Gran Turismo Sophy: Sony AI x Polyphony Digital

  • Thread starter Magog
  • 1,719 comments
  • 194,641 views
There's one other corollary here I forgot about:
It's impressive if it pans out as they are theorizing it (though perhaps not as they are explaining it). The main question to take from here is how it can scale down. Someone mentioned a TAS as a point of comparison and I can see that, but even moreso I'm reminded of AI in fighting games up until... ~1996 or so? You'd get a couple of rounds where you could actually compete in MKII or Street Fighter The Movie The Game or Art of Fighting 2; but then the game decided the computer would react to your inputs... perfectly, instantly, to the exact frame that you made them; at which point the actual winning strategy was always to find something to cheese them on. One needs look no further than Forza to see AI where it covers tracks so perfectly that it's basically just hopeless to even bother trying to play them on the highest difficulty.
I've gone to the bat for GT5's AI in the past, and I still think it was (originally) pretty good for a base to build off of. But the thing that let it down the most is that despite having a pool of ~1000 cars or whatever to choose from to create race fields in single player, PD's selections for individual events were frequently abjectly terrible. Thinking of things like the Nurburgring 4 hour race where you'd have 3-4 cars that were decently matched and would fight you (you'd have to sandbag yourself, granted) and generally turn consistent laps. Then there was one car that was far and away faster than everything else and you'd hope it wouldn't be in the race (IIRC it was the Amuse S2000 GT1?); and then the entire rest of the AI pool would be completely hopeless and not even matched well with each other.



So, at the end of the day, if you were really lucky, you'd have... 1/3rd of the field actually be fun to race against. And that's been something that's been a major issue since PD stopped having the rubber banding AI starting with GT4; and something that even if the AI was as terrible as it was in GT4 you could have worked with to get good races out of (though not with the chase the rabbit crap or the "slow down to let you catch up" GT6 stuff) if more care was put into that part but if remains on the backburner no amount of good AI will fix.
 
Last edited:
I hear you, could it be not a complete loss of load but an incremental reduction of it, depending on inertia? Zero contact points and we would be talking BTTF2 DeLorean lol Maybe they're trying to allude to the perfection of riding the very ragged edge of grip under unique circumstances. Breaking while turning left suggests front left have the highest load and rear right the lowest? and if this cause an oscillation in the car you can probably attenuate it if you're really adept with the throttle on exit. That kind of driving would be useless except for qualifying, proper tyre murder for a novice too! Probably the reason we struggle as videogame drivers is we don't get the inertial feedback an F1 driver gets its all by eye / tyre squeel sfx for us (hopefully haptics might change this).
It's why I think it's a mistranslation based on Kaz talking about balance of load and load distribution, rather than 'no load'.

Trail braking doesn't also necessarily increase tyre wear, if its done correctly, it's the same for left foot braking. Done correctly it can actually be better on the tyres overall as it allows you to best balance the load across the four tyres.

In the case of tail breaking, think of it more as overlapping the braking and steering phase to ensure you are using the available limit of each tyre as far into the corner as possible. This allows you to use it while racing, not just limiting it to qualifying.
 
Last edited:
Very interesting seeing how they did this. I'm guessing it's akin (but much more advanced) to something like Ant Colony optimization.

I also didn't quite understand the difference between each agent. Why was one faster considering they use the same tech?

They need to better understand slipstream; they're running too close all the time negating the advantage. Wonder how they'll integrate sophie with the existing AI system in GT7 or will it replace it entirely.
 
How am I supposed to play GT7 and its probably slow AI when I just saw that Sophy AI? This could change the whole game with a good AI. Sure, it's too fast and needs obvious improvements but still...
 
I wonder if this could be used to determine a more accurate PP system for cars. If you have something you know can drive a perfect lap repeatedly and iterate a set-up each time, could be good for stuff like BOP calculations too.
 
One easy way to introduce fallibility into the AI control is to introduce errors or inaccuracies or perturbations in the model's output.

A deceptively simple and likely surprisingly rich way to do this might be to not generate control output at a fixed 10 Hz. If this is jittered between a range of say multiples of game physics ticks, loosely averaging around 10 Hz, there might be some useful and believable imperfections. There might also be some rare but hilarious jankiness too, everyone wins.

The pace will still be white hot on average though, so adding noise to the output as well will make the model give itself something to get its teeth into every tick. It's also possible to prevent it from remembering what it's learned, or prevent it from learning outright.


