Yeah, yeah. I'm not giving up on self-driving cars, but it's true that things haven't been going as swimmingly as I'd thought they might. It turns out to be a hard problem!
Then again, my old friend Atrios is stuck in Philadelphia. I'm in Paris and just got back from a glorious day at the French Open. So who really has the better of it right now?
Yes, I was really there.This is Henri Cochet, one of the "Four Mousquetaires" (really) of French tennis fame in the 1920s.In a third-round match at the new Simonne Mathieu show court, Jil Teichmann of Switzerland (in the near court) played Victoria Azarenka of Belarus. Teichmann had a huge cheering section and eked out a win by tiebreaker in the third set.
55 thoughts on “OK, sure, but . . .”
Brett
2017 was Peak Self-Driving Car Hype. I remember that was when Uber and others basically thought it was just around the corner, and we'd have self-driving taxis in 5 years. There's not that much optimism anymore, and Waymo in particular seems to struggle at getting their cars to operate outside of where they've trained them in Phoenix.
It still is coming, and the technology is getting better. But there's more skepticism that we will see it before 2030.
On a similar topic, I thought this essay from a guy at a now-defunct Self-Driving Truck company was fascinating. He talks well about both the technical and business obstacles to bringing this technology to market, namely-
1. It's one thing to have a self-driving feature that can work once in a while, far harder to have it reliably work over and over again.
2. Investors were much more interested in hype and the promise of super-big returns than a pretty profitable business. They also basically expected founders to lie to them and promise the world in new features that largely didn't work.
realrobmac
To me and the other skeptics, your #1 has always been obvious.
Getting a car to follow well laid out lanes on a controlled access highway on a sunny day with no construction or other weirdness going on is a pretty easy problem really. But it's all the other stuff that is complicated and a real self-driving car needs to work 100% of the time. 99.9% is nowhere near good enough.
Will we ever have self driving cars that really work 100% of the time. I think we probably will be it's a long long way off and will probably take a major infrastructure effort, with various tech assistants to self driving distributed all around, embedded in roads, on poles and street signs, with a dedicated communication network for self driving cars to communicate with one another and so on.
I know it's not the same thing but I have a relatively new car equipped with LIDAR and all kinds of driving assist features, and the number of times this thing misfires is just incredible. Plus it doesn't work on rainy days.
The notion that pure self driving cars without some special assist from a network like this will simply work reliably on their own will probably die over the next few years.
Special Newb
Yes, more likely that the road will drive the car tbh.
DButch
I also have adaptive cruise control - fortunately it works in even moderately heavy rain - using radar. I haven't tested our new 2022 Leaf in tropical deluge level rain though. There are a lot of other detection and warning features that make driving easier and safer. I discovered that the emergency braking was capable of being quite aggressive when somebody passed me on the left, cut in front of me, and then realized he was too close to the car ahead and slammed on his brakes. My Leaf slammed on the brakes too, and just ahead of my reaction time.
Nissan does make it very clear that it's NOT autopilot though. If you take your hands off the wheel for any length of time, it will get very "unhappy", start the emergency flashers going, then start slowing down all the way to a dead stop. The sales guy had me try it on a deserted stretch of road during the test drive. Same thing happens with adaptive cruise control on - if the cars ahead of you start slowing down it will also slow down - right to a dead stop. If they start up again within (IIRC) a few seconds, the Leaf will start moving. If not, you have to apply the "gas" yourself.
Brett
I guess technically people don't work 100% of the time as drivers (most accidents are from driver error), so you wouldn't need it to be perfect. Just much better than regular drivers.
PostRetro
The reality is that self-driving car technology is already here and enhancing parts of the driving experience. Whether it is adaptive cruise control or adaptive braking, the cars are taking over parts of the driving experience. Real-time traffic and navigation integrated with parking availability, and points of interest are simply the behavior training that provides a higher comfort level for fully autonomous vehicles.
rrhersh
Yabbut you are making a distinction between self-driving car technology and self-driving cars. Neither the hype nor the skepticism was ever about adaptive cruise control or adaptive braking. It was about getting into the car and taking nap while the car took you to your destination.
wvmcl2
I always thought that 2030 was a more realistic date of when we might see self-drivers in common use, and at this point I think I would bump that to 2032-35.
