Almost-unbeatable AI Comes To Gran Turismo 7 (arstechnica.com) 34
An anonymous reader shares a report: Last year, Sony AI and Polyphony Digital, the developers of Gran Turismo, developed a new AI agent that is able to race at a world-class level. At the time, the experiment was described in a paper in Nature, where the researchers showed that this AI was not only capable of driving very fast -- something other AI have done in the past -- but also learned tactics, strategy, and even racing etiquette. At the time, GT Sophy -- the name of the AI -- wasn't quite ready for prime time. For example, it often passed opponents at the earliest opportunity on a straight, allowing itself to be overtaken in the next braking zone.
And unlike human players, GT Sophy would try to overtake players with impending time penalties -- humans would just wait for that penalized car to slow to gain the place. But in the intervening year, Sony AI and Polyphony Digital have been working on GT Sophy, and tomorrow (February 21), GT Sophy rolls out to Gran Turismo 7 as part of update 1.29, at least for a limited time. Until the end of March, players can try their skills against Sophy in the GT Sophy Race Together mode in a series of races with increasing difficulty levels. There's also a one-versus-one match where you race Sophy in identical cars, so you can see how much slower you are than the AI.
And unlike human players, GT Sophy would try to overtake players with impending time penalties -- humans would just wait for that penalized car to slow to gain the place. But in the intervening year, Sony AI and Polyphony Digital have been working on GT Sophy, and tomorrow (February 21), GT Sophy rolls out to Gran Turismo 7 as part of update 1.29, at least for a limited time. Until the end of March, players can try their skills against Sophy in the GT Sophy Race Together mode in a series of races with increasing difficulty levels. There's also a one-versus-one match where you race Sophy in identical cars, so you can see how much slower you are than the AI.
Not surprising (Score:4)
Re:Not surprising (Score:5, Interesting)
On top of that, mastering a track through repetition will provide the AI a big advantage. A human being could race the same track a number of times a day to exercise, before their results worsen (fatigue, boredom, burn-out, etc). The AI could exercise 24/7 forever, constantly improving. That's before we even get to accelerated training, where everything is sped up 10 times and so on.
Re:Of course it's not surprising (Score:4, Insightful)
mastering a track through repetition
Oh people. Everyone forgets that the "AI" (or let's just call it "The Game") controls not just the opposing drivers, but the game physics as well. The code in the opposing driver logic has access to the code in the game physics logic. You think the opposing player logic is actually visually interpreting the road? God no. The opposing-player logic is querying a bezier curve, then looking up what the exact coeficient of static friction is between its car's tires and that road.
This doesn't even begin to mention that opposing-player logic (you see how I'm avoiding the use of the term "AI" - it's on purpose) has, for generations, been designed to tweak or outright break the game mechanics to suit itself. And if you think this particular opposing-player logic isn't ruthlessly cheating, you trust the designers far more than I do.
Not according to the paper linked to the article, (), unless I missed an important point. They are only feeding in visual information, the AI is on a separate machine, not running as part of the game code itself: [researchgate.net]
"As is common, the agent was given a static map defining the left and right edges and the center line of the track. We encoded the approaching course segment as 60 equally spaced 3D points along each edge of the track and the center line (Figure 1(b)). The span of the points in any given observation was a function of the current velocity so as to always represent approximately the next 6 seconds of travel. The points were computed from the track map and presented to the neural network in the agent’s egocentric frame of reference."
Re: (Score:2)
mastering a track through repetition
Oh people. Everyone forgets that the "AI" (or let's just call it "The Game") controls not just the opposing drivers, but the game physics as well. The code in the opposing driver logic has access to the code in the game physics logic. You think the opposing player logic is actually visually interpreting the road? God no. The opposing-player logic is querying a bezier curve, then looking up what the exact coeficient of static friction is between its car's tires and that road.
This doesn't even begin to mention that opposing-player logic (you see how I'm avoiding the use of the term "AI" - it's on purpose) has, for generations, been designed to tweak or outright break the game mechanics to suit itself. And if you think this particular opposing-player logic isn't ruthlessly cheating, you trust the designers far more than I do.
Not according to the paper linked to the article, (AI Paper [researchgate.net]), unless I missed an important point. They are only feeding in visual information, the AI is on a separate machine, not running as part of the game code itself:
"As is common, the agent was given a static map defining the left and right edges and the center line of the track. We encoded the approaching course segment as 60 equally spaced 3D points along each edge of the track and the center line (Figure 1(b)). The span of the points in any given observation was a function of the current velocity so as to always represent approximately the next 6 seconds of travel. The points were computed from the track map and presented to the neural network in the agent’s egocentric frame of reference."
Sorry, I seem to fail at posting links...
Re: (Score:2)
I'm curious how it would perform if it had an artificial limit put in place such that it mimicked the average human reaction times.
I doubt you'll have to wonder very long. Watching AI destroy humans in competition is boring as hell.
Not exactly the revenue generator gamemakers want, unless AI is going to somehow open it's wallet.
Works even better with... (Score:5, Funny)
There's also a one-versus-one match where you race Sophy in identical cars, so you can see how much slower you are than the AI.
This works even better with the PlayStation camera attached and pointed at you, then you get an on-screen readout that uses advanced AI-driven facial analytics to show you an exact percentage decrease in your self-esteem and bonus rage-against-the-machine meter!
Racing Etiquette (Score:4, Insightful)
Racing etiquette? That sounds like a weakness to me.
Re: (Score:1)
In real racing there are factors an AI won't consider. Cost of colliding with another car. Political consequences. Fines. Penalties. AIs tend to "race cleanly" and humans don't.
AIs also don't think "outside the box" so no outside passes while hugging the outside wall to take enough places to advance to the next stage. Humans do.
