Support NeoGAF

[Thirty7ven]

“Sophy is very fast, with lap times better than expected for the best drivers,” he says via a translator. “But watching Sophy, there were certain moves that I only believed were possible afterward.”

In contrast to board games that AI has mastered, such as chess or Go, Gran Turismo requires continuous judgments and high-speed reflexes. It’s far more complex than action games like Starcraft or Dota and demands challenging driving maneuvers. A Gran Turismo ace must balance pushing a virtual car to its limits and wrestling with friction, aerodynamics, and precise driving lines with the subtle dance of trying to overtake an opponent without unfairly blocking their line and incurring a penalty.

Gerdes says the techniques used to develop GT Sophy could help the development of autonomous cars. Currently, self-driving cars only use the kind of neural network algorithm that GT Sophy employed to keep track of road markings and perceive other vehicles and obstacles. The software controlling the car is handwritten. “GT Sophy’s success on the track suggests that neural networks might one day have a larger role in the software of automated vehicles than they do today,” Gerdes writes.

GT Sophy mastered Gran Turismo through hours of practice. As with other recent feats of AI gameplay, this involved training an algorithm known as a neural network to improve its operation of the game’s controls by providing positive and negative feedback. The approach, known as reinforcement learning, is inspired by the way animals respond to success and failure in the real world. Although it is decades old, the method has come to the fore in recent years, thanks to more sophisticated algorithms, more powerful computers, and more copious training data.

GT Sophy is the first AI capable of beating professional esports drivers in such a realistic, high-speed game. In a series of races held in July and October 2021, Sophy bested the best Gran Turismo drivers, including Miyazono.

Peter Wurman, director of Sony AI America, says mastering Gran Turismo is as important an AI landmark as dominating chess or Go. Wurman cofounded Kiva Systems, which developed the shelf-moving robots that transformed Amazon’s warehouses. Because a Gran Turismo driver must understand how to beat other drivers without incurring penalties for unfair play, Wurman says GT Sophy points the way to robots that learn how to interact with humans in more sophisticated social settings. “With board games, you have a long time to decide what you're going to do between each move,” Wurman says. “Real-time interaction is what we do every day.”

“Sophy takes some racing lines that a human driver would never think of,” he says. “I think a lot of the textbooks regarding driving skills will be rewritten.”

More at the link, pretty exciting.

 

[ethomaz]

Sophy name seems like Sony + Polyphony child.

Imagine putting that AI in GT7 people will go mad.

 

[SLB1904]

Finally. I was waiting for this

Edit: no gameplay. Ffs

 

[Pantz]

Sophy will get smashed into the first corner wall like all the rest.

 

[Anos Voldigoad]

Looking forward to the very first sophy gameplay

 

[Bwesh]

...far more complex than action games like Starcraft or Dota and demands challenging driving maneuvers.

I highly doubt this. Dota needs constant calculations and focus on where you are, where your allies and enemies are, your own cooldowns, your enemies cooldowns, buyback statuses, what items your enemies are building, and more. There's also more on-the-fly decision making in Dota than Gran Turismo.

 

[IntentionalPun]

Interesting stuff.

Would be interesting to note if this is due to some "cracks" in the "simulation" or if these sorts of techniques would actually work on a real track with real cars.

Typically the "lines" that drivers use in these games are meant to mimic those used in real life.

But honestly.. from a machine learning standpoint... not sure how interesting this is lol Folks working on car AI aren't trying to make cars win races on controlled track conditions.

 

[Thirty7ven]

I highly doubt this. Dota needs constant calculations and focus on where you are, where your allies and enemies are, your own cooldowns, your enemies cooldowns, buyback statuses, what items your enemies are building, and more. There's also more on-the-fly decision making in Dota than Gran Turismo.

Sounds more like you're imagining a lone car on the track. Going 260km/h with a bunch of other cars at the same time, and trying to overtake, not getting overtaken, while accounting for possible penalties and reacting to human player actions?

 

[SLB1904]

I highly doubt this. Dota needs constant calculations and focus on where you are, where your allies and enemies are, your own cooldowns, your enemies cooldowns, buyback statuses, what items your enemies are building, and more. There's also more on-the-fly decision making in Dota than Gran Turismo.

That's not accurate at all. Racing you have to make decisions under of 10th of a sec.

 

[Aenima]

From what i understand this AI Sophy is not to be used in the game, it would be impossible to beat her driving fair. It seems they are using GT to train the ai for use in real life cars and other robotics.

I also heard this AI will soon change her name to Skynet.

 

[alucard0712_rus]

B-Spec mode will be awesome with good AI

 

[Raven117]

I highly doubt this. Dota needs constant calculations and focus on where you are, where your allies and enemies are, your own cooldowns, your enemies cooldowns, buyback statuses, what items your enemies are building, and more. There's also more on-the-fly decision making in Dota than Gran Turismo.

LOL. Been drinking that koolaid that DOTA is the most complex and real time fast decision making than anything, huh?

 

[Kuranghi]

That's not accurate at all. Racing you have to make decisions under of 10th of a sec.