My feeling has been that the legal, political, organizational issues would hold it back much more than technological issues would. That's still true, but it also seems the technology issues are more challenging than a lot of people thought at first.
Paris is indeed perhaps the most beautiful and livable city in the world. It is my favorite city and the best part is that you don't need a car—self-driving or otherwise.
wvmcl2
Besides having the Metro, Paris is now very cyclable as well - last time I was there a few years ago I cycled all over the city. It was fantastic, but wouldn't have been safe to do twenty years ago.
Yes, I must agree that Paris is getting better all the time. I also must confess to having badly misjudged Anne Hidalgo, who is proving to be the best mayor ever.
This country has a major party that’s irreversibly and utterly committed to neoliberalism. They’ve had power several times in the past decades. Look around and tell me that they get things done.
No, it didn't "turn out" to be a tough problem.
It was always a tough problem. Some people knew that way back when, and some people just glossed over the obvious difficulties and plowed ahead making ridiculous predictions..
dspcole
Wow! I’m envious!!
Of you in Paris, not the car….
George Salt
Admit it, Kevin. You got suckered by a world-class conman. Here's a pretty good article that explains what happened:
My optimism about self-driving cars never had anything to do with Elon Musk.
iamr4man
Have you seen this:
“ Waymo will begin operating self-driving taxis in San Francisco without a person behind the car’s vestigial wheel. In a March 21 blog post, it said it was “ready to begin introducing the Waymo Driver in fully autonomous mode” and called it “a major step on our path to deploying a fully autonomous commercial service.” https://qz.com/2146123/waymo-self-driving-cars-will-go-fully-driverless-in-san-francisco/
MontyTheClipArtMongoose
Wither our robot future?
George Salt
Poor Elon. Now that he's outed himself as a MAGA-loving Trumpkin, all of his erstwhile progressive fanbois have disowned him. Elon Who? Never heard of the the guy!
rick_jones
The article of Kevin’s to which Atrios links was singing the praises of Waymo. Pretty sure Waymo was not Musk…
Forget Tesla. Forget Uber. Forget Ford and GM and Toyota. All of those guys will have driverless cars eventually, but they’re far too prone to talk too big and deliver too little.
Waymo is just the opposite. They’re the company that’s been working on driverless technology the longest. They’ve driven by far the most miles. They have the most sophisticated modeling and test software. And they don’t make endless promises they can’t keep. They speak softly but carry a big stick. If they say they’re opening a real live driverless car service by the end of the year, then they probably are. And it will probably work.
MontyTheClipArtMongoose
Waymo: a Man & his dream.
Crissa
Technically, Tesla drives more miles - just not in automated mode. They have 100x more units in testing and 1000x more units collecting data and shadow driving.
But Waymo is doing it now, with actual units in level 4 autonomy.
cld
Well I spent the day doing the laundry, then the dishes, then I thought I'd take a shower but I didn't have the strength, then I had a migraine and laid in bed for a few hours, then I tried to read and it came back, and now you tell me I'll never have a self-driving car and you're at the French Open.
Steve_OH
When a human driver encounters an entirely novel scenario, say, rounding a curve on a twisty two-lane road in the Rocky Mountains and seeing a zebra straddling the yellow line, that driver is likely to react in a reasonable, if not necessarily optimal, way.
The current generation of (sort of) self-driving cars, will, in effect, panic, and the outcome is much less likely to be reasonable. This is because of the way that they're trained. The problem with seeing true self-driving cars anytime soon (and, to be honest, with machine learning in general) is that no one knows how to train them to handle novel scenarios. At all.
I'm sure that the problem will be solved at some point, but because we really have not the slightest clue how to solve it at this time, I don't expect to see true self-driving in my lifetime.
rick_jones
When a human driver encounters an entirely novel scenario, say, rounding a curve on a twisty two-lane road in the Rocky Mountains and seeing a zebra straddling the yellow line, that driver is likely to react in a reasonable, if not necessarily optimal, way.
