ESTEEMStream.News

ESTEEM Center for Equity in Science, Technology, Engineering, English & Math

ESTEEMStream.News

ESTEEMStream.News

The Rocket League AI that Teaches Itself

Meet Nexto, the Rocket League AI with 250,000+ hours of training
The+Rocket+League+AI+that+Teaches+Itself

The Psyonix-added Rocket league bots are anything but good. They drive around the field aimlessly, and it seems like they’re there to bully you, not help you. Eventually, the Rocket League community decided to do something about it. They created an amazing bot named Nexto. Nexto is a machine learning bot with over 250,000 hours (28 years) of training that taught itself to play the game. It learns by running hundreds of sped up game simulations at once. This helps it predict the possible outcomes for each movement. It has run enough simulations to bring it up to Grand Champion, the second highest Rocket League rank.

The bot learns using something called machine learning. Machine learning is when you assign point values, or rewards, to doing certain things, like touching the ball, or winning a game. You can also take points away for certain things, like losing a game. This way the bot will learn what, and what not to do, without human interference. However, this system makes it a little bit of a ball hog.

The bot trains by playing many sped up games against itself at the same time to accelerate learning. This way, it can try different things and learn how to repeat them and defend against them until it can defend against everything. Eventually, with enough training, the bot will be able to defend against everything. It has played so many games, that right now, it has over twenty-eight years of training. The game has only been released for 8 of those 28 years. Personally, if I had 28 years of free time, I think I would be better than just Grand Champion.

The bot has some things it is good at, and some things it’s not good at. The bot is very skilled at defense, and attacking in 1v1s, because neither team has a teammate. However, once it goes into 2v2s, its performance falls a little. Because the bot does not have a hive mind with both cars, the rewards are different for each car. This means they will play forward almost constantly, even though one bot is enough to handle 2 of most ranks. This means if you can get the ball past the bots (key word: if), it will occasionally be an empty net.

 

Related Stories

https://wiki.rlbot.org/

https://gizmodo.com/rocket-league-ai-nexto-gaming-1850006117

https://store.epicgames.com/en-US/p/rocket-league

https://rlbot.org/

https://github.com/Rolv-Arild/Necto

 

Take Action

Nexto was created for Rocket League using the RL Bot GUI at rlbot.org. You can find Nexto at the Necto Github Repository.

More to Discover