Discussion in 'transport' started by roryer, Jun 13, 2014.
Pulling out into a busy lane of traffic where people in the lane give signals such as a wave or flash of the lights to permit you to join, other times they do not.
Merging with another lane of traffic at an obstruction where two lanes make one, sometimes it is possible to go one side followed by another like a zip, but sometimes vehicles will not give way forcing a new plan.
Following a slow tractor and trailer on an unmarked muddy country road in the twilight where it cannot be certain oncoming vehicles will have their lights on. Working out a safe place to overtake.
If all cars were self-driving then all these situations would cease to be a problem as the cars could talk to each other and agree a plan.
But there aren't plans for all to be all self driving are there?
Anyhow there will always be motorbikes and tractors, and pedestrians, and cyclists to consider.
I would opt for a hybrid which permits self driving as an option, because I enjoy driving.
On the first one they might not be able to react to a wave or flashed lights but they will react to drivers slowing to let you out which is what is actually needed to create the space for you to pull out.
The middle one, the car will react accordingly, just like a human, expecting that people/cars will merge in turn but when someone forces their way through, the car or a human will react the same - brake and check if the person/car behind has left a gap for you to pull into instead.
The last one, the cars use radar I think so would be far better at spotting an oncoming vehicle without lights. Working out a safe place to overtake is a matter of physics and a computer will be much quicker and more accurate than a human. This one I think is absolutely not a problem at all, automated cars would be better at this than humans I think - more accurate information.
Weltweit, what is it about your decision making process that you think is impossible to replicate? What information are you using that is impossible for a machine to use? What brand new situations are you encountering that are so unlike anything you have previously experienced as to require new approaches to be developed?
For another example, there is country driving and there is London driving. Driving in London, drivers will pull out when there is not a gap, their actions create the gap. You notice someone from London in the countryside because they seem to be driving incredibly aggressively, but that is how they drive in the city. I don't think computers will be able to cope with that.
I don't have your trust in sensors, nor the computers they are connected to.
I live near one of Britain's larger forests. At night I stick to 45 because of the likelihood of a deer running out into the road, they do, how would a computer register this risk and act accordingly?
From which perspective? The computer being pulled out on by a driver from the city creating a gap? This is a simple problem solved by detecting an object moving into your path and braking. A computer can probably react more quickly than a human would.
From the other perspective of needing to push their way out into traffic in a city? This is more challenging clearly as cars are likely to be rightly programmed to be cautious but as more and more cars become autonomous and start talking to each other, the problem will be solved - you don't need every car to be autonomous, just enough that they'll create spaces for each other often enough to not get stuck.
I have more trust in sensors and the computers they are connected to, than I do to human sensors and the brains they are connected to. A computer will be continually scanning it's surroundings with radar and visual sensors (including using IR at night which means the cameras can see things we can't and possibly stuff like heat sensing even), likely they will be able to detect the deer running out into the road (or the possibility thereof) far in advance of when a human would. They will react more quickly to it happening and be able to consistently brake and control the car where humans sometimes fuck it up. The knowledge that deer often run onto particular sections of road can be programmed into a computer - and if it is signed then those signs are on a database and it will be known to the car and be able to reduce it's general travelling speed accordingly. Just like you do.
Lots of drivers hit deers each year. Sometimes automated cars might too (only a question because they don't exist yet so no empirical evidence). The question becomes which will do it less. I'd bet on the automated vehicles.
I have 35 years driving experience, on a wide variety of vehicles, on a wide variety of roads, I just think it very unlikely programmers can code for every eventuality which I and other drivers have already come across and navigated successfully.
As you probably already know recently a Tesla hit a truck and trailer which pulled across the highway in the US, killing the driver. It was equipped with radar yet didn't detect the truck in time to activate the brakes.
Every sensing system has advantages and limitations, there is no one sensor sees all solution.
A human can visually assess sheep sitting or standing on the verge of a road and make a judgement as to whether they are going to remain eating grass there or wander into the path of the vehicle. An autonomous vehicle's IR sensor could only detect that there was a sheep there. And that is assuming the Auto vehicle had IR capability.
Sure - but humans crash every day on the motorway too. Will it happen less with automated vehicles? Are their sensors better or broader, reactions quicker, distractions fewer? I would say yes to all of those - and the more and more vehicles become automated and start talking to each other, the safer they will become - had both the tesla and the truck been automated and talking to each other, the truck wouldn't have pulled out on the tesla.
So many of the decisions are matters of physics and computers handle that better than any human will.
A human might think the sheep wasn't going to move, not slow down, get it wrong and hit the sheep. A car might always slow down and never hit the sheep. One driver might choose to always slow down just in case, whilst others decide not to. In the event the sheep does run out, startled by something, the human who has decided not to slow down may have already disregarded the sheep as a hazard and not being paying attention. The computer will always pay attention to the sheep. The computer is much more likely to be able to consistently reach the optimum high speed to never hit the sheep whereas a human will often go slower than may be needed to stop if the sheep does decide to run out.
I don't really see a problem there tbh, much of it depends on how you - as a human - think it's important to always avoid the collision, what probability of not avoiding it are you happy with and how good are you at judging the likelyhood of that grazing sheep running out, compared to how a computer is likely to be programmed to deal with the situation. Risk assessment is something humans are not very good at generally so if computers can have information in the way google sucks it up then I would say they'll probably be better at it. I'm sure they will be situations around the margins but if 99/100 situations are an improvement and 1/100 isn't then it's an improvement. even if that's 51 and 49/100 it's improving.
