Gran Turismo Sophy: Sony AI x Polyphony Digital

[Tornado]

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.

 

[Famine]

I hope something has been lost in translation.

I think so - the translator said something about normally only loading the two front wheels in cornering, which made me go

I think they're trying to convey the idea that rather than loading two wheels and unloading two, top tier drivers (and Sophy) can manage weight transfer to limit the unloading on... I guess the inside front maybe? ... but in extremely simple terms.

I think...

 

[MroczMR]

In the rest of the quote, Yamauchi says that this behaviour is similar to top-tier real-world drivers like Verstappen and Hamilton, and somehow she's learned these same techniques.

So it is reproducible for a human.

Or... <puts on tinfoil hat> Verstapeen and Hamilton are in fact super advanced humanoid robots with advanced AI placed there by reptilians, and GT Sophy just exposed both F1 teams! <puts down tinfoil hat>

 

[Magog]

I think so - the translator said something about normally only loading the two front wheels in cornering, which made me go

I think they're trying to convey the idea that rather than loading two wheels and unloading two, top tier drivers (and Sophy) can manage weight transfer to limit the unloading on... I guess the inside front maybe? ... but in extremely simple terms.

I think...

I was really curious about what they meant by that too. Braking will transfer load to the front and turning will transfer weight to the inside so I guess they mean they are pushing all but the outside rear wheel to the limit?

 

[ShiftingGears]

I gon't get this? I mean, why I couldn't do the same as a human? As long as Sophy is only allowed to use the regular possibilies of a steering wheel, ****ft paddles and pedals...the outcome must be the same or still reproducable for a human. Its not a "Deep Blue" who just runs 100 millions of calculations per seconde to win in chess. Racing is a bit different

Howevery, I would be curious how Sophy reacts to greyish unfair, dirty driving which is not worth to get a penalty.

Its total nonsense and I hope that it didn't come from his usual translator.

I was really curious about what they meant by that too. Braking will transfer load to the front and turning will transfer weight to the inside so I guess they mean they are pushing all but the outside rear wheel to the limit?

You have it the wrong way around, the weight transfer goes to the outside. The inside gets unloaded.

 

[br1x]

I think so - the translator said something about normally only loading the two front wheels in cornering, which made me go

I think they're trying to convey the idea that rather than loading two wheels and unloading two, top tier drivers (and Sophy) can manage weight transfer to limit the unloading on... I guess the inside front maybe? ... but in extremely simple terms.

I think...

I believe he was simply referring to trailbraking, as "normal" drivers may brake (2 front wheels loaded) and start turning after (2 outside wheels loaded) but she learned trailbreaking by herself at some point.

Not that trailbraking is anything only Hamilton and Verstappen, but ever single professional racing driver and very likely every GT Sport player above A DR uses.

 

[GTalaune]

To me it looks like the AI rides the brake too much when cornering, it's weird. I think ABS default is a bit broken tbh, and allows for exploit, especially by AI. ABS on weak feels better, more "real", even if slower.

 

[Scaff]

I think so - the translator said something about normally only loading the two front wheels in cornering, which made me go

I think they're trying to convey the idea that rather than loading two wheels and unloading two, top tier drivers (and Sophy) can manage weight transfer to limit the unloading on... I guess the inside front maybe? ... but in extremely simple terms.

I think...

Yep - I did the exact same when I read it.

I was really curious about what they meant by that too. Braking will transfer load to the front and turning will transfer weight to the inside so I guess they mean they are pushing all but the outside rear wheel to the limit?

Outside, not the inside, you load the outside when turning.

 

[Kazuya720]

wheel, ****ft paddles and pedals..

Why does shi-ft or should I better say gear-changing-paddles got out asterisked ?? Is "shift" a curse word??

 

[Famine]

Is "shift" a curse word??

No, but that isn't what you typed. Quick clue, "shift" only has one "t", not two either side of the "f".