Though not from the zebra’s point of view I suspect…
ScentOfViolets
Sometimes known in the biz as "Nicholas Cage is a ripe avocado." The almost-canonical example is training a network to recognize handwritten digits 0-9 on datasets comprised of the same ... and then presenting it with the letter 'A'. What does it get mapped too? There are only ten possibilities, non of them, siginificantly, a letter of any type.
J. Frank Parnell
Well we have passed the peak of irrational exuberance for self-driving cars and are now well into the trough of disillusionment. We will see how long the continued development to a useful product takes, but it might not be in Elon Musk's lifetime.
pjcamp1905
Yeah, I believe I told you it was hard a few years ago, along with all the other things you think AI is just on the cusp of. Those are also hard.
The overall problem is that it is mostly machine learning. Machine learning is amazing but it is basically just remembering what happened in the past. It is amazing how much can be done just using memory, but it should be obvious that not everything falls in that category.
Self driving cars made rapid advances when they were learning how to do those things which require only memory. Now they've hit the things that don't and development has slowed dramatically.
golack
The most promise will be for hybrid systems. Self-driving trucks in a convoy led by a human driver--in overnight runs to avoid traffic, and drivers pick them up for the "last mile' or 10.
Crissa
We could do this now - the hurdles are regulatory and manufacturing.
ScentOfViolets
If the network is just memorizing, we tend to call this 'overfitting' 🙂
KenSchulz
+1
golack
But how did Kevin get to the French Open? I will venture to guess it was not a self driving car.
kgus
FWIW, I see -- here in Phoenix -- at least one Waymo car on the streets every day. Never saw so many before this spring (and I drive a lot).
Anandakos
Deflect much, Kevin?
golack is exactly right: drivers for long-distance trucks are THE biggest problem our supply-chain fuster-cluck faces right now. If a driver-team could take ten 54-footers from LA to Tucson, another to Midland, a third to Jackson and a fourth to Atlanta or Charlotte that would get TEN times the stuff to Spartanburg in the same amount of time that a single driver team takes today.
No, it wouldn't solve the GHG issue -- only trains can do that, but they aren't fast and reliable enough for a lot of shipments. They are getting better but the rails know that those truck-trains are coming very soon, so they don't want to invest the money to triple-track every mile between LA and Sierra Blanca (UP) or Clovis (BNSF) and double track the rest of the routes to Atlanta. That's what it would take to APPROACH the speed and reliability of trucks.
Zephyr
As several mentioned above humans are pretty good at dealing with ambiguity and unique situations that computers just can't deal with. The other day I was reminded of this seeing a delivery van deal with an unusual overhanging metal ramp in a parking area. Most of the ceiling of the place had plenty of height, but this one metal edge was hanging down and would have acted like a can opener for any tall enough vehicle that drove close to it on the edge of the driving lane. It wasn't marked at all. Sure, in an ideal world this thing would have not been there or covered with bright colored warnings or something, but it wasn't. The van driver was able to drive past it with maybe 6 inches to spare. Would a self-driving van have been able to do that? Or, the other day I had to slow up and swerve around some construction worker on the edge of the road who had extended a long, barely visible level out across the driving lane. What caught my attention from a distance was this guy just standing near the road in a place where people don't normally stand, even though off the road. Would a computer recognize that? These daily driving problems are just endless.
Salamander
My sf movie group just watched and discussed Blade Runner. Back in 2019, we were supposed to have "replicant" humanoids who were indistinguishable from regular humans, only stronger, faster, and smarter (and which we cheerfully enslaved). We were supposed to have massive migrations off-planet to other worlds, made ready for us by those replicants. We were supposed to have flying cars!
But they didn't drive themselves.
Salamander
Oh, and they didn't have cell phones in "Blade Runner", either. It's funny what writers will think is "futuristic" at any given time.
Austin
Who could’ve guessed that self driving cars would be so difficult? I mean, except all the academics who’ve studied the subject for the last half century of course.
