Can Humans Learn From AI in Racing?

The most important benchmark is missing in action:

Will AI-learning be able to benefit and improve on real track, still respecting virtues of driving behavior to competitors, track lines, etc.??

Otherwise I'm afraid folks will return calling our sim world 'just another video game'.

Been real track competing late 80ies/start 90ies, and for me real track and sim is different worlds, but still closer than the arcade games.

AI learning should NOT let us turn into arcade gaming style, that will be probited on real track.

Today? Learn how to cheat, yah.

I think humans can learn from racing against AI that is as fast or faster than they are, but they cannot learn from AI that violates track limits, such as what I observed in the Sophy video I saw.

Sure, the time was fast but the ability of the computer to make corrections in steering, braking and acceleration much, much faster than a human made the lap simply impossible. The behavior of the AI in setting that lap should not be celebrated, but viewed as a failure as it failed to mimic a human.

When Sophy can set a fast lap using the very same limitations of a human then I would call it a success.

Where AI opponents have helped me is in preparing for races. For me, specifically, running a practice race prior to the iRacing Daytona 24 this year helped me identify lines which I could use to protect my position, or even pass. In the actual race I was able to use a line I found going into T1, which allowed me to protect my position at the beginning of my stint, and later it allowed me to make passes into T1 that helped our team gain positions.

Identification of the T1 line, and another line after the bus-stop and into T1, later allowed our team to protect our position from P1 (we were P2 and P1 was trying to lap us), for over an hour. P1 was faster than us on the oval but we were faster in the infield. Our hope was that P1 would try to overtake going into T1 on the only line he had, which was an untenable line due to the line we were able to exploit, and have an off that could put him into the wall. P1 did, indeed have to go off-line but not enough that we could gain a "huge" advantage on him.

Yet...due to his inability to make a safe pass, we were able to hold him off, remain on the lead lap and with less than ten minutes left the P1 driver disconnected. Unfortunately, we were ultimately unable to pass him while they were in the pits, but instead of winning by nearly a lap, P1 crossed the line only a few seconds ahead of us.

IMO, this is how AI can be beneficial to sim racers.

2112 said:

Sure, the time was fast but the ability of the computer to make corrections in steering, braking and acceleration much, much faster than a human made the lap simply impossible. The behavior of the AI in setting that lap should not be celebrated, but viewed as a failure as it failed to mimic a human.

When Sophy can set a fast lap using the very same limitations of a human then I would call it a success.

Click to expand...

Sophy is actually technically more limited than humans in terms of inputs, capped at 10 Hz, while humans in theory can achieve 60 Hz. And Sophy actually operates the wheel rather than just sending directions to the game without going through a peripheral, people present during the last Salzburg live event say during the Sophy demo there, all Sophy cars were affected a rig and you could see the steering wheels move on their own. And apparently, Sophy makes very smooth inputs and very small corrections, if at all.

This has been going since 2019.​

The Indy Autonomous Racing Challenge at CES 2022

Luciano Santos Filho said:

Lol, this move by Hizal is basically the sim version of Hungary '86

Click to expand...

Piquet's pass on Senna is the one you mean I think? Well, if the AI can also teach manners and composture to some humans, I'm all for it!

I’m surprised this is even a question. AI supremacy in chess was solved long ago, and the human players have been learning from the AI ever since, to up their games against other humans.

Now simracing is a much harder analogue-type problem vs the discrete nature of chess moves, so it was always going to take longer. But it will happen - perhaps already has with Sophy. It’s only a matter of processing resources and good AI tools.

I suspect that dialed-down AI opponents will at some point be a far better experience than online against humans, with all the positives and none of the negatives. There will always be a role for competition of course - human nature is to test ourselves against each other, but many of us get our competitive kicks from real life (me included) and will stick with the bots in our hobby.

Yes we can (ok it was easy), plus I would say you MUST learn from ai trajectory and behavior to know how to beat it : this is what most people with a (working) brain do Especially when you discover a track for the first time, following AI train is the best to learn the curves and brake points.

As I see it humans are in danger of soon getting constructed by AI.

Fun aside, we all are already programmed to behave and do our work in a way that it suits the needs of AI/software. The design of a progam, the working software in the bureau defines what we can do and what not , and how we must do it. Like language structure,ks this feeds also back on our way how we think, feeds back on our creativity, or lack of, and it makes individuality a mass-produced quality that is not individual at all.

M D Gourley said:

Usually find myself following the AI when racing on a new track, for me it is a great way of learning.

Click to expand...

Yep, I do this too. Set AI skill at a low level and treat it like an open track day. Follow them around at a relaxed pace, learn the track, racing line, breaking points, etc......it's really fun actually

AI should learn from us instead visa versa

I think there's an opportunity overlooked where a sim with AI can teach you to drive better. I haven't seen a virtual coach yet. It looks at how you drive and gives feedback. For example, that you always turn in too early or too late in a bend, brake too early or too late or accelerate too early or too late when exiting the bend. I think it's really easy for a computer to be able to see that and give you feedback. Only then would we really be able to learn from a computer.

Yes Behaviour , but the DB noises are not polite

If the grass and tire physics would be right, you never would have made that corner.
With all the load on the outer wheels that should have been dirty and cooled down because of the grass strip it been riding just before.

So Sony needs to address grass & dirty tires in to the physics loop.

Anyhow better AI is always welcome, would save me a lot of laps testing.

No. At least until they realize that the IA has to operate with the same limits as a human. The IA can analyze the car behaviour hundreds of times per second with a level of precission that outmatches a human by several orders of magnitude, the IA makes always a perfect choice hundreds of times per second, the IA don't have reaction times, the IA don't have input lag of any kind, the IA has infinite level of precission on it's inputs every single time millions of times per lap, the IA can modulate the pedals and steering wheel inputs orders of magnitude more accurately than a human.

