Train neural networks to drive by adjusting reward functions and evolutionary parameters, then watch your AI navigate custom tracks and campaign missions. The game offers hands-on tinkering with neural weights and fitness metrics, making it especially appealing to those curious about machine learning mechanics rather than traditional racing gameplay.
About this game
Train AI to drive! Create your own neural network and watch it evolve from chaos to mastery with the power of evolutionary algorithms and natural selection. Fine-tune the learning process to craft the ultimate AI driver, and test your creation in campaign challenges or on custom-built tracks.
Players, last 68 days
Peak 54 · Low 16
Daily peak concurrent players, sampled once a night from Steam.
What players are saying
▲ Recommended2 hrs
Great piece of software, just wish you could modify what gives the network rewards, for example right now the ai prioritizes slowly getting to the goal over speed, but i want it to be fast. Let me make it so the score is higher the faster it goes.
It's really fun and a good learning tool. Something I would like to see is the ability to directly edit the weights of a car during training to see the effect it has on the car.
Edit: The dev has since added the ability to directly edit weights. Nice!
As someone who really enjoys messing with machine learning I'm having a great time with this. My only note is I would love a feature to let the boundary of the generation be laps instead of time. Makes it hard to measure metrics that are avgs against one another on longer tracks. Maybe even leaderboards for the various lineages in a given session?
To the modders: I love what you're doing, and I don't mind the reverse engineering. I'm glad you chose my game, and it's amazing to see what these cars can really do. Why: the community raised a fair point, and this is where I landed. On the main leaderboards everyone should have the same tools, even if they're limited. I never wanted this game to become staring into graphs. It started as a learning tool and grew into a game with a great community. Please tag modded networks properly. Thanks to everyone who does. Untagged external tools count as cheating and mean a permanent ban. I hope this works for both sides. Feedback welcome. How it works (1.2.1) Modded networks are hidden on leaderboards by default, and everyone below them moves up. Tick Show modded networks to see them. They're shown by default if your own entry is modded. Hall of Fame, Track of the Week and the Discord leaderboards count only non-modded networks.
{STEAM_CLAN_IMAGE}/45280890/7883f4743bd715bd98bb0890b1966fa79a1438bd.png{STEAM_CLAN_IMAGE}/45280890/f26cb24483f7cd64b473c79215c6a59cbb512c08.png 🚀 NEW FEATURES Adaptive Simulation Speed (MAX): Simulation speed changes based on framerate up to 8x. Speed button now cycles 1x → 2x → MAX → 0.5x Chart Generation Range: Show only the last 10, 50, 100 or 500 generations in the training chart many generations default to last 100 🛠️ IMPROVEMENTS Performance: Faster physics with many cars on track Faster Loading: Menus load networks faster; race opponents and bundled race tracks are much smaller 🐛 FIXES Fixed: High RAM usage in charts when training on random tracks Fixed: Mouse wheel zooming the camera while scrolling or using sliders in panels Fixed: Various minor bug fixes
Final results All rounds: https://www.youtube.com/playlist?list=PLM4dUSrwOPCg Congratulations to the winners, and thank you to all 20 entrants! Winners: I'll contact you by email in the next few days. Please check your spam folder.
Posts come from Steam's official announcements feed.