Austin
I wonder if this unable-to-predict-every-possible-encounterable-problem problem is why we still have pilots in planes…? 🤔
KenSchulz
Precisely. It has always struck me as odd that so much effort was put into automating driving, instead of first tackling much easier problems like trains and ships. (Automated takeoffs and landings of aircraft in a range of conditions is already solved, but, as you say, the pilots are there to deal with the unusual). But of course it was because the automobile market is so large; nobody was going to get obscenely rich selling automation for shipping or rail.
royko
As you often find, it's solving that final 10% or so of a problem that makes engineering challenges difficult. Unlike Atrios, I do think we'll get there, but I think it will be a decade or two further out than you, Kevin. I also think it will also need to come with some civil re-engineering to make roads more self-driving-car friendly.
Have fun in Paris!
Goosedat
Bourgeois competition is glorious.
KenSchulz
Being an engineering psychologist, I have long had questions about how self-driving cars will be integrated into the automotive-transportation system. For example: Other than traffic jams/road work, when were you last on a road where the average speed of traffic was at or below the posted limit? Will self-driving cars be allowed to exceed limits? To match the average speed of other traffic? The 75th-percentile speed? What speed will be allowed when there is no other traffic nearby? For another example: how aggressively will self-driving cars be allowed to drive? Will drivers have control over that? Will aggressive drivers continue to drive their cars, leaving already-careful drivers as the market for automation?
George Salt
And how will human drivers behave when 50% of the cars on the road are autonomous vehicles? Will people habitually cut off autonomous vehicles, knowing that an autonomous vehicle will always break? Will teenage knuckleheads start playing "chicken" with autonomous vehicles just for kicks?
You will have to have some roads where you have to drive strictly by the laws, where self-driving cars can go, and some roads where you can drive like you drive now, where self-driving cars will not go.
Considering the amount of detectors that self-driving cars must have, anybody that drives badly on the strictly-legal roads will be caught immediately.
The non-strictly-legal roads will probably become crazies only in short time.
rrhersh
My other example is from my daily commute, on semi-rural roads, typically one lane each direction with a solid double yellow line down the middle. On trash day you find yourself behind a truck stopping every few feet. In practice this is less of a problem than one might think. You edge the nose of your car around to see what is on the other side, then zip around the truck. The trash guys typically will signal if there is oncoming traffic you can't see, but the path of wisdom is to wait until the truck has crested the hill or gotten through the curve. Regardless, this is all technically illegal, as it requires crossing that double line. What would a self-driving car do? I have asked this question many times and never gotten a straight answer.
cheweydelt
This kind of good natured ribbing is what I love most about the blogosphere.
Special Newb
Just say it. "I was wrong and Atrios was right." YOU CAN DO IT!
rrhersh
"It turns out to be a hard problem!" I wrote a blog piece back in either 2015 or 2016 asking a series of awkward questions about how self-driving cars would work. My questions weren't answered, but a lot of people were eager to assure me that I was a luddite. Most of the questions are still unanswered. It was obvious at the time that the enthusiasts were reporting that 95% of the problem had been solved, and that solving the remaining 5% was obviously a trivial matter.
Yikes
I have had a Tesla with the FSD suite since 2019.
Oh, its solvable, and will be solved. I mean, as to when, I have not clue. From the improvements I have seen over the past three years none of them have been predictable in the sense of "if today the car can do X, then three months from now or three years from now it will be able to do Y."
The problem is not the "driving part" - my car already drives on FSD better than I do, its the recognition part.
There is currently the tech argument going on between cameras, radar, and lidar, all plus GPS. I am not qualified to comment.
However, the huge problem of converting camera input to recognizable objects is proceeding nicely. Once all the objects are recognized having the car respond appropriately is not going to be a problem.
I used to wonder whether human "anticipation" of driving events was going to be unsolvable, as FSD mostly relies on "reacting" rather than anticipating. But seeing how well a computer, who never gets tired or distracted, can react has caused me to not wonder about this any more
Self driving cars won't need to anticipate, basically. Humans need to anticipate to be a good driver because our reactive abilities are limited.