Its got nothing to do with reaction time, its about making extremely complex intelligent decisions, in Dota/LoL that could be running away to buy something, engaging, supporting someone else who just turned up, chasing, staying put. In GT it has to avoid the other cars, its not following a live changing economy as the game/race goes on.

There's only so much that could happen in a race since its a linear journey round a set track where the speed of opponents will be within a predictable range, I'm not saying there aren't other factors to consider in GT outside of the immediate logistics of the cars around you, but I think its fair to say its nowhere near as complex as a free moving open map where any player can be anywhere and decide to retreat or advance for many different reasons.

I'd say the decisions that could be made by an AI in a high-level game of DotA/LoL would change at such a pace that the AI could often never make a decision that wasn't too out of date to be useful, just because it can't reasonably predict what the opponents will be doing that far in the future. Compared to a racing game where you can predict what the opponent will do further into the future because there are so many less avenues of decision the opponent can go down.

 

[Guilty_AI]

Interesting stuff.

Would be interesting to note if this is due to some "cracks" in the "simulation" or if these sorts of techniques would actually work on a real track with real cars.

Technically they are using these "cracks" on the simulation to develop the AI. However, considering they're simulating a real car in a real track, it shouldn't be impossible to use the developed AI for our own matrix, as long as the car and track are the same ones used during the training.

Well, the concept isn't really anything new, you can find plenty of people on youtube doing similar things.




Its using the same methodology that was used for this sophy.

 

[REDRZA MWS]

I’m still waiting for the PS3’s folding at home to cure cancer.

 

[Spaceman292]

Sophy will get smashed into the first corner wall like all the rest.

I wouldn't mind smashing Sophy

 

[SLB1904]

Its got nothing to do with reaction time, its about making extremely complex intelligent decisions, in Dota/LoL that could be running away to buy something, engaging, supporting someone else who just turned up, chasing, staying put. In GT it has to avoid the other cars, its not following a live changing economy as the game/race goes on.

There's only so much that could happen in a race since its a linear journey round a set track where the speed of opponents will be within a predictable range, I'm not saying there aren't other factors to consider in GT outside of the immediate logistics of the cars around you, but I think its fair to say its nowhere near as complex as a free moving open map where any player can be anywhere and decide to retreat or advance for many different reasons.

I'd say the decisions that could be made by an AI in a high-level game of DotA/LoL would change at such a pace that the AI could often never make a decision that wasn't too out of date to be useful, just because it can't reasonably predict what the opponents will be doing that far in the future. Compared to a racing game where you can predict what the opponent will do further into the future because there are so many less avenues of decision the opponent can go down.

Dude you clearly don't know nothing about racing. What ignorant statement lol

 

[Guilty_AI]

That's not accurate at all. Racing you have to make decisions under of 10th of a sec.

This kind of reaction time doesn't really matter much for an AI. Its the decision making process thats troublesome.

The AI can perceive a incoming wall pretty fast no problem, and it has more than enough leeway to react in time, being a machine and all. The real problem here is how we teach the AI that it is supposed to break when it sees said oncoming wall.

In that sense, competing with other humans is definitely a much more complex task, since humans are far more unpredictable than a racing track. As i showed in the video earlier, the methodology is 'simple' enough that you can find people on youtube applying it to different games for educational purposes, or even just for the heck of it.

 

[a multicultural team of v]

This is pretty impressive, I mean they are essentially teaching an AI Newtonian physics simply by trial and error. Brake point, break locking, tire grip, the effect of tire wear, the effect of having 10 gallons of gas instead of 12, etc all need to be figured out, again by trial and error. Granted, it's not like a human does any of this precisely, but we are able to understand how they all relate and intuitively comprehend how they all effect each other because we can actually think. This AI is figuring it out by simply driving a billion times and seeing which one produces the best outcome.

I would also really like to see two of these AI race each other. They would both essentially be fighting for the same "optimal" path, so I wonder what kind of decisions would be made when both AIs have to react to each other doing the same thing. This AI might actually prove what starting position is quantitatively the best and most likely to produce a race winner, because it would be the only variable in the experiment.

 

[SLB1904]

This kind of reaction time doesn't really matter much for an AI. Its the decision making process thats troublesome.

The AI can perceive a incoming wall pretty fast no problem, and it has more than enough leeway to react in time, being a machine and all. The real problem here is how we teach the AI that it is supposed to break when it sees said oncoming wall.

In that sense, competing with other humans is definitely a much more complex task, since humans are far more unpredictable than a racing track. As i showed in the video earlier, the methodology is 'simple' enough that you can find people on youtube applying it to different games for educational purposes, or even just for the heck of it.

Yeah but this ai is beating the best players in the world. These guys are constantly in the world finals. And if you see the races they are very close races. That's impressive

 

[Guilty_AI]

Yeah but this ai is beating the best players in the world. These guys are constantly in the world finals. And if you see the races they are very close races. That's impressive

Its impressive, but its important to note they're beating those players in time score. They're basically driving an empty track and making the best time possible, other players aren't something it needs to consider, unlike what you'd see in a Dota 2 or Chess AI.

