Behind the
little robot.

The real wiring map

MaleCNS v1.0, Janelia Research Campus / FlyEM and collaborators. 166,700 neurons and 25,582,938 connections. The supplied flybrain matrix is a signed, normalized derivative, losslessly reordered into outgoing-column storage and gzip chunks. No connections are pruned and no weights are quantized.

MaleCNS project · CC BY 4.0 data license · Preserved source license and attribution

The simulation and learning

The simplified leaky integrate-and-fire browser model is adapted from flybrain 0.1.0 by alextitonis, MIT licensed. The connectome's biological weights stay fixed. An added 379-input, 64-hidden-unit, 12-output neural readout learns using coach-labeled examples, Adam, and learner-state corrections. A frozen run uses that readout to select every button; it is not a video or a coach-driven replay.

The exhibit does not show biological learning in a fly, consciousness, or proof that a fly understands a bike. The dots are real group summaries in a readable layout, not anatomical locations. Button votes are not calibrated confidence. Gold means the original stage time target; it is not a claim of a world record.

The original game and robot

Revere Bike's original game rules and Canvas artwork come from the supplied Revere reference project. Races use stock Brown Beauty at level 1, with zero upgrades, and retain the original special stages and unicorn presentation.

The tracked Island Explorer robot and its articulated rig come from the supplied Island v4.1 reference. Only the character is reused, not the Island world. Its visor transmission is disabled for rendering cost, and authored arm/finger poses hold the added unbranded retro gamepad. Gamepad design is inspired by classic two-button controllers; this exhibit is not affiliated with Nintendo.

Revere and Island game/character materials are not included in the reusable brain download, and this exhibit does not grant new reuse permissions for those reference assets.

Rendering, sound and chapter cards

Local Three.js r185 modules render the original robot and curved CRT. Postprocessing uses RenderPass, UnrealBloomPass and OutputPass. All sound effects, the quiet melody and room hum are procedural WebAudio. Sound is optional and requires your gesture.

Fifteen chapter illustrations were generated with the explicitly authorized ImageGen CLI fallback using the current robot/CRT screenshot as a continuity reference. Source PNGs, prompts, hashes and review records are kept with the project under art/intermissions; deployed chapter cards are optimized WebP files. Generated scenes contain no UI, readable text, logos or watermark.

Your experiment

Download the complete brain kit

Archive size and SHA-256 verification record. The kit includes the full dataset, reusable model/loader/adapter/readout, a worker-based example and licenses. Export your learned controller separately from the arcade notebook.