Hacker News Reader

Comment by Groxx | Hacker News Reader

>they cannot be perfect

Two things:

1) neither are people; far from it. (you already knew this though, so...)

2) I wholly agree... as long as other people are driving. Get rid of people's irrationality / drunken behavior on the road, and automated car driving becomes a much easier problem. A computer can take more into account than a human, and decide faster.

The transition is almost guaranteed to occur at some point, unless we reach mass transit for everyone prior to that (and who drives those?). It has to start somewhere, and that start will probably be harder than at any other point along the road.

View on HN

2) I wholly agree... as long as other people are driving. Get rid of people's irrationality / drunken behavior on the road, and automated car driving becomes a much easier problem. A computer can take more into account than a human, and decide faster.

Even if you have near perfect data about where the other cars are there's always uncertainty about other objects on the road: tree branches, deer, a child running after her puppy. I'd speculate that it's better to rely on a vision system that can recognize those as well as cars well enough to be safe instead of treating different classes of objects differently.

True, but once you know what other cars are doing, and get data from cars ahead, you should be able to identify things like that before you even encounter them.

A tree falling on a car while it's driving: relatively unavoidable. Cars hitting that car: far less likely with computers. They don't get "shocked", or hit the wrong pedal, and have faster reaction times. And pile-ups due to an accident are often far more deadly than the accident itself, especially in busy areas where automated driving could make its biggest improvement in safety.

Deer / kids / unexpected obstacles: can be detected prior to them even entering the road by cars ahead, and flagged for caution for cars behind. IR cameras will spot living things for you, motion / blockage detection will spot things near / in the road. Once such a thing is flagged, it's reasonable to assume that other cars would leave more room between each other and around the obstacle, and probably slow down. All things which people should do, but only actually do rarely, and never collaboratively except at extreme cost and increased risk over something organized.

> They don't get "shocked", or hit the wrong pedal, and have faster reaction times.

Yes they do. You could perhaps argue they do so less.

OK, kernel panics aside.

Unless they're using a random number generator seeded with outside data, all computer operation can be called deterministic and 100% predictable. Because it's 100% repeatable given identical preconditions. People? Not so / not currently doable.

If the wrong operation for a situation were performed, it'd be due to a higher-level decision going wrong, not the command. Meanwhile, people can "hit the brakes" (high-level command) and instead slam the accelerator (command not carried out). And fully believe they hit the brakes, and that the car accelerated uncontrollably.

Even discounting buggy software; think hardware, wiring. It is quite possible that a malfunction will take place (they do, constantly).

I could go for "fully computerised traffic would be significantly safer than human traffic", but absolutes about the former being unfailing are just silly.

I think we'll need multiple processors thinking independently and reaching consensus before the gas pedal is pushed and letting any single processor hit the brakes (and disengage the gas, of course). Also good would be to allow other cars to request that a car stop, perhaps even communicating with the traffic control network to make it stop.

And of course, there will be the Android app the pedestrian can use as a last resort to avoid beng run over that stops any cars heading for the phone. :-)

Reminds me of a 0.1-version of this application that I heard about some years ago: some guy saw a fairly fast-moving heading toward a red light with pedestrians crossing with the driver looking down at something. Guesstimating the car was about to be too close to stop in time, the guy threw his cell at the car and got the driver's attention.

it was self defense, officer

Even still, let's imagine they're performing using a pseudo-random number generator, and let's imagine they're using a trained neural network at some part of the program. The training (genetic or back-propagation based) didn't account for a very specific scenario, and the tool mis-classifies it, telling the computer to do the wrong thing.

If you're suggesting that we know what a computer will do on every single given input, we don't. That's in a class of unsolvable problems, as it requires the halting problem to be solved.

[Added the following section after being down-voted]

A "shock", in the deterministic computer world, could be an incorrect classification resulting in an incorrect action. The results are similar to when a human is shocked and doesn't act correctly.

Is the scenario I've painted limited to neural nets? No, any trained classifier, even any program at all may have "minor" slips like this. (When dealing with an analogue world, there's a ridiculously large number of situations that may exist). But you know what? Humans don't perform perfectly; and a computer that's been shown to perform the correct final* response 100% of the experimental time, and 99.9997% of the real time is competent at the activity, and trustworthy.