 

[Kazuya720]

Or... <puts on tinfoil hat> Verstapeen and Hamilton are in fact super advanced humanoid robots with advanced AI placed there by reptilians, and GT Sophy just exposed both F1 teams! <puts down tinfoil hat>

You totally convinced me! Now that's a believable explanation! but... what means this "puts on tinfoil hat" ?? Does this improve to think more logical? I have to try this, rightnow! by myself

@Famine: my bad!

 

[MroczMR]

You totally convinced me! Now that's a believable explanation! but... what means this "puts on tinfoil hat" ?? Does this improve to think more logical? I have to try this, rightnow! by myself

@Famine: my bad!

I put it on to block 5G waves which can cause covid-19...

 

[AlexWilmot]

I was really impressed by the AI presentation, obviously needs some tweaks, the Porsche race was super aggressive.

Be interesting to see what settings are available for it once its brought in a future update.

 

[Strittan]

Ok, I have to admit it’s quite impressive after all. I don’t know how to implement it into the game though, and what use it will have, apart from humiliating me I guess.

As far as I’m concerned, it needs to learn pit stop strategy, how to conserve tyres and how to avoid accidents, and it needs to learn these things in 400+ cars on 90+ track layouts. I think it will take a while.

Edit: Also, PD better improved the off-tack physics for GT7, because going off-road as consistently as Sophy does and still be that quick is not right.

 

[Griffith500]

It's simply learned how exploring track limits can generate faster lines.

It then randomly stumbled upon the necessary control inputs that makes the line viable and stable, and reinforced this by exploring around that with minor permutations. It also learned how to recover from input mistakes in the process, as a recovery will be more highly rewarded than a spin. Anything rewarded is solidified in the model and will be expressed again under similar circumstances in practice.

The sheer number of trial and error attempts are mind boggling and this is one of those cases where humans' inability to see behind their own scope, in terms of scale or time or large numbers in general, can make it seem like witchcraft or, indeed, iffy.

 

[Muller_cj]

Yep, as other teams have done in the past, in most cases the implementation into a commercial product is the tricky bit, as an AI that takes the perfect action every time isn't actually that entertaining. Adding in that 'human fallibility then often breaks it.

Indeed, but that's not a valid reason for the AI to not have it.

The AI made mistakes in the Races, last night, mistakes any player would make, So I'm struggling to see your argument. What we have in most games now is an AI on Rails

Will it fully make it into GT7 probably not, but I suspect some parts of it are in the game already like the Virtual Lap for PP values

 

[MroczMR]

I wonder another thing about this AI. If they could hypothetically put it to other race sim, Assetto Corsa Competizione for example, and train it will it start to learn other racing game? Asking just for curiosity.

 

[Magog]

I was really impressed by the AI presentation, obviously needs some tweaks, the Porsche race was super aggressive.

Be interesting to see what settings are available for it once its brought in a future update.

It's quite clearly a work in progress but that they were able to tweak it from losing to winning in a real race against human opponents in just 4 months shows how much progress has been made and how malleable the simulation is.

 

[Dave A]

The AI made mistakes in the Races, last night, mistakes any player would make, So I'm struggling to see your argument. What we have in most games now is an AI on Rails

Will it fully make it into GT7 probably not, but I suspect some parts of it are in the game already like the Virtual Lap for PP values

The AI in GT Sport makes mistakes.

The Sophy AI we saw learned and improved with every race, it if did something that led to a mistake, it didn't do it agan, if it did something it could do better, it'd do it better the next time. It only makes mistakes while it's learning. When @Scaff said "human falibility" that's completely different. We don't make a mistake once, learn from it and never repeat that mistake again. If we did, we'd all become perfect at what we do, it'd only be a matter of time. We certainly become better, but we never become completely consistent or perfect.