I would also really like to see two of these AI race each other. They would both essentially be fighting for the same "optimal" path, so I wonder what kind of decisions would be made when both AIs have to react to each other doing the same thing.

This is an important point you brought up that some people might get confused about. The AIs in this type of methodology are driving alone in the track. They aren't really adapted to react to other drivers. If you put them against each other, they'd either have to be ghosts or they would just crash.
NVM, this specific one was trained to race against other people

 

[Thirty7ven]

Yeah but this ai is beating the best players in the world. These guys are constantly in the world finals. And if you see the races they are very close races. That's impressive

Aside from all the physics involved, of which there basically none to account in LoL, there’s also a matter of strategy to win a race that involves defensive and offensive positioning without penalties. The aim here is for the A.I to behave like a professional racer, it’s not just about perfect lines in a time trial.

 

[Thirty7ven]

Its impressive, but its important to note they're beating those players in time score. They're basically driving an empty track and making the best time possible, other players aren't something it needs to consider, unlike what you'd see in a Dota 2 or Chess AI.

Watch the video.

 

[SLB1904]

Its impressive, but its important to note they're beating those players in time score. They're basically driving an empty track and making the best time possible, other players aren't something it needs to consider, unlike what you'd see in a Dota 2 or Chess AI.

What? Did you watch the video?
They made a tournament lol

 

[GymWolf]

So this is basically driveavatars 2.0?!

 

[[Sigma]]

@17:50 is the first race.

 

[SLB1904]

So this is basically driveavatars 2.0?!

 

[BlackM1st]

I wouldn't mind smashing Sophy

 

[GymWolf]

Sorry i didn't followed this thing from very close because i don't care for driving games but i like IA stuff.

 

[NickFire]

Sounds cool, but I'm not sure I want it. I'm gonna lose plenty of races without it already.

 

[Thirty7ven]

Sorry i didn't followed this thing from very close because i don't care for driving games but i like IA stuff.

Drivatars as I understand were just ghosts so to speak, they were mimmicking somebody's recorded lap. This is an actual A.I.

 

[Oddvintagechap]

Like I thought they are using this to train their future cars AI.

Sony is going to need a City diving sim mode in their games, maybe that gathers google maps data.



It’s amusing, I guess.

 

[Guilty_AI]

What? Did you watch the video?
They made a tournament lol

Watch the video.

Yeah, they really are running against other players : 'P (though they did have a problem with it crashing into other players at first as expected).

Still, its not good to compare it to the AIs of the likes of Dota 2 and Chess, because those still require a lot more of "human" decisions in the mix while a race still relies more on reaction time (which is something AIs can vastly exceed humans at just by its nature)

 

[CobraAB]

Looking forward to getting it on with Sophy.

 

[itsjustJEFF]

I never thought it would be some scientist in a lab that would end up creating true AI, it'll be some game developer or a hobby programmer that inadvertently creates it. Unknowingly unleashing that evil upon us... LOL

EDIT: Sorry, I didn't give context to the hobby programmer. He only wanted to create a digital "companion"... for company

 

[Elios83]

Yeah I expected it was about this.
And they confirmed that Sophy is coming to GT7 in the future.

 

[sncvsrtoip]

 

[Thirty7ven]

Yeah, they really are running against other players : 'P (though they did have a problem with it crashing into other players at first as expected).

Still, its not good to compare it to the AIs of the likes of Dota 2 and Chess, because those still require a lot more of "human" decisions in the mix while a race still relies more on reaction time (which is something AIs can vastly exceed humans at just by its nature)

I'm gonna listen to the A.I experts that actually work on the field and teach at university, but you're entitled to your opinion.

 

[12Goblins]

hm, I'm not convinced this is "far more complex" than starcraft or dota AI

 

[alucard0712_rus]

@17:50 is the first race.

Now that's impressive AND fun at the same time. It's actually more fun to watch how AI drives than real drivers lol.

 

[//DEVIL//]

Sophy will get smashed into the first corner wall like all the rest.

Omg I laughed so hard at this lol

 

[Zimmy68]

Translation, Sophy will crash into everyone at the first turn to take 1st place.

 

[nikos]

Sophy's hacking.

 

[Fess]

They say ”Gran Turismo”, is this AI in GT7?
Either way this is amazing.

 

[Thirty7ven]

They say ”Gran Turismo”, is this AI in GT7?
Either way this is amazing.

Will come in an update.

 

[Calling 911 Melts Beams]

sophy's choice 2: grand autismo

 

[Tommi84]

 

[Guilty_AI]

I'm gonna listen to the A.I experts that actually work on the field and teach at university, but you're allowed to your opinion.

Regarding that... who exactly was the "AI expert" that said it was more complex than these other AIs. Its on the article but i don't remember seeing anyone in the video saying that.

 

[Spaceman292]

I'm just picturing her with a crash helmet. And fat tats.

 

[onQ123]

@17:50 is the first race.

What I got from this is that we need a live commentator for all online games