Screenshot of Rocket League game featuring a sports car flying and hitting a large ball on a field
Courtesy of Psyonix

Cheaters Hacked an AI Bot—and Beat the Rocket League Elite

Expert players of the popular vehicular soccer game have been caught out by lesser players using a superhuman bot built on cutting-edge machine learning.

Last week, Reed Wilen, an elite gamer who uses the handle “Chicago” in Rocket League, a popular vehicular-soccer game, encountered a strange and troubling new opponent. The player seemed like a novice at first, moving their rocket-powered vehicle in a hesitant and awkward way. Then they caught and balanced the ball perfectly on the hood of their car, and dribbled it with superhuman skill towards the goal at high speed.

Not only was the other driver clearly a bot—it was also ridiculously good. “It is very confusing to play against,” Wilen says. “Its perfect dribbling would cause havoc on almost every player.”

Wilen is one of a number of elite Rocket League players to have recently encountered the bot in competitive play. It is not yet good enough to beat all comers, but it can play to a high level, allowing less skilled players to cheat their way to a higher ranking.

Rocket League is frenetic and extremely tricky to play. Each player controls a car capable of impossible acrobatics inside an arena where gravity and physics are apparently set to ludicrous mode. The objective is to use your vehicle to maneuver a giant ball past your opponent and into their goal, a task that requires considerable skill and patience. Sometimes two players work together as a team, making huge leaps, desperate parries, and accidentally colliding, all while trying to anticipate and counter their opponents' own antics.

Top Rocket League players will often launch their cars through the air to move the ball toward the goal, but Wilen says the bot he faced appears to have been trained specifically to carry it on the ground. “The bot doesn't really flip around too often and doesn't jump in the air,” he says, apparently because it hasn’t been programmed to, or learned how to do so. “Instead, it waits for the ball to come down, where it catches it on top of the car and performs a perfect dribble towards the opposing team’s net,” Wilen says.

The bot that Wilen and others have come up against is called Nexto. It picked up the ability to dribble and score using an artificial intelligence approach known as reinforcement learning, which has underpinned research breakthroughs that let computers master other difficult games such as Go and Starcraft. The technique has also been applied to more practical areas, including chip design and data center cooling in recent years. Reinforcement learning entails creating a program that can perform a task at a basic level and improve by responding to feedback as it practices.

The company behind Rocket LeaguePsyonix, part of Epic Games, allows players to deploy bots to practice against. In 2020 it made an application programming interface (API) available to help developers build bots more easily. Last April, a group of Rocket League enthusiasts with coding skills announced RLGym, an open source library for building reinforcement-learning bots for Rocket League. Later in the year, the group released several open source AI bots—including an especially skilled dribbler called Nexto.

Rolv, a member of the collective that built Nexto, who provided only a first name and says they work in AI, says the AI bots are not designed to work in competitive play, only as practice opponents. But apparently someone hacked Nexto to allow it to play in a human’s place. Rolv says the collective has several more advanced bots in the works, including one that can learn from watching human gameplay. The group is now reconsidering whether to release those more powerful bots to to prevent cheaters from making use of them as happened with Nexto.

Sergey Levine, an associate professor at UC Berkeley who studies reinforcement learning, says the situation roiling the Rocket League elite reflects how rapidly AI tools are becoming more accessible. He adds that it may be possible to detect bots like Nexto using machine learning, but that this is still an emerging science. “One way to detect RL agents is with other RL agents,” Levine says. “Applying this requires running training against the bot thousands of times.”

Another Rocket League bot developer, who knows the developers of Nexto and declined to give a real name but uses the handle Zealan on Discord, says the game is a fascinating challenge from an AI perspective. To have a chance in play against humans, bots must anticipate the result of actions many seconds ahead. “Nexto is already superhuman in some situations,” says Zealan, who has an interest in machine learning. “Trust me, in a couple years, there will be top level beyond-pro Rocket League bots.” 

Epic declined to comment but told PC Gamer that it is working on ways of detecting and blocking bots like Nexto. That could start a game of cat and mouse, with cheaters using more advanced bots and more complex methods to avoid detection, and the company responding in turn with new countermeasures. Psyonix later issued a statement saying that it had banned a number of accounts found to have been using the bot. The company said that it had introduced a new way for users to report suspected cheating, as well as new bot countermeasures, although it did not provide further details.

“I’m interested in how the situation will evolve,” says Ted Xiao, an AI researcher who has been following the episode with interest. “I’ve heard players are sharing the exploits that Nexto fails against, and I suspect the bot abusers don’t have the technical chops to fine-tune the agents further, so maybe there will be an impasse for now.”

Wilen, the elite player surprised by Nexto , says the situation may become more pressing if the bot learns to fling itself through the air with the ball in addition to dribbling on the ground. “I've heard that it will be soon, and when that happens, it’s not gonna be a fun sight for other Rocket League players.”

Updated 01-19-2022, 5.25 pm EST: This article has been updated with information from a statement released by Psyonix.