Gigapixel

Self-play works for driving the way it worked for Go: let policies practice against each other for billions of rounds and good behavior emerges without anyone recording a single human trajectory. The catch, until now, was that the fast self-play simulators fed their cars a bird's-eye vector map — a god's-eye list of boxes and lanes. A real car doesn't get that. It gets pixels from a camera.

Where the imitation-trained car plows into traffic at speed, the self-play car yields so cautiously it gets stuck — it learned defense, not mimicry.

Gigapixel closes that gap by rendering the simulated world into an ego-centric, camera-like image fast enough to keep self-play running at scale — about 50,000 driving steps per second on a single rentable GPU, roughly a thousand times faster than the photorealistic renderers researchers usually reach for. The trick is refusing photorealism: the world is drawn as bare cuboids for cars, thin strips for lanes, and small spheres for traffic lights. The bet is that geometry, not texture, is what a policy needs — and that speed buys more than prettiness.

A policy trained this way, from pixels and self-play alone, edges past its human-imitation baseline on a closed-loop test track where the car must actually drive rather than match logged trajectories. The failure mode is a virtue: where the imitation-trained car plows into adversarial traffic at speed, the self-play car yields so cautiously it sometimes gets stuck — it learned defense rather than mimicry.

The honest limit is baked into those bare shapes: a world of boxes and spheres has no debris, no weather, no strange lighting, so a car raised in it still needs paired sim-and-real data before it meets a real road. What the work shows is narrower and real — that the scale advantage of self-play, long stranded on abstract maps, can now reach the pixel input a shipped camera stack actually runs on.