2017 was Peak Self-Driving Car Hype. I remember that was when Uber and others basically thought it was just around the corner, and we'd have self-driving taxis in 5 years. There's not that much optimism anymore, and Waymo in particular seems to struggle at getting their cars to operate outside of where they've trained them in Phoenix.
It still is coming, and the technology is getting better. But there's more skepticism that we will see it before 2030.
On a similar topic, I thought this essay from a guy at a now-defunct Self-Driving Truck company was fascinating. He talks well about both the technical and business obstacles to bringing this technology to market, namely-
1. It's one thing to have a self-driving feature that can work once in a while, far harder to have it reliably work over and over again.
2. Investors were much more interested in hype and the promise of super-big returns than a pretty profitable business. They also basically expected founders to lie to them and promise the world in new features that largely didn't work.
To me and the other skeptics, your #1 has always been obvious.
Getting a car to follow well laid out lanes on a controlled access highway on a sunny day with no construction or other weirdness going on is a pretty easy problem really. But it's all the other stuff that is complicated and a real self-driving car needs to work 100% of the time. 99.9% is nowhere near good enough.
Will we ever have self driving cars that really work 100% of the time. I think we probably will be it's a long long way off and will probably take a major infrastructure effort, with various tech assistants to self driving distributed all around, embedded in roads, on poles and street signs, with a dedicated communication network for self driving cars to communicate with one another and so on.
I know it's not the same thing but I have a relatively new car equipped with LIDAR and all kinds of driving assist features, and the number of times this thing misfires is just incredible. Plus it doesn't work on rainy days.
The notion that pure self driving cars without some special assist from a network like this will simply work reliably on their own will probably die over the next few years.
Yes, more likely that the road will drive the car tbh.
I also have adaptive cruise control - fortunately it works in even moderately heavy rain - using radar. I haven't tested our new 2022 Leaf in tropical deluge level rain though. There are a lot of other detection and warning features that make driving easier and safer. I discovered that the emergency braking was capable of being quite aggressive when somebody passed me on the left, cut in front of me, and then realized he was too close to the car ahead and slammed on his brakes. My Leaf slammed on the brakes too, and just ahead of my reaction time.
Nissan does make it very clear that it's NOT autopilot though. If you take your hands off the wheel for any length of time, it will get very "unhappy", start the emergency flashers going, then start slowing down all the way to a dead stop. The sales guy had me try it on a deserted stretch of road during the test drive. Same thing happens with adaptive cruise control on - if the cars ahead of you start slowing down it will also slow down - right to a dead stop. If they start up again within (IIRC) a few seconds, the Leaf will start moving. If not, you have to apply the "gas" yourself.
I guess technically people don't work 100% of the time as drivers (most accidents are from driver error), so you wouldn't need it to be perfect. Just much better than regular drivers.
The reality is that self-driving car technology is already here and enhancing parts of the driving experience. Whether it is adaptive cruise control or adaptive braking, the cars are taking over parts of the driving experience. Real-time traffic and navigation integrated with parking availability, and points of interest are simply the behavior training that provides a higher comfort level for fully autonomous vehicles.
Yabbut you are making a distinction between self-driving car technology and self-driving cars. Neither the hype nor the skepticism was ever about adaptive cruise control or adaptive braking. It was about getting into the car and taking nap while the car took you to your destination.
I always thought that 2030 was a more realistic date of when we might see self-drivers in common use, and at this point I think I would bump that to 2032-35.
My feeling has been that the legal, political, organizational issues would hold it back much more than technological issues would. That's still true, but it also seems the technology issues are more challenging than a lot of people thought at first.
Paris is indeed perhaps the most beautiful and livable city in the world. It is my favorite city and the best part is that you don't need a car—self-driving or otherwise.
Besides having the Metro, Paris is now very cyclable as well - last time I was there a few years ago I cycled all over the city. It was fantastic, but wouldn't have been safe to do twenty years ago.
Yes, I must agree that Paris is getting better all the time. I also must confess to having badly misjudged Anne Hidalgo, who is proving to be the best mayor ever.
