A Gateway to the Future: Neural AI in Simracing

[paracletus]

Google/Deepmind AI got mentioned earlier - it is incredible for sure. Taught itself to wipe the floor with any chess grandmaster in no time; not even Stockfish has a chance against it now.

More to the point - could it teach itself how to race? Completely different obviously, but it is possible if the resources were put into that. Trouble is it would be unbeatable and you'd never have a chance - a bit like Lewis Hamilton in the mercedes

 

[2stains]

, I would say RF2 is leading by an example in terms of the awareness and positioning. It still has fundamental flaws with slip streaming and overtaking though.

Not to mention the on/off throttle behavior instead of easing on of the throttle through corners

 

[delpinsky]

Geoff Crammond did it best with Grand Prix 4, some 17 years ago.

Not only! He created the very first stewards moving the cars off track... thing never seen after his saga, sadly. He was ahead of his times!

 

[Bryan O'Connor]

It sounds like you are under the naive impression that a RL pro driver dont have to bother with setup knowledge.
No driver would be hired in as example F1 if he was only able to give some borderline useless info like "I have understeer in T1 and oversteer in T14".
As an example Kevin Magnussen was prefered by the McLaren team against some other junior drivers because he was able to give very precise useable feedback to the pit.
And you can be sure that even the drivers in the bottom end F1 teams is able to discuss and understand setup parameters on a reasonable level.
Otherwise they would be without a job.
Only taxi drivers can be hired without knowledge about what is under the hood.

Being able to give detailed and precise feedback is a lot different to being able to set up a car.
Magnussen might be able to detect tyre and suspension movements to a miniscule degree through a fast turn - but would he know which set of springs in the back of the garage to fit to the car for him to be able to take that corner flat?
He may eventually after being given a set previously from the engineers but he wouldn't know without prior knowledge - which is what Sims assume all players have.

 

[StefanoCasillo]

Google/Deepmind AI got mentioned earlier - it is incredible for sure. Taught itself to wipe the floor with any chess grandmaster in no time; not even Stockfish has a chance against it now.

More to the point - could it teach itself how to race? Completely different obviously, but it is possible if the resources were put into that.

yeah about the resource thing. Wikipedia suggests that DeepMind Technologies, the company acquired by Google (for $400 million.. ouch) to develop the Deepmind AI counts 700 employees as of 2017.
That's almost twice the size of the biggest studios working on the biggest racing games.
And I'd argue that having credible AI racing on the limit and reacting to an immense amount of inputs variables is way more complex than a game of chess.

 

[StefanoCasillo]

Being able to give detailed and precise feedback is a lot different to being able to set up a car.
Magnussen might be able to detect tyre and suspension movements to a miniscule degree through a fast turn - but would he know which set of springs in the back of the garage to fit to the car for him to be able to take that corner flat?
He may eventually after being given a set previously from the engineers but he wouldn't know without prior knowledge - which is what Sims assume all players have.

exactly.. it's actually waaaaay more complicated than that and we can already get a feel for that in current sims that simulate aero-heavy racecars.
He might know what kind of spring he wants, but would he know how that spring is going to impact ride heights and thus aero load in every section of a track?

 

[David Wright]

Why I absolutely do not believe in the PC ressource thing is because good old NR2003 used the full physics for the AI cars!
You could even create setups for the AI drivers just like you could for yourself.

.
Do you have a source for this? I am curious because GPL certainly doesn't use the full physics for the AI cars - nothing like. Also sims which don't use "full" physics still have AI setups just like you could do for yourself.

 

[NIXXXXX]

Live for Speed had learning AI 15+ years ago. And it used the same physics as the player as far as I remember.

LFS was way ahead of its time.

 

[David Wright]

LFS was way ahead of its time.

But the AI was crap in LFS.

 

[BrunoBæ]

Magnussen might be able to detect tyre and suspension movements to a miniscule degree through a fast turn - but would he know which set of springs in the back of the garage to fit to the car for him to be able to take that corner flat?

Maybe or maybe not.
There is probably a reason drivers drive the car and engineers get oily hands adjusting your springs
But we are discussing the communication between drivers and engineers.
Not if there is a division of labor in a team - in real life.

My point is that its pretty naive to believe that just because you in a sim dont have a clue and/or dont want to mess with setting your virtual car up - then its in accordance with how it is in the real racing world.

And you can be sure that even the drivers in the bottom end F1 teams is able to discuss and understand setup parameters on a reasonable level.

 

[gambolmaker]

You said in the opening post that the AI will be much more human. But all I read is about how they react to humans. It doesn't sound like the AI will make human mistakes. Those would and will have to still be coded... even if you have talent files, like older sims had, and there's a percentage for mistakes in there, and some random generator tells the AI "now make a mistake", what will that mistake be? Will it make any sense? How will the AI know what a realistic mistake at that specific point on the track is?
What it sound like is that you'll get robots who might be able to react to humans, but who will be perfect on every lap, occasionally making weird mistakes out of nowhere. And nowhere does it say how they will race against each other, which for me is also a big, big part of realistic AI.

I recently did a race in AMS. I was battling another car, got by him. Then he cut a corner to overtake me again. Actually cut a corner. This wasn't a human, this was the AI. I've had AMS' AI make moves to one side and overtake me on the other side when I block them. I'm actually pretty happy with the state of the AI at the moment, at least in some games.

 

[BrunoBæ]

.
Do you have a source for this? I am curious because GPL certainly doesn't use the full physics for the AI cars - nothing like. Also sims which don't use "full" physics still have AI setups just like you could do for yourself.

