Sony AI exploring “what we can do” with Gran Turismo Sophy and its new “fourth skill area” – strategy

Waiting inside a busy Amsterdam Theatre, the tension ahead of a title-deciding race was palpable. But before the on-track action began, punctuated by competitive racing and a contentious ending, Kaushik Subramanian sat on a sofa to discuss an extended research project.

This was significant because whenever the Senior Research Scientist at Sony AI appears, Gran Turismo players pay close attention.

The Gran Turismo World Series World Finals event has a yearly tradition of announcing exciting new content and game developments.

Last week’s event was no exception, where an electric Hyundai and a naturally aspirated V12 single-seater were driven within Gran Turismo 7 for the first time.

Sony AI representatives were present to showcase a significant development to its formidable Gran Turismo artificial intelligence agent.

First discussed publicly nearly three years ago, and currently in the midst of a roll-out within Gran Turismo 7 (two more compatible tracks were updated earlier this month), behind the scenes, Sony AI has been working on expanding Gran Turismo Sophy’s capabilities beyond short races.

It now has a newfound, work-in-progress, ability – strategic manevouers.

“As you probably saw with the initial GT Sophy vision paper, we are trying to explore what we can do with the AI agent and one of the questions that came up was to try to see if we can maybe take a step towards trying to teach it about racing strategy,” explained Subramanian to Traxion.

“In the past, when you’ve seen papers that we have published, a lot of it has been largely about car control, racing skills and maybe embracing etiquette. But, we haven’t really looked at strategy as much.

“So we asked ourselves, ‘Can we take a step in that direction?’

“Working closely with [Gran Turismo developer] Polyphony Digital we are trying to see how we can train GT Sophy to become capable in that space.”

On the aforementioned stage, earlier in the weekend it transpires that 12 of the Gran Turismo World Series drivers had conducted a test race where they weren’t actually driving.

It reminded me of an old game mode, B-Spec, last seen in 2013’s Gran Turismo 6, where the player acts as a team manager or strategist, as opposed to being behind the wheel.

For many, a fundamental flaw with B-Spec was that the driver could be a little… let’s say, erratic. 

GT Sophy, on the other hand, has proved to be adept at humanistic behaviour, something that, in theory, could translate into a more engrossing management-style gaming experience.

To get to the exploratory stage as seen in The Netherlands, using the Red Bull X2019 Competition around the Lago Maggiore venue, the Sony AI team ran through some significant tweaks.

“For this particular exhibition, we did, of course, have to make some changes, in particular, allowing it to be able to handle tyre wear, fuel consumption and aspects of those that are important for strategy,” explained Patrick MacAlpine, Research Scientist at Sony AI, to Traxion.

“But furthermore, as a way to increase engagement and allow strategy to be tested, was to allow humans to be able to dynamically change how it potentially was using the tyres.

“Players were able to, for instance, change the pace of the car, and that was primarily controlling how much tyre it was using.”

In the Dutch exhibition, players could request GT Sophy to switch between one of five levels of outright pace, plus an additional aggressive mode – this can influence overtaking but can also lead to mistakes like losing control or the AI driving off the track. They also selected the time the it would visit the pitlane.

The signature on-screen emoticons displayed how the agent felt about their requested actions.

“Previously we’ve focused on racing car control, sportsmanship and racing tactics. But this is a new higher level, fourth, general skill area for GT Sophy,” continued MacAlpine.

“For GT Sophy to react accordingly, one primary update has been to the input features to the agent. We had to tell it how much fuel was remaining and how much tyre wear on each corner there was.

“Plus, we also had to update some of our reward functions so that it would be sensitive enough to know about consumption and wear, allowing it to be potentially penalised for using too much of one or the other.”

“We wanted to make sure that the agent would experience scenarios we might see in a race in terms of the distribution of tyre wear. For example, you may see more tyre wear on one side of the car than the other.”

So, when will it be in Gran Turismo 7 for those at home to test? Could there soon be an additional management-style mode using GT Sophy?

Good question. Sony AI is not committing to that happening, yet. For now, this is an additional strand to its ever-expanding exploratory research – which currently includes the creation of a second type of agent that can ‘see’.

“We don’t have any timeline for such sort of support or release like this,” responds Subramanian.

“We definitely imagine that you could have an agent with these capabilities. It would make it easier for players to try a few different strategies, see how the race will play out over say 10 or 15, laps, and then iterate on that process.

“But regarding whether players are going to get access, again, there’s no fixed timeline.

“Looking at the feedback online, we can see players want more of the technology in different parts of the game and we’re going to work with Polyphony Digital to see what we can do for that purpose.”

Which begs the question, what’s next? Subramanian claims that learning from the World Series driver feedback is the first priority before the strategy elements evolve further.

“This is an exploration completely on the Sony AI side,” he says.

“We had a research question that we wanted to answer about GT Sophy learning things about strategy, and of course, Polyphony Digital has a lot of experience with racing strategy, and so they guided us in this process.

“One thing that stands out is we want the response of the AI to match the player’s intuition and expectations. These players have raced so often, they know the craft really, really, well. For one of the buttons, which is an aggressive mode for overtaking, we think it intuitively makes sense. but the question is, does it match every player’s expectations?

“For some yes, for others, maybe less so and so we’d like to bridge that gap next, making sure when a player selects that option, they know what to expect.”

“For me, seeing what happened at the exhibition, it’s really exciting to see other ways that we can get GT Sophy to interact and engage with players,” concludes MacAlpine.

“That’s a new, exciting, research area and frontier that we’re exploring with this work.”