116 replies to this topic

[jjcale]

There was a thread on a self driving car made by google some time ago ... well things have move on a bit. The boffins are now working on cars that can learn tracks all by themselves (apparently) and react to changing car and track conditions.



So ...how long till we have the racing world's equivalent of Deep Blue where the computer is able to beat the best humans?

[Risil]

Whatever, they're basically all robots anyway.

[Snic]

Interesting topic.

Driving is actually a sport where a computer could excel in as there are only 3 variables - steering, throttle and brakes.

In theory a computer could always keep a car at the limit of wheelspin and locking tyres - and do this much more consistently than a human could, if the sensors are accurate enough. In a straight line with equal cars, there's no doubt a computer could out drag a human.

In theory a car would also be able to pick up when the car starts drifting faster than a human too, and hence correct faster. However the skill is in how much to correct, which is very very instinctive.

Now when it comes down to comparing it to Deep Blue is where the similarity fades. All these 'intelligent' computers are actually incredibly stupid, they simply learn from what humans do and eliminate errors along the way, but if the're ever put against something that they haven't seen before they become completely redundant. Deep Blue was made up of the memory of a million games of chess; and was one of the most powerful super-computers ever built at the time. And that's to play a game where the board, pieces and rules are always the same.

Driving in the wet, where the grip is changing at every corner, requires creativity and adaptability - something computers aren't known for. It would require huge amounts of data collection to get anywhere near that level of instinct, which I doubt would ever happen. And besides that, the processing power it would to drive at the limit is beyond anything Deep Blue do, if it ran on Windows it would probably crash () at the first corner

Edited by Snic, 10 November 2012 - 11:55.

[superdelphinus]

'instinct' is still a human reaction to data though, it's not magical

[Red17]

A computer can learn to ago around, but it will never be as effective as humans
Just look how hard it is to make a proper AI racer in a virtual environment, let alone in the real world.

[techspeed]

The problem with computer controlled cars is that they drive to the cars limit. The computer lost at Thunderhill Raceway because the human driver could drive beyond the cars limit.
http://www.bbc.co.uk...nology-20165345

[lbennie]

The problem with computer controlled cars is that they drive to the cars limit. The computer lost at Thunderhill Raceway because the human driver could drive beyond the cars limit.
http://www.bbc.co.uk...nology-20165345

no they can't. Then that wouldn't be the car's limit.

[prty]

The problem with computer controlled cars is that they drive to the cars limit. The computer lost at Thunderhill Raceway because the human driver could drive beyond the cars limit.
http://www.bbc.co.uk...nology-20165345

That's just sensationalistic, what is happening is that at the moment humans are closer to the limit than the computer. Just give it more time, and the situation will change.

And I think right now the main problem for automatic cars is to know exactly where they are, and react to external changes. The driving technique itself can easily exceed that of humans right now, just look at traction control or ABS.

Edited by prty, 10 November 2012 - 14:07.

[Aqualung]

The problem with computer controlled cars is that they drive to the cars limit. The computer lost at Thunderhill Raceway because the human driver could drive beyond the cars limit.
http://www.bbc.co.uk...nology-20165345

I'm not sure you understand what the word "limit" means.

[Lights]

Interesting topic. I can see thing going somewhere but I don't see them beating humans ever.

[Wander]

Of course they will beat humans once they find a way to program them to drive optimally all the time, but that will still take decades.

[Sausage]

Just a matter of time. Computers (software basicly) are much more reliable than humans if programmed correctly. Also a software driver in F1 could effectively simulate things like traction control and such. Not that that would be legal since you'd need all kinds of sensors wich are currently not allowed, but I suppose it won't be able to drive at all without sensors obviously. For road cars this is of course not an issue, wich is more the aim of this and current research.

Btw "Deep Blue" that also took quite some time. In the 80's the best chess programs could beat the best amateurs at chess, but it took the gradual hardware upgradings to beat GM's. It only uses other games from humans as opening book though for the rest it adapts and does calculations. Modern chess software is way beyond that though, even freeware chess-engines run on your own (decent) computer can beat or draw super-GM's.

[Dmitriy_Guller]

A computer can learn to ago around, but it will never be as effective as humans
Just look how hard it is to make a proper AI racer in a virtual environment, let alone in the real world.

