Why Self-Driving Cars Are Still a Long Way Down the Road

moon_unit2 writes "Technology Review has a piece on the reality behind all the hype surrounding self-driving, or driverless, cars. From the article: 'Vehicle automation is being developed at a blistering pace, and it should make driving safer, more fuel-efficient, and less tiring. But despite such progress and the attention surrounding Google's "self-driving" cars, full autonomy remains a distant destination. A truly autonomous car, one capable of dealing with any real-world situation, would require much smarter artificial intelligence than Google or anyone else has developed. The problem is that until the moment our cars can completely take over, we will need automotive technologies to strike a tricky balance: they will have to extend our abilities without doing too much for the driver.'"

Comments

Don't have to be perfect, just better

[JDG1980]

This writer makes a fundamental mistake: believing that if full driverless technology is not perfect or at least near-perfect, it is therefore unacceptable. But this is not true. Driverless technology becomes workable when it is better than the average human driver. That's a pretty low bar to clear. I know all of us think we're above-average drivers, but there are a lot of really bad drivers out there, and even a flawed automatic system could do a better job.

Re:Don't have to be perfect, just better

[icebike]

I know all of us think we're above-average drivers, but there are a lot of really bad drivers out there, and even a flawed automatic system could do a better job.

That depends entirely on the failure mode.

"Fail to Death" is actually acceptable to society as a whole as long as the dead person held his fate in his own hands at some point in the process. This is why we get so incensed about drunk drivers that kill others. A person doing everything right is still dead because of the actions of another. But if you drive under that railroad train by yourself, people regard it as "your own damn fault".

When the driverless car crosses the tracks in front of an oncoming train it will be regarded differently. Doesn't matter that the driver was a poor driver, and had a lot of fender benders. Most of those aren't fatal

In spite of that, I believe Google is far closer than the author gives them credit for. They have covered a lot of miles without accidents.

Granted, we don't know how many times there were errors in the Google cars, where one part of the system says there is nothing coming so change lanes, and the human or another part of the system notices the road is striped for no-passing and prevents it. Google is still playing a great deal of this pretty close to the vest.

Re:Don't have to be perfect, just better

[Jeremi]

However, how would it react to sudden situations at high speed? According to the article and everything I know so far... not well.

In principle, at least, an automated system could react better than a human to sudden emergency situations... because a computer can process more input and faster decisions than a human can, and also (just as importantly) a computer never gets bored, sleepy, or distracted.

Dunno if Google's system reaches that potential or not, but if not it's just a matter of improving the technology until it can.

Re:Don't have to be perfect, just better

[adamstew]

Having self driving cars would eliminate the "drunk driving" cause of traffic fatalities, as well as all of the "distracted driving" fatalities as well. That is, according to your own numbers, 47% of all driving fatalities.

Also, according to Wikipedia The speed that someone was driving was listed as the cause for 5% of fatal crashes, and "driving to fast for the road conditions" was listed as another 11% of fatal crashes. A self driving car would also, probably eliminate almost all of these crashes...so there is another 16% of all fatal crashes.

In fact, there was a study done in 1985 that concluded that the human factor was SOLELY responsible for 57% of all fatal car crashes. It also found that the human factor was entirely or partly responsible for 93% of all fatal crashes...these are the situations where there may have been bad road conditions or other outside factors, but the driver didn't react in the best way to avoid an accident. These numbers are almost 20 years old, but they are probably still representative of the current stats.

A self driving car should eliminate nearly all of the 57% of crashes where the human factor contributed solely. It should also eliminate a healthy portion of the remaining 36% of fatal crashes where the human factor was a contributor as the car can be programmed to respond in the best possible way to a huge number of road conditions.

Based on these numbers, I believe it is reasonable to say that self driving cars should eliminate about 75% of all fatal crashes. Technology and machinery is also, when compared to the human factor, extremely reliable...particularly when it is designed correctly. I have no problems saying that self-driving cars will eliminate 75-90% of all fatal crashes.

I am also certain that there will be some outlier situations where, if the driver had been in control the entire time, that a fatal accident would have been avoided...The technology failed and the car drove over a cliff, or the signal sent to cars about a railroad crossing didn't activate, etc, etc... but those incidents would be offset by a HUGE margin with the number of incidents that the self driving cars prevented. These same arguments were made against seat belts! There have probably been several examples where someone who was wearing there seat belt drove in to a lake and drowned because they couldn't get free of the car because of the seatbelt. But it is proven that seat belts save FAR many more lives than they cost.

