Gran Turismo Sophy: Sony AI x Polyphony Digital

  • Thread starter Magog
  • 1,719 comments
  • 194,655 views
I can't stand the negativity of some members here. Poly and Sony try something and all these members want is minimising the project. We all know how bad the AI is in GT. Can we just enjoy this announcement even if it won't make to GT7 day one.
....agreed. Not just negative but myopic IMO.

They've said they have plans to implement the technology in GT7 and I see that as good news. Anything to help improve the AI in the game or future iterations of it.

To me, this appears to be a big step in the right direction for AI in games so I'm excited to see how the new Sony AI division fares as they continue development, begin implementing the logic in more games, etc.

Just the fact that they chose GT7 to showcase and implement the tech (to me) is exciting. ....no idea why some just want to crap all over it especially those really dissatisfied with the existing AI. You would think this should be encouraging news on that front vs. just little tweaks to what is already there (which I think most of us are expecting).

I'll take a wait and see approach personally. I do hope what they have learned with the Sophy project can be implemented sooner rather than later into GT7.
 
And it seems the A.I will have their own unique personality.
sophy.jpg
 
The biggest issues with this entire experiment is if Polyphony do intend to apply this to later games (and GT7) then how does it get applied? Watching the videos, you see that they have pace in droves, but not that much applicable race craft. Much more brute force then anything. Likewise, if it gets implemented, how does the racing change? Because if it's still the same old chase the rabbit stuff as GT has relied on for years now...now you have replaced rabbits with cheetahs on the highest end. And I don't know about anyone else, but I don't consider good and stimulating AI to be one that effectively cheats by getting a 40 second lead alongside massive amounts of pace.

If it does get implemented in GT7, then at what point does it get put in? Because if it turns into another Track Creator situation where it gets put in at way too late of a point to actually be useful, if it even comes at all, then it's guaranteed that it doesn't come to PS4 players. And since they decides that cross gen is the way to go instead of fully next gen...you are going to piss that section of the userbase off.

So much of this could have been solved if they realized the AI was bad and hired someone to fix it and make it much more human instead of basically using a sledgehammer to break into a chestnut. And really, all this does is raise more questions about how it actually gets implemented in real world scenarios, which opens up more problems with GT's moment to moment racing.

That is, if it even gets implemented in GT7.
 
I take it you are equally full of praise for the other developers and teams that have already done so?

I'm much more guarded, as applying machine learning of this type to a commercial product has had mixed results and that from companies that have either been doing it longer or with a bigger budget.


Edit to add. Found it, one of the questions MS asked years ago (original FM years ago) about developing AI for games is what is the actual end goal

One that always takes optimal actions

or

One that delivers best entertainment value

It's still a valid question and GT Sophy, right now, seems to be in the first camp.


I think they needed to just put in the leg work to achieve something, before they fine tune it. The research and development can now be used for B Spec, and filling Sport lobbies if they need to. Heck, even normal open lobbies could have the ability to add AI for a more competitive feel.

The question is though, do they need to create complete optimization of the AI first, or would it be more beneficial to add in the margin of error from the beginning?

So, as far as I understand, it’s fast but dumb. Nothing to be the least bit excited about in other words, as expected.
Why not? The AI in GT has never been fully optimized. This opens the door for endless possibilities to increase the value of the product. In addition to a learning environment for a large portion of employees.
 
I haven't been able to watch the entire video, but it seems there are only two scenarios here.
If they have not explicitly stated that GT7 AI will not be improved, GT7 AI will be significantly better than previous AI.
Logic: it would be a major PR own goal if it would not be improved. PD is aware of the criticism of the earlier AI. It would be completely unnecessary to publish Sophy now and it will predictably raise expectations about GT7 AI, so any communications manager worth his substantial pay would understand that if there is no major improvement in GT7 AI, this is the worst time to publish it.
Or, if they have explicitly said it will not be in GT7 at launch, I would expect the AI to be improved asap.
So, good news!
 
Last edited:
The biggest issues with this entire experiment is if Polyphony do intend to apply this to later games (and GT7) then how does it get applied? Watching the videos, you see that they have pace in droves, but not that much applicable race craft. Much more brute force then anything. Likewise, if it gets implemented, how does the racing change? Because if it's still the same old chase the rabbit stuff as GT has relied on for years now...now you have replaced rabbits with cheetahs on the highest end. And I don't know about anyone else, but I don't consider good and stimulating AI to be one that effectively cheats by getting a 40 second lead alongside massive amounts of pace.

