RE: Robot car closes on driver track times
Discussion
SlipStream77 said:
ScoobieWRX said:
They'll need a lot more than 3 computers in the end to mimic the reactions, thoughts and experience of a human.
Computers may be able to super multi-task all day long every day but they don't have the deft and instinctive touch of a professional human driver, nor can they make an on the spot decision based upon the driving conditions, steering feel and sensation a pro driver uses to action and control inputs.
Computers may be able to super multi-task all day long every day but they don't have the deft and instinctive touch of a professional human driver, nor can they make an on the spot decision based upon the driving conditions, steering feel and sensation a pro driver uses to action and control inputs.
The decisions made by an onboard computer are likely to be limited only by the quality of the data sent to them from sensors and the software algorithms.
Sensors on F1 cars record 1000's of channels, some of these are recorded over 1000 time a second. Using these sensors to monitor aspects such as g-levels, suspension deflection, wheel speeds and employing a range of high res. imaging hardware, it would be entirely possible for a computer to drive the 'perfect lap' within its own constraints.
The only issue I can imagine would be that you'd need a very fast processor for dealing with the imaging side.
A while ago, when an automatic landing system was being developed for civil aircraft, during testing it was shown to have a problem. It was too good - it landed consistently in the same place and was wearing the runway. They had to introduce a random factor to the landing position.
Computers will lap faster than drivers, it's just a matter of time until someone develops the system.
I wonder if we will see multiple classes at Le Mans for AI?
They have been doing it years as far as I know.
Obviously with enough sensors an F1 car robot driver could drive as well as a real human F1 driver, or better. I just think the computing power and programming (probably via lots of teaching a neural network) to achieve that are still way out of our range.
You only need to realise how stupid computers can be when junk in = junk out.
Neural networks are great when you can control a set of variables easily but managing the huge sets to have an autonomous F1 driver?
I think by the time we can make neural networks work that effectively to let them do complex tasks like drive an F1 car like a real driver would, the day we can all put our feet up and let machines run the world for us while we kick back and relax
Dave
DannyScene said:
Trains are always something i wondered about, just how neccesary are the drivers?
Lad I went to school with has driven trains for 20-odd years (all over the UK including the Underground in London, Metro in Newcastle and mainline trains) and he reckons train drivers are about 50% "make the passengers feel safe", 40% "give the train companies someone to blame in an accident" and 10% "make the system work"All of that last 10% is expensive stuff to get right automatically - it requires large-scale investment and high levels of maintenance and planning, which private rail companies are afraid-of - so they use humans to bodge the gap between the system (trains, points, rails, stations, scheduling etc.) and what people expect from it (they're part of the 20 in the 80/20 rule).
If a plane were to suffer electronic failures, the systems are designed to keep control surfaces etc. working and allow the pilot to land - full sized jets can be glided into land - full size jets can be landed in rivers etc.
If anything goes wrong in a train - it just stops, that doesn't require human intervention
McSam said:
There's almost no question of ever making a "better" race driver than a person.
On a track solo, a computer could outdrive a person now - it's just a matter of investment in something I can't work-out why you'd invest in.On a track full of other computer drivers - that's doable too, tho you'd stuggle to teach them to 'race' in any sort of interesting way (you'd need a "John Cleland" subroutine for the max entertainment value)
On a track full of humans - computers wouldn't stand a chance because we couldn't teach them to deal with so much randomness
The things which we do well and computers do badly are often subtle tho. If we were facing each other and a threw a ball at you, your brain would make a staggering amount of calculations about likely weight/direction/curve and automatically move your hand to meet that ball. For a computer to do the same with cameras and a robot arm is staggeringly complex.
Thing is - I'm not sure reading conditions and adjusting a line around a track and factoring-in speeds for corners etc. is that complex.
For the people who said "let's not teach them stuff which will make them take-over"...
Back in 2008, DARPA had a robot which did this - they're now producing this thing for military applications, it can follow troops and navigate terrain carrying heavy loads
https://www.youtube.com/watch?v=W1czBcnX1Ww
Right now they're making this - and soon, it will come looking for you - and you're screwed...
https://www.youtube.com/watch?feature=player_embed...
We may-as-well have cool robots which drive cars and hoover and mow the lawn - but the miltary will make their killing machines whether we do that or not...
