 seconds.
• Space (or Start) starts a round and moves on to the next layout, P pauses, R (Retry) retries the layout, N (New layout) skips to a new one.

What you see

• Eye view. The robot's two cameras look 55° to either side, so straight ahead is the white line between the two panels. Each dot is one of FLYNN's photoreceptor columns, sampling its camera at that point. This is the robot's entire view of the world.
• Goal arrow. The direction of the goal relative to your heading (FLYNN senses it as "wind" through its antennae neurons). There is no distance.
• Bump lamps. The touch sensors on the left and right of the robot's head.

How it works

FLYNN's side is a recording of the trained network driving in the original MuJoCo physics simulation, so its runs are exactly what the network did. Your robot is simulated in your browser, with its motion, collisions and eye view calibrated against that simulation. The two robots never meet: each drives its own copy of the arena, so racing a recording is the same as racing the live network. Times are simulated seconds.

Paper: FLYNN: Robust Neural Network for Robot Navigation using Fly Brain Topology (Wang & Chen, 2026). Code: github.com/ben-gitdev/fly-gym.

Connectome data: FlyWire FAFB v783, CC BY-NC 4.0.