The Sophy AI does look great, but it's certainly not ready to be packaged into a commerical product yet, no one would ever beat it once it's learned enough. So they need to find a way to scale it, add in some inconsistency and from what I saw, improve it's race craft. I'm also not sure if the AI has ay compensation for reaction times, it takes a human time from something happening to react to it, the AI was reacting to things almost immediately from what I could see.

You dont want to go from one extreme of AI that doesn't pose a challenge to decent players, to AI the best players in the world can't beat. They may already have a way of scaling it, but from what Kaz said it's most certainy not ready yet. It is very promising though.

 

[Scaff]

The AI made mistakes in the Races, last night, mistakes any player would make, So I'm struggling to see your argument.

I have to disagree, what we saw were mistakes based on an AI attempting to opptomise the rules its been given.

Human falability is the last characteristic I would give them.

It doesn't overdrive the car, screw up braking points, get pressed by opponents, etc.

What we have in most games now is an AI on Rails

I can think of a good number that really, really don't have that.

Will it fully make it into GT7 probably not, but I suspect some parts of it are in the game already like the Virtual Lap for PP values

PD and Sony AI have been quite clear that we don't.

 

[Dave A]

Will it fully make it into GT7 probably not, but I suspect some parts of it are in the game already like the Virtual Lap for PP values

Just to address this seperately, I don't think it is in GT7 at launch, infact they were clear that they hope it will make it to GT7 in some form at some point.

In addtion, I don't think it would ever be used to determine a cars PP value (and it shouldn't). Because it's a learning AI, you could calcualte a cars PP with it, then without changing anything, recalculate the PP and get a higher number due to the AI learning the car better and putting in a better lap.

You'd be much better with a fixed AI than a learning AI for something that needs to be used to produce a consistent measurement, it could be a relatively very simple AI to be honest.

 

[evldave333]

Neither do I, as it's complete nonsense and I hope something has been lost in translation.

The only way for you to only have load only on two or three tyres would be for the other one or two to have fallen off or you're lifting a wheel (as in tripoding an FWD or in older RWD cars managing to lift a front under acceleration).

As for allowing a car to brake and turn not being possible for human beings, umm trail braking?

I really do hope this is just a 'lost in translation' moment.

This doesn't sound right load can surely go in four directions LRFB. The best drivers and Sophy can attenuate it with throttle and break adjustments at the moment they feel it... the good ole ass gyroscope. Saying something is a nonsense might be a mistake it just might be really unconventional, and something new to add to our skill sets. I think there must be more to learn or we would all be Lewis Hamilton.

 

[Silver Arrows]

Again, Sophy in isolation is a cool experiment, but the actual feasability of it needs to be shown. And to be honest, it's less about whether it can be implemented (it feels like it will, but in a much longer time frame then what Kaz and some of the more optimistic people on this forum think, certainly not during GT7's time of life) but whether it can be scaled down to the average player, and whether the over-aggressive and blatant track cutting can be mitigated for mass consumption.

Again, the TAS comparison is apt and correct. In isolation, a cool showcase of what could be done, but ultimately isn't achieveable (or in this case, beaten) by 0.1% of humans. But unless major aspects can be iterated upon and mitigated, then it will stay that way. Nor does it really solve the short term problem of the AI being (likely) bad now.

 

[Scaff]

This doesn't sound right load can surely go in four directions LRFB. The best drivers and Sophy can attenuate it with throttle and break adjustments at the moment they feel it... the good ole ass gyroscope. Saying something is a nonsense might be a mistake it just might be really unconventional, and something new to add to our skill sets. I think there must be more to learn or we would all be Lewis Hamilton.

It's nonsense to say a car, which has four contact points, can only have load at two or three of them (with the exceptions for three that I stated, but don't apply here).

The only way for a point to not have load is for it to not be in contact with the road.

Which is also why the comment raised @Famine eyebrows and why I said its most likely been lost in translation.

Just as the comment about braking while turning being reserved for the like of F1 champions, is once again I hope a translation issue, as trail braking isn't exactly an uncommon technique to use.

