Evolution Behind the Wheel
This car has no map. It sees the world through nine distance whiskers and feels its own speed — that is the entire input to a 16-neuron net that outputs only steer and gas. Nothing here was taught by backpropagation: the brain was evolved by a genetic algorithm. Switch circuits and watch the same champion adapt — that is what the per-track adapters buy you.
Overview
The driver is a feed-forward network of just sixteen hidden neurons. Its only senses are nine forward range-finders — like a bat's echolocation — plus its current speed. No global view of the track, no racing line, no rules: only what the whiskers report, frame by frame.
Methodology
The weights were found by neuroevolution — a genetic algorithm that mutates and selects whole networks by lap fitness, never computing a gradient. One frozen base brain is specialized per circuit by LoRA-style adapters: small low-rank weight deltas, one per curriculum level, so a single champion masters Monaco through Ironcliff.
Applications
Neuroevolution for control where gradients are unavailable, sim-to-real transfer of policies learned in simulation, and curriculum learning — bootstrapping hard tasks from easy ones with cheap per-task adapters instead of retraining from scratch.