ROBOT BRAIN
ARCADE
A LITTLE ROBOT.
A VERY BIG BRAIN.

AUTONOMOUS PLAY · REAL NEURAL ACTIVITY

Small robot.
Big learning energy.

A real fruit-fly brain map, a robot with a gamepad, and fifteen wild rides.
Watch it see, decide, press, and learn.

CH. 01
NOW PLAYING

Revere Bike

Warming up
Setting up the arcade… Full brain: 138.6 MB download. A desktop browser is recommended.
RIDE TIME 0.00s
GOLD TARGET —
YOUR SESSION BEST —
READY TO ROLL
LEARNING CIRCUITTry, race, keep or undo
CERTIFIED REPLAYVerified champion, same controller
LEARNING CIRCUIT · AUTONOMOUS OPTIMIZERREADY TO TEST THE CHAMPION COPY
TRY 0 · GEN 0
STAGE0 / 15
BONUS0 / 6
CURRENT TOTAL--
ACCEPTED BEST--
LAST TRY--
VS ACCEPTED--
GOLD STAGES0 / 15
MUTATION SIZE0.012

The certified champion copy is your accepted controller.

Waiting for the first candidate change.

Try one small button-decider change → race all 15 stages → keep it only if it is faster and earns 15 Golds. Otherwise, undo it. The fruit-fly wiring stays fixed, and nothing changes during a race.

Progress has not been saved yet.
CERTIFIED REPLAY READY FOR CERTIFIED REPLAY
RUN --
STAGE 0 / 15
BONUS ROUTES 0 / 6
RUNNING TOTAL --
LAST REPLAY --
FASTEST REPLAY --
REPLAY DELTA --

The robot will start with the certified champion.

All 15 stages including six bonus routes. Lower total time wins. The verified controller never changes.

LIVE SIMULATION, NOT A VIDEO ISLAND EXPLORER · ORIGINAL RIG

Saved bests are separate for each controller.
THE BIG IDEA / 01

Practice is allowed to be messy.

First, a learner tries things and updates its button-decider. Then a previously trained controller takes a fresh ride. The labels always tell you which one is driving.

YOUR NEXT EXPERIMENT

Take this brain
home with you.

The complete connectome, not a tiny demo.
Build something wonderfully weird.

Download the brain kit ↓ Full dataset + reusable JavaScript + source credits
The brain notebook · what is actually learning?

A map, not a tiny conscious rider

This is a simplified spiking simulation using a real fruit-fly connectome. A connectome is a wiring map. This exhibit does not prove that a fly understands a bicycle or thinks like a person.

The character is the original Island explorer robot. The biology behind its controller is insect-derived; the robot itself is not an insect.

Read the original FlyLab experiment for the first public account of this project.

What changes with practice?

24 road signals stimulate visual-neuron groups. We measure 379 features from the full network and feed them to a small 379 → 64 → 12 button-decider. Only that added decider's weights change.

Practice uses coach-labeled examples and Adam training. A trained run uses a frozen neural checkpoint. No coach selects its buttons.

The frozen champion earned Gold in all 90 tests across 15 stages. Read the verification record. These tests are evidence, not a promise that every future experiment will win.

Make your own experiment

Electric-signal dots are group summaries arranged for readability, not anatomical positions. “Loss” measures disagreement with teaching examples; lower loss does not guarantee a faster ride.

Waiting for dataset verification.

Watch the button-decider change · save your experiment

64 thinking units, 12 button combinations

The top row shows real hidden-unit activity. Below it, each little square is an actual learned connection to a button combination. Purple and green mean opposite weight signs; brighter means stronger. Only this added network learns.

Waiting for the first neural choice.

Your learner is saved on this device when browser storage is available.

Sound & comfort settings

Sound starts only after your click. Settings are remembered when storage is available.

WELCOME TO THE ARCADE

You watch. The robot plays.

1. Learning Circuit: clone the champion, try a tiny button-decider change, race every stage, and keep it only when it is faster with 15 Golds.

2. Certified Replay: the verified champion is frozen, so the same seed produces the same ride and time.

3. Practice: a separate learner updates from coach examples. Its loss and updates do not describe Learning Circuit.

4. Follow the signals: road → brain activity → button votes → fingers → bike. A gallops, B boosts, and the D-pad leans.

You can pause, change playback speed, inspect the brain, or stop the autonomous loop. Playback speed never changes the simulation score.