Why driving is hard—even for AIs

Despite promises of "soon," the infrastructure to support the driverless future isn't there yet.

Read the whole story

 

[Sonzaisuru]

Wouldn't a wolf have been a more appropriate impediment on the way to grandmother's house?

 

[Nonapod]

Speaking as someone who hit a deer with his brand new car little more than a month ago (doing $5k in damge), I often wonder how certain unpredictable events can be dealt with. In my case the deer bounded out into the road from a dense forest right in front of my car.

To deal with an animal suddenly running into the road properly it would have to be detected first obviously. How difficult do dense trees make such a detection? I think it's easier to detect something in motion than to detect something that's standing still but could be capable of springing into motion like a deer.

 

[l27]

To me driving is so dynamic. Until AI can think as dynamically as I can full autonomous driving isn't going to become a reality outside tightly controlled environments or areas. I can't imagine what a self driving car would do in some of the situations I've been in. We've also seen what can happen when AI gets it wrong.

I also think there will have to be a larger tipping point. Where you'll see "self driving cars only" zones to remove some of the dynamic situations out of the equation. But you still have pedestrians and other movable objects and weather, so I don't know. Eventually when all cars on the road can communicate with each other and the roadway infrastructure quickly and fully, then I'll trust full autonomous vehicles. Until then I would want some level of control.

We're not there yet but I can't wait until we are.

 

[thetay48]

Another factor would be that as long as there are traditional “dumb” cars out there AIs are at a disadvantage since they can’t predict human stupidity. Noticing cats in the road or a car door being opened is one thing but a person deciding to cut across several lanes of traffic to reach an off ramp in 10-15 seconds might be too sudden and unexpected for an AI. Or at least AIs in their current state

 

[GFD]

Discussion of V2V/X/I/whatever communications always makes me worry about security. Like, if a city's network here is compromised and malicious actors start broadcasting incorrect information, what measures would be in place to prevent collisions or traffic jams from occurring? Or even if the system has some sort of software bug. Obviously the car can just observe its surroundings as they do now to see what's actually physically happening, but it would be making at least some decisions based on this information it's receiving, or it wouldn't be useful to have it in the first place. I've never seen discussions about how it would be handled before, so my brain is just full of nightmare scenarios about it.

 

[Dr Gitlin]

Discussion of V2V/X/I/whatever communications always makes me worry about security. Like, if a city's network here is compromised and malicious actors start broadcasting incorrect information, what measures would be in place to prevent collisions or traffic jams from occurring? Or even if the system has some sort of software bug. Obviously the car can just observe its surroundings as they do now to see what's actually physically happening, but it would be making at least some decisions based on this information it's receiving, or it wouldn't be useful to have it in the first place. I've never seen discussions about how it would be handled before, so my brain is just full of nightmare scenarios about it.

It's actually a topic we've explored here at Ars in the past, and will continue to explore in the future.

 

[Frosty Grin]

If we're at the point where V2V is necessary, is it smart to keep on with individual transportation when most people travel in the same direction on the same road? You might as well use your phone to communicate your destination and it will automatically pick out a bus or a shared taxi for you.

 

[Fatesrider]

For a vehicle to even approach autonomy, it has to understand the instant-to-instant changes in its immediate environment and what they mean. It has to know how to react. And it has to know the important nearby things that don’t change—like where houses and trees are.

No, actually, for a vehicle to approach autonomy, it needs to anticipate what is going to happen, not from moment to moment and reacting in that fashion.

A machine acts from moment to moment, and that doesn't always work out very well (see the other Ars article about driverless vehicle woes). The fact is, people don't generally react from moment to moment (only in emergencies, really) and tend to look ahead and anticipate what's coming. It's when they start driving moment to moment that they get into trouble.

I mean, I'm not an AV AI, I've only been driving for about 50 years. I only have about a million and a half miles of driving distance, and I've been in a few accidents (literally all of them due to someone else's stupidity where physics was against avoiding a collision). My last ticket was for speeding when I was 21 in a speed trap in South Carolina. So I may not have a clue about how AV AI's SHOULD work. I only know what humans do.