It's all moot since this model as is won't make the game, so any scalability tricks can be designed directly into the model instead.
 
No tyre degradation from running the curbs all the time ? No slip ?

Kinda weird !

Or maybe this is just in arcade mode
Tyre wear was turned off for these races so no, there wasn't any degradation for Sophy to worry about.

I'd guess that's a later stage of development.
 
I've just sat through the video again and I still can't get my head around what SOPHY can do with the car. I can watch a top human driver setting an exceptional lap and believe what I'm seeing but with SOPHY, it feels like I'm seeing GT's physics engine foibles laid bare and exploited.

It's bloody clever stuff but the way they put a lap in looks like a parody of a human's efforts.
 
I wonder if this could be used to determine a more accurate PP system for cars. If you have something you know can drive a perfect lap repeatedly and iterate a set-up each time, could be good for stuff like BOP calculations too.
Unlikely as it's a learning AI, so even without changing anything on a car the PP would get higher as the AI learns. You just need a simple AI to provide a consistent PP value.
 
Tyre wear was turned off for these races so no, there wasn't any degradation for Sophy to worry about.

I'd guess that's a later stage of development.
oh ok, than the ai will be driving more normal when it's final, because now it would be totally unrealistic, maybe first 2 laps in the lead but after that,... I'm realy looking forward to drive myself :)
 
I've just sat through the video again and I still can't get my head around what SOPHY can do with the car. I can watch a top human driver setting an exceptional lap and believe what I'm seeing but with SOPHY, it feels like I'm seeing GT's physics engine foibles laid bare and exploited.

It's bloody clever stuff but the way they put a lap in looks like a parody of a human's efforts.
This is exactly what takes away from the excitement for me. It’s bloody fast, but not very realistic, and I thought a more realistic AI was what we were all wishing for, no?
 
This is exactly what takes away from the excitement for me. It’s bloody fast, but not very realistic, and I thought a more realistic AI was what we were all wishing for, no?
More realistic and more competetive, as of that demonstration the Sophy AI is too agressive, but if you could scale it back so people can set a difficulty of thier choosing and sort out a few niggles with the aggression and race craft it could become something special. But making it user friendly in those ways isn't easy.
 
Last edited:
Fraga, Miyazono, Gallo, etc.. are kind of like robots too or beyond human, alien and still somehow find ways of improving.

Both are highly unrealistic yet we see it with our own eyes what is possible. The drivers reactions were like our reaction to them, "I didn't know you could do that".

The science community seems pretty impressed with this too. So I think unrealistic may not be term to apply here. It's our perception that's limited.
 
The science community seems pretty impressed with this too. So I think unrealistic may not be term to apply here. It's our perception that's limited.
Which is strange to me because another research study was done last May, and didn't seem to get much recognition. Doing essentially the same exact thing in GT Sport, which I assume this research was baselined from.
 
More realistic and more competetive, as of that demonstration the Sophy AI is too agressive, but if you could scale it back so people can set a difficulty of thier choosing and sort out a few niggles with the aggression and race craft it could become something special. But making it user friendly in those ways isn't easy.
This. I don’t know how you’d teach an AI to take it down a notch or two, without making it behave weirdly. Maybe by implementing two separate penalty systems? The one for the AI being more strict than the one for us players, or something, I don’t know.
 
Last edited:
This. I don’t know how you’d teach an AI to take it down a notch or two, without making it behave weirdly. Maybe by implementing two separate penalty systems. The one for the AI being more strict than the one for us players, or something, I don’t know.
You can do a bad job and just make it brake earlier than it needs to, accelerate later and not use full throttle when it can, or you can try to cap the AI's learning at certain increments. Both have inherant problems associated with them.
 
Great stuff, looking forward to this. If I had to make a guess, I would say this will arrive as a special mode with fixed preset conditions, similar to the Hi-Fi from GT1. I see this as a nice tool for pro drivers to train themselves on one track and with one car but a bit difficult to make it work in standard career races, with so much track and car combinations available.
 
Which is strange to me because another research study was done last May, and didn't seem to get much recognition. Doing essentially the same exact thing in GT Sport, which I assume this research was baselined from.
Honestly I don't know enough about science (and the business behind it) to know what would be considered strange.

I just noticed a lot articles yesterday from places that don't normally talk about GT. And their takes on it gave me the impression this was different.

