Gran Turismo Sophy: Sony AI x Polyphony Digital

  • Thread starter Magog
  • 1,719 comments
  • 195,778 views
GT Sophy is only part of it. I think we will see other Sophys soon.

Sony just aquired Haven Studios. Only 1 source but in this article it mentions Sony AI being implemented. I've also read that this AI tech was already used in enhancing graphics in the Miles Morales game.


The Sony Ai blog is really informative - https://ai.sony/blog/blog-019/

for instance, Sophy works for it,

Once an experiment begins, the agent learns to drive through trial and error. Understandably, it struggles at the start. With an hour of training, the agent is able to complete laps, albeit slowly and carefully. This progression continues, and after eight hours of training, the agent is completing laps at times similar to the default AI built into the game. After a day, the agent is near the top 10 percent quantile of the best human players. It takes another eight to 10 days of training to shave off the last few seconds and milliseconds and reach consistent superhuman performance without any racing penalties.
 
Last edited:
It learned to do that itself after training using a huge array of very expensive hardware for hundreds/thousands of hours. That's why it's not particularly exciting, because it's still a long way from being reality in a regular $70 video game.

If this was just the regular AI programmed to be that fast without any machine learning people might be more excited as that'd mean it'll be easily a available soon but right now it's essentially vaporware.

There is no clear roadmap laid out for how this system becomes reality for a consumer, not just a video on YouTube. It could and is likely years away, with vague decelerations that they want to bring it to the game.
The smarted computers in the world and its AI uses machine learning. Including ai in Tesla and other automated real cars. The ai your looking for is limited.
 
The smarted computers in the world and its AI uses machine learning. Including ai in Tesla and other automated real cars. The ai your looking for is limited.
The AI I'm looking for is one that actually exists in a game. I don't really care how they achieve it, but my point was what we're seeing here is still a long way from a final product. It's still just Sony AI hooking into GTS via their API. It's not something anyone can just fire up a regular build of GTS and race against.

I'll be excited when PD show us a video of a GT game with real conpetive AI just working as a regular feature in the game anyone can use. We seem to still be a long way from that.
 
Everything the AI did in the video was legal to the current track limits. Meaning it currently isn’t cheating and If you were fast enough (impossible) you could replicate this.
bro all 4 wheels went off the track over the rumble strips

Cheating
 
even if it still finds exploitable tracks doing that.
Yes that's exactly what PD is doing, exploiting their own game :banghead: this AI is a joke if this is the end result... but fortunately it probably isn't the end result cause who needs an AI that not even the fastest player in the world can beat...
Almost seems like PD is trying to test how you can get the best possible time in their game including exploiting track limits, but for what purpose??? :S I have no clue...

bro all 4 wheels went off the track over the rumble strips

Cheating
I think he means legal within the track limits of the game, the real life track limits and the game track limits don't line up 100% (that's obvious unless they've really disabled penalties etc)
 
Last edited:
On the Sony AI blog they give details on the process. It's all within track limits. The calculations are based on machine learning with those limits. Trial and error is important to the process.

To get to A+ level on 1 track takes 10-12 days or 240 hours. If you put in 240 hours of practice you could achieve the same, theoretically.
 
On the Sony AI blog they give details on the process. It's all within track limits. The calculations are based on machine learning with those limits. Trial and error is important to the process.

To get to A+ level on 1 track takes 10-12 days or 240 hours. If you put in 240 hours of practice you could achieve the same, theoretically.
Not if the inputs from the AI are inhuman, something very common with TAS. They can input frame perfect inputs that a human never could. Especially with a wheel, a human couldn't turn their wheel to a perfect degree every corner.
 
Not if the inputs from the AI are inhuman, something very common with TAS. They can input frame perfect inputs that a human never could. Especially with a wheel, a human couldn't turn their wheel to a perfect degree every corner.
It starts from zero. It takes approximately 1 hour to lap cleanly, 8 hours for Sophy to match the in-game AI.

Sophy is a learning simulator. It may eventually develop the skill to be perfect but it's simulating human learning.
 
It starts from zero. It takes approximately 1 hour to lap cleanly, 8 hours for Sophy to match the in-game AI.

Sophy is a learning simulator. It may eventually develop the skill to be perfect but it's simulating human learning.
Not sure how that relates to what I said. You said that with enough practice a human could match Sophy theoretically, I'm saying that it wouldn't be the case if the AI has learnt to enter inputs at a level of speed and accuracy that no human ever could.

Again this is like TAS, there are many games where certain tricks/moves can only done by a bot, not a human. A human can't press buttons 50 times in a row exactly 13ms apart each time, for example. Maybe Sophy is doing something similar to manage to drive on the grass and not spin, where a human couldn't. We don't know.
 
We don't know.
We do, actually. It's discussed in the Nature article that Sophy operates at 10 Hz.