If the only thing out there on the road were other AI's, then I don't think they'd have the problems (at least to the simplistic degree they do with respect to, say, merging in dense traffic) they currently seem to not be able to deal with. But as long as there are other humans doing the driving around them, they really need to be programmed to anticipate things and act in advance of a condition rather than deal with a condition only when it happens.

From my reads about how they operate, I'm not at all sure they have any ability to do that. And doing that is a key component to safe driving.

 

[Bongle]

Despite the thousands of annual deaths, the human-driven car is really a bit of a marvel. Using two barely-spaced apart optical sensors located inside the car (thus partially obscured by car structure) with a relatively narrow field of view, humans average 100M miles per fatality.

Waymo is on its way there, but they still have to get to 10x their current mileage without anyone dying.

 

[firsttimecaller]

If we're at the point where V2V is necessary, is it smart to keep on with individual transportation when most people travel in the same direction on the same road? You might as well use your phone to communicate your destination and it will automatically pick out a bus or a shared taxi for you.

a lot of this talk has me inventing a bus.

 

[cognizant_ape]

Speaking as someone who hit a deer with his brand new car little more than a month ago (doing $5k in damge), I often wonder how certain unpredictable events can be dealt with. In my case the deer bounded out into the road from a dense forest right in front of my car.

To deal with an animal suddenly running into the road properly it would have to be detected first obviously. How difficult do dense trees make such a detection? I think it's easier to detect something in motion than to detect something that's standing still but could be capable of springing into motion like a deer.

Sounds like a job for infrared.

 

[Maltz]

While I agree 100% with everything the article says, I find it interesting that this is some sort of revelation for some people who seem to think we're just a few years away from having fully autonomous cars that can go nearly anywhere. Or even talking about a time when we could ban manual driving entirely. Maybe I'm just being curmudgeonly, but I don't think we're anywhere as close to commonplace self-driving cars (outside of perhaps interstates and other well-controlled areas) as is generally believed.

 

[ryanja]

Another factor would be that as long as there are traditional “dumb” cars out there AIs are at a disadvantage since they can’t predict human stupidity. Noticing cats in the road or a car door being opened is one thing but a person deciding to cut across several lanes of traffic to reach an off ramp in 10-15 seconds might be too sudden and unexpected for an AI. Or at least AIs in their current state

Came here to say something similar. Creating an AI that gets good enough at predicting human unpredictability (i.e. stupidity) to trust it to have control over a vehicle on the road shared by countless unpredictable human sources is a world different than making an AI assistant who knows to turn on your kitchen lights when you get home or remind you to pick up eggs at the supermarket.

I wonder if someone has plugged in enough data to make an educated guess on what percentage of vehicles would need to be fully AI controlled for them to be considered safe without a reasonable chance of an accident... Like, do we need 50% of the vehicles to be fully automated (removing half of the human factors and having enough of them on the road sharing data about likely reckless human drivers), or does it need to be higher, like 75% or 99% of all the vehicles?
(editing for 4pm spelling errors)

 

[NoMorePosting]

"As smart as a car may be, it needs an equally smart (if not smarter) infrastructure around it"

No, it doesn't.
Humans rely almost entirely on vision to drive, and cameras already have way better resolution than human vision. So theoretically cameras with nothing external to the car are enough. (Which is where I think Elon Musk is coming from with his comment about not needing lidar).

Driving isn't a hard task - humans can do it moderately well after all, the problem is that there are thousands of simple sub-tasks / edge cases, all of which have to be recognised, understood, programmed, tested and used at the correct time.

The processing power we have currently in modern processors is pretty amazing - so as usual its all down to software.

The real problem I see is the bogus use of the term Artificial Intelligence and Deep Learning for the rudimentary software currently available.

