Gran Turismo's Sophy AI beats a world record on the Nordschleife

Article Discussion
sauber1.jpg
The infamous AI developped by a joint effort of Sony and Polyphony Digital comes back under the spotlight today.

A video was just posted today on the official Sony Ai youtube channel, showcasing the virtual driver beat the world record previously established by Igor Fraga on the most difficult of the Lewis Hamilton Time Trial Challenges: a lap of the Nürburgring Norschleife with the Sauber C9 on GT Sport.


While Lewis Hamilton's time was a very respectable 5:40.622, Igor Fraga, FIA GTC World Champion and Formula 3 driver, achieved the best time ever recorded by a human player with an incredible 5:26.682 lap. But that time is now good for the 2nd step of the podium only, as Sophy improved on it by a whopping 3.7 seconds margin, with only 5:22.975 to complete the lap.

The lap is quite spectacular to watch, although some of the lines taken will fuel up controversy: Sophy doesn't hesitate to use all the space the game allows before disqualifying the challenger, even if that means having most of the car on the grass. To clarify, the game takes the "2 wheels on track" rule to an extreme, so as long as 2 wheels still touch the tarmac, you're all good.

sauber2.jpg


Is this the only reason why Sophy is so fast ? According to high level Gran Turismo Esports players, the AI seems to still be stuck with the automatic gearbox as it uses 1st gear through Wehrseifen, the slowest corner of the track. This is still believed to be inefficient in that situation, which means Sophy could still unlock more potential if taught how to use a manual gearbox and shift at ideal points.

Kazunori Yamauchi has previously affirmed in an interview that the addition of Sophy in GT7 was planned for a future update, although no date estimation was given. More than just being an impossible to beat opponent, the goal is to make Sophy a true partner which would act as a teacher and a friend to players. Apparently, the AI can also be saved into different images at different states through its learning process, and provide opponent behaviors reflective of a wide variety of driver archetypes, from unsecure beginners and amateurs to experienced high split racers. If these things end up being inplemented to GT7, it should not only provide a new 3 chilis level of difficulty, but also improve the general behavior of the AI through the lower difficulty levels. But with the high amount of content in the game, we can wonder how much time will be needed to make Sophy truly ready to tackle all the cars / tracks / weather combinations possible.
  • Haha
  • Like
  • Angry
Reactions: Beezer215, NinjaDesignz, Jojosan and 10 others

Original source

About author
GT-Alex
Alexandre Tonini
Global motorsports enjoyer, long time simracer, Gran Turismo veteran, I've been driving alongside top drivers since the dawn of online pro leagues on Gran Turismo, and qualified for the only cancelled FIA GTC World Tour. I've left aside competitive driving in 2020 to dedicate myself to IGTL, a simracing organisation hosting high quality events for pro racers and customers, to create with friends the kind of events we wished we could have had. We strive to provide the best events for drivers and the best content for viewers, and want to help the simracing scene grow and shine further in the global esports scene.

More in Gran Turismo

Comments

E
The AI just demonstrates how little Gran Turismo hast do to with being a simulation that it claims to be, no wonder they deactivated the comments on youtube.
 
P
One day it will be perfect, with Sony or with Microsoft or with whoever else, one day...
But that day has not arrived!!!!
And when it comes, what will it prove?
Like the day the man who multiplied 23623587485 by 1254844563 was beaten by a calculator, what did that prove?
That an electronic circuit goes much faster than neurons.
Or that a baseball bat is stronger at returning a ball than a palm.
Let's compare what is comparable, in the meantime, from my point of view, Sony covers itself with ridicule by producing this kind of demonstration.
 
R
Too bad people are focusing on the onrealistic physics and track limits of GT7 and completely missing the point of what this new ai implementation can bring to racing games in general. Off course this adaptive learning ai is still very new and has to be developed way more, but people have proved it has potential and when all issues are ironed out this ai can bring offline racing to a new level with way more human ai behaviour. If this eventually works in GT7 it can be used in every racing game.

If you read about the way Sophy is implemented, you wil understand that this is the way you want the ai is to be implemented in future racing games. If you don’t, all you see is a group C car exploiting track limits with way too simple offtrack physics.
 