So, no. The IA don't operate under the same reality as the human being, the human being can only get frustrated of not being able to even keep up with the IA in a single corner. This is as stupid as trying to beat excel doing calculations, the human can't win.

TRASGU said:

No. At least until they realize that the IA has to operate with the same limits as a human. The IA can analyze the car behaviour hundreds of times per second with a level of precission that outmatches a human by several orders of magnitude, the IA makes always a perfect choice hundreds of times per second, the IA don't have reaction times, the IA don't have input lag of any kind, the IA has infinite level of precission on it's inputs every single time millions of times per lap, the IA can modulate the pedals and steering wheel inputs orders of magnitude more accurately than a human.

So, no. The IA don't operate under the same reality as the human being, the human being can only get frustrated of not being able to even keep up with the IA in a single corner. This is as stupid as trying to beat excel doing calculations, the human can't win.

Click to expand...

GT-Alex said:

Sophy is actually technically more limited than humans in terms of inputs, capped at 10 Hz, while humans in theory can achieve 60 Hz. And Sophy actually operates the wheel rather than just sending directions to the game without going through a peripheral, people present during the last Salzburg live event say during the Sophy demo there, all Sophy cars were affected a rig and you could see the steering wheels move on their own. And apparently, Sophy makes very smooth inputs and very small corrections, if at all.

Click to expand...

Yes, you can learn from watching/following AI cars, just as you can from watching/following human controlled cars. Learn little bits from as many sources as possible (including what not to do) and add to your driving anything that suits your style while improving your lap time and consistency.

As for learning cheating from AI - actually it teaches us a lesson that every F1 driver and team knows. Push the limits of every rule, exploit track limits until warned (then complain that you haven't) and use every loophole in the regulations. Sophy's lap around the Nordschleife was unrealistic and ridiculous - it couldn't be done in the real world, but within the context of the game, all it did was exploit the limitations of the game's physics as well as the game's lax interpretation of track limits. Much the same as humans do in iRacing. I don't like it, but until they close the loopholes, it's not cheating.

GT-Alex said:

Sophy is actually technically more limited than humans in terms of inputs, capped at 10 Hz, while humans in theory can achieve 60 Hz. And Sophy actually operates the wheel rather than just sending directions to the game without going through a peripheral, people present during the last Salzburg live event say during the Sophy demo there, all Sophy cars were affected a rig and you could see the steering wheels move on their own. And apparently, Sophy makes very smooth inputs and very small corrections, if at all.

Click to expand...

Humans can't achieve 60 input corrections per second, that's just marketing pr statement bs based in probably some misleading fact interpreted in a way favourable to their interests.

It is easy to put to their claim to test with real life data: just do a lap while recording telemetry, now in the middle of a straight push the clutch so the car can keep moving and now press and depress the throttle as fast and as many times as you can in a second, finish the lap and now watch in the telemetry how many times you did a cicle of press and depress. Hint: even 10 times per second isn't achievable at all, while the IA can even modulate multiple inputs to inhumane accuracy 10 times per second.

Now think about being capable to also do the same with the brake pedal and steering: you can't, the IA can because it has such a processing speed that it can multitask like a human never could, the throttle inputs can be as acurate as 1/65536 of the throttle complete range (16 bit, the maximum that the USB HID protocol allows).

And not only that, the thottle position progression can be as smooth as it gets, try to do 4 throttle modulations in 1 second and now watch in the telemetry how imprecise your inputs are. Now imagine all the driving inputs in every axis done simultaneusly with the same level of accuracy and repetitivity all lap long without a single positioning mistake.

And now the elephant in the room, the IA still can think orders of magnitude faster than any human, the IA can make incredibly accurate predictions of the car behaviour (it is a computer making math calculations after all, the human is just trying to predict what could happen), the IA can perceive changes in car behaviour orders of magnitude smaller than a human being, the IA can gather data from the car behaviour instantly, the IA don't have reaction times.

The human when is braking and a wheel is about to start to lock needs some time to recognize the situation, more time to think about if he needs to make an input to correct it, more time to think about how much, and more time to send the signal to the muscles and to them overcoming the resistance of movement of the pedal, and then the inputs are inaccurate so the pedal is going to end up overshooting or undershooting the value that the human had in mind, so now he needs to compensate for the error again.

The IA simply can recognize the situation up to 1 physics calculation cicle before a wheel ends up locking up under braking, and still has time to spare to make a correction with an inhumane level of accuracy and smoothness with no delay at all.

This IA is a fascinating conceptual experiment, and with some small changes in the algorithm so the IA would ended up respecting the reasonable track limits i could be even a tool for find the ideal setup of a car and best line around a track automatically. It is also fascinating from the point of view of visualizing how a perfect ideal lap would look like.

A human can't equal the performance of a computer/machine, we invented them to overcome our own human limitations. So if we want to use them as a learning tool, then those computers/machines are going to need to have the same limitations as a human being.

As long as Sophie can be turned on and off at will and her maneuvers fall within the bounds of the rules, she's welcome, as she can help make the aliens even more alien.

Although I really don't care; I'm going to be as clumsy at the wheel as ever, Sophie or no Sophie...

wish it was on pc last gran turismo game i played was gt5 on the ps3 back when i was like 6 or 7 with my little brother from 2011 to 2013 before my ps3 bricked itself but cant justify spending basically a grand for a ps5 controller gt7 and ps plus just for 1 game :/

Beezer215 said:

Anyone want to buy this horrible mess from me? When are they going to start offering refunds for GT7?

Click to expand...