Build. Evolve. Compete.
Evolve a car from scratch or put your best creation up against drivers from around the world.
Evolve a population of cars, generation by generation, until one is fast.
Submit your best chassis — the server re-runs it on official roads to confirm the score.
Race verified champions from around the world and climb the rated leaderboard.
The full loop
Five ideas tie the whole game together — from a blank population of shapes to a ranked global record.
You don't design a car directly — you tune a genetic algorithm and let it evolve one. Each generation, a population of random-ish triangle chassis with wheels drives across a procedural road; the cars that get furthest are bred (crossover + mutation) into the next generation. Fitness is simply distance travelled.
Training happens entirely in your browser, so before a car counts for anything competitive, the server re-simulates your best genome itself on official roads to confirm the score. That verified genome becomes your champion — the one thing "Compete" and the Gauntlet both use.
Your champion is matched against another player's (or, if the pool is thin, a house bot) on a private road neither of you has seen. Furthest distance wins and ratings move on a standard Elo scale.
This one isn't a match you start — it's automatic. Every time you verify a champion, it's also replayed solo against this week's shared public road and your distance is posted to the Gauntlet leaderboard. There's no separate "enter" step; re-verify anytime to try to beat your own score before it resets Monday.
Everyone starts at 1000 rating. It moves with every ranked race and slowly decays if you go quiet for more than a week, so it always reflects recent form. Every so often the season ends: the top 10 are frozen into a permanent Hall of Fame, then every rating is compressed back toward the middle — not wiped — so the new season starts closer together instead of a full re-grind.
Before you tune the panel
The Hyperparameters and Genome panels have a lot of dials. Here's what each one is actually doing.
Heavier (denser) triangles low in the chassis lower the center of mass, which keeps a car from flipping on steep tiles. Wild, spiky shapes can reach further per bounce but tip over more easily.
More wheels add stability and grip but also weight to haul around. Wheel placement matters as much as count — evolution has to discover both together.
High elitism keeps your best cars safe but can stall progress; high mutation explores wilder shapes but can wreck a good design. Selection bias controls how strongly winners get to breed. There's no fixed best setting — it depends on how stuck the population looks.
The road starts flat and gets steeper and choppier the further you go, so late-run distance is a much harder test than early distance. A car that stalls at tile 50 every run probably needs a stability fix, not just more generations.
GenCars is a small genetic-algorithm sandbox: you don't drive, you breed. Evolution runs entirely in your browser; the arena and weekly Gauntlet are backed by a server that keeps every verified score honest by re-running it itself. No two roads a champion faces in the ranked arena are known in advance to either player.
Evolve AI cars, then race your champion against the world.
Cars are polygon chassis drawn as triangles with wheels at vertices, driven by motors. The population races; furthest cars are selected, crossed over and mutated to breed the next generation. Both seeds make a run fully reproducible.
Ranked · ruleset v1
Your champion, rating and verified race history are tied to your account.
No races yet.
Watch AI cars evolve from scratch using a genetic algorithm — each generation, the fastest cars breed the next wave.
Train a car in the solo sandbox — tune mutations, body shape and wheels until you've built something fast.
Sign in with Google, verify your best car in the Arena, and race it against real opponents for an Elo rating.