Ideally, you want the AI in a racing game to be good enough that the race is a challenge, but perhaps not quite so unbeatable that it’s a fool’s errand to try to win.
That has been harder than it sounds to get right—the older players among us will remember the AI in earlier installments of GT, which mostly stuck to a preprogrammed line with little concern for where you, the human player, wanted to put your car. Microsoft’s long-running Forza franchise actually has its origin in a neural net experiment, and more recently, its developers used each human player to train an AI agent that would populate their friends’ games, albeit with imperfect results.
Sony AI introduced Sophy to the world in 2022, publishing a paper in the journal Nature explaining how the company trained the AI to be able to beat the very best human GT players in the world. In late February of this year, a time-limited feature added a series of four-versus-one matches against Sophy set to different difficulty levels, as well as a match where the AI is completely unleashed—and almost completely unbeatable, at least to us mortal players.
“One of the challenges that really sticks out was trying to find a good balance between having the Sophy car be respectful of opponents around it—so have good sportsmanship—but at the same time being competitive in a race and making sure that it tries to take advantage of positions,” explained Kaushik Subramanian, senior research scientist at Sony AI.