Nature
The core actions of the agent were mapped to two continuous-valued dimensions: changing velocity (accelerating or braking) and steering (left or right). The effect of the actions was enforced by the game to be consistent with the physics of the environment; GT Sophy cannot brake harder than humans but it can learn more precisely when to brake. GT Sophy interacted with the game at 10 Hz, which we claim does not give GT Sophy a particular advantage over professional gamers or athletes.
Nature
Our agent ran asynchronously on a separate computer and communicated with the game by means of HTTP over wired Ethernet. The agent requested the latest observation and made decisions at 10 Hz (every 100 ms). We tested action frequencies from 5 Hz to 60 Hz and found no substantial performance gains from acting more frequently than 10 Hz. The agent had to be robust to the infrequent, but real, networking delays. The agent’s action was treated the same as a human’s game controller input, but only a subset of action capabilities were supported in the GT API. For example, the API did not allow the agent to control gear shifting, the traction control system or the brake balance, all of which can be adjusted in-game by human players.
Nature
GT Sophy took actions at 10 Hz, which was sufficient to control the car but much less frequent than human actions in GT. Competitive GT drivers use steering and pedal systems that give them 60 Hz control. Whereas a human can’t take 60 distinct actions per second, they can smoothly turn a steering wheel or press on a brake pedal. Extended Data Figure 2b, c contrasts GT Sophy’s 10-Hz control pattern to Igor Fraga’s much smoother actions in a corner of Sarthe.
 
Last edited:
We do, actually. It's discussed in the Nature article that Sophy operates at 10 Hz.
A BOT operating at 10Hz has a major advantage over a human using a controller that inputs 60Hz. That BOT can utilize all 10Hz 100% of the time, meaning that it's reacting to everything happening 10 times each second, the entire time it's active. A human cannot react at 10Hz. That BOT can go full rotation left to right instantly (10 times each second). Can a human with a wheel do that?
 
Can a human with a wheel do that?
We already know a well implemented A.I. is better than the best human drivers. The real challenge is going to be figuring out how to massage artificial intelligence into what I would call artificial competence, where bots have character traits and flaws, make errors that humans would. We all want an A.C. that would allow us to adjust Sophy to suit our own skill levels, roughly, so we can have a competent race, not lose every race.
 
We already know a well implemented A.I. is better than the best human drivers. The real challenge is going to be figuring out how to massage artificial intelligence into what I would call artificial competence, where bots have character traits and flaws, make errors that humans would. We all want an A.C. that would allow us to adjust Sophy to suit our own skill levels, roughly, so we can have a competent race, not lose every race.
That's what we all want. It was the point of my original post on the topic. Who cares about an AI ripping through a track by itself. I want to see how it performs in a real world situation with other competitors on the track. My reply to Jordan was about the article saying a human with a 60Hz controller, somehow has an advantage over Sophy using 10Hz input. Which is completely false. Without proof that nothing was altered to the physics or handling (real time telemetry), that Sophy video is useless.
 
That BOT can go full rotation left to right instantly (10 times each second). Can a human with a wheel do that?
The Bot can't do that. It says in the text Jordan quoted that "the agent’s action was treated the same as a human’s game controller input", the game controller input is smoothed and speed sensitive.
 
The Bot can't do that. It says in the text Jordan quoted that "the agent’s action was treated the same as a human’s game controller input", the game controller input is smoothed and speed sensitive.
Take that article with a grain of salt. Show me the telemetry. That car went from full Dukes of Hazard, to Michael Schumacher in the same corner without flinching. I'm not buying it.
 
The think Sony need to sort out is getting more PlayStation 5 in I still can’t get it
If you're after one sign up to stock check alerts, I'm getting alterted to available stock on an almost daily basis, I just don't want a PS5 yet and haven't bothered to unsubscribe.
 
Take the Nature article for peer review with a grain of salt... says random person of the internet.
Sorry. Meant to say "I take that article with a grain of salt". You believe whatever hype you want to about Sophy. All I said was show me proof of what you did with telemetry and put other cars on the track.
 
Sorry. Meant to say "I take that article with a grain of salt". You believe whatever hype you want to about Sophy. All I said was show me proof of what you did with telemetry and put other cars on the track.
That's the point of the article. They publish their findings. Now experts review it.

Taking it with a grain of salt is redundant. You're expecteded to approach it with skeptisim. Then read the article. Not avoid reading the article. That's not taking it with a grain of salt, that just avoiding reading the article.
 
That's the point of the article. They publish their findings. Now experts review it.

Taking it with a grain of salt is redundant. You're expecteded to approach it with skeptisim. Then read the article. Not avoid reading the article. That's not taking it with a grain of salt, that just avoiding reading the article.
The expression of "take it with a grain of salt" is thinking something is most likely untrue or incorrect. I read the article that they claim is all facts, and I think a lot of it is made up BS, or they are leaving out key information. Why are you expected to approach it with skepticism?

take something with a grain of salt

idiom US (UK take something with a pinch of salt)

to not completely believe something that you are told, because you think it is unlikely to be true:
You have to take everything she says with a grain of salt, because she tends to exaggerate.
 
Last edited:
The expression of "take it with a grain of salt" is thinking something is most likely untrue or incorrect. I read the article that they claim is all facts, and I think a lot of it is made up BS, or they are leaving out key information. Why are you expected to approach it with skepticism?

take something with a grain of salt

idiom US (UK take something with a pinch of salt)

to not completely believe something that you are told, because you think it is unlikely to be true:
You have to take everything she says with a grain of salt, because she tends to exaggerate.
I appologize. It sounded like you didn't read it.
 
This is what Sophy's telemetry looks like compared to Igor Fraga.

From the Nature article:

Screen Shot 2022-07-14 at 2.20.53 PM.png
 
Back