Re:Don't have to be perfect, just better

[jonesy16]

I have to respectfully disagree. All of the situations that you've mentioned thus far are well within the realm of possibilities given current and near term technological advancements. Human beings will always be limited to 1) their imperfect memory of the route being driven assuming they've driven it before and 2) their sole source of input which is the visual electromagnetic spectrum. Driving doesn't require true AI, in my opinion. There are only so many things that can happen from a programmatic standpoint and it really boils down to collision avoidance. You have a route and a volume of space that you occupy along that route at any given time. Either something (an object, person, animal, etc) is going to occupy the same volume at the same time or it isn't. Collision detection is very easy to program, and the technology is sufficiently advanced at this point to be able to detect objects both big and large and make real-time assessments to determine the action that leads to the best chance for survival of both the object and the car. Those calculations are performed by a computer operating much faster and with near-instantaneous reaction time compared to its human equivalent that has to spend time deciding whether it's best to accelerate, brake, swerve (or a combination of those) and then perform the muscle actuations to initiate that action.

Remember, too, that a computer system can have access to near-perfect data such as GPS records for the route, as well as other object/road input systems beyond just the visual spectrum. That's not to say the right this minute we have a perfect set of data for every road, but certainly for the majority of traveled roads we have a pretty complete picture which could be used to provide the car's route in the absence of visual feedback. When it snows here, 4 lanes turns into 2 because humans can't see the lane boundaries, but that's not a limitation for a computer system programmed with the road trajectory to within inches. You may be able to interpolate where the road is 300 feet in front of you, but someone not familiar with the area might not. I drive on some country roads around here where, during a blizzard, you have no feedback about the roads location except for the random house every 1/4 mile. If I didn't know that road was perfect straight, I'd be off of it in no time. An advanced optics system can see further, clearer and more completely than ANY human.

I'm not trying to say that this whole process is trivial and there are reasons why it will take a long time to develop and implement. But I don't believe that there are any reasons which can't be overcome with present day technology.

Maybe AUTO drive only express lanes that cars will

[Anonymous Coward]

Maybe AUTO drive only express lanes that cars will go at high speed at near bumper to bumper

Answer

[WilyCoder]

Because Skynet, that's why.

Re:What's wrong with Google cars

[compro01]

What happens next?

The driving computer sees the 4 feet of water ahead using the cameras/radar and stops because it determines the water is too deep to ford?

Assisted driving first, insurance second

[EmperorOfCanada]

My prediction is that people will be so resistant to just letting go of the steering wheel that the major car companies will give up with that route and pursue having super assisted driving. That is basically cruise control on steroids. Already companies like Mercedes have cruise control that will maintain a safe distance from the car in front, matching their slower speed or even emergency braking if needed. Other cars will do what they can from having you change lanes and side-swiping another car. So I suspect that all the robot driver technology will end up holding your hand more and more. Technically you will be the driver but the robot will be ready to prevent stupidity and also react when you don't. After a while it will finally reach a point where you can just take your hands off the wheel (the car will probably bleat plaintively) and the car will maintain speed and the lane. But nobody will call it robotic driving.

But then the breakthrough will be that some company that has crossed some critical line of self-driving capability will say that full liability insurance is included with the price of the car. Potentially they will even cover all insurance short of trees falling on the car and whatnot as they will be sure the car can't cause an accident and that with all the cameras and sensors that some other fool can't blame you or your car if they are the cause of the accident.

At this point my money would be on cars finally being marketed and sold as robotic self-driving cars. Shortly after this the tidal wave will wash away all the non-robotic cars as being a dangerous menace. The key here is that most cars by this point will be largely capable of being autonomous or very close to autonomous with only antiques being the hold outs.

But, and the big but, is that some robotic car will drive off a cliff or into a train or whatever and that single incident or small collection of incidents (and their Youtube videos) will get everyone saying, "Those things are death traps, I'll never let the car drive." This will temporarily postpone the inevitable but going from 35,000 US annual road deaths to 35 will be too much reality for foolish people to fight for long.