If it does get implemented in GT7, then at what point does it get put in? Because if it turns into another Track Creator situation where it gets put in at way too late of a point to actually be useful, if it even comes at all, then it's guaranteed that it doesn't come to PS4 players. And since they decides that cross gen is the way to go instead of fully next gen...you are going to piss that section of the userbase off.

So much of this could have been solved if they realized the AI was bad and hired someone to fix it and make it much more human instead of basically using a sledgehammer to break into a chestnut. And really, all this does is raise more questions about how it actually gets implemented in real world scenarios, which opens up more problems with GT's moment to moment racing.

That is, if it even gets implemented in GT7.
You think about GT7 when it's more about gaming in general. As far as I know, track creator wasn't feature on the front page of Nature? Do you really believe Sony spent time and money with some of its best engineers for a feature of GT7? It happens to use Gran Turismo as a platform but the concept is far beyond that IMO.
 
think they needed to just put in the leg work to achieve something, before they fine tune it. The research and development can now be used for B Spec, and filling Sport lobbies if they need to. Heck, even normal open lobbies could have the ability to add AI for a more competitive feel.

The question is though, do they need to create complete optimization of the AI first, or would it be more beneficial to add in the margin of error from the beginning?
Two major questions exist for me around implementing it, and why I personally think it's a way off.

The first is scalability, what we've seen is useless to anyone other than the quickest GT drivers on the planet. Teaching it to be slower isn't just turning a dial down.

The second is making it engaging and enjoyable to drive against, with a degree of human falability.

Those two factors have been issues with both the current AIs on the racing market that have been developed via machine learning. With the second being the main one that breaks it.

Right now if it were AI in a FPS, it would never miss, and that's not my idea of fun.
 
Last edited:
from what I have seen..they will need down-tune ai 🤣🤣🤣 if a champion can't beat it...then how the hell us regular racers could do anything...haha begginer ai for 10.000 cr. man some car would get forever to have 🤣🤣🤣🤣 they can't be serious
 
You think about GT7 when it's more about gaming in general.

Considering that the experiment was performed entirely within GT Sport, and was shown and desired to have effects within the series itself, of course I am going to comment on it within a GT context. Why wouldn't I?

As far as I know, track creator wasn't feature on the front page of Nature?

No, but Track Creator was built and hyped around the same outline, that it would potentially revolutionize the GT series. It came basically at GT6's end of life, and no one cared, especially considering it was a shell of the form that Kaz hyped it up as.

Time tells whether Sophy follows the same fate.

Do you really believe Sony spent time and money with some of its best engineers for a feature of GT7?

Until it's shown in a context that is not GT, then it is a feature and hope that is contained within GT only, and should be commented on such merits.
 

"What really stands out is GT Sophy's performance against human drivers in a head-to-head competition. Far from using a lap-time advantage to outlast opponents, GT Sophy simply outraces them. Through the training process, GT Sophy learnt to take different lines through the corners in response to different conditions. In one case, two human drivers attempted to block the preferred path of two GT Sophy cars, yet the AI succeeded in finding two different trajectories that overcame this block and allowed the AI's cars to pass (Fig. 1)."

d41586-022-00304-2_20106388.png
 
Cool thing about this is we will be the guinea pigs at some point, that will be an interesting race day indeed! perhaps it will only manifest as ghost initially for Hamilton style challenges.
 

"What really stands out is GT Sophy's performance against human drivers in a head-to-head competition. Far from using a lap-time advantage to outlast opponents, GT Sophy simply outraces them. Through the training process, GT Sophy learnt to take different lines through the corners in response to different conditions. In one case, two human drivers attempted to block the preferred path of two GT Sophy cars, yet the AI succeeded in finding two different trajectories that overcame this block and allowed the AI's cars to pass (Fig. 1)."

d41586-022-00304-2_20106388.png

Impressive stuff, Sophy. 👏
 
Those two factors have been issues with both the current AIs on the racing market that have been developed via machine learning. With the second being the main one that breaks it.