Last edited:
TheCatweazle
Rusty911 said:
Too bad people are focusing on the onrealistic physics and track limits of GT7 and completely missing the point of what this new ai implementation can bring to racing games in general. Off course this adaptive learning ai is still very new and has to be developed way more, but people have proved it has potential and when all issues are ironed out this ai can bring offline racing to a new level with way more human ai behaviour. If this eventually works in GT7 it can be used in every racing game.

If you read about the way Sophy is implemented, you wil understand that this is the way you want the ai is to be implemented in future racing games. If you don’t, all you see is a group C car exploiting track limits with way too simple offtrack physics.
Click to expand...
Sorry, Rusty911, but you seem to be missing the point.

If this is the future of racing games, I will no longer play them.

Through thousands of repetitions, Sophy learned to exploit the physics engine's weaknesses to the fullest, delivering a lap which was as close to perfection as the game allows for. If this holds true for all future racing games, it means that you, the human player, need to (a) know and understand all possible strengths and weaknesses of a game's physics engine, (b) need to know with utmost certainty how the weaknesses can be exploited, and (c) need to exploit said weaknesses to the fullest for EVERY SINGLE SECOND OF EVERY SINGLE LAP OF A RACE. If you cannot do that, the AI will drive circles around you.

You need to achieve perfection. If you can't do that, you're screwed.

I don't know about you, but I play computer games to be entertained, to have a good time, sometimes even to enjoy seeing the AI mess things up. But the most important thing is, I need to be convinced that I might have a chance to win. I need to feel that if I didn't win, it was up to me and my own shortcomings that I failed, not due to an overpowering AI that can only be defeated by achieving that which humans simply can't achieve.

The future you mentioned will give me exactly that, an AI which is absolutely unbeatable.

And if that's the future, I'm out the door.
 
R
TheCatweazle said:
Sorry, Rusty911, but you seem to be missing the point.

If this is the future of racing games, I will no longer play them.

Through thousands of repetitions, Sophy learned to exploit the physics engine's weaknesses to the fullest, delivering a lap which was as close to perfection as the game allows for. If this holds true for all future racing games, it means that you, the human player, need to (a) know and understand all possible strengths and weaknesses of a game's physics engine, (b) need to know with utmost certainty how the weaknesses can be exploited, and (c) need to exploit said weaknesses to the fullest for EVERY SINGLE SECOND OF EVERY SINGLE LAP OF A RACE. If you cannot do that, the AI will drive circles around you.

You need to achieve perfection. If you can't do that, you're screwed.

I don't know about you, but I play computer games to be entertained, to have a good time, sometimes even to enjoy seeing the AI mess things up. But the most important thing is, I need to be convinced that I might have a chance to win. I need to feel that if I didn't win, it was up to me and my own shortcomings that I failed, not due to an overpowering AI that can only be defeated by achieving that which humans simply can't achieve.

The future you mentioned will give me exactly that, an AI which is absolutely unbeatable.

And if that's the future, I'm out the door.
Click to expand...
Don't worry mate, that lap on the Nordschleife is just a display of what the ai is capable of. Important is to remember that this way of implementing the ai is working and will lead to more human ai behaviour. For the average joe like you and me, there is an ai level slider to set the ai to match our pace.
 
S
Rusty911 said:
Too bad people are focusing on the onrealistic physics and track limits of GT7 and completely missing the point of what this new ai implementation can bring to racing games in general.
Click to expand...
That's because it's irrelevant as long as it's limited to games with that level of realism.
 
RedBull48
Neilski said:
Part of the problem with some racing games is that the track limits are unpredictable. I've experienced this in AC a lot (on Kunos tracks and mod tracks), and you can't even see the limits until you exceed them, on a once-per-lap basis (unless you install extra mods to monitor your on-track status). So it's beyond tedious to learn the limits precisely so you know where you can push your luck. On something as long as Nords, it's just really not practical at all for normal players.

