Sony’s Gran Turismo AI thrashes the world’s best drivers

Sony’s brand-new artificial intelligence laboratory revealed its first breakthrough technology on Wednesday — GT Sophy, an AI capable of superhuman reactions and airtight racing strategy to beat the most skilled human drivers. It’s the cover star of this week’s edition of the science journal Nature, but don’t go looking for it at launch in Gran Turismo 7The next month.

Sony states that GT Sophy’s success is not due to the fact that the AI can execute complex decisions at lightning speed but rather because it can race three lines. Gran Turismo Sport Tracks down to the millimeter. Motorsport, noted Michael Spranger, Sony AI’s chief operating officer, also relies upon etiquette — hard and aggressive driving that still plays fair and observes the spirit of rules as much, if not more than, the letter of what is legally allowed.

Due to the peculiarities of racing etiquette it can be extremely challenging to code an AI for a classic racing video game that drives hard, but fair. Kazunori Yamauchi (creator of Gran Turismo) and chief executive officer of Polyphony Digital said GT Sophy was designed to respect competitors as well as to act in a manner that would be respected by human drivers.

“The agent should be a friend, a comrade, a buddy to human beings, an agent that people can feel sympathy with,” Yamauchi said, through a translator. “Also, the agent can stimulate the emotion of people, so that the agent and human beings can mutually respect each other.”

In a racing demonstration following a half-hour presentation, four GT Sophy bots raced against four Gran Turismo esports competitors — Tomoaki Yamanaka (2021 TGR GT Cup champion), Takuma Miyazono (2020 Nations Cup world champion and 2021 runner-up), Ryota Kokubun (2018 Nations Cup Asia/Oceania Champion), and Shotaro Ryu (runner-up, 2019 Japan National Esports champion, youth division). GT Sophy cars were placed in 1, 3, 5, 5 and 7th place on the grid. Ryu, Yamanaka (2018 Nations Cup Asia/Oceania Champion), Takuma Miyazono (2020 Nations Cup world champion and 2021 runner-up) and Ryu (runner-up, 2019 Japan National Esports champion, youth division). Eight raced on Autodrome Lago Maggiore. This is a fictional circuit. Gran Turismo SportPorsche 911 RSR Type 991.

GT Sophy Rouge, the pole-sitter, led wire-to-wire and won by 5.8 seconds over Yamanaka; Rouge’s fastest lap was 1:54.373, more than two seconds faster than Yamanaka’s best at 1:56.422. Both are extremely dominant margins given the competition. Yamanaka pulled within 1 second of Rouge midway through the second lap. But the constant pressure from GT Sophy Lavande who started and finished third made Yamanaka incapable of making any serious overtaking efforts.

Rouge was very aggressive at several points, so much that observers speculated that the human-racing stewards would report the AI to track officials for breaking the limits. Rouge evidently took every allowed millimeter of track, and no penalty was issued. Yamanaka was almost overtaken by Lavande in turn 7, but he closed the door and focused on his defense, giving the race to Rouge. The first four on the grid all finished in order, with Kokubun and Ryu taking fifth and sixth after GT Sophy Emeraude smacked the wall on lap 2 — a racing risk that seemed to affirm GT Sophy’s sophistication more than disprove it.

Yamauchi, in a Q&A following the demonstration, acknowledged that developing an unbeatable AI might be a technical achievement, but not much fun for everyday players. “In any sense, GT Sophy will always understand the surrounding environment and the conditions, and that includes the level of the players, too,” Yamauchi said. “So I’m sure that, ultimately, we will be able to provide joy and fun as GT Sophy races with people.”

GT Sophy joins Gran Turismo 7, This will go live on PlayStation 4 & PlayStation 5 March 4. However, the AI will become available through an upgrade. Yamauchi and Sony didn’t give a window for when that update would come.

But to give an example of the kind of heat Sophy can bring, if necessary, Yamauchi explained that the AI has mastered a type of cornering that most humans wouldn’t conventionally attempt, much less ever be taught.

Typically, racing drivers are taught to brake in a straight line under a “slow in [to the curve]Fast out [exit]” philosophy. “Gran Turismo Sophy doesn’t do that, necessarily,” Yamauchi said. “When Gran Turismo Sophy goes into a curve, it actually brakes as it turns into the curve. Usually when you go into a curve, the load is only on the two front tires; but Gran Turismo Sophy’s case is that you have the load on three tires, two in the front and one in the rear as well. The car can brake while turning.

“We notice that, actually, top drivers such as [seven-time Formula One world driver’s champion]Lewis Hamilton [2021 world champion] Max Verstappen actually are doing that, using three tires, going fast in and fast out, all these things that we thought were unique to GT Sophy,” Yamauchi said.

It points both to how much GT Sophy has learned — over 45,000 hours of machine learning, in fact — and also how much more it can discover.