Gran Turismo 7 [PS4, PS5]

[vspectra]

"What really stands out is GT Sophy's performance against human drivers in a head-to-head competition. Far from using a lap-time advantage to outlast opponents, GT Sophy simply outraces them. Through the training process, GT Sophy learnt to take different lines through the corners in response to different conditions. In one case, two human drivers attempted to block the preferred path of two GT Sophy cars, yet the AI succeeded in finding two different trajectories that overcame this block and allowed the AI's cars to pass (Fig. 1)."

 

[chris1515]

https://sonyai.github.io/gt_sophy_public/

Some videos

EDIT: From a Pdf someone have access via his university:

Sophy maintained two lists of their state features: one for cars in front
of the agent and one for cars behind. Both lists were ordered from closest
to farthest and limited by a maximum range.
We trained GT Sophy using a new deep RL algorithm we call quantile
regression soft actor-critic (QR-SAC). This approach learns a policy
(actor) that selects an action on the basis of the agent's observations
and a value function (critic) that estimates the future rewards of each
possible action. QR-SAC extends the soft actor-critic approach36,37 by
modifying it to handle N-step returns38 and replacing the expected value
of future rewards with a representation of the probability distributions
of those rewards39. QR-SAC trains the neural networks asynchronously;
it samples data from an experience replay buffer (ERB)40, while actors
simultaneously practice driving using the most recent policy and continuously
fill the buffer with their new experiences.

 

[snc]

"What really stands out is GT Sophy's performance against human drivers in a head-to-head competition. Far from using a lap-time advantage to outlast opponents, GT Sophy simply outraces them. Through the training process, GT Sophy learnt to take different lines through the corners in response to different conditions. In one case, two human drivers attempted to block the preferred path of two GT Sophy cars, yet the AI succeeded in finding two different trajectories that overcame this block and allowed the AI's cars to pass (Fig. 1)."

yah but also made some skatchy lines with drifting, with high tyre wear they would kill it in few laps but best was using pit lane in le mans haha (how it was not penalty ;d)

https://twitter.com/i/web/status/1491469951173120005

 

[PSman1700]

One day multiplayer will be just as good against AI/bots. If that translates to FPS and other genres that is. Atleast you'd never have to worry about empty servers again

 

[Arwin]

There are also real race drivers who race on GTS, David Perel is South African who followed his dream and up to last year was racing GT3 Ferrari's. He started out with GTS just to keep his reactions and driving skills up, then eventually migrated to iRacing and ACC and then even made a comparison video of Real Racing vs ACC vs GT Sport vs iRacing at Spa, one of his favourite tracks.

He also had a hand is getting ACC where it is now as he and other GT3 drivers gave input to the developers which helped improve the game.

I anticipate his videos for GT7 and hopefully get a comparison video for that too.

Yeah I saw that one too. He has a bit lower opinion of GT than others but then he was pleasantly surprised.

 

[Unknown Soldier]

This was very interesting. First video out of them all where I see the human lead the ghost for close to half the lap (other than the Le Mans race). Still, the bot taking corners where no human would, driving two wheels off on the grass which is extremely hard to catch as a human. If they can limit it from doing so, I would like to see the time Sophy would then do. We basically see the same thing happen at Seaside where the bot goes two wheels off, destabilizes the car and yet is able to catch it and turn through faster than the humans.

Humans most likely can replicate it but it's not a natural line as Gallo showed in his final run versus the AI.

I'd also like to know if they are racing with strong penalties. We usually race with strong penalties and I can see where possibly the AI would get penalties. That first corner in Emily's video on her last run you can see the AI has three wheels off track unless the pink is still considered track(which I believe it's not).

Usually in Daily race if you have three wheels off you get a penalty.

 

[chris1515]

https://arxiv.org/abs/2103.14666

https://arxiv.org/abs/2111.06449

https://ieeexplore.ieee.org/abstract/document/9372847

 

[Karamazov]

I wonder if the AI can get frustrated and rage quits if it loses.

 

[PSman1700]

I wonder if the AI can get frustrated and rage quits if it loses.

If they have different tiers, otherwise it cant even loose

 

[Karamazov]

AI vs AI !

 

[Unknown Soldier]

I wonder if the AI can get frustrated and rage quits if it loses.

No, it just seems to bump you off the road if you are in front. In that race versus the Japanese team, the one AI(Blanque I think it was) bumped Ryo off and he lost around 5 secs time getting back on. They gave the AI 1 second penalty. Great penalty system they have for the competition. At least give the AI a penalty that fits the crime i.e. 5sec or 10sec

Mind you it's not just AI racers. Humans are just as culpable for penalties.

Asses like KIMI Velocini should get permanent bans. Not even a penalty during this race. They need to have a look at the system again.

 

[Arwin]

It’s a matter of what data you feed it and what your goal is. If that’s what you want, it’s pretty easy to achieve. That is also why they have different training goals - be good at time trials, be good at racing/overtaking etc.

 

[RobertR1]

"What really stands out is GT Sophy's performance against human drivers in a head-to-head competition. Far from using a lap-time advantage to outlast opponents, GT Sophy simply outraces them. Through the training process, GT Sophy learnt to take different lines through the corners in response to different conditions. In one case, two human drivers attempted to block the preferred path of two GT Sophy cars, yet the AI succeeded in finding two different trajectories that overcame this block and allowed the AI's cars to pass (Fig. 1)."

lol the "human player" line there is so shit.

 

[Unknown Soldier]

lol the "human player" line there is so shit.

Yeah, you would hug the inside and make the other racers go the long way around. Might not be the fastest way but at least your opponent should still be behind you if you had good traction out of the corner.

 

[snc]

 

[PSman1700]

I thought ray tracing was in-game?

 

[London Geezer]

I thought ray tracing was in-game?

nope

 

[Arwin]

No everywhere but in the actual races (4k60) in replays etc. Although we don’t have a definitive word on the supported graphics modes yet.

 

[Pete]

Someone in the YT comments noted that the red car on the left at 12:33 is actually 2D. Good catch, and a funny optimization for essentially a loading screen.

 

[Pete]

This was very interesting. First video out of them all where I see the human lead the ghost for close to half the lap (other than the Le Mans race). Still, the bot taking corners where no human would, driving two wheels off on the grass which is extremely hard to catch as a human. If they can limit it from doing so, I would like to see the time Sophy would then do. We basically see the same thing happen at Seaside where the bot goes two wheels off, destabilizes the car and yet is able to catch it and turn through faster than the humans.

Humans most likely can replicate it but it's not a natural line as Gallo showed in his final run versus the AI.

Sophy’s corner entry under heavy braking looked incredible (when it wasn’t hip-checking the humans into oblivion), at least in the Japanese team video. As you said, I wonder if humans can learn from Sophy’s lines or if that kind of (w)reckless abandon requires reflexes and precision that we can’t reliably match.