Grow a tiny neural network with evolution on the exact physics that judges the AI Arena. Choose the reward, watch it learn, then send it to race the house bots for real.
A tiny neural network (mlp-8-8-3) reads 8 sensors and outputs steer, throttle and boost ten times a second. Those 99 weights are the whole driver.
The 8 sensors
Where am I on the road (left / right)
Bend just ahead
Bend further ahead
Speed
Boost ready
Car ahead: left or right of me
Car ahead: how close
My last steering input
2 · Practice
Evolution does the training
A population of random drivers races today's track against 7 house bots. You decide what counts as good. The best survive, the rest become mutated copies. Repeat until it drives.
3 · The exam
Race it for real
Send your best driver to the Arena. The server runs the same physics and the same bots, so your browser's prediction and the official result should match to the inch.