The actual prospect of autonomous cars?

[Gear300]

What are the prospects of us having the technology sometime in the second-half of century?

I've had this question for some time since I've never tackled actual engineering, but I figured I would ask anyway. Technologists can sometimes sound overly optimistic about things. The likes of "machine learning" does not happen on people's laptops, especially since you can run matlab code which gets very slow once the row size grows larger than on the order of 10,000. The only commercial architectures I've heard that do feasible machine learning are distributed or cloud architectures in which the computing resources are scaled larger than the data involved. Past that, computers are not much smarter than their compilers, or their "grammar organ" (although it seems to be the case that their compilers are at least as competent as human grammar organs; even so, there is lots to language that grammar does not cover).

But can machines limited to their current grammar-memory-&c architectures really solve the riddle of identifying

as a credible stop sign? If not, autonomous fleet and metro would require significant infrastructure enforcing.

When viewed like this, we cannot decouple actual machine intelligence from the hardware or engineering. The reason I asked for autonomous cars is namely that if we do not have autonomous cars by, say, half-way through next century, then it kind of makes one wonder what went on in the meanwhile. If one's technology optimism were in navigation technologies and Earth-space stuff, then innovation in autonomous driving might somehow precede advances in aviation and rocketry (our current rocketry is inefficient from what I hear).

Technology optimism has been reasonable so far. Sure enough, we have not yet reached Mars, and ironically, oil alone would not have taken us to Mars anyway. But we instead unlocked the Pandora boxes of the gene and logic computing &c, which is no small feat. Any opinions?

Gear300

 

[anorlunda]

It might not be 100% black and white. Highway officials have long called for driverless cars on restricted highways. That's very different than driverless everywhere, all the time.

 

[Gear300]

I guess limited access highways would make sense. Well dang, I used to be sure we would have self-driving cars by the end of century. I guess I'll have to look into it more (I have zilch engineering experience).

Amir Hamid

 

[anorlunda]

Many modern airliners are able to take off, travel, and land fully automated. But nobody is suggesting that we take pilots out of the cockpit. It is as much a social issue as a technical one.

 

[FactChecker]

Many modern airliners are able to take off, travel, and land fully automated. But nobody is suggesting that we take pilots out of the cockpit. It is as much a social issue as a technical one.

I don't think I would call it a social issue. The technology would need to be very mature and reliable before it should be completely depended on. There are a million special situations that I can think of. I have serious doubts about automatic systems staying in a lane or finding parking places when snow covers everything. What about driving on a road where the surface is being repaired? What about dodging potholes? What about seeing a board with nails sticking up? etc. etc. etc.
On the other hand, a system that is an assistant to the driver is much easier to imagine. (I would love to have adaptive cruise control or a lane departure warning system in my car right now. )

 

[anorlunda]

Here's the challenge. California freeway officials said that they must provide for a future with 3x as many vehicles per hour on the existing freeways.

No new roads, no road widening. Congestion on California highways is already infamous. There is no way to attain 3x while allowing any human drivers on the road. Even emergency vehicles must fit into the automated pattern.

True, that's a restricted case, but it is real.

 

[Baluncore]

Only humans need to read road signs. Autonomous cars do not need to recognise road signs. The autonomous vehicle can download the local transport maps from the web when they need it. The permitted domain of autonomous vehicles can be defined on the map. The domain can be increased or reduced simply by changing the local transport map. Speed limits and road signs should be embedded as transponders in the road, while traffic light status should be transmitted independently, without any need to see the light. Transponders on vehicles would enhance collision avoidance.

While in an autonomous zone, a human driven car should enforce speed limits, traffic lights and stop signs. Older cars would need only a transponder to identify the vehicle, and that would provide audio warnings to the driver.

 

[Vanadium 50]

What is the difference between "prospect" and "actual prospect"?

 

[anorlunda]

Speed limits and road signs should be embedded as transponders in the road,

Not needed. My GPS already knows the speed limit from database data on every street.

 

[FactChecker]

Not needed. My GPS already knows the speed limit from database data on every street.

My GPS does not always have me on the correct road when two roads are side-by-side (like an interstate and its access road). I would prefer something embedded in the road.

 

[Baluncore]

My GPS already knows the speed limit from database data on every street.

My GPS gets updated with changes only every 6 months. Roadworks with temporary speed restrictions need a transponder immediately. Likewise, until all cars are autonomous, accident scenes could be controlled better by transponders.

 

[Office_Shredder]

My GPS does not always have me on the correct road when two roads are side-by-side (like an interstate and its access road). I would prefer something embedded in the road.