 

[el_oscuro]

Another factor would be that as long as there are traditional “dumb” cars out there AIs are at a disadvantage since they can’t predict human stupidity. Noticing cats in the road or a car door being opened is one thing but a person deciding to cut across several lanes of traffic to reach an off ramp in 10-15 seconds might be too sudden and unexpected for an AI. Or at least AIs in their current state

Came here to say something similar. Creating an AI that gets good enough at predicting human unpredictability (i.e. stupidity) to trust it to have control over a vehicle on the roar shared by countless unpredictable human sources is a world different than making an AI assistant who knows to turn on your kitchen lights when you get home or remind you to pick up eggs at the supermarket.

I wonder if someone has plugged in enough data to make an educated guess on what percentage of vehicles would need to be fully AI controlled for them to be considered safe without a reasonable chance of an accident... Like, do we need 50% of the vehicles to be fully automated (removing half of the human factors and having enough of them on the road sharing data about likely reckless human drivers), or does it need to be higher, like 75% or 99% of all the vehicles?

Even if all vehicles had it, the AI would still have to deal with children, pets, animals and all kinds of other random unpredictable hazards.

 

[WereCatf]

While I agree 100% with everything the article says, I find it interesting that this is some sort of revelation for some people who seem to think we're just a few years away from having fully autonomous cars that can go nearly anywhere. Or even talking about a time when we could ban manual driving entirely. Maybe I'm just being curmudgeonly, but I don't think we're anywhere as close to commonplace self-driving cars (outside of perhaps interstates and other well-controlled areas) as is generally believed.

Me, I'll be surprised if fully autonomous cars can handle a real, proper Finnish winter in anything less than 20 years.

 

[NoMorePosting]

Despite the thousands of annual deaths, the human-driven car is really a bit of a marvel. Using two barely-spaced apart optical sensors located inside the car (thus partially obscured by car structure) with a relatively narrow field of view, humans average 100M miles per fatality.

Waymo is on its way there, but they still have to get to 10x their current mileage without anyone dying.

And in all road conditions and speeds, without causing road rage with their timidity.

 

[WereCatf]

Humans rely almost entirely on vision to drive, and cameras already have way better resolution than human vision. So theoretically cameras with nothing external to the car are enough.

Cameras work only if they can see. Higher resolution doesn't do jack shit, if the camera's vision is obscured by e.g. heavy snowfall.

 

[ThinkingMore]

In addition to the 'cat runs across the road' list of unexpected random events, there is another side to driving with other people and that is the anticipation and understanding of intent. In a very simple case like when I slow down to let a person enter the highway in traffic, I may flash my lights or even wave the other driver on, they see me and proceed. An AD car will have a hard time with such a simple thing. If you pay attention, there are tons of similar 'driver to driver' interactions we have as a normality.

A challenge of AD tech is and will be driving along side humans.

 

[ryanja]

Another factor would be that as long as there are traditional “dumb” cars out there AIs are at a disadvantage since they can’t predict human stupidity. Noticing cats in the road or a car door being opened is one thing but a person deciding to cut across several lanes of traffic to reach an off ramp in 10-15 seconds might be too sudden and unexpected for an AI. Or at least AIs in their current state

Came here to say something similar. Creating an AI that gets good enough at predicting human unpredictability (i.e. stupidity) to trust it to have control over a vehicle on the roar shared by countless unpredictable human sources is a world different than making an AI assistant who knows to turn on your kitchen lights when you get home or remind you to pick up eggs at the supermarket. Smart cars, smart roads?

I wonder if someone has plugged in enough data to make an educated guess on what percentage of vehicles would need to be fully AI controlled for them to be considered safe without a reasonable chance of an accident... Like, do we need 50% of the vehicles to be fully automated (removing half of the human factors and having enough of them on the road sharing data about likely reckless human drivers), or does it need to be higher, like 75% or 99% of all the vehicles?

Even if all vehicles had it, the AI would still have to deal with children, pets, animals and all kinds of other random unpredictable hazards.

Great call. I imagine that there could be a sensor network along the sides of the roads that can alert vehicles to children, pets, animals, trees, utility poles, or anything else that may come upon the road. I think that'd be a pretty neat alert for the AI, something like "hey, an object just entered from behind a tree 100 meters up, proceed with caution".