That's wildly optimistic of humans, and wildly pessimistic of computer algorithms, which have evolved way past just following human step-by-step instructions. And the problem with AI racers in computer games is two-fold: one, the hardware limitation, and two, the lack of resources computer game makers have at their disposal.

[tomjol]

If someone could figure out a way to make money from a self-racing car I'm sure this would happen tomorrow.

Edited by tomjol, 10 November 2012 - 15:50.

[Lights]

Just a matter of time. Computers (software basicly) are much more reliable than humans if programmed correctly. Also a software driver in F1 could effectively simulate things like traction control and such. Not that that would be legal since you'd need all kinds of sensors wich are currently not allowed, but I suppose it won't be able to drive at all without sensors obviously. For road cars this is of course not an issue, wich is more the aim of this and current research.

Btw "Deep Blue" that also took quite some time. In the 80's the best chess programs could beat the best amateurs at chess, but it took the gradual hardware upgradings to beat GM's. It only uses other games from humans as opening book though for the rest it adapts and does calculations. Modern chess software is way beyond that though, even freeware chess-engines run on your own (decent) computer can beat or draw super-GM's.

I get your point but F1 isn't exactly chess. Chess is calculating options, while a large part of driving an F1 car is feeling. Of course, it's possible to simulate that but then still I don't see them beat human F1 drivers.

[Risil]

I don't think a computer programme would be as good at pulling in the sponsorship either.

[tomjol]

I get your point but F1 isn't exactly chess. Chess is calculating options, while a large part of driving an F1 car is feeling. Of course, it's possible to simulate that but then still I don't see them beat human F1 drivers.

It's only about feeling because it's a human doing it.

[Dmitriy_Guller]

It's only about feeling because it's a human doing it.

Exactly. Physical feelings and intuition aren't magic, they're just human (and suboptimal) ways of processing information.

[Shiroo]

go 30 years back, and computer were way inferior to human, now Watson Computer beats best people in answers for most complicated questions. It can answer on 70% of questions, where humans to around 65% of these.

in 10-20 years, they will be superior (100% rate, and cost only 1milion, and will diagonse people for their illness)

So it wont be long enough when drivers will become rather useless. Though it will never be AI controlled car. Why? Coz F1 is about WDC, what's the point of WDC then if all AI will drive car to the limit, lapping all the time with best times possible atm. It will be only Constructor championship then.



On many fields computers already beat human beings, and tat was rather unthinkable 20-30 years ago.

[Lights]

It's only about feeling because it's a human doing it.

Like I said, it's possible to simulate, in other words find another way for handling. I don't doubt the possibility but I doubt the success.

[Tenmantaylor]

It's definitely possible to program computer to drive fast on the limit as long as it is given enough data and programmed to understand it. In the racing sim Live for Speed the AI cars drive using the same physics as the users car and given enough experience they can get really quick. Not match the aliens but beat most people who havnt practiced much. This is rare in most racing games, even in GPL the AI cars had 'magic grip' values to help them stay on track. In Lfs, part of the AI algorithm of the AI cars allowed it to perfectly control oversteer reading the grip and weight dist. values. This same algorithm was also used as a driving aid for keyboard and mouse drivers who were at a disadvantage compared to wheel drivers. After a while World Records were being set by keyboard drivers and it turn out the counter steer algorithm was proving a huge advantage so it got banned!

[tomjol]

Like I said, it's possible to simulate, in other words find another way for handling. I don't doubt the possibility but I doubt the success.

What we call 'feeling' a computer can do with a few sensors and some relatively simple maths, and it can do it a hell of a lot faster and more accurately than any human could hope to do.

[Fatgadget]

Interesting topic. I can see thing going somewhere but I don't see them beating humans ever.

Think again. Program a 'puter with the perfect lap and it will replicate it time and again! Already a computer has beaten a human at chess...

[turssi]

Think again. Program a 'puter with the perfect lap and it will replicate it time and again! Already a computer has beaten a human at chess...

Yeah, but on a trace track Deep Blue will get its tires slashed and radiators smashed in a jimmy.

[Fatgadget]

Yeah, but on a trace track Deep Blue will get its tires slashed and radiators smashed in a jimmy.

Umm..Suspect if programmed appropriately it can also return the favour - No?

[BetaVersion]

Snic's post is brilliant.

A perfect artificial intelligence system (computer with a perfect algorithm and connected to very accurate sensors) would react to wheelspin and etc much faster than any human could ever do.