Sure, an AI gearbox will beat a human with MT everytime, because the human MAY get to the perfect shift point but the AI WILL get there. The human may, through many iteration
Re: (Score:2)
In Gran Turismo you get penalized for un-sportsmanlike behaviour. Things like deliberately colliding with other cars for advantage, abusing the track limits, weaving around dangerously, and so forth.
If you do it too much your rating drops and you get to race against less skilled players.
Re: (Score:2)
No, no, he didn't slam you, he didn't bump you, he didn't nudge you... he *rubbed* you. And rubbin, son, is racin'.
Harry Hogge
Re: (Score:2)
It's essentially the prisoner's dilemma. If the human players operate a tit-for-tat strategy then it can be preferable not to defect.
Re: (Score:2)
Racing etiquette? That sounds like a weakness to me.
Yeah, I guess when you get rid of all those meatsacks filled with blood on the racing track, you can afford to not give a flying fuck about anything in front of you.
Re:Racing Etiquette (Score:4, Informative)
In a real race, there are actual rules on racing etiquette (notwithstanding what you might have seen in the Ricky Bobby movie). A driver who violates the rule will get black flagged by the marshals, which means their race is over. So observing those rules is required to win the race.
Re: (Score:2)
In a real race, there are actual rules on racing etiquette (notwithstanding what you might have seen in the Ricky Bobby movie). A driver who violates the rule will get black flagged by the marshals, which means their race is over. So observing those rules is required to win the race.
A race car driver who breaks the rules, risks their own death and the death of other human drivers. THAT is why rules exist. The race can be "over" in a very permanent way otherwise.
We wouldn't give a shit about rules if we didn't give a shit about humans surviving. This is the entire reason entertainment has grown from Romans killing in the Colosseum every weekend to UFC earning billions by not causing death in the octagon.
Reminds me of the thing I had with the Matrix (Score:1)
Re: (Score:2)
Not that the Matrix is without logical flaws (and most great sci-fis have at least some), but the people agents were generally shooting at were being controlled from outside the Matrix.
So you could say the agents were trained in how to combat issues arising purely within the Matrix but didn't re-train their AI to combat freed humans with their cheat codes.
Maybe that was even a feature not a bug, because each generation of The One was meant to pop up, have some success vs the agents, then die.
Re: (Score:2)
Well, from a purely "technical" point of view, the free humans in the Matrix were using cheat codes. That's pretty much how they could be missed.
Re: (Score:2)
Yeah, so this reminds me of one of the main problems I personally had with The Matrix. If the agents in the Matrix are algorithmically aware of exactly how everything happens within said matrix, how would they ever miss when firing a shot?
Because the weapons being used to "fire shots" are weapons that exist within a world full of imperfect rules FOR humans. That's why. Recall what the Agent spoke about when talking about the first iterations of the Matrix. How they were a bit too perfect and were ultimately not accepted by the human batteries powering it all, forcing the machines to make an imperfect world.
I'm rather surprised you missed this.
Re: (Score:2)
I think Morpheus and his team were abberations that didn't fit said algorithms and that's why Agent Smith and his team sought to suppress them. Their actions couldn't be predicted and as such, the shots fired were less effective.
Re: (Score:2)
My main problem with the Matrix was they were terrible.
Re: (Score:2)
The matrix? Yeah, great movie... too bad they never made any sequels, I bet they would have been great!
so it does not need to cheat like other race ai's? (Score:2)
so it does not need to cheat like other race ai's?
Re: (Score:2)
I was just thinking that. "Unbeatable" computer controlled opponents sure ain't something new. This is usually achieved by giving them some sort of "unfair" advantage, make their vehicles faster through corners or make physics not apply to them, at least fully, so to give an illusion of a "fair" race while keeping the challenge up by essentially putting these cars on rails.
But it sounds like they want to finally give the computer controlled cars exactly the same physics and characteristics that you get and
Re: (Score:2)
All in all it doesn't seem that difficult given the characteristics of a vehicle are known, the characteristics of a track are known, and there is only 1 perfect line through a course. You could very easily make an unbeatable computer player just with that information, without having to resort to Mario Kart 64-style cheating.
So I guess the real difference here is in how the car reacts to other players. But even Mario Kart 64 had computer opponents that played in ways the article describes, like "waiting to
What? No rubber band ai? (Score:2)
I'll miss having a vehicle catching up to me on a straightaway like it's got a couple of P&W J58 engines strapped to it.
All that effort ... (Score:1)
and the cars still sound like the small appliance section of a Harvey Norman store on Saturday afternoon.
Re: (Score:2)
That's just from the lame mufflers they put on.
pfft (Score:3)
Try playing Doom on 'nightmare' or about half the games since computers first had them.
Making 'the computer' insanely good has long been basically trivial as far as problems go.
The skill comes in making a bot that isn't superhuman. Personally, I think the bots that are in Teamfortress classic are darn good.
Grand Turismo is a Simpler Task (Score:2)
I think it's a matter of time before AI becomes better than human drivers, but Grand Turismo is a much simpler task than the real world, or even more advanced racing sims like i-racing.
In the real world, you can't use the exact same inputs lap after lap and expect the same result because the environment is changing around you. The tires are getting warmer/colder and degrading over time (in a series like F1, you also have to use different tire types over the course of the race). The track itself is changing
Why AI? (Score:2)
Regular algorithmic drivers are already impossible to beat in racing game. Always has been. It is just too easy to make them drive on the right curve, right speed.
Re: (Score:2)
Does the AI still cheat ... (Score:2)
Does the AI still cheat with rubberbanding or will this actually be turned off?
Will the AI get penalties when they blindly ram into players?