 

[Stochastic]

There are a lot of challenges with V2V and V2I communication. For V2V, a standard needs to be established that allows interoperability between different vehicles. There are a lot of open questions. What information, precisely, will be transmitted? Will the standards evolve over time as the capabilities of autonomous vehicles change and improve? Will very old AVs be barred from the road if they can no longer communicate with the latest communication protocols?

V2I faces a whole different set of issues. How will this infrastructure be secured? Will the infrastructure be managed in a centralized way (kind of like the way all stoplights in LA are synchronized) or will each piece of infrastructure work independently? What happens if power is lost and the infrastructure stops functioning temporarily? Will AVs be dependent on this environmental infrastructure such that they can't operate in rural areas or developing countries?

None of these problems are intractable of course, but they will likely require deep partnerships between the private and public sectors to solve.

 

[NoMorePosting]

Humans rely almost entirely on vision to drive, and cameras already have way better resolution than human vision. So theoretically cameras with nothing external to the car are enough.

Cameras work only if they can see. Higher resolution doesn't do jack shit, if the camera's vision is obscured by e.g. heavy snowfall.

My god, how do humans drive in reduced visibility?

Do they slow down or something?

 

[Stochastic]

Humans rely almost entirely on vision to drive, and cameras already have way better resolution than human vision. So theoretically cameras with nothing external to the car are enough.

Cameras work only if they can see. Higher resolution doesn't do jack shit, if the camera's vision is obscured by e.g. heavy snowfall.

At the risk of sounding glib, a simple solution to that would be for self-driving cars to only operate when visibility is sufficiently high. Also, Google is already working on that problem.

 

[Frosty Grin]

Me, I'll be surprised if fully autonomous cars can handle a real, proper Finnish winter in anything less than 20 years.

Global warming will take care of that.

 

[jdale]

But the car also needs to know the local speed limit and whether it should adjust for the weather, which it also needs to understand. Some of that information may come from remote resources, like highly localized weather information and a machine-understandable municipal map (which, by the way, will probably look like an XML file and not what you’d understand a map to look like). Because high bandwidth and low latency are going to be key, that’s a job for a 5G network.

Why high bandwidth and low latency? The car should not depend on getting a map RIGHT NOW. It should already have the map. The map should be downloaded well in advance, in anticipation of being needed. It does not need to be up to the second. If it waits till the last second to get the map, sometimes there will be a delay and it will not happen.

If there are other cars around when your car takes action to avoid the cat, your car should be able to communicate with them, too, so they can understand instantly whether they need to brake, swerve, or accelerate; otherwise, your car might hit them instead of the cat. To make that work, you need vehicle-to-vehicle communication (called V2V), which will require an interoperable industry standard that doesn’t yet exist.

We're decades away from 100% participation in V2V. That means every car needs to be able to handle interactions with other cars based on the same unreliable signals used by human drivers, e.g. turn signals and brake lights. V2V may be helpful but it can't be relied on. I would also argue that V2V signals should never be highly trusted -- they will be one component of the information received by a car, alongside maps and various types of sensor data, but they will be the easiest to manipulate maliciously. Malicious in this case could be trying to cause you harm, but is more likely to be little stuff like delaying you so another car can go through the intersection without waiting. This is not a technical problem that can be solved with better sensors or certificates, it is one of human nature as long as the car is being used (directly or indirectly) by humans.

 

[JeffFromAbove]

Is this an ad for 5G? Why would a car need 5G to know the weather? In fact, why wouldn’t a car be able to discern weather the same way we do by looking at the environment?

It always perplexes me when someone insists that AI requires lidar, connectivity, and super-human traits in general when humans don’t need any of those things in order to drive. There is no reason to suspect cars can’t learn to make do with the same inputs and likely process those inputs at speeds that are orders of magnitude faster than a human. At the very least, it seems naive to assume that they can’t without more evidence.

 

[ThinkingMore]

I have a couple of kids of learner’s permit age, and it’s my fatherly duty to give them some driving tips so they won’t be a menace to themselves and to everyone else.

