Gran Turismo Sophy: Sony AI x Polyphony Digital

  • Thread starter Magog
  • 1,719 comments
  • 194,663 views
I will quote myself:
"I don't recall Polyphony promising and promoting a feature this much and then not following through. So even if you have a very valid point this time I am optimistic but I don't expect anything this year, the answers are simple to vague to make me think otherwise."

25 people from SONY AI worked on Sophy, several press events and lots of PR all around the world on social media platforms. It will be a huge defeat if it never makes it into GT7
No, because that team has already had the success that they wanted. They demonstrated a method of training an AI to drive at a level comparable to top human players. It isn't primarily intended as a means to make AI for GT7, or those researchers would have been working for Polyphony directly. It's a generic AI project that happens to use Gran Turismo.

If it makes it into Gran Turismo or even leads to improvements in GT AI then great, but that's not their goal. I think the marketing putting Gran Turismo so much in the forefront has blinded some people to the fact that the involvement of Gran Turismo is largely incidental. The research would have worked as well with basically any similar game, it just happened to be Gran Turismo because Sony backed research project (presumably).
 
It will be a huge defeat if it never makes it into GT7
They had many years to develop modified AI for GT7 without this deep learning.
And for potentional customer it's better to reveal GT Sophy in near future after GT7 release when PD could clearly say exact date when Sophy will be implemented, in which state, for which modes etc.
Current situation is that many of us will finish GT7 driving school and many career events in first 3-6 weeks after release. Without Sophy it will be traditional GT experience, I'm sure.
 
I will quote myself:
"I don't recall Polyphony promising and promoting a feature this much and then not following through. So even if you have a very valid point this time I am optimistic but I don't expect anything this year, the answers are simple to vague to make me think otherwise."

25 people from SONY AI worked on Sophy, several press events and lots of PR all around the world on social media platforms. It will be a huge defeat if it never makes it into GT7
Not just 25 people but some of the best AI researchers in the industry and publicated in Nature. Now, we'll see what Poly will do with this AI agent and I'm not optimistic at all:lol:
 
I think it bears repeating that Sophy as we've seen it cannot be used for opponent AI.

The opponent AI must have a different basis in order to ship with the game, but even patching won't make Sophy compatible as is.

Because it will be on a different basis, the adaptive performance scaling can be made without worrying about specific training regimes for Sophy.

As we all know, the main issue with GT's AI is the implementation, not the model. I can recall times where the AI performed close to expectations, before the lifting off behaviour and the like and especially the rare times we had control over AI opponent selection to avoid the rabbit issue.

So once again it just comes down to options.
 
I think it bears repeating that Sophy as we've seen it cannot be used for opponent AI.

The opponent AI must have a different basis in order to ship with the game, but even patching won't make Sophy compatible as is.

Because it will be on a different basis, the adaptive performance scaling can be made without worrying about specific training regimes for Sophy.

As we all know, the main issue with GT's AI is the implementation, not the model. I can recall times where the AI performed close to expectations, before the lifting off behaviour and the like and especially the rare times we had control over AI opponent selection to avoid the rabbit issue.

So once again it just comes down to options.
It could easily be adapted by making them use less of the track, or apply power, steering and braking more slowly.
 
I think it bears repeating that Sophy as we've seen it cannot be used for opponent AI.

The opponent AI must have a different basis in order to ship with the game, but even patching won't make Sophy compatible as is.

Because it will be on a different basis, the adaptive performance scaling can be made without worrying about specific training regimes for Sophy.

As we all know, the main issue with GT's AI is the implementation, not the model. I can recall times where the AI performed close to expectations, before the lifting off behaviour and the like and especially the rare times we had control over AI opponent selection to avoid the rabbit issue.

So once again it just comes down to options.
I read this post 2 times, and i am no sure i get what you are trying to say?
 
Sophy, as it has been developed and demonstrated, cannot be used for opponent AI in GT7 retail in people's homes using consumer hardware.

The adaptability people ask for doesn't need to be trained.

It just isn't the concern here, because the major issues with GT's AI have little to do with its pace and everything to do with PD wanting to provide accessibility, without providing the option to sidestep it.

It is interesting what they eventually do with it, I still think it'll be a driving aid.
 
Sophy, as it has been developed and demonstrated, cannot be used for opponent AI in GT7 retail in people's homes using consumer hardware.

The adaptability people ask for doesn't need to be trained.