Exactly. This isn't the silver bullet that people think it is, the thing that they can finally use to rub it into those who (rightfully) criticize the AI the series has used.

All this is, is an experiment. An experiment that, seeing from my eyes, is basically unapplicable until the two factors that you mentioned are addressed, and shown.

And who knows when that day comes? For now, the AI in GT7 (from what has been shown in pre-SOP footage) is still bad and still more or less the same that has been a problem for the series in general. It can absolutely be fixed, by human hands, yet this is such a roundabout and potentially even worse screwing up of the problem if it ever comes.
 
Exactly. This isn't the silver bullet that people think it is, the thing that they can finally use to rub it into those who (rightfully) criticize the AI the series has used.

All this is, is an experiment. An experiment that, seeing from my eyes, is basically unapplicable until the two factors that you mentioned are addressed, and shown.

And who knows when that day comes? For now, the AI in GT7 (from what has been shown in pre-SOP footage) is still bad and still more or less the same that has been a problem for the series in general. It can absolutely be fixed, by human hands, yet this is such a roundabout and potentially even worse screwing up of the problem if it ever comes.
Let's call it Garlic then, it will keep the bloodsuckers away that prey on anything bad about the AI (or lack of) that is a given with GT series.
 
I haven't been able to watch the entire video, but it seems there are only two scenarios here.
If they have not explicitly stated that GT7 AI will not be improved, GT7 AI will be significantly better than previous AI.
Logic: it would be a major PR own goal if it would not be improved. PD is aware of the criticism of the earlier AI. It would be completely unnecessary to publish Sophy now and it will predictably raise expectations about GT7 AI, so any communications manager worth his substantial pay would understand that if there is no major improvement in GT7 AI, this is the worst time to publish it.
Or, if they have explicitly said it will not be in GT7 at launch, I would expect the AI to be improved asap.
So, good news!
Exactly! This gives me hope that the AI will be significantly improved in GT7 from the start.
 
Thee sheer amount of ego that had to have been behind this.
So instead of actually trying to fix an issue that's has been plaguing the series for decades they would rather involve a multi bilion dollar corporation (surely that is the only way) only to come up with basically a proof if concept for previously unheard concept of fast AI without any clear road map for an actual implementation into the game. If this isn't the ultimate Kaz/PD move then I don't know what is.
 
the AI in GT7 (from what has been shown in pre-SOP footage) is still bad and still more or less the same that has been a problem for the series in general.
This is a bit of a reach, Kaz has stated the AI has been worked on and improved for GT7, lets see... It might even, dare I say it... be good! The future AI endeavour shows the ethos at work behind the scenes at Polyphony, there's nothing wrong with striving for perfection.
 
Thee sheer amount of ego that had to have been behind this.
So instead of actually trying to fix an issue that's has been plaguing the series for decades they would rather involve a multi bilion dollar corporation (surely that is the only way) only to come up with basically a proof if concept for previously unheard concept of fast AI without any clear road map for an actual implementation into the game. If this isn't the ultimate Kaz/PD move then I don't know what is.
--PD only have 1 guy working for AI = bad
--PD help developed an AI that would benefit Gran Turismo but also self driving cars in the future = also bad

180452045_1541145856227130_4613797948972569208_n.jpg
 
Last edited:
Two major questions exist for me around implementing it, and why I personally think it's a way off.

The first is scalability, what we've seen is useless to anyone other than the quickest GT drivers on the planet. Teaching it to be slower isn't just turning a dial down.

The second is making it engaging and enjoyable to drive against, with a degree of human falability.

Those two factors have been issues with both the current AIs on the racing market that have been developed via machine learning. With the second being the main one that breaks it.

Right now if it were AI in a FPS, it would never miss, and that's not my idea of fun.
Making it life like is probably the biggest hurdle. But as far as turning it back, it might be a little more involved but not implausible. I've seen other games discuss how AI or catch up is engineered, not specially just more optimized driving, but rather lowering the weight of a vehicle, or increasing grip. Perhaps they would even increase power. But all of this could be reversed to delay the reactions or inputs of the AI and give it more of a lifelike response to scenarios, and put it on par with the human competition. Then from whatever variables work, a slider could be engineered to provide X level of driving in a race, or per AI in a custom race.
 
Back