AI Can Now Also Outrace Human Champs in the Videogame 'Gran Turismo' (scientificamerican.com) 40
Scientific American reports:
To hurtle around a corner along the fastest "racing line" without losing control, race car drivers must brake, steer and accelerate in precisely timed sequences. The process depends on the limits of friction, and they are governed by known physical laws — which means self-driving cars can learn to complete a lap at the fastest possible speed (as some have already done). But this becomes a much knottier problem when the automated driver has to share space with other cars. Now scientists have unraveled the challenge virtually by training an artificial intelligence program to outpace human competitors at the ultrarealistic racing game Gran Turismo Sport. The findings could point self-driving car researchers toward new ways to make this technology function in the real world.
Artificial intelligence has already conquered human players within certain video games, such as Starcraft II and Dota 2. But Gran Turismo differs from other games in significant ways, says Peter Wurman, director of Sony AI America and co-author of the new study, which was published this week in Nature. "In most games, the environment defines the rules and protects the users from each other," he explains. "But in racing, the cars are very close to each other, and there's a very refined sense of etiquette that has to be learned and deployed by the [AI] agents. In order to win, they have to be respectful of their opponents, but they also have to preserve their own driving lines and make sure that they don't just give way."
To teach their program the ropes, the Sony AI researchers used a technique called deep reinforcement learning. They rewarded the AI for certain behaviors, such as staying on the track, remaining in control of the vehicle and respecting racing etiquette. Then they set the program loose to try different ways of racing that would enable it to achieve those goals. The Sony AI team trained multiple different versions of its AI, dubbed Gran Turismo Sophy (GT Sophy), each specialized in driving one particular type of car on one particular track. Then the researchers pitted the program against human Gran Turismo champions. In the first test, conducted last July, humans achieved the highest overall team score. On the second run in October 2021, the AI broke through. It beat its human foes both individually and as a team, achieving the fastest lap times....
"The lines the AI was using were so tricky, I could probably do them once. But it was so, so difficult — I would never attempt it in a race," says Emily Jones, who was a world finalist at the FIA-Certified Gran Turismo Championships 2020 and later raced against GT Sophy.... "Racing, like a lot of sports, is all about getting as close to the perfect lap as possible, but you can never actually get there," Jones says. "With Sophy, it was crazy to see something that was the perfect lap. There was no way to go any faster."
The article notes that Sony AI is now working with Gran Turismo's developer (the Sony Interactive Entertainment subsidiary Polyphony Digital) to potentially incorporate a version of their AI into a future update of the game. "To do this, the researchers would need to tweak the AI's performance so it can be a challenging opponent but not invincible..."
Artificial intelligence has already conquered human players within certain video games, such as Starcraft II and Dota 2. But Gran Turismo differs from other games in significant ways, says Peter Wurman, director of Sony AI America and co-author of the new study, which was published this week in Nature. "In most games, the environment defines the rules and protects the users from each other," he explains. "But in racing, the cars are very close to each other, and there's a very refined sense of etiquette that has to be learned and deployed by the [AI] agents. In order to win, they have to be respectful of their opponents, but they also have to preserve their own driving lines and make sure that they don't just give way."
To teach their program the ropes, the Sony AI researchers used a technique called deep reinforcement learning. They rewarded the AI for certain behaviors, such as staying on the track, remaining in control of the vehicle and respecting racing etiquette. Then they set the program loose to try different ways of racing that would enable it to achieve those goals. The Sony AI team trained multiple different versions of its AI, dubbed Gran Turismo Sophy (GT Sophy), each specialized in driving one particular type of car on one particular track. Then the researchers pitted the program against human Gran Turismo champions. In the first test, conducted last July, humans achieved the highest overall team score. On the second run in October 2021, the AI broke through. It beat its human foes both individually and as a team, achieving the fastest lap times....
"The lines the AI was using were so tricky, I could probably do them once. But it was so, so difficult — I would never attempt it in a race," says Emily Jones, who was a world finalist at the FIA-Certified Gran Turismo Championships 2020 and later raced against GT Sophy.... "Racing, like a lot of sports, is all about getting as close to the perfect lap as possible, but you can never actually get there," Jones says. "With Sophy, it was crazy to see something that was the perfect lap. There was no way to go any faster."
The article notes that Sony AI is now working with Gran Turismo's developer (the Sony Interactive Entertainment subsidiary Polyphony Digital) to potentially incorporate a version of their AI into a future update of the game. "To do this, the researchers would need to tweak the AI's performance so it can be a challenging opponent but not invincible..."
Aimbots (Score:3, Insightful)
Re: Aimbots (Score:4, Insightful)
Re: (Score:3)
Not intelligently (Score:4, Insightful)
Artificial intelligence has already conquered human players within certain video games, such as Starcraft II
It didn't "conquer" human players with intelligence, it clicked faster than human players, and we've known for years that you don't need intelligence to do that.
Google's Starcraft program was more advertising than research.
Re: (Score:3)
It didn't "conquer" human players with intelligence, it clicked faster than human players, and we've known for years that you don't need intelligence to do that.