It just isn't the concern here, because the major issues with GT's AI have little to do with its pace and everything to do with PD wanting to provide accessibility, without providing the option to sidestep it.

It is interesting what they eventually do with it, I still think it'll be a driving aid.
Jimini Cricket. That’s what I’ve been thinking PD would do with it. Similar to programming a Safety Car guiding players around a course as in GT4.

To add, since that study was showing how much that AI learned, can PD build a program that understands how each driver learns during the licence tests? Therefore, the program will adapt to each player. It wont be a true AI like Sophy and it might be more complicated than traditional Beginner, Intermediate and Professional choices. Not necessarily a rubber band program, but more like adaptive cruise control.
 
Last edited:
Jimini Cricket. That’s what I’ve been thinking PD would do with it. Similar to programming a Safety Car guiding players around a course as in GT4.

To add, since that study was showing how much that AI learned, can PD build a program that understands how each driver learns during the licence tests? Therefore, the program will adapt to each player. It wont be a true AI like Sophy and it might be more complicated than traditional Beginner, Intermediate and Professional choices. Not necessarily a rubber band program, but more like adaptive cruise control.
So, Drivatar?
 
So, Drivatar?
I don’t remember it, like what‘s been mentioned in Forza(So long ago). If that’s what I’ve described, then, yes. However, I’m not sure how PD programmed the Skyline safety/pace car to lead players around a lap of the Nordschliefe.

Something else, mentioned, when tuning cars. There will be virtual lap calculations to adjust PP. Can that translate to AI doing those virtual laps? PD have AI lapping a circuit, to determine penalties. Can that be implemented for AI?
 
Sophy, as it has been developed and demonstrated, cannot be used for opponent AI in GT7 retail in people's homes using consumer hardware.

The adaptability people ask for doesn't need to be trained.

It just isn't the concern here, because the major issues with GT's AI have little to do with its pace and everything to do with PD wanting to provide accessibility, without providing the option to sidestep it.

It is interesting what they eventually do with it, I still think it'll be a driving aid.

But it will not be an issue to run Sophy on a PS5 the real question is if the PS4 can handle it. The adaptability would make sense to be done with AI training sine that would use the strength from the Sophy AI to build possible 100s of AI drivers with all kinds of "personalities".
 
But it will not be an issue to run Sophy on a PS5 the real question is if the PS4 can handle it. The adaptability would make sense to be done with AI training sine that would use the strength from the Sophy AI to build possible 100s of AI drivers with all kinds of "personalities".
It won't be a problem. One simple solution: Cloud computing. You arleady need to be connected to internet for save game to work, why not make computing of Sophy on their servers? You will just see driving output on console which won't be any more demanding than standard AI car or other players.
 
But it will not be an issue to run Sophy on a PS5 the real question is if the PS4 can handle it. The adaptability would make sense to be done with AI training sine that would use the strength from the Sophy AI to build possible 100s of AI drivers with all kinds of "personalities".
It's nothing to do with which console to use and everything to do with how to generate a feed of "data" for each AI opponent to react to, just as it (Sophy) has been trained to do.

Therefore, Sophy as-is is not suited for use as opponents in a deliverable game. It's much more likely to be a driving aid in the vein of a B-Spec Bob taking over or perhaps giving you a nudge in the right direction as you drive. Because the game is already generating a feed of data for you to react to...

The cloud is not the answer either, as usual.


That is unless they train it to use a more compact data form instead, like the pre-existing AI already does. There's no reason the machine learning techniques couldn't be applied with no need to change the opponent AI model they have used to date. It will just be trained to "deploy" that model in different ways according to the learning outcomes :)

As for AI adapting to the player, PD have been collecting player usage data since the PS3, I think they will continue to do so.
 
It won't be a problem. One simple solution: Cloud computing. You arleady need to be connected to internet for save game to work, why not make computing of Sophy on their servers? You will just see driving output on console which won't be any more demanding than standard AI car or other players.
Kinda, but cloud computing isn't free. Sony/Polyphony may not want to pay for a bunch of extra hardware and bandwidth just to supply AI to their game. It's a solution, but it's not necessarily as simple as you make it sound for companies whose primary purpose is to make money.
 
Kinda, but cloud computing isn't free. Sony/Polyphony may not want to pay for a bunch of extra hardware and bandwidth just to supply AI to their game. It's a solution, but it's not necessarily as simple as you make it sound for companies whose primary purpose is to make money.