Back in 2008, DARPA had a robot which did this - they're now producing this thing for military applications, it can follow troops and navigate terrain carrying heavy loads
https://www.youtube.com/watch?v=W1czBcnX1Ww
Right now they're making this - and soon, it will come looking for you - and you're screwed...
https://www.youtube.com/watch?feature=player_embed...
We may-as-well have cool robots which drive cars and hoover and mow the lawn - but the miltary will make their killing machines whether we do that or not...
Terminator X said:
Rubbish in rubbish out. Until they can "think" it's not gonna happen. We don't even understand the human brain yet so we're fookin miles away from intelligent robots. Thank f
k I say as it will end in tears ...
TX.
Apt username.k I say as it will end in tears ...TX.
Give it 30 years and we will have fully autonomous robots (think I, Robot), less if we crack the practicalities of quantum computing.
405dogvan said:
The things which we do well and computers do badly are often subtle tho. If we were facing each other and a threw a ball at you, your brain would make a staggering amount of calculations about likely weight/direction/curve and automatically move your hand to meet that ball. For a computer to do the same with cameras and a robot arm is staggeringly complex.
True, but http://www.youtube.com/watch?v=-KxjVlaLBmk 3 years ago!Contigo said:
Cotty said:
How about if they replaced the ring taxi drivers with robots, would you go for a ride?
Is it any different than getting on a plane where only about 5 minutes of the journey is actually human controlled? Even for health and safety the ring cab would have to have a human at the wheel in case things went wrong with the automation. I find it quite annoying if a car parks itself. But that's a totally different league than this.
Now just wait for the first madman to hack the software... This is plain madness.
A thought occurs to me.
At the age of 23 Vettel became the youngest driver ever to win the World Drivers' Championship.
Thats 23 years invested in him becoming arguably the best driver of his era.
If 23 years were to be invested in a computer based system with necessary funding and the same level of focus and determination I could easily see it being better
At the age of 23 Vettel became the youngest driver ever to win the World Drivers' Championship.
Thats 23 years invested in him becoming arguably the best driver of his era.
If 23 years were to be invested in a computer based system with necessary funding and the same level of focus and determination I could easily see it being better
I'm touched by the faith you lot have in the ability of a human.
In reality a human is a fairly old piece of electrical kit. The wiring in a human works at nowhere near the speed of electrical cable, and the processing power of a human brain cell is nowehere near the speed of even a pretty average computer.
The problem is that the two aren't easy to compare because they work on two totally difierent system architectures. The whole of human history have been to develop machines that fit in with the human condition. Cars are one example. Why does a computer require a car to get around? Its nonsensical.
The human brains only advantage over a current computer is parallel processing ability. Its a scale issue, and if the world keeps turning them sooner or later the gap will close to nothing.
In reality a human is a fairly old piece of electrical kit. The wiring in a human works at nowhere near the speed of electrical cable, and the processing power of a human brain cell is nowehere near the speed of even a pretty average computer.
The problem is that the two aren't easy to compare because they work on two totally difierent system architectures. The whole of human history have been to develop machines that fit in with the human condition. Cars are one example. Why does a computer require a car to get around? Its nonsensical.
The human brains only advantage over a current computer is parallel processing ability. Its a scale issue, and if the world keeps turning them sooner or later the gap will close to nothing.
Driving around a track, with no other competitors, seems like exactly the sort of problem a computer would be ideal at solving. It is pretty close to some of the things that have been mentioned, like autoland. Humands are much better at some things, like language analysis, emotions, complex learning, art - but very precise optimisation is not one of them.
If you take your average contemporary car, stuff like ESC already intercepts a lot of what is happening with the car before the driver even knows about it. When the driver does know about it, it is all seat of the pants feeling stuff - imprecise, late, and ultimately a poor approximation of a sensor, except perhaps in the visual sense. The decision making is based on a combination of these blunted inputs, prior experience and guesswork. A human is also subject to emotional interference, fatigue and other unhelpful stimuli.
The fastest route around a track is a mathematical formula, not art. Any art comes from the human struggle to approximate the same thing.
If you take your average contemporary car, stuff like ESC already intercepts a lot of what is happening with the car before the driver even knows about it. When the driver does know about it, it is all seat of the pants feeling stuff - imprecise, late, and ultimately a poor approximation of a sensor, except perhaps in the visual sense. The decision making is based on a combination of these blunted inputs, prior experience and guesswork. A human is also subject to emotional interference, fatigue and other unhelpful stimuli.