I noted at least 5 incidents in the Sophy lap where the car was way beyond the white line with all four wheels, and was only *maybe* still "on" the kerbing. I guess the kerbing must be considered part of the track in this game?
Yesterday in Austria, it was clear that (at Spielberg at least) F1 doesn't consider the kerbing to be part of the track, since cars exceeding the white line with all four wheels were penalised. I can never remember which formulas allow which combinations of the rules (lines vs kerbing etc.).

Honestly, the off-track antics downgrade the whole performance from impressive to silly in my eyes.
Click to expand...
Regarding F1, what makes this particularly confusing is the FIA's inability to have any sort of consistency. Last year they would regularly use a combination of curb and white line to define track limits. Now it seems only the white line which to me makes the most sense. However to compound the issues they rotate stewards so perception changes race to race. This is a large part of why the drivers are growing soo frustrated.
 
Last edited:
GT-Alex
Don't forget the learning AI is running on multiples interconnected consoles. The whole process won't be running in your living room anytime soon. What you'll most likely end up with on the client side is a set of images of the AI code taken at different steps through the learning process, or an overall image with sliders to adjust, or both. You could even make different personalities by changing the reward / punishment triggers.
 
Italya
I think this is just pushing it as far as it'll go, the AI isn't even finished yet and I would be very surprised if this is what you'd get racing against it lol
b)b said:
Lol GT Sophy, your cheating AI buddy.
We don't want out of this world AI making impossible laps. We want realistic and believable AI.
Click to expand...
 
vettelsepp
In real life, this lap would have ended after a few corners with this driving style over the curbs.
 
Zoltan76
Rusty911 said:
Sophy raced according to the same physics and track limit rules as every human GT7 player. GT7 apparently does not punish the racer very hard by going off track like other games do like AC, RF2, AMS2, etc.
Click to expand...
And it says everything about the fantastic physics of GT. ROTFL.
It is an arcade game.
 
Last edited:
sc4dr
I'm sorry, but I am just totally unimpressed. Kinda like fake news.

Is RD turning into a GT fan boy page?? Haha roll eyes....

I prefer to just wait for an online server that has real people in it. Nothing else comes close. just one mans opinion.
 
Last edited:
T
It is really interesting as a programming/physics exercise. But that's it, that IA exercise ignores a lot of factors in order to consider it worthy of being feasible as a learning tool.

It drives in the way a trained computer would handle a car in a game, a real car can't instantly apply the microseconds inputs from the IA to the hydraulic cylinders of the brakes, nor appying microsecond inputs that generate instant engine torque variations, in real life there is inertias and lag induced by electronical, mechanical and chemical reasons.

The IA actuates the pedals like a PWM, it means that the pedals are pushed completelly on and off in incredibly small amounts of time, several orders of magnitude faster than any human could ever do. You can't learn from hundreds of 0 to 100% pedal pushes per second, you can't even visualize it properly.

The IA also doesn't have any kind of reaction time included, any human being has a reaction time in every input he does and when something unexpected happens like for example a loss of grip over a bump or grass the human needs some time to think if something needs to be done, what needs to be done, and how much.

Humans overcome this very limitation by predicting the car behaviour based on the experience of the previous laps in both inputs and car behaviour. The drivers with the best car control are people that can recognize that the car is going to missbehave a split second faster than an average person, and thus predict what the car is going to do.

And even with instant reaction time the arms and foots need some time to take the nervous impulse to the arms and legs and to start to move them, additionally the arms and legs also have inertia and internal frictions.

What I'm trying to say is that this particular AI even while on paper it is impressive as a concept, it is almost useless as a learning tool.

It takes unrealistic lines over the kerbs and grass because a computer can afford to make several thousands of micro corrections per second, it can sense changes in car behaviour several orders of magnitude smaller than a human being ever could sense and basically lives in a realm in where input lag and reaction time is inexistent.

And also this demo don't show is how the IA adapt to changes or if it can learn a new track like a human being does without several hundreds of thousands of iterations, or how it adapts in changing weather finding a new wet line that differs from the dry line in where the IA was trained.

I hope that they keep developing the IA from a human perspective with real life limitations to inputs, reaction times, sensory limitations being taken in account. At that point i would become a learning tool and even could be used to automatically find the best car setup or to calculate race strategies. The possibilities are almost endless.
 
