A toy game learning in your browser. A shared 3 → 6 → 3 neural controller chooses each saber’s position and readiness. Score-based evolution improves 45 real weights, with no prerecorded actions.
The large red and teal nodes show the real controller twice, once per arm. Their size and pathway strength follow its activations and hits. The twelve layered lobes, small neurons and their traveling signals are illustrative anatomy, with invented topology and paced timing. No biological or connectome fidelity is claimed.
1× cycles actual simulation and learning steps. ↶ restores the policy with the best score on a fixed held-out validation set. Reset starts new learning from the same seed. Sound is optional, synthesized locally. Drag the brain to orbit it.
Red and blue targets need the matching saber near their center and ready at impact. The percentage is validation hit accuracy on the toy task.