The fastest route around a track is a mathematical formula, not art. Any art comes from the human struggle to approximate the same thing.
trashbat said:
The fastest route around a track is a mathematical formula, not art. Any art comes from the human struggle to approximate the same thing.
That is true, right down to using quantum mechanics to solve for probability of grip levels at every single point on the track surface.For instance, newtonian inputs require a constant for things like coefficient of grip, but the real track surface isn't homogenous. It's made up of a fine patina of details. A slug might have run over a part of the track and your tyre might run 1mm wider one day vs the last and then the car has to adjust to that.
Trillions of variations will exist so we have to average things out, or solve with very complex solutions to give us probabilities of grip levels etc.
In any case, a computer could drive as well as a human when the AI is so good they are in essence a human brain that can interpret and learn, but also have the strength of the perfect recall and recording of events they learnt from, and much more precise outputs of control changes.
But until the AI is that good, in my view, then a human will be able to defeat the averages that computers depend on to work. By the time you provide the data to the computer to drive so fast that it's so accurate for the entire course, it'll change the next day, or over the course of the day.
Neural network AI is obviously the solution but someone needs to 'breed' it and test it for fitness and let it evolve.
It'd be interesting to see if this robot is using anything like that, or is just using direct programming algorithms based on physics inputs and a database of track data.
If you could simply give it the boundaries of the track driving surface and let it go I'd be impressed, but I bet it's not that good!
Dave
Mr Whippy said:
That is true, right down to using quantum mechanics to solve for probability of grip levels at every single point on the track surface.
For instance, newtonian inputs require a constant for things like coefficient of grip, but the real track surface isn't homogenous. It's made up of a fine patina of details. A slug might have run over a part of the track and your tyre might run 1mm wider one day vs the last and then the car has to adjust to that.
Trillions of variations will exist so we have to average things out, or solve with very complex solutions to give us probabilities of grip levels etc.
Yes, but once down at this level, it only matters once you are aiming towards a combination of infinitely optimal lap time and wringing 100% performance from the car - and a human driver, however good, is so far from either of those. Humans don't have all the information either, so will cock it up in an unexpectedly slippy corner, be too cautious at the next, etc.For instance, newtonian inputs require a constant for things like coefficient of grip, but the real track surface isn't homogenous. It's made up of a fine patina of details. A slug might have run over a part of the track and your tyre might run 1mm wider one day vs the last and then the car has to adjust to that.
Trillions of variations will exist so we have to average things out, or solve with very complex solutions to give us probabilities of grip levels etc.
Mr Whippy said:
In any case, a computer could drive as well as a human when the AI is so good they are in essence a human brain that can interpret and learn, but also have the strength of the perfect recall and recording of events they learnt from, and much more precise outputs of control changes.
Quite. Add to that instant refinement of technique; average wheel slip reaches a threshold so the computer backs off immediately, whereas a human might wait until they end up too far on the grass or get yelled at by the team manager to ease off.Mr Whippy said:
But until the AI is that good, in my view, then a human will be able to defeat the averages that computers depend on to work. By the time you provide the data to the computer to drive so fast that it's so accurate for the entire course, it'll change the next day, or over the course of the day.
Neural network AI is obviously the solution but someone needs to 'breed' it and test it for fitness and let it evolve.
It'd be interesting to see if this robot is using anything like that, or is just using direct programming algorithms based on physics inputs and a database of track data.
If you could simply give it the boundaries of the track driving surface and let it go I'd be impressed, but I bet it's not that good!
Dave
I don't think it requires neural networks or anything of that complexity to beat a human. Because the environment is largely static, weather being the obvious exception, GPS and some fairly static sensor-driven algorithms are probably enough for a good pass at it; vision processing and environmental sensors would add further benefit.Neural network AI is obviously the solution but someone needs to 'breed' it and test it for fitness and let it evolve.
It'd be interesting to see if this robot is using anything like that, or is just using direct programming algorithms based on physics inputs and a database of track data.
If you could simply give it the boundaries of the track driving surface and let it go I'd be impressed, but I bet it's not that good!
