Gran Turismo Sophy: Sony AI x Polyphony Digital

  • Thread starter Magog
  • 1,719 comments
  • 191,255 views
The Key is in the wording "Breakthrough " which is not used unless you have managed to do something new, and if they have had a Breakthrough, then what ever it is, it's ready to use, They also stress AI Gaming Division.

So it's with an AI Gaming Division and theh have developed something new, So it's something to do with Gaming Ai, in partnership with a game that uses AI, when the gaming company is about to luanch a new game, If it was a tie in marketing, it would be marketed similar to Brembo announcement.
 
The Key is in the wording "Breakthrough " which is not used unless you have managed to do something new, and if they have had a Breakthrough, then what ever it is, it's ready to use, They also stress AI Gaming Division.

So it's with an AI Gaming Division and theh have developed something new, So it's something to do with Gaming Ai, in partnership with a game that uses AI, when the gaming company is about to luanch a new game, If it was a tie in marketing, it would be marketed similar to Brembo announcement.
Let's not pretend that marketing people aren't keen to misuse language wherever possible. Just because they chose to use the word "breakthrough" doesn't mean that it's actually a breakthrough in any real world sense. And just because you've had a breakthrough, it absolutely does not mean that it's ready for widespread use. Discovering atomic fission was a breakthrough, it didn't mean that they could suddenly make nukes the next day.
 
There is nothing to prove
Codemasters AI have improved significantly in the last couple years. They can be a bit overzealous, especially in F1 2021 (but then again.. so can humans, and their aggressiveness normally sees them take the damage instead of you) - but considering they're AI and aren't operating on the same physics as the player, they offer really good close racing, most of the time leaving you unharmed.
While I do agree, comparing F1 2021 to GT7 is not really fair.

In F1 2021 you have essentially 1 car with slightly varying performance specs through the different teams and upgrades in career and such. That is considerably easier to get to an acceptable level than the 400+ different cars in GT7 with far more influential upgrades than in F1.
 
The Key is in the wording "Breakthrough " which is not used unless you have managed to do something new, and if they have had a Breakthrough, then what ever it is, it's ready to use, They also stress AI Gaming Division.

So it's with an AI Gaming Division and theh have developed something new, So it's something to do with Gaming Ai, in partnership with a game that uses AI, when the gaming company is about to luanch a new game, If it was a tie in marketing, it would be marketed similar to Brembo announcement.
You highly underestimate the power (and deception) of marketing.
 
Last edited:
There is nothing to prove

While I do agree, comparing F1 2021 to GT7 is not really fair.

In F1 2021 you have essentially 1 car with slightly varying performance specs through the different teams and upgrades in career and such. That is considerably easier to get to an acceptable level than the 400+ different cars in GT7 with far more influential upgrades than in F1.
F1 does have far less variety in car performance and upgrades than something like GT, but I think you're underselling it a bit. The difference between cars and tyres is substantial, as is the number of upgrades. Add in a fairly granular damage system, track evolution and tyre degradation, ERS/DRS, weather changes, pit stops, flags, safety cars, and a wide range of AI settings. The games have also had F2 and historic cars available. You start to see that there's still a lot of variables that the AI system has to take into account. There's a lot going on in an F1 race that isn't going on in your standard 3 lap dash in GT.

Expanding the AI system from something like F1 into a game like GT would be much more complex, but the bulk of the heavy lifting is arguably done. I don't think anyone reasonably expects the AI in an expansive game like GT7 to directly match the AI in single make or limited field games like F1 or ACC, the opportunities for fine tuning a game like GT7 just don't exist. But it'd be nice if it was at least on the same page.
 
F1 does have far less variety in car performance and upgrades than something like GT, but I think you're underselling it a bit. The difference between cars and tyres is substantial, as is the number of upgrades. Add in a fairly granular damage system, track evolution and tyre degradation, ERS/DRS, weather changes, pit stops, flags, safety cars, and a wide range of AI settings. The games have also had F2 and historic cars available. You start to see that there's still a lot of variables that the AI system has to take into account. There's a lot going on in an F1 race that isn't going on in your standard 3 lap dash in GT.

Expanding the AI system from something like F1 into a game like GT would be much more complex, but the bulk of the heavy lifting is arguably done. I don't think anyone reasonably expects the AI in an expansive game like GT7 to directly match the AI in single make or limited field games like F1 or ACC, the opportunities for fine tuning a game like GT7 just don't exist. But it'd be nice if it was at least on the same page.
Nah, I don't think "the heavy lifting" is done at all. I think quite the opposite, the easy part is done: AI for 1 car with slight variations driving against 1 car with slight variations. (or 2 cars with no variation for the F2 plus the historic cars which had very limited use in the past, I think only some overtaking challenges and time trials? Correct me if I'm wrong)

The difficult part would AI for 400 cars with quite big variations, racing probably about 100 different cars with quite big variations on considerably more tracks as well. If you consider endurance races and the weather system, the points you made for F1 are (or can be) actually largely the same for GT. I truly believe creating AI for a game like Forza or GT is on a whole different level compared to ACC or F1.

