‘Gran Turismo’ Champ Who’s Beat Some Of The World’s Best Isn’t Human

Researchers at Sony AI have created a bot that is exceptionally good at playing the racing simulation video game Gran Turismo. So good, in fact, that it’s capable of beating some of the world’s best players—and it did.

Sophy, as the AI is called, is not only fantastic at racing a car, but also at devising tactics and strategy to get to the finish line, as well as still respecting the etiquette and sportsmanship that comes with racing. 

These autonomous AI agents were trained through a novel deep reinforcement learning platform developed by Sony AI, Polyphony Digital, and Sony Interactive Entertainment, according to a press release. 

The AI agents were trained on three car and track combinations in Gran Turismo: Sport, with elements such as track positions, starting speeds, and opponent skill level randomized. 

Polygon reports that more than 45,000 hours of machine learning were put into the training, with a study on the subject being published in the journal Nature. 

Their training was put to the test in a presentation: races between four Sophy bots and four Gran Turismo esports competitors. Tomoaki Yamanaka, Takuma Miyazono, Ryota Kokubun, and Shotaro Ryu all have credentials like Gran Turismo Cup Champion and Nations Cup Champion, for example. 

After racing the Autodrome Lago Maggiore course in the Porsche 911 RSR Type 991, Sophy ended up with 1st, 3rd, 5th, and 7th positions on the grid, followed by the four players. 

“Gran Turismo Sophy is a significant development in AI whose purpose is not simply to be better than human players, but to offer players a stimulating opponent that can accelerate and elevate the players’ techniques and creativity to the next level,” Hiroaki Kitano, CEO of Sony AI, states in the press release.

There’s “still room for further improvement,” according to the firm. There are plans to incorporate Sophy into future iterations of the Gran Turismo, perhaps as a formidable opponent for (real) players to train against. 

[via Ars Technica and Polygon, images via Sony AI]