A new branch within Sony imaginatively named Sony AI has partnered with Polyphony Digital on a paper which explores the possibility of teaching artificial intelligence how to drive cars in Gran Turismo based on visual cues. The paper, authored by Ryuji Imamura, follows a similar document published in 2020, whereby a computer was taught how to drive at superhuman speeds through physical interactions with a track.
In this instance, the study focused on the key information that players use in order to master Gran Turismo games: the track itself, the heads-up display, and force feedback. The methodology is extremely complicated, but the AI was eventually able to achieve a lap time within the top 10 per cent or so of real-life players, and about 4.5 seconds faster than the average time posted by human players. Not bad at all, then, for a computer program.
Of course, there’s no suggestion that any of this learning will be applied to Gran Turismo 7, although it’s exciting that Polyphony Digital is participating in this kind of research. Poor AI has been a criticism levelled at past games in the simulation racer series, and it’s something that will need to improve if its single player experience is to match up with its best-in-class online offerings.