Hey David you know probably that your question is some kind of a rhetoric one - right?
You know a question that normally dont need or dont really have an answer with new content inside.
Because none of us will ever get real inside information of what was/is inside the NR2003 code - so we can only use the rumours and hints that have been fooling around since 2003.
And even if I actually could quote Papyrus or as example Kaemmer of this then nobody would know if it was just hype.

Your question is reasonable but my answer is as it is.

CatsAreTheWorstDogs: To the best of my knowledge the AI code in GPL and NR2003 is 2 different things.

 

[delpinsky]

yeah about the resource thing. Wikipedia suggests that DeepMind Technologies, the company acquired by Google (for $400 million.. ouch) to develop the Deepmind AI counts 700 employees as of 2017.
That's almost twice the size of the biggest studios working on the biggest racing games.
And I'd argue that having credible AI racing on the limit and reacting to an immense amount of inputs variables is way more complex than a game of chess.

Since Google created Stadia, their Cloud based gaming platform: 4K@60fps HDR, surround audio and so on.. why Kunos don't have some chat with them, in order to develop a new generation of racing simulators, using their AI know-how: Assetto Corsa 2.0 - The Real AI Driving Simulator

 

[TylerDurden4321]

Hehe, they're certainly not good in recent outings that's for sure! I'd like to think the Drivatars in Forza 1 (2005) were a lot better because I was on the team that developed them but maybe my rose tints are too strong ;-) We split the problem into 3 at the time: 1) learning to control the car based on a full physics simulation 2) learning racing lines and 3) "race craft". It was the racecraft aspect we didn't crack at all and that was 100% programmed. That's still the biggest problem and I don't think we'll see any great strides there for a while TBH

No, maybe the system is not that bad... under the hood.
To me, personally, one of the big mistakes in the concept is that drivatars adapt to the set difficulty. I can get one of my fastest friends, who I can beat 1 time out of 10 (if that) on New Driver difficulty and trash him or at the Unbeatable difficulty and then still beat him, unless he's in the Top3 drivers where at least the pole driver usually drives far ahead of the field (and no qualifying to start next to him).

Also the way the campaign races are set up in the Forza games - 3 laps, start in 16th place - makes all players drive in completely unrealistic aggressive ways. Then there is the rewind option and the dynamic racing line that are used by most players, so the differences in braking points, throttle and lines through corners are minimal. Having a mode where you train your drivatar with those unrealistic equalizers turned off, putting you in random conditions (FWD, RWD, AWD, light car, heavy car, modern slicks, bias ply road tires, aero cars, low power momentum cars, open wheelers, against 1 strong opponent against many slow opponents, on different tracks, in the dry or the rain,etc.) could improve them a lot imo. As an incentive you could earn more drivatar credits the longer you have trained your drivatar. Also, the campaign races would need to be redesigned to take advantage of that.
The way the game is structured now, it's impossible to know if Drivatars are actually terrible or just terrible because they're not really given a chance to be any good.

PS. that tire degradation and temps as well as pit stops altogether are a non-issue in the current FM games is also not helping.

 

[enduser282]

So much for pc being the master race!

Its obvious AI has been neglected because of pub server racing, IP servives are very common we might as well utilize that instead of extensive & costly AI programming, pub racing good right? Uhm, depends who u ask.
Tho iracing has taken it seriously, i see ppl still have mixed feelings for it and then they are also developing AI. Why is that, they developed such a platform and still turn to AI eventually?

Advanced AI is very welcome but i didnt see anything really impressive in that motogp vid. Does anybody tell AI to crash in 1st corner, its how well its sensing the surrounding and not so much about neural networks isn't it? Faster lap times is a complete gimmick because we dont know if its using player physics and you also need for it to be adjustable. If i compete with the pros i ll never win thats no good for a game, its just hype atm.
A problem i have in rf2 is not the AI overall lap time but the fact that it goes very slow mid corner and i hit them, or i have to keep a long distance which makes it very difficult to follow close and overtake as AI ejects out of the corner with super duper grip.

Another thing is if u got mixed types of racing in a sim vs a close and controlled enviroment like moto gp championship.
Eg. a GT cuts over curbs as much as possible while OW does generally the opposite, which creates issues in current bunch of sims. While alternative racing lines may exist in the engine they might not be implemented, this is track (mod) related and in neural AI u must teach it all the same... so there is that.
I think scripting can cover most racing scenarios pretty well but if neural AI can get the developer there faster its a very strong tool.
If i slow down to tweak some ui setting and AI sits behind me then its not even about human behaviour, if anything computers should have instant reflexes and should recognize immediatly when u r off pace (after a threshold is set for the AI to read). So what i see is a sensing problem and not human like behaviour. Say u take all of Hamilton's carrier track by track, imprint it in an AI which is blind, add a human player as well as the rest of the grid and u got a grand failure.

 

[DriveClubMaster]

DrivATARS™

Those are great, aren't they... huh? ...No? ...ANYone??

Drivatars are far from perfect, very often rather dumb. Its a decent concept that cant work well. this is the new thing. Forza 7 has very annoying drivatars, have it on PC

 

[Korky]

No, maybe the system is not that bad... under the hood.

You make several really good points which come back to the compromise nature of games and the way they are produced. The producer / designer are focussed (particularly in "collector driven" games like Forza) on the game experience as a whole and have tight (marketing driven) time constraints, testing constraints etc and so as long as the AI is "good enough" and they have some buzz words to fling around for the press then it's "job done". Realistically this is the only pragmatic solution because you'll never keep everyone happy - especially these days where everyone on Reddit is a deity appointed expert on everything

I would stress that I left Forza behind after v1 so I'm not sure how much of the original system remains but I definitely see some core concepts / behaviours under the mess! What I do know is that I absolutely hated the AI in FM7 and mostly play R3E, AC(C), rF2 etc these days and mostly in solo hotlap mode because I find AI racing frustrating for the most part.