Disclaimer: I love machine learning, I love neural networks, and I see great usefulness for them in this world; however, I also believe that the perfect combination of variables could yield unusual results. Is it unlikely? Incredibly. Possible? yes.

* Consider a situation where a child is playing with a ball. The computer may not recognize that anomaly on the side of the road as the child and the ball, or even as something to be avoided. If the computer recognizes the anomaly as something that needs more information; then tells the car to slow down until it can acquire that information, then recognizes it as something that needs to be avoided and does so, successfully, safely, and with sufficient space for the conceivable outcomes that may have resulted with this child; then, while the computer couldn't discern the object at first, it did in the end and adapted to the situation appropriately. The "final response" was correct.

[edit: rather than simply down-vote me, please tell me where I'm wrong so I can learn.]

Wow. I don't even know where to start on that one, because that is about 100% the opposite of true.

I work on autonomous driving robots. Teams I have been on in the past won extremely competitive autonomous vehicle competitions. Our robots crashed all the time. ALL THE TIME!

There is nothing deterministic about how autonomous systems work. In fact our algorithms were by definition random number generators! We randomly generated "hypotheses" and then had heuristic tests that figured out which was the most likely trail and then chose it. It worked well, 99% of the time. But that 1% was a killer.

Mind you, since we were winning we were doing better than the other solutions which tried to be more deterministic and use maps or things like that. We were basically stateless which reduces some of the feedback associated with getting messed up in some of your estimations.

The problem is all at the low level, not the high level.

Give a robot a perfect map and it is easy to find a way around safely. The problem comes determining what the hell the camera is seeing (lasers are a lot easier, but not as flexible and more expensive). That is why the bulk of the effort is focused on computer vision not path finding, path finding algorithms are simple.

Sure actuating the pedals is fairly trivial, but to call autonomous systems "deterministic" is scary to someone who knows how fickle they are. I'm a huge advocate of self-driving cars, but we are not there yet, and when we get there we need to do a pile of peer review on the code (which better be open source). You never know what race conditions will blow it all up.

It's deterministic. You can run a fully-accurate simulation of your entire setup in a computer, feed in the current input data, and get the precise output in the simulation that the real one would get.

That's essentially the definition of deterministic. If your machine is powerful enough, you can even do it before the real one completes its operations. Anything deterministic is predictable in this manner.

That you can't predict it, or that you don't have a computer powerful enough to predict it has nothing to do with something not being deterministic, nor predictable. Even a massive, PRNG-influenced neural network is deterministic.

------------

On an aside for people who are thinking of quoting the halting problem at me:

The halting problem requires that the detector returns an answer, and it must be true or false and accurate, or the proof through paradox fails. Therefore, it cannot simulate the operation of what it's detecting, because there could be an infinite loop which would cause it to run without end. You also cannot merely abort it after a while, as that would be either a 3rd return state ("unknown") or an inaccurate response.

Ok yes, if I give the robot the same exact pictures and random generator seeds, it is deterministic.

But in the field, even if it is stationary the pictures are never the same from frame to frame and everything relies on a lot a randomness, and random noise.

Simulators are almost a waste of time because the real challenge is handling that noise, NOT dealing with the information extracted after the noise. Once you have a clear view of the world, everything it is easy, it is getting the eyes to actually understand anything that is hard.

For many of those below, a rebuttal, and some more meta discussion:

I have made no claims that computers can make flawless decisions. I make no claims they can detect / avoid everything avoidable. I don't know where you're getting the idea that I've said this, unless you're dragging in past battles with other people - everywhere I've claimed computers would do better I've included qualifiers like "likely".

The point I was trying to make is that the instructions of a computer, barring external phenomena which cannot be accounted for, are fully deterministic and accurate. If the high-level decisions are made incorrectly, then the end result will be incorrect, but all down the path everything will be accurately performed.

Not so with humans (just look at someone with epilepsy, or at your own memory), or at least not in the minds of some people, which is the only reason I include it. And, if nothing else, we're far further from accurately simulating an entire human body / neural net than from simulating another computer.