I also have a permit aged future road warrior, just starting to log her required hours. As I ride in the passenger seat, the challenge to me is to decide if I need to warn her of an upcoming hazard, and how long I need to wait to see if she recognizes it. I don't want her to learn to just react to everything I see.

When I feel I must say something, I ask "what do you see?" instead of "do you see that stopped car?". That way I know if she recognizes it herself. I ask that a lot.

 

[Legatum_of_Kain]

That’s without mentioning how to handle mechanical breakdowns and system failures.

 

[Celery Man]

The thing that Waymo et al have completely missed is that driving is fundamentally a form of communication between people. It's a social activity that is greater than the sum of our senses delivering information about the environment to our brains. Even minor things like eye contact play an important role in a human's ability to safely navigate alongside other human drivers.

 

[ottovonblotto]

Another factor would be that as long as there are traditional “dumb” cars out there AIs are at a disadvantage since they can’t predict human stupidity.

And those cars will always be there, and legal to drive (at least that's my prediction).

Look at the world of aviation - it's *still* perfectly legal to fly a plane without radio, a transponder, or even lights.

And of course your autopilot always makes the simplifying assumption that there is never anything in the way.

 

[foofoo22]

Well if a small animal runs in middle of road, the best thing to do is hit it (unless you can definitely stop). You swerve and most of time you're going to be the one hurt rather than the small animal.

I know my choice which I would rather be hurt...hope the person programming the car AI feels the same.

 

[Don Reba]

As smart as a car may be, it needs an equally smart (if not smarter) infrastructure around it.

It does not need it. It could benefit from it. In the ways the following examples show. If that infrastructure was required for self-driving vehicles to function, we would not have them driving around at present.

 

[Davinchy]

Seems to me that google has solved the “unpredictable human” aspect of this. Track and catalog every car on the road and put them in a database just like they do for advertising. Now you know that bob in his f150 likes to run red lights and is showing up on your camera in front of a red light.

 

[a nonny moose]

rock slides and flash floods could be eliminated as hazards if we level out the land masses to, oh, say, plus/minus 10 feet and train the weather to just water when and where we need it.. would be good to even out the length of days while we're at it.

 

[Ansible42]

You know what we are super good at making, and can easily make them better and build them very quickly; trains! We have known how to make them autonomous for like 30+ years.

The only drive (lol) for self driving cars is marketing and profits, not engineering. Though as a cyclist I will trust computers more than humans.

 

[Foiler]

The streets alone aren't even kept up around here - potholes everywhere. Street paint is so eroded and worn through you can barely even see them. I don't see even basic markings for self-wrecking cars to be maintained. It would take $1,000,000,000,000's (trillions) of dollars just to make them marginally driveable. Do you think small towns will? Uh, no. Be prepared to pay a big chunk of your income just to streets - if you can even find enough people to work on them... all just for these cars? Not going to happen. I'm trying to be reasonable here, but some people have this pie in the sky mentality.

Some of the turn lanes here are so sharp that if you were to not touch the lines, you would have to swerve so hard your passengers would hit the door/window. Humans drive over the lines all the time. What are one of these cars going to do? Hang on! Oh, then when it's slick, it will slip into oncoming traffic or over a kid/pedestrian. Slam on the brakes to make the turn? Good luck wit that.

 

[jukes_]

Part of the problem is defining and measuring success.

We seem to have this expectation that we want strictly safer outcomes from autonomous driving. People seem unwilling to tolerate any system where the autonomous vehicle makes some mistake that a normal human wouldn't have, even if the overall statistics of autonomous vehicles is much safer.

This seems like it makes it really hard to gain acceptance, given that it's both hard to measure mistakes that the autonomous vehicle "prevented" and that people have a very bad estimation of how good their own driving is.

 

[RickyP784]

I can just imagine the gridlock when the great power outage of 2075 hits: all the infrastructure needed to autonomously pilot vehicles is offline and nobody knows how to drive/has a steering wheel anymore.

Just think of the solar flare of 2124 that fries all of our electronics and satellites!