This. The number of times you're going down a highway and it thinks you got off an exit. Or you get off an exit and it doesn't know. Getting 30 vs 60 mph wrong in either direction would be a disaster.

I think the ai in the car could help correct gps errors like this. The camera can see if you actually took an exit

 

[FactChecker]

I think the ai in the car could help correct gps errors like this. The camera can see if you actually took an exit

That's a good point. I don't know what the current capabilities are, but it seems like this problem must have already been solved if they are testing self-driving cars now.

 

[Baluncore]

If you tag the road with transponders, then you do not need GPS, nor AI to fix the problems with GPS.

 

[Office_Shredder]

If you tag the road with transponders, then you do not need GPS, nor AI to fix the problems with GPS.

Then you're at the mercy of the local authorities maintaining the road tags as things change and they break.

Sorry, the transponder for the speed limit broke here. Everyone needs to just sit tight until a new one is installed.

 

[Baluncore]

Then you're at the mercy of the local authorities maintaining the road tags as things change and they break.

Sorry, the transponder for the speed limit broke here. Everyone needs to just sit tight until a new one is installed.

How often is a road closed because a speed limit sign has been damaged?

Transponders only need to be the size of an RFID tag or a credit card. The replacement would be very simple using 5 minute epoxy. Street signs get damaged in accidents, and vandalised. Signs take longer to replace than a transponder.

Transponders are very reliable. If they were unreliable, then put in two, a few metres apart.

 

[anorlunda]

I think we're just arguing a tit-for-tat back and forth between friends and foes of driverless cars.

To me, the most relevant question is the standard of performance. Are we holding machines to a different standard than humans?

I remember well my experiences as a firefighter, trying to protect an accident scene from Neanderthal human drivers. Once I had to smash the windshield of a driver to get his attention enough to stop. There were victims lying in the road in front of him. That's the status quo for human drivers.

 

[FactChecker]

I think we're just arguing a tit-for-tat back and forth between friends and foes of driverless cars.

To me, the most relevant question is the standard of performance. Are we holding machines to a different standard than humans?

I remember well my experiences as a firefighter, trying to protect an accident scene from Neanderthal human drivers. Once I had to smash the windshield of a driver to get his attention enough to stop. There were victims lying in the road in front of him. That's the status quo for human drivers.

I think there is great value in self-driving cars if there is an option for a driver to override it. But one of the recent tests was a car with no steering wheel. That seems like a mistake. A couple of years ago my iPhone navigation told me to turn the wrong way on a one-way street and when I turned the other way it kept directing me around the block to the same wrong turn. Those things need to be corrected before the possibility of driver control is removed.

 

[anorlunda]

A couple of years ago my iPhone navigation told me to turn the wrong way on a one-way street and when I turned the other way it kept directing me around the block to the same wrong turn. Those things need to be corrected before the possibility of driver control is removed.

Once again, do you have an objective standard of perfection to be applied to both humans and machines? Endless anecdotes will get us nowhere.

 

[FactChecker]

Once again, do you have an objective standard of perfection to be applied to both humans and machines? Endless anecdotes will get us nowhere.

For what design option? There are many possibilities in a complete spectrum of levels of automation. Completely automatic with no possibility of driver override? A passive driver always monitoring an active automatic system? A passive automatic system constantly monitoring an active driver? An automatic assistant for subtasks?
I think that progress will happen (are happening) in steps of increasing automation as the technologies continue to develop to improve safety.

 

[FactChecker]

Once again, do you have an objective standard of perfection to be applied to both humans and machines? Endless anecdotes will get us nowhere.

In completely automatic systems, every anecdote identifies a system design flaw that potentially represents hundreds or thousands of malfunctions. I have no idea how many people would have been sent the wrong way down a one-way street or circling endlessly around the block if steering wheels are removed.

 

[Baluncore]

Once again, do you have an objective standard of perfection to be applied to both humans and machines?

"Perfection" is 100% perfect, and not 1% less.
In the same way that something cannot be 99% "unique".

 

[Office_Shredder]

How often is a road closed because a speed limit sign has been damaged?

Never, but if you think that's because the sign never gets damaged, then I guess we're at an impasse.

 

[gmax137]

Once again, do you have an objective standard of perfection to be applied to both humans and machines?

Good question.

Do we have an agreed to "purpose" for the autonomous system? What is the rationale? Is it because the human driver system used now is deficient? (how? too many accidents, too many drunk drivers?). Or is the autonomous system just "too cool" to not do it?

I think we need a "problem statement" before the pros and cons of hypothetical solutions can be discussed.

 

[anorlunda]

I think we need a "problem statement" before the pros and cons of hypothetical solutions can be discussed.