Exactly. I’ll believe an “AI” conquers human players when not only is it denied all internal game states, but must acquire its data through two crappy stereo cameras, two low fidelity microphones, a crude IMU, and haptic feedback through a humanoid form with over a hundred degrees of freedom that’s the only interface to the game.
It did use video cameras (Score:1)
https://www.ai.sony/publicatio... [ai.sony]
The last author of this paper from 2021 is the same as the author of the 2022 Nature paper.
If you want an explanation of how it is done, see my comparison to Amazon's Deep Racer, which uses a crappy camera in a $500 robot, or a virtual camera in a simulation:
https://slashdot.org/comments.... [slashdot.org]
Re: Not intelligently (Score:3)
Re: (Score:2)
Actually if you develop that idea more, it could actually be done, since Starcraft is customizable.
The difficult part is how to allow players to define the behavior, since most players are not programmers.
Re: (Score:2)
You don't read articles do you? AlphaStar restricted the number of mouse clicks it could register to 22 non-duplicated actions every five seconds of play, to align it with standard human movement. In this Gran Turismo AI too, the speed of reaction was slowed down to be human-like.
Re: (Score:2)
AlphaStar restricted the number of mouse clicks it could register to 22 non-duplicated actions every five seconds of play, to align it with standard human movement.
Yeah that's a lie. The lie is that it aligns with standard human movement.
I watched the AlphaStar matches, and saw it do super-human micro, and commit hilarious strategic mistakes a human would never make.
Re: (Score:2)
Ok, fair enough. I take it back.
Re: (Score:2)
Btw a serious improvement in the computer micro is that the computer is extremely precise in all its clicks. For example, it can move a unit out of the way a fraction of a second before it gets hit.
Re: Not intelligently (Score:2)
IRL is the challenge. (Score:2)
Computer can beat humer at computer car driving. DUH.
Call us back when a computer can beat a human driving on a real road (not a perfect racetrack in a virtual world).
Re: (Score:2)
The training strategy for AlphaStar went through a few iterations. When using their population-based “AlphaStar League” approach they showed that the AI did in fact use a lot of macro strategy, and it’s micro optimizations did not exceed those of professional human players.
In its games against TLO and MaNa, AlphaStar had an average APM of around 280, significantly lower than the professional players, although its actions may be more precise. This lower APM is, in part, because AlphaStar starts its training using replays and thus mimics the way humans play the game. Additionally, AlphaStar reacts with a delay between observation and action of 350ms on average.
https://deepmind.com/blog/arti... [deepmind.com]
I think it’s quite fascinating, actually.
Re: (Score:2)
280 APM isn't a human limit because of EAPM. A human can't make 280 accurate clicks in a minute.
it’s micro optimizations did not exceed those of professional human players.
Did you watch the games? This is outright false.
Re: (Score:2)
AI in the Starcraft program has artificial limits on clicks per minute. To level ground a bit. It won by being smarter and more efficient not by clicking faster.
Re: (Score:2)
To level ground a bit.
It clearly didn't work, they didn't do it correctly as I've mentioned in other comments. It was obvious if you watched the games.
It won by being smarter and more efficient not by clicking faster.
Why do you think this? Did you even watch the games?
Future road rage (Score:2)
Re:Future road rage (Score:5, Funny)
Re: Future road rage (Score:2)
No video (Score:3)
Not a single video in any of those useless links.
don't waste your time (Score:2)
don't waste your time looking at computers playing video games.
They did you a favor by not adding links.
Champs? How about Chimps? (Score:2)
https://www.youtube.com/watch?... [youtube.com]
In the real world? (Score:2)
Seriously? Playing a videogame about cars and driving a car in real life are two entirely different problems.
Re: (Score:2)
can it beat the dup reposting on Slashdot? (Score:1)
can the AI beat the dup reposting on Slashdot?
https://games.slashdot.org/sto... [slashdot.org]
How to beat AI (Score:1)
Just put a banana is its tailpipe [youtu.be]
Ultrarealistic? (Score:4, Insightful)
by training an artificial intelligence program to outpace human competitors at the ultrarealistic racing game Gran Turismo Sport
Since when was Gran Turismo considered "ultrarealistic"?
Re:Ultrarealistic? (Score:4, Funny)
by training an artificial intelligence program to outpace human competitors at the ultrarealistic racing game Gran Turismo Sport
Since when was Gran Turismo considered "ultrarealistic"?
When it is compared to Need For Speed and Mario Kart.
Re: (Score:3)
Re: (Score:2)
>> Since when was Gran Turismo considered "ultrarealistic"?
Since we drive real cars on real roads.
Racing games (Score:1)
Re: (Score:2)
It was not achieved because nobody cared about this BS of a computer playing computer games.
um... (Score:2)
I should hope so. Or else the programmer really sucks
Emulate the physical input (Score:1)
yeah no shit (Score:2)
when you factor out raw reaction time (where the computer obviously has terrific advantage over meatbags), auto racing is basically just solving an optimization problem of combining the interaction of the car's characteristics and the track to generate a path that minimizes lap time.
this sort of problem is what computers were invented to solve; you just replace "car" with "primary fission trigger" and "track" with "radiation channel and secondary fusion containment" and "minimizes lap time" with "maximizes
I play for the game itself, not for victory. (Score:1)