The only problem of PCs is when things gets unpredictable and out of the algorithm, such as when the track have unpredictable grip levels.

A compute simply follows a method while a top driver can kind of "improvise" to different circunstances. I think that'a why a computer will never be as good as a top human driving in "uncontrolled enviroment".

Just think of Button and Hamilton. Jenson is the more robotic guy following his method while Lewis is the guy with that "extra talent" to react to what is unpredictable(out of the algorithm)

Edited by BetaVersion, 10 November 2012 - 19:55.

[Nick Planas]

Imagine the podium interview after it won a race

[Fatgadget]

Imagine the podium interview after it won a race

Something along the lines of.... Id like to thank my designer...my programmer ..my engineer etc etc...not much different to what we hear today.

[BetaVersion]

Just a matter of time. Computers (software basicly) are much more reliable than humans if programmed correctly. Also a software driver in F1 could effectively simulate things like traction control and such. Not that that would be legal since you'd need all kinds of sensors wich are currently not allowed, but I suppose it won't be able to drive at all without sensors obviously. For road cars this is of course not an issue, wich is more the aim of this and current research.

Btw "Deep Blue" that also took quite some time. In the 80's the best chess programs could beat the best amateurs at chess, but it took the gradual hardware upgradings to beat GM's. It only uses other games from humans as opening book though for the rest it adapts and does calculations. Modern chess software is way beyond that though, even freeware chess-engines run on your own (decent) computer can beat or draw super-GM's.

are you sure about that?

[Brandz07]

how would a computer handle changeable conditions like a wet track?

[Lights]

What we call 'feeling' a computer can do with a few sensors and some relatively simple maths, and it can do it a hell of a lot faster and more accurately than any human could hope to do.

Yeah it's all so simple, yet game after game is being brought out with buggy AI that can't match human drivers without unrealistic grip.

[bonjon1979a]

Imagine the podium interview after it won a race

We have kimi, there's no need to imagine...

[Lights]

Think again. Program a 'puter with the perfect lap and it will replicate it time and again! Already a computer has beaten a human at chess...

As been said, chess is far from F1.

[rijole1]

If someone could figure out a way to make money from a self-racing car I'm sure this would happen tomorrow.

Yeah, that's the problem.
I think people would be bored after race or two.
No viewers, no money

Edited by rijole1, 10 November 2012 - 22:07.

[rijole1]

Imagine the podium interview after it won a race

But seriously, It would of course be a podium for programmers

[Koen]

Maybe I'm over optimistic but I think it's only a matter of getting enough resources (money and some talented people) for research and development.
But it's something so unimportant and unprofitable (why not develop some new bombs and guns instead) that it won't happen in the near future, even if it's possible.

[Ross Stonefeld]

how would a computer handle changeable conditions like a wet track?

Same way a human does? Detects that there's a change in the weather, backs off, and contantly recalculates the available grip.

[spacekid]

how would a computer handle changeable conditions like a wet track?

Traction control and ABS? These things already exist, and are examples of how a computer program is able to outperform a human. Also humans are programmed to an extent - racing drivers learn from a young age the correct motor control to control the car, control sliding and throttle/brake response. They also learn from experience things like when the racing line changes in wet conditions. Hamilton or Schumi didn't know these things the first time they drove, they built up the experience over time the same way evolving computer algorithms do.

This sort of thing doesn't surprise me, and I'm sure that a computer algorithm could be written that could drive a lap of an F1 car faster than a human. I agree though that the programming would be very difficult for actual racing, but I don't see it as impossible. Just imagine one of these things developing its algorithm driving thousands of laps over a period of 20 years in a range of different track conditions.

Its an interesting tech demonstration but I don't see that people will ever be interested in watching a computer controlled vehicle as its the human element that excites us.

[tomjol]

Yeah it's all so simple, yet game after game is being brought out with buggy AI that can't match human drivers without unrealistic grip.

Yes, because video games as we all know are the pinnacle of engineering.

[Lights]

Yes, because video games as we all know are the pinnacle of engineering.

That's the point, why would this be more interesting, enough to be of importance that it receives a large enough investment. It's not like the world is eagerly waiting for the first computer race.

[Bloggsworth]

'instinct' is still a human reaction to data though, it's not magical

No it isn't, and it isn't simply reaction to data either; what distinguishes the experienced driver is the anticipation of future data input and the ability to prepare the reaction in advance, all done without conscious thought.