That said it can absolutely be better than it is and it should be better, it is not impossible just a lot more difficult than those games. They have had a lot of time to make improvements, it just feels they are so minimal if they even exist at all. One can't expect it to be perfect or even as good as F1 but it should be progressing a lot faster than it is.
 
Nah, I don't think "the heavy lifting" is done at all. I think quite the opposite, the easy part is done: AI for 1 car with slight variations driving against 1 car with slight variations. (or 2 cars with no variation for the F2 plus the historic cars which had very limited use in the past, I think only some overtaking challenges and time trials? Correct me if I'm wrong)
Without being someone who is actually making AI systems on the reg it's hard to say what is easy and what is hard. But Polyphony has limited resources applied to the problem, so we can probably assume that they're going to do the easy parts first. With only one or two people, there simply isn't time to tackle hard stuff and leave easy stuff on the table. So we can make some assumptions about what is easy and what is hard based on what this small team has chosen to focus on.

They've done getting the part of getting the AI to work consistently on a large number of cars, so that would suggest that it's relatively easy.
They haven't done the part of getting the AI to drive quickly and with somewhat natural behaviours on even one car, which would suggest that it's relatively hard.

If it was easy to get a good AI but hard to adapt it to many cars, we'd see a few cars in Gran Turismo have excellent AI and the rest be bad. That would be functionally similar to what we have now, but you'd have at least a handful of good cars.
 
Without being someone who is actually making AI systems on the reg it's hard to say what is easy and what is hard. But Polyphony has limited resources applied to the problem, so we can probably assume that they're going to do the easy parts first. With only one or two people, there simply isn't time to tackle hard stuff and leave easy stuff on the table. So we can make some assumptions about what is easy and what is hard based on what this small team has chosen to focus on.

They've done getting the part of getting the AI to work consistently on a large number of cars, so that would suggest that it's relatively easy.
They haven't done the part of getting the AI to drive quickly and with somewhat natural behaviours on even one car, which would suggest that it's relatively hard.

If it was easy to get a good AI but hard to adapt it to many cars, we'd see a few cars in Gran Turismo have excellent AI and the rest be bad. That would be functionally similar to what we have now, but you'd have at least a handful of good cars.
I think it was exactly their goal to make the AI consistent first and foremost, with a comparable level of performance throughout all cars and tracks. They surely did not want lightning fast and aggressive AI in one car/track combination and slow and rail-like AI in another.

And if you only look at GT, that may be an assumption, but if you look throughout racing games overall what you mostly see is either a few cars with good AI or a lot of cars with not so good AI. So based on that the difficulty is neither making it fast and somewhat natural for very few cars or consistent through many cars but the combination of both.
 
As someone who understands Japanese I feel disappointed in myself for not immediately getting the play on word for "AI".

AI - artificial intelligence
AI (愛) - love in Japanese hence the heart logo
Wait till it sets in that AIBO is short for love robot. :sly: (not really)
 
Unveiling Gran Turismo Sophy : An AI Breakthrough

I’m thrilled to announce Sony AI's very first AI breakthrough - Gran Turismo Sophy which is featured on the cover of Nature magazine’s Feb 10th issue.


Nature
Gran Turismo Sophy is a revolutionary racing AI agent that has learned to master the PlayStation® game and driving simulator, Gran TurismoTM (GT) Sport. Through training, Gran Turismo Sophy acquired racing skills to challenge and outrace some of the world’s championship-level GT drivers.

Please check out a special video message from Kenichiro Yoshida, Chairman, President and CEO, Sony Group Corporation, about this exciting project.



This breakthrough has been one of the key challenges Sony AI has worked on since its inception in April 2020. Over the past two years, our team has been working closely with Polyphony Digital (PDI) - the creator and developer of the Gran Turismo franchise - and the cloud gaming team at Sony Interactive Entertainment (SIE) to develop Gran Turismo Sophy. Gran Turismo Sophy learns through novel deep reinforcement learning techniques to race at the highest level. It is trained using the cloud gaming infrastructure managed by SIE for massive scale training.

Gran Turismo is not just a game but also a realistic driving simulator. Over the years, PDI has created one of the most realistic racing experiences and brought the core experience of racing to PlayStation. This means that each car and each track are modeled to simulate the real-world experience of driving in terms of visuals, audio, and most importantly, dynamics. This realism of Gran Turismo is precisely what makes it a great challenge for AI research.