 

[regnar]

I get where it's learning to drive the course & push the limits of physics, but we've seen it try to protect it's position from being passed, it like it wants to win, come in 1st place, wheres that coding coming from & why couldn't there be human interjection along the way to remove bad habits.

 

[Bojador]

I will have the chance to speak with Kazunori Yamauchi and the Sony AI team about Gran Turismo Sophy in the coming days — let me know if you guys have any specific questions about Sophy that you'd like answered. 👍

Will we have AI bots in the online mode for daily races and custom lobbies to fill up grids?

 

[Samcastor]

I get where it's learning to drive the course & push the limits of physics, but we've seen it try to protect it's position from being passed, it like it wants to win, come in 1st place, wheres that coding coming from & why couldn't there be human interjection along the way to remove bad habits.

You can try reading my post a few pages ago where I try to state in simple terms the approach to such optimizations and how these behaviors are developed. I can try to answer something if it is not very clear. There already is human interjection to remove bad habits. It gets penalized in its evaluation for making contact, getting off-track penalties, etc. and these punishments (note, that due to convention for algorithms, these are called rewards in the literature even if they are affecting the evaluation in a negative manner) were added to get certain behavior and get closer to racing etiquette.

I wonder another thing about this AI. If they could hypothetically put it to other race sim, Assetto Corsa Competizione for example, and train it will it start to learn other racing game? Asking just for curiosity.

Never played ACC, so obviously a bit of a gap here. Sophy itself is built around Gran Tourismo, but the approach itself should be robust enough. They use model-free learning, meaning they did not model the car and driving physics for Sophie to rely on, it relies on things it perceives and the feedback, just like a human does. So yes, you should be able to use the algorithm and the general implementation and use it in similar racing games. It will likely still need to be tailored to the game to weed out unforseen behaviors and will obviously need to be coded in such a way that it can interact with that game in particular.

 

[eran0004]

What’s really insane about all this, everything they programmed for sophy and the questions about is sophy doing something good or bad, these could have been implemented in the game for players.
Seriously. Rewarding players over a lap, letting a player know if a turn was good or bad, braking early or too late. The coaching. It should be in the game since day one.

It has been in the game since Gran Turismo 1

It’s basically just the lap time, when it comes to the pace. For race craft I assume they penalise the AI for contact as well.

 

[evldave333]

It's nonsense to say a car, which has four contact points, can only have load at two or three of them (with the exceptions for three that I stated, but don't apply here).

The only way for a point to not have load is for it to not be in contact with the road.

Which is also why the comment raised @Famine eyebrows and why I said its most likely been lost in translation.

Just as the comment about braking while turning being reserved for the like of F1 champions, is once again I hope a translation issue, as trail braking isn't exactly an uncommon technique to use.

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).

 

[05XR8]

I get where it's learning to drive the course & push the limits of physics, but we've seen it try to protect it's position from being passed, it like it wants to win, come in 1st place, wheres that coding coming from & why couldn't there be human interjection along the way to remove bad habits.

Yeah. I get that too. I guess it’s do to the nature of racing. I see it as being sly. Racers like Shane Van Gisbergen are slick like that. Not dirty, but move in ways not expected.

At Maggiore, I take that extreme inside line at T2. Not to pass a car ahead, but to optimise my line. On the run down to Dragon Trail Seaside hairpin, I don’t run on those rumble strips. As Tom was saying, it’s a bit of a gamble the car may become unsettled, when braking and preparing to turn. However, I see many players that that liberty.

I race as clean and fair as possible. The way Sophy races is more to optimise all of the road. If it’s not penalised for certain lines, it’s going to continue the quickest, most optimal way around the circuit.
Again, the way Sophy kept taking that T2 at Lago, it calculated when the car ahead wasn’t in on that trajectory on the extreme right of the curb. Humans rarely make the move for the pass, but Sophy kept positioning the car for that line.