Driving school.

Reshape the road and watch the drivers adapt.

options

preparing the drivers

Loading

how it learns

First attempt: — · Best: — · Time found: —

A small neural network controls steering and a signed pedal command: brake, coast, or accelerate. It observes lateral position, heading, speed, bearings, and the curvature of the road ahead so it can prepare for corners. It has access to local track geometry; it is not learning from camera images.

The 410 network parameters were learned offline from 90,110 examples across 80 generated circuits, using a conservative controller with the same grip limits. On this page, 16 mutated networks try your track each generation. The best survives. The physics and scoring stay fixed; the network weights change. There is no scripted improvement or promise that every track will work.

Practice runs in a browser worker, using your device. All sixteen candidate attempts and the current survivor are simulated, then replayed together. Each visible round is one real generation; selection waits until its replay ends. Playback speed changes presentation only. The green survivor is retained unless a challenger scores better. Crashes end an attempt; completed laps rank above partial attempts, then faster laps win. Reduced-motion mode shows completed rounds without animation.

Reshaping pauses practice and keeps the last selected network weights. Applying the shape tests that same driver on the new circuit and starts a fresh comparison. Drawing a new circuit, choosing another circuit, or changing the starting driver resets to the selected starting model.

This is a simplified grip-limited driving model. Braking, acceleration and cornering share a limited tire-force budget; turning too hard at speed makes the car run wide. There is no prescribed racing line: any path within the road can win if it completes the lap faster. The model does not simulate suspension, tire temperature, gears or detailed tire slip. Close or overlapping road sections can confuse the local observations. Draw a roomy loop without crossings.

Follow a driver to inspect its actual pedal commands: rust for braking, gray for coasting, and green for acceleration. The line is its recorded trajectory, not a precomputed optimum.

Read the model ↗ · Inspect the pretrained weights ↗