Dave
I don't know the whole answer to this question, but a lot of it hinges on the leeway between acceptable speed and disaster. Suppose you can take a given corner at 60mph but will come off the track if you try it at 65mph. In human terms there may be little feeling between the two, but in sensor terms there may be lots of predictors flagging that it is going to go badly before the commitment to the 65mph attempt is inescapable - subtle changes in traction etc, some learnt values setting off alarms and so on. Beyond the basics of autonomous control, this is probably the most difficult bit to get right, but it's nothing like as difficult as many computational problem solving applications that have already been satisfactorily dealt with.
SlipStream77 said:
A while ago, when an automatic landing system was being developed for civil aircraft, during testing it was shown to have a problem. It was too good - it landed consistently in the same place and was wearing the runway. They had to introduce a random factor to the landing position.
e.g. This was human precision, not computer.
I am not a civil pilot, so if there are any out there I stand to be corrected.
trashbat said:
I don't think it requires neural networks or anything of that complexity to beat a human. Because the environment is largely static, weather being the obvious exception, GPS and some fairly static sensor-driven algorithms are probably enough for a good pass at it; vision processing and environmental sensors would add further benefit.
I don't know the whole answer to this question, but a lot of it hinges on the leeway between acceptable speed and disaster. Suppose you can take a given corner at 60mph but will come off the track if you try it at 65mph. In human terms there may be little feeling between the two, but in sensor terms there may be lots of predictors flagging that it is going to go badly before the commitment to the 65mph attempt is inescapable - subtle changes in traction etc, some learnt values setting off alarms and so on. Beyond the basics of autonomous control, this is probably the most difficult bit to get right, but it's nothing like as difficult as many computational problem solving applications that have already been satisfactorily dealt with.
I agree generally, but I don't consider that intelligence or adaptability.I don't know the whole answer to this question, but a lot of it hinges on the leeway between acceptable speed and disaster. Suppose you can take a given corner at 60mph but will come off the track if you try it at 65mph. In human terms there may be little feeling between the two, but in sensor terms there may be lots of predictors flagging that it is going to go badly before the commitment to the 65mph attempt is inescapable - subtle changes in traction etc, some learnt values setting off alarms and so on. Beyond the basics of autonomous control, this is probably the most difficult bit to get right, but it's nothing like as difficult as many computational problem solving applications that have already been satisfactorily dealt with.
If the system was suddenly used at a different track where tarmac had a different nature and adaptation systems were calibrated on the other tarmac it could all end it tears.
Having a week to lap a circuit recursively improving performance isn't really intelligence, it's just a program with some feedback logic from sensors.
With the right algorithm you'd just be using determination of slip angle/ratios and tyre temps probably, and just let the car keep pushing it until it's running them all at the edge of the envelope... the line round the circuit is probably a given so stick to that at all costs.
But the fact you can't just change a variable like the shape of the track, or the weather, and the system adapt reliably, is why I don't see much or any value in it beyond a novelty.
If you wet the track and the car has to go out again for 50 laps starting from low speed again to 'find' where it can go faster and slower then it's not so good vs a driver who can see puddles or get a rough feel much more quickly through intuitive observation and thinking beyond basic parameters.
I think we will see computers being better at everything eventually, even better at us from technical thinking through to creative processes. We will literally be inferior brains in bags of mostly water that have given birth to an intelligence that will out-last us.
What that means for us is beyond my knowledge... I just keep thinking back to "The second renaissance" video from The Animatrix...
Hmmm
Dave
Mr Whippy said:
I think we will see computers being better at everything eventually, even better at us from technical thinking through to creative processes. We will literally be inferior brains in bags of mostly water that have given birth to an intelligence that will out-last us.
What that means for us is beyond my knowledge... I just keep thinking back to "The second renaissance" video from The Animatrix...
Dave
Maybe not. The human brain, despite having a smaller theoretical processing capability, is still capable of producing superior results to computers in any area where the answer is not based on solving mathematical formula's created and pre-programmed into the computer by humans.What that means for us is beyond my knowledge... I just keep thinking back to "The second renaissance" video from The Animatrix...
Dave
Perhaps the future lies in 'connecting' increased artificial memory storage and simple additional processing to the vastly superior organic computers we all have.
Gassing Station | General Gassing | Top of Page | What's New | My Stuff