Viewed cynically, I wondered why these publications were giving a PS game free advertising. And assuming they wouldn't. Came to the conclusion this was scientific.
 
Which is strange to me because another research study was done last May, and didn't seem to get much recognition. Doing essentially the same exact thing in GT Sport, which I assume this research was baselined from.
The Sophy demonstration was also in GT Sport, the difference is this is a culmination of that, we're now seeing the AI leaning to race not only the clock but also different permitations of itself and on top of that, human drivers.
 
Last edited:
You can do a bad job and just make it brake earlier than it needs to, accelerate later and not use full throttle when it can, or you can try to cap the AI's learning at certain increments. Both have inherant problems associated with them.
Yeah, and then there’s conserving tyres and knowing when to go in for fuel. All this, applied to 400+ cars and 90+ track layouts. Very time consuming I’d say.

Just imagine how long it’d take for it to learn one good pit strategy in one car on a 20 lap La Sarthe time trial, let alone race alongside 15 other cars.
 
If Hamilton was an AI, he wouldn't dress like he does. :D
Actually this is reptilian board of fashion which decide what Hamilton wears. They met every 6 day of 6 month at 6.pm and decide whole wardrobe of Hamilton for full year...


Ok that was the last one... i Don't want to get ban for offtopic... reptilians are watching :nervous:
 
Fraga, Miyazono, Gallo, etc.. are kind of like robots too or beyond human, alien and still somehow find ways of improving.
The issue is that these players are still ultimately human, and are still bound by pre-set rules and regulations and by the actions that other human players take. Sophy, by comparison, is blazing fast yes, but incredibly aggressive even compared to the most daring drivers, and in some regard, will blatantly cut track limits and take frankly bonkers lines in order to better themselves.

Another thing that crossed my mind with regards to Sophy is...how will it actually apply when done in the real world, specifically online? I have a feeling one of the goals is to use Sophy as a potential way to fill online lobbies that aren't filled (IE, Halo Infinite or Battlefield 2042) but how would Sophy react to penalties from blatant contact or especially cutting track limits?
 
It's trivial to disable the learning though so it's actually a great idea.
if You disable the learning entirely, you might get very odd results when it comes up against a car/track/tune/settings combo it hasn’t practiced before.

In addition, what if it’s learnt to be more proficient at RWD cars over FWD cars at the point the learning is disabled? To be honest, I see a plethora of possible issues with using it to get a consistent PP rating.

It’s a lot simpler to just use a simple AI to do a simple task consistently rather than try to dumb down a complex one.
 
Last edited:
Now there are the Ai's that many asked for.
From what I saw, it reminded me of the game Assetto Corsa1 or PCars Gotty and very recently Dirt5.

When I saw in GT6 in Sierra a Fiat 500 (V6 Abarth engine) running away like a fool... (great track that would be so good to appear in GT7 and therefore in GTSport).

In the GT3 Porsche race, both Japanese, who are overtaken by two AI, got out of their line, and they took advantage. Amazing! Apparently they create adrenaline, especially to PRO players, and by the way I would like to see Sophy Orange with the best Tomahwak Nurb TT,,, or in Rally!
 
This. I don’t know how you’d teach an AI to take it down a notch or two, without making it behave weirdly. Maybe by implementing two separate penalty systems? The one for the AI being more strict than the one for us players, or something, I don’t know.
Put limits on it's reaction time (old drivers), put less weight in to being aggressive, punish it less for mistakes, limit the learning runs etc etc there is lots of adjustments to create 100s of different AIs with strengths and weaknesses. Mix and match the variables with different car types that creates AIs that are faster with certain car types. Do the same for tracks. The possibilities are endless.
 
Put limits on it's reaction time (old drivers), put less weight in to being aggressive, punish it less for mistakes, limit the learning runs etc etc there is lots of adjustments to create 100s of different AIs with strengths and weaknesses. Mix and match the variables with different car types that creates AIs that are faster with certain car types. Do the same for tracks. The possibilities are endless.
The possibilities certainly are endless, it's just easier said than done without getting pretty wonky results. But hopefully it's a task they are up to. If they pull this off it has the potential to be a long term game changer. But machine learning AI isn't new, and it's notoriously difficult to make it feel natural and scalable and not make it do strange things at the same time. Time, technology and techniques move on of course so it's an exciting project.
 
Last edited:
Back