Well there are unlimited ways to express the problem. But post #6 contains one of them. No direct replies to that challenge.

 

[FactChecker]

Well there are unlimited ways to express the problem. But post #6 contains one of them. No direct replies to that challenge.

Is it expected that self-driving cars can meet that challenge? I know that experts think it will allow faster traffic flow, but getting 3x more traffic flow in LA seems optimistic to me.

 

[anorlunda]

Is it expected that self-driving cars can meet that challenge? I know that experts think it will allow faster traffic flow, but getting 3x more traffic flow in LA seems optimistic to me.

Cars heading for the same exit could be coupled in a train. That reduces the number of independent vehicles. The number of exits is strictly limited. A train could also keep going with one or more failed cars in the chain. Individual cars in the train could not change their minds or ask for a sudden unplanned rest stop.

I don't claim that is an actual proposal. I'm just trying to illustrate that we might have to re-think the paradigm to accomplish major change.

 

[gmax137]

Well there are unlimited ways to express the problem. But post #6 contains one of them. No direct replies to that challenge.

No new roads, no road widening. Congestion on California highways is already infamous. There is no way to attain 3x while allowing any human drivers on the road.

OK, but call me skeptical that the autonomous system can run at 3X the speed. Or maybe it can: looks like that freeway is going only about 12 mph (that's 20 kph or 5.4 m/s). Still, running the cars that close together at 35 or 40 mph would be kind of terrifying. I'd rather be in a train.

EDIT - written while other posts in progress. By "train" I mean rail transit.

 

[FactChecker]

Still, running the cars that close together at 35 or 40 mph would be kind of terrifying. I'd rather be in a train.

I would hate to be in that when a car got a flat tire.

 

[Office_Shredder]

OK, but call me skeptical that the autonomous system can run at 3X the speed. Or maybe it can: looks like that freeway is going only about 12 mph (that's 20 kph or 5.4 m/s). Still, running the cars that close together at 35 or 40 mph would be kind of terrifying. I'd rather be in a train..

This is a false choice i think. If your traffic flows more smoothly you won't get jams that cause cars to pile up on the road like that in the first place. At an example if your road can exit 1k cars per minute and 1.1k cars per minute enter the road, after an hour you have 6k cars that need to sit somewhere while they wait in line to exit. You only need to improve the exit speed by 10% to have 0 sitting traffic.

 

[gmax137]

If your traffic flows more smoothly you won't get jams that cause cars to pile up on the road like that in the first place.

That's the beauty of trains: there's one controller (the train engineer) for all of the throttles and all of the brakes. Auto traffic has to contend with each individual driver braking almost at random, and the reaction times stack up, inevitably leading to backups.

I'm slowly convincing myself that autonomous cars, under some kind of central control, could have potential...

 

[Baluncore]

It is necessary to consider the period of transition from human to autonomous control. There will be a significant period while mixed traffic is using the same infrastructure. The end point must be reached via a safe series of functional transitions.

 

[swampwiz]

I think the development will play out such that there will be 2 modes of operation - one for main arterial roads and one for non-arterial rodes. The main arterial roads will be well-mapped, and even some transponder devices could be planted along them to aid in navigation, and so the operation can be at full speed. The non-arterial roads will be mapped to some extent, and so the operation will be quite slow (like 15 mph) just to be safe. I can see new subsivisions being built (or existing ones that are self-contained being upgraded) with transponders to make operation there be at the regular speed.

I fully expect to own such a driverless car by 2030.

 

[StatGuy2000]

I would suggest this clip from EconTalk, featuring computer scientist Melanie Mitchell specifically discussing the prospects of self-driving cars, as part of a broader discussion from her book, "Artificial Intelligence: A Guide for Thinking Humans" (which I've read and highly recommend):



Her conclusion (from the book, and from the brief clip here) is that true trustworthy fully self-driving cars are still a long way away, as such a system must be able to react to unpredictable events that they could not have been able to be trained on from training and test cases.

 

[anorlunda]

I would suggest this clip from EconTalk, featuring computer scientist Melanie Mitchell

Her opening statement includes a questionable assertion.

"Problems that humans can solve because they have knowledge. Let's say. But machines can't solve because data is not knowledge."

I also reject her main point that self driving cars get rear ended too often because they stop for objects on the road that are difficult to identify. I say that human drivers are at fault if they make snap decisions to run over some objects. That plastic bag on the road might contain a kitten. A child's ball rolling toward the road might be followed by a child. So if I stop for any object in or near the road and get rear ended, the collision is not my fault. Ditto for an AI driver.