What Makes Gran Turismo Sophy a Unique AI Challenge?
The first important point is that racing requires drivers to control physical systems at their limits. Racing is essentially trying to drive cars at the edge of control or just beyond. Estimating braking points, finding the best line, searching for grip on the track to maximize speed and control are by itself very interesting Machine Learning problems, but racing means that you are not alone on the track. Other drivers impact the dynamics of the car through drafting effects that can help reach speeds higher than the actual top-speed of the car. But they can also drastically increase the braking distance due to reduced down force.

Then there are all the tactical aspects which are built on top of the highly dynamic control problem. Racers have to find lines to pass their opponents taking into account the opponent's reaction to the overtake attempts. That turns racing into a complex back and forth between opponents including defend and pretend behaviors as well as pushing the limits on the racing lines, braking and acceleration points.

Last but not least, racing drivers need to adhere to specific rules that govern how far drivers can skid off the track and who gets blamed when a collision occurs. Stewards look at critical race situations and determine potential time penalties for drivers. The rules are imprecisely specified and open to interpretation. That makes them difficult to encode and difficult to optimize for directly. There is also a concept of fair play in racing. Racing should be engaging and fun. Collisions should not be used as a tool to win races. At the same time, racers need to drive aggressively to win. Finding the right balance is a real challenge.

All these things together distinguish Gran Turismo Sophy from earlier AI agents in chess, shogi, Go and other AI breakthroughs such as AlphaStar and OpenAI Five for real-time, multiplayer strategy video games. Chess, shogi, Go are perfect information games that are solved by AI systems without the need to master real-world physics, but with a strong focus on strategies in discrete board games. Even AlphaStar and OpenAI Five, while technically dealing with in-game physics - are focussed on strategic aspects and are not attempting to master real-world physics. Now, Gran Turismo is all about trying to capture real-world physics. Tactics, strategy and race etiquette are important but they need to happen while the car is speeding along at its physical limits.

More than an AI Challenge - An AI that aims to deliver new gaming and entertainment experiences
For Sony AI, Gran Turismo Sophy is not just a scientific AI challenge. It is also an important step in unleashing human imagination and creativity with AI - which is our mission. From the very beginning, our focus has been to partner with creators, and in this case PDI, the makers of Gran Turismo, to develop a game AI agent that could not only race competitively against the top Gran Turismo drivers, but one that would be fun for the top drivers to race against. And with Gran Turismo Sophy, we are making great steps towards achieving this goal. In the development of Gran Turismo Sophy we regularly interacted with top drivers to test out the latest versions. One of the most memorable moments was in a time trial race when Emily Jones, FIA Gran Turismo Championships 2020 World Finalist, was talking about how she was inspired by Gran Turismo Sophy to try out new racing lines that she hadn't seen before.


The Incredible Team and the Future
Gran Turismo Sophy is an extremely exciting project that we could only do because of incredible collaboration and teamwork across the globe.



The Sony AI team from around the globe watching the racing match which took place in Japan

I am extremely proud of the teams that worked on Gran Turismo Sophy and created this achievement. We can announce Gran Turismo Sophy today because of a unique collaboration between Sony AI, PDI, and SIE. This type of collaboration is truly unique to Sony and demonstrates how we can utilize AI to inspire creators and gamers alike and Gran Turismo Sophy is only the beginning of what’s more to come. Sony AI and PDI will explore how to integrate GT Sophy into future releases of the Gran Turismo series.

See More Info & Look out for more from Sony AI. Please visit our Gran Turismo Sophy website, read our paper in Nature, and view the video on the Making of Gran Turismo Sophy.

Finally, we have additional flagship projects in the areas of Gastronomy, Imaging & Sensing and AI Ethics so please stay tuned for more news from Sony AI!

Link to the PDI Blog
Link to the SIE Blog
Link to the Press Release
 
Last edited:

But as was already mentioned, other AIs have learned to set very quick lap times on their own, in games or even in real cars on real tracks. To prove GT Sophy's skills, it went up against some of the world's best GT: Sport players. The first human-versus-AI test happened in July 2021, when Emily Jones, Valerio Gallo, and Igor Fraga were asked to beat GT Sophy's lap times in a time trial. Despite being able to see a ghost of the AI's lap, none were able to beat it.

Interestingly, GT Sophy had identified some racing lines that had eluded even the world's best human players. Jones was bested by the AI's lap time at Dragon Trail, 107.964 seconds versus 106.417 seconds. "There were certain corners where I was going out wide and then cutting back in, and the AI was going in all the way around, so I learned a lot about the lines," she said. "And also knowing what to prioritize. Going into turn 1 for example, I was braking later than the AI, but the AI would get a much better exit than me and beat me to the next corner. I didn’t notice that until I saw the AI and was like, ‘OK, I should do that instead.’"
 
I guarantee that if this is ever actually implemented, that it basically leads to the same events that we have now, except the rabbits now are cheetahs.

How much money was dumped into this where hiring a few programmers with experience in racing game AI could have sufficed? It seems like such a roundabout way to potentially fix a problem.
 
Back