Pedestrian collision avoidance technology is already available in the Volvo S60.

I didn't know that. That's pretty awesome :D

Except it doesn't actually work. http://www.engadget.com/2010/09/27/volvo-pedestrian-avoidanc...

I agree computers will be able to drive much better than humans, and this transition will occur at some point--the question is, when? Is the 8 years mentioned in the article anywhere close to realistic? I don't think so.

Not only does the technology have to become better in the aggregate than humans, it has to become so much better there's aren't any crash situations where human drivers perform better. There also can't be many crash situations when the computer doesn't significantly outperform a human. Otherwise both public sentiment against the cars, and legal liability will be too great.

Imagine: whenever there's a software bug and an automated car crashes in any situation where a human might not have, the GoogleCar manufacturer will get sued, the local paper will write an article about how Google killed someone, etc. And then, once a year, there will be another story about how crashes are down 2%... that's a lot of lives saved, but which article do you think will get more attention? And do you think "our cars usually make driving safer, just not the night of October 10th when John was killed" will be a winning defense?

I can't wait for this technology, I think it really will deliver all the benefits the article mentions long-term. But it's a long way off (I hope I'm wrong). I give a lot of credit to Google for investing in it now--this is one of the things I read and think "don't be evil" is more than just marketing-speak.

Keep in mind that there are already millions of lines of code running in modern cars. The worst-case failure mode of all that code would certainly result in loss of life, and legal liability on the part of the manufacturer/insurer/etc. So really the question is just how much automation we'll see in automobiles in the future, not if.

The Toyota Pedal case certainly points towards liability demanding less code.

If there was a automation feature that prevented accidents in 90% of the situations where a human would crash but caused an accident in a situation that was 10% as likely to occur as a human crash, the result would still huge liability since the manufacturer would "definitely" be at fault in the latter case.

Suppose a robot driver is 1/10th as likely to hit and kill a person as a human driver. It would still be something that would happened. And when it did, the results would very messy since the law presently sees any accident through the lens of finding a person at fault. When a robot kills (and a robot will kill if they are "out there") is the fault that of the owner or the manufacturer or the programmer?

I assume the programmer works for a company, and the manufacturer is simply building the device. If the problem was related to faulty components, then I would assume the manufacturer. A flaw in the software or design? I would then contribute it to the company that designed the bot. If the device is improperly used then it would be the owner. I think it all comes down the specifics inside the context.

All of the sorting-out approaches you mention would certainly come into play.

Normally, the programmer is in the clear but obviously there would be exceptions - if you could prove they intentionally introduced code that make the car do something bad, they would be liable. Now introduce a plaintiff's lawyer to muddy-up the definition of these terms.

Let me point out that determining liability/culability/etc in a messy accident is difficult task when you've got just two human decision makes involved. It's decided by jurors who aren't necessary geniuses but probably think they understand human reactions and decision. Introduce two or three more decision makers into the discussion and you've got a messy process indeed.

We saw how far the faulty Toyota accelerator pedal charge went. A big factor in the reactions to evidence was the spookiness of "drive by wire" (I think we've gone over how the actual cause was driver fault but consider how the complexity of the system made a lot of people less likely to believe this).

I mean, it's terribly ironic. Driving in general is one of the most dangerous activities a modern person can engage in. If robots drove our cars the way we drive our cars, the protests would be huge. In a sense, the way we've socially normalized driving is through the huge insurance and liability complex. This assures people that when someone drives wrong, they will be punished somehow. A similar or even significantly small number of accidents without person to punish would be unsatisfying socially.

And I suppose we'll have to get rid of bicycles and pedestrians, too, in this perfect world where intuition of human behavior is unnecessary to safely pilot a vehicle on the road?

Far from it. If cars were less dangerous to bikers and pedestrians, people could actually use bike lanes safely, and cross at walk signals.

Besides. I never said nor implied "perfect". Merely better, and being better is easier without other human drivers, who can far more quickly take a bad (or good) situation and make it worse. Bikes don't go 60+ mph and weigh over a ton. Neither do pedestrians. And very few of either of those carry large quantities of gasoline.

Reply on news.ycombinator.com