And once again, represents a roundabout, honestly sledgehammer to open a chestnut way to approaching a problem that could be easily fixed if Polyphony...applied the same ideals in making the sounds better to better AI.
 
It's not a feasible solution and it's unlikely to be what PD have planned for Sophy. Technically it's possible, though...

I think this all comes from a misunderstanding of how "the cloud" and "machine learning" have been used for AI in the past.

You have to think about how to do that in a computationally- and bandwidth-conservative way, so that it works for all players with some kind of always on internet (and especially those with capped data tariffs).

Then it's clear that the in game AI would use a standard model trained before release, and any subsequent "learning" would be bulk across all players based on lightweight feedback from the game. Any immediate adaptiveness would already be trained in the model during development, all backed up by drawing on a larger knowledge base stored in the cloud.
 
This is the discussion thread for an article on GTPlanet:

How “Gran Turismo Sophy” Actually Works: More Details on Polyphony Digital and Sony AI’s New Technology

Polyphony Digital recently unveiled Gran Turismo Sophy, a new artificial intelligence driving system which will appear in Gran Turismo 7. The technology was developed in collaboration with a 25-person team at Sony AI, utilizing the latest advancements in machine learning. The team’s research was published in Nature and GT Sophy was tested against (and defeated!) some of the best Gran Turismo drivers in the world at a live event in Tokyo last year...
 
Last edited:
So it is only based on internal game data as presented, interesting.

Which means it's really a case of nailing down the training. And that granted is no small feat.


But still as expected Sophy will only be used for one AI "agent" when running on the console, which unfortunately means it is limited in scope and of course actual excitement, but nonetheless still interesting.

Based on what Kaz said, we can speculate carefully about what the usage will be:

  • A training tool for players; perhaps a ghost, or assisted input, training player etiquette etc.
  • A single BSpec Bob; online against other Bobs, or to race against yourself, or to test strategies faster.
  • A spectator to learn more itself, to help PD develop Sophy further; watching the best racers to learn etiquette and tactics, could it also be used to learn when to issue penalties?


To be clear this tells me the ordinary opponent AI will be an evolution of what's come before.
 
Last edited:
Nice, you were able to get the scalability question in there. That was my major concern and in hindsight, it's obvious they would include a slower version lol. No matter how much training, many people (me included) wouldn't be able to lap at the level Sophy got to.

In addition to the implementations mentioned, hopefully Sophy will be available simply as AI you can race at any time.
 
So it is only based on internal game data as presented, interesting.

Which means it's really a case of nailing down the training. And that granted is no small feat.


But still as expected Sophy will only be used for one AI "agent" when running on the console, which unfortunately means it is limited in scope and of course actual excitement, but nonetheless still interesting.

Based on what Kaz said, we can speculate carefully about what the usage will be:

  • A training tool for players; perhaps a ghost, or assisted input, training player etiquette etc.
  • A single BSpec Bob; online against other Bobs, or to race against yourself, or to test strategies faster.
  • A spectator to learn more itself, to help PD develop Sophy further; watching the best racers to learn etiquette and tactics, could it also be used to learn when to issue penalties?


To be clear this tells me the ordinary opponent AI will be an evolution of what's come before.
Where does it say that in the article though? We've already seen multiple Sophies with different "personalities" used in a race against the top players. Plus the article mentions they will race with you.

“Sophy will probably appear in front of the player in three forms,” Yamauchi-san explained. “As a teacher that will teach driving to players, a student that will learn sportsmanship from players, and as a friend to race with. I wouldn’t rule out the possibility of a B-Spec mode, where the player is the race director and Sophy is the driver.”
 
Last edited:
GT is famous for his legendary poor AI.
Now, you tell me it could become the best.

It's good to to see after the success of GTSport online mods, with this GT7 era, GT works a lot into the offline part.
 
Sophy would definitely make B-Spec more bearable than using Bob again.
Idk, honestly sounds like that is very, very far away. B-Spec with giving it instructions like different pace, having tire degredation and fuel usage, even having it drive over 400 cars on all tracks under different weather conditions when they are currently at 3 cars on 3 tracks in 1 condition. And as I understand it from their comments on the drawbacks of their machine learning technology, they can't use any of those 3 cars or tracks to apply it to some to degree to other cars, tracks or conditions. Seriously sounds like years away.
 
Back