GT-Alex
TRASGU said:
It is really interesting as a programming/physics exercise. But that's it, that IA exercise ignores a lot of factors in order to consider it worthy of being feasible as a learning tool.

It drives in the way a trained computer would handle a car in a game, a real car can't instantly apply the microseconds inputs from the IA to the hydraulic cylinders of the brakes, nor appying microsecond inputs that generate instant engine torque variations, in real life there is inertias and lag induced by electronical, mechanical and chemical reasons.

The IA actuates the pedals like a PWM, it means that the pedals are pushed completelly on and off in incredibly small amounts of time, several orders of magnitude faster than any human could ever do. You can't learn from hundreds of 0 to 100% pedal pushes per second, you can't even visualize it properly.

The IA also doesn't have any kind of reaction time included, any human being has a reaction time in every input he does and when something unexpected happens like for example a loss of grip over a bump or grass the human needs some time to think if something needs to be done, what needs to be done, and how much.

Humans overcome this very limitation by predicting the car behaviour based on the experience of the previous laps in both inputs and car behaviour. The drivers with the best car control are people that can recognize that the car is going to missbehave a split second faster than an average person, and thus predict what the car is going to do.

And even with instant reaction time the arms and foots need some time to take the nervous impulse to the arms and legs and to start to move them, additionally the arms and legs also have inertia and internal frictions.

What I'm trying to say is that this particular AI even while on paper it is impressive as a concept, it is almost useless as a learning tool.

It takes unrealistic lines over the kerbs and grass because a computer can afford to make several thousands of micro corrections per second, it can sense changes in car behaviour several orders of magnitude smaller than a human being ever could sense and basically lives in a realm in where input lag and reaction time is inexistent.

And also this demo don't show is how the IA adapt to changes or if it can learn a new track like a human being does without several hundreds of thousands of iterations, or how it adapts in changing weather finding a new wet line that differs from the dry line in where the IA was trained.

I hope that they keep developing the IA from a human perspective with real life limitations to inputs, reaction times, sensory limitations being taken in account. At that point i would become a learning tool and even could be used to automatically find the best car setup or to calculate race strategies. The possibilities are almost endless.
Click to expand...
I'm sorry but most of what you say there is inaccurate.

Regarding the way inputs are activated, I'm pretty sure it can't do a PWM instant 0-100%, since the way the inputs work in the game itself has a built in delay when you use pad with the cross and square buttons for throttle and brake. While the actuators below the buttons are on-off, you can very clearly see the inputs have some kind of progression there, so the AI can't force the pedals to actuate faster than that.

Moreover, Sophy's inputs frequency is limited to 10 Hz, which means it can't be as smooth as humans. Regarding reaction times, they were measured at about 23-30ms for Sophy VS 250-300ms for professional athletes, so they added an artifical delay to Sophy. Papers say they've tested 100ms, 200ms and 250ms, although we don't know excatly what kind of delay is applied to this run, it was definitely run with some.

Sophy's ability to learn under the same restrictions as humans is the whole point of the project. It's a neural network, self adapting to whatever you throw at it. But trial and error is unavoidable (just like for humans by the way), and a neural network is just too heavy to throw into consumer hardware, so state images will be what we'll get.
 
Last edited:
B
I mean what's the point of a similar run? the car is absolutely unreal in terms of physics. while on grass at full throttle the car barely moves. I had gtsport for few months, sold it with the ps4 an bought a PC, installed AC, CM and mods and now we can talk about physics
 
R
I definitely don't want AI be that fast, I want easy victories in these kind of arcadish driving games. There is enough stressfull sims out there.
 

Latest News

Article information

Author
Alexandre Tonini
Article read time
2 min read
Views
23,813
Comments
68
Last update
Join the discussion

Share this article

What's needed for simracing in 2024?

  • More games, period

  • Better graphics/visuals

  • Advanced physics and handling

  • More cars and tracks

  • AI improvements

  • AI engineering

  • Cross-platform play

  • New game Modes

  • Other, post your idea

Results are only viewable after voting.
See comments…
Back
Top