[Skinnyguy]

I think this should be introduced in F1 soon. Then Lewis and Fernando could talk about how they´re "fighting an engineer, not a driver" without talking bullshit.

[icecream]

No it isn't, and it isn't simply reaction to data either; what distinguishes the experienced driver is the anticipation of future data input and the ability to prepare the reaction in advance, all done without conscious thought.

i guess it's arguing semantics. when you say 'the anticipation of future data', well yes, but this anticipation occurs due to current data. we can't see into the future of course, but our brain decides what will probably happen based on what is already known. a computer racer can do exactly the same thing, and will need to if it is to be competitive.

with or without conscious thought doesn't really matter. it's still your brain making calculations (either in the background, or in the foreground).

[Dmitriy_Guller]

No it isn't, and it isn't simply reaction to data either; what distinguishes the experienced driver is the anticipation of future data input and the ability to prepare the reaction in advance, all done without conscious thought.

Anticipation is a way to cope with slow reaction times. If you can react instantly to what's happening, there is no need to anticipate what's about to happen.

[halifaxf1fan]

The cars would probably achieve the fastest way to complete the race distance but the races would be rather boring without all the human error!

[prty]

No it isn't, and it isn't simply reaction to data either; what distinguishes the experienced driver is the anticipation of future data input and the ability to prepare the reaction in advance, all done without conscious thought.

And guess that, a computer can do that better than humans too. Just to give you an idea, there are algorithms right now that can predict with a 75% of accuracy if the answer to a question is going to be 'yes' or 'no', just by looking at body language of the participants in a conversation.

[CSquared]

Anticipation is a way to cope with slow reaction times. If you can react instantly to what's happening, there is no need to anticipate what's about to happen.

This is interesting stuff, and also stuff where people can be arguing past each other. I think anticipation is not about reaction time or being able to calculate without data, it's about how to understand and calculate the currently-available data. The car yaws a certain degree on turn in, indicating you're gonna miss your apex and slow your exit if you don't fix it in time. The car ahead of you rides the curb and twitches a certain way, indicating it's about to spin across your path. This stuff is about physics - momentum, friction, acceleration. Reaction time can't overcome the physics. It's all about how you calculate, or anticipate if you want, what the current physics say about what the physics will be 10, 100, or 1000 milliseconds from now. Getting the robot to sense all that data correctly, make good estimates/guesses for the data that's missing, and calculate what it all means is extremely challenging, but once we get down to the basic physics calculations, the computer can do that more accurately and faster than a human.

Edited by CSquared, 11 November 2012 - 02:44.

[pingu666]

well the computer will be mostly better at traction, because it would have reasonable tc, maybe even crazy good tc. and braking it would have ABS. it would also possibly have the stability control, single wheel braking etc, if the rules allowed it too...

working out the racing line on the fly would be harder, but it could be set by a human, or it could learn it

it would need to combine radar, video, laser range finder? gps, plus the telemetry style sensors

[tomjol]

As many have pointed out, this is a simple case of physics. Yes there is a certain amount of "racing intelligence" which would need to go alongside that, but that's really just some more physics and a "best option" calculation.

Frankly I don't think anything about this is difficult. The only slightly tricky thing would be engineering in a reliable element of "randomness" to ensure that, say in a field of these cars, they didn't all choose the same bit of track and thus conflict with each other, endlessly making the same adjustments to avoid a car making the same decisions.

That's the point, why would this be more interesting, enough to be of importance that it receives a large enough investment. It's not like the world is eagerly waiting for the first computer race.

It wouldn't, I can't think of any reason why it might be worth doing as more than a one-off to showcase the technology, hence I don't think it will happen.

But I maintain that that's the only reason why not.

Edited by tomjol, 11 November 2012 - 08:46.

[sergeym]

well the computer will be mostly better at traction, because it would have reasonable tc, maybe even crazy good tc. and braking it would have ABS. it would also possibly have the stability control, single wheel braking etc, if the rules allowed it too...

working out the racing line on the fly would be harder, but it could be set by a human, or it could learn it

it would need to combine radar, video, laser range finder? gps, plus the telemetry style sensors

I think there should be a way to formally define the best racing line - and computer will simply try to follow it, making corrections based on its sensor input. About human "feeling" and "inctinct" - i think it is just our way to make fast decisions based on limited data. In theory computer may be able to use more complete data to judge the situation.