BigTom I don't share your confidence or faith in computers or sensor systems.
I remember an engineer friend coding a program to do something quite simple in a car. The code was massive, absolutely humungous, and it wasn't anything like an engine system or ABS or anything mission critical.
And then there are those imponderables.
You are driving an auto 4x4, a big vehicle, on a busy motorway, there is a crash in front of you which puts a small car, with 4 occupants, suddenly in your lane - too close to stop for. In the outside lane a truck which has also come to a stop.
Your Auto car has ABS so it can still steer while under maximum braking but if it hits the small car the passengers will be badly injured or killed, while if it swerves into the back of the truck you may be injured. What will the auto car do? What would you as a human driver do?
Worth saying that all cameras have IR capability (the light sensors pick up infra red naturally), so all cars will have this. It's not needed to detect sheep, except at night when there isn't light in the human visual range (IR = night vision, infra red light our eyes can't see, the greeny-grey nightvision mode cameras have). Heat sensing probably what you mean but as long as the car has clear sight it will be able to distinguish sheep from other types of objects in the same way it can distinguish cyclists from pedestrians from trees etc. Heat sensing is useful when you don't have visual sight of eg: (hot) deer running through woods amongst (cold) trees.
I don't share your confidence or faith in humans or our sensor systems
I, as a human, would react later and slower to the situation, possibly braking less sharply, and therefore hitting either the car or the truck at a greater speed leading to a higher likelyhood of death or serious injury whichever way I happened to react. We might not like the thought of having to pre-decide in this kind of situation rather than just reacting and therefore removing our own responsibility towards the crash (which there isn't really one but people will feel it) but regardless of the decision made as to swerve or not, the outcome has a higher chance of being better if the computer is handling it.
In either case, probably the simplest thing. Your lizard brain dominates and doesn't have time to make value judgements.
I believe the latest S Class Mercedes has radar assisted braking. I would expect features like that on cars before full autonomy.
And then there was the claim last week that someone had hacked a Tesla, gaining control of the brakes remotely.
In no time at all, the algorithm will have 1000s of years of driving experience on every variety of vehicle and every variety of road.
I don't think you understand how machine learning works. It isn't programmed in by a programmer. It learns from the accumulated experience of every vehicle it is operating.
Or rather, it learns by observing all the human drivers operating their semi-autonomous cars.
Self driving cars may soon be a thing of the past what with people carrying drones and self flying cars. I heard about self flying cars being developed in Portlad Oregan on the BBC this morning but can't find much in the way of links.
This supersized drone will fly you to work (or anywhere)
Self-flying cars in the sky? - CBBC Newsround
Assuming the same program is present on every variety of vehicle and road which is unlikely.
Oh it is programmed, do you think some kind of empty computer just starts to take inputs and learn, the computer runs a program to enable it to operate, and learn, and that program is constructed by humans. One may be programmed by Google, another by Volvo, another by Toyota, another by Tesla etc etc .. and I doubt they will talk one to another.
And that is after they capture and interpret inputs from their various in car sensor systems which they will use to make real time safety critical decisions on driving.
Different companies will have their own systems that probably won't talk to each other, but they will each have many thousands of vehicles on the roads, all learning together. Every company will very quickly have a data set way larger (and more accurate/reliable) than the one you've accumulated in 35years.
Your optimism is commendable
Whilst stuff about sensors and how cars react might be optimistic, I really don't see why it's optimistic to think that each of eg Google, Tesla, Apple, Ford etc will have thousands upon thousands of automated cars on the roads and they will all be learning together. 35 yrs of driving experience actually isn't 35yrs because you're only driving for part of the day, say it's a couple of hours each day average, that means ~800hrs/yr or ~28,000 hours of driving. If google only has 28 cars, they will reach your level of experience in about 18months. 280 cars and it's less than 2 months, 2,800 cars and that's just 5 days. You only need a very small fleet of cars to get the hours in, this isn't a matter of optimism, just maths. Tesla have sold 140,000 cars since they started so even the smaller manufacturers will get those numbers pretty quickly. Imagine how many Ford or VW would have if they choose to go alone and not purchase a google/apple/uber system to run their cars off.
"Whilst stuff about sensors and how cars react might be optimistic,"
This is what I think you are optimistic about, well that and machine learning being as easy to do as you and others seem to think, and the whole project being put into practice in the time frames mentioned by some in this thread.
I think you're being a mix of technologically pessimistic and naive to the bulldozing nature of following the money.
For me the problem with the whole concept is the issue around responsibility.Normally when you accept a ride in a car you are in a position to make some judgement at least of the competence of the particular driver.If you are killed in a sense you are the author of your own misfortune.I don't see how that works when you are jumping into the passenger seat in effect beside a robot.And as for the process of machine-learning how many people is it okay to sacrifice in order that these driver-less car programs achieve maximum efficiency?Could you not even apply the same argument to government of a country and suggest that machines,being more efficient would be likely better at it in the long run as long as we were prepared for them to make a few errors as an unfortunate but inevitable part of the learning process??
I am technologically pessimistic, yes, but partly because I have seen how slow the automotive industry is to innovate with inventions that have a lot of testing behind them. The automotive industry is, or has been, risk averse, and the supplying components industry features massive testing regimes for even the simplest component. Then we have the concept that above all this we are going to add a computer system to drive the car collaborating with HAL9000 at base. Will the computer system be able to detect when a tyre is down on pressure like a driver can, will the computer be able to .. sorry I have to go, will be back later.
Separate names with a comma.