Neoliberals get shit done, you dingus.
This country has a major party that’s irreversibly and utterly committed to neoliberalism. They’ve had power several times in the past decades. Look around and tell me that they get things done.
You're arguing with a satire account.
Good point.
No, it didn't "turn out" to be a tough problem.
It was always a tough problem. Some people knew that way back when, and some people just glossed over the obvious difficulties and plowed ahead making ridiculous predictions..
Wow! I’m envious!!
Of you in Paris, not the car….
Admit it, Kevin. You got suckered by a world-class conman. Here's a pretty good article that explains what happened:
"When I First Saw Elon Musk for Who He Really Is"
https://slate.com/technology/2022/05/elon-musk-tesla-twitter-fables.html
My optimism about self-driving cars never had anything to do with Elon Musk.
Have you seen this:
“ Waymo will begin operating self-driving taxis in San Francisco without a person behind the car’s vestigial wheel. In a March 21 blog post, it said it was “ready to begin introducing the Waymo Driver in fully autonomous mode” and called it “a major step on our path to deploying a fully autonomous commercial service.”
https://qz.com/2146123/waymo-self-driving-cars-will-go-fully-driverless-in-san-francisco/
Wither our robot future?
Poor Elon. Now that he's outed himself as a MAGA-loving Trumpkin, all of his erstwhile progressive fanbois have disowned him. Elon Who? Never heard of the the guy!
The article of Kevin’s to which Atrios links was singing the praises of Waymo. Pretty sure Waymo was not Musk…
Waymo: a Man & his dream.
Technically, Tesla drives more miles - just not in automated mode. They have 100x more units in testing and 1000x more units collecting data and shadow driving.
But Waymo is doing it now, with actual units in level 4 autonomy.
Well I spent the day doing the laundry, then the dishes, then I thought I'd take a shower but I didn't have the strength, then I had a migraine and laid in bed for a few hours, then I tried to read and it came back, and now you tell me I'll never have a self-driving car and you're at the French Open.
When a human driver encounters an entirely novel scenario, say, rounding a curve on a twisty two-lane road in the Rocky Mountains and seeing a zebra straddling the yellow line, that driver is likely to react in a reasonable, if not necessarily optimal, way.
The current generation of (sort of) self-driving cars, will, in effect, panic, and the outcome is much less likely to be reasonable. This is because of the way that they're trained. The problem with seeing true self-driving cars anytime soon (and, to be honest, with machine learning in general) is that no one knows how to train them to handle novel scenarios. At all.
I'm sure that the problem will be solved at some point, but because we really have not the slightest clue how to solve it at this time, I don't expect to see true self-driving in my lifetime.
Though not from the zebra’s point of view I suspect…
Sometimes known in the biz as "Nicholas Cage is a ripe avocado." The almost-canonical example is training a network to recognize handwritten digits 0-9 on datasets comprised of the same ... and then presenting it with the letter 'A'. What does it get mapped too? There are only ten possibilities, non of them, siginificantly, a letter of any type.
Well we have passed the peak of irrational exuberance for self-driving cars and are now well into the trough of disillusionment. We will see how long the continued development to a useful product takes, but it might not be in Elon Musk's lifetime.
Yeah, I believe I told you it was hard a few years ago, along with all the other things you think AI is just on the cusp of. Those are also hard.
The overall problem is that it is mostly machine learning. Machine learning is amazing but it is basically just remembering what happened in the past. It is amazing how much can be done just using memory, but it should be obvious that not everything falls in that category.
Self driving cars made rapid advances when they were learning how to do those things which require only memory. Now they've hit the things that don't and development has slowed dramatically.
The most promise will be for hybrid systems. Self-driving trucks in a convoy led by a human driver--in overnight runs to avoid traffic, and drivers pick them up for the "last mile' or 10.
We could do this now - the hurdles are regulatory and manufacturing.
If the network is just memorizing, we tend to call this 'overfitting' 🙂
+1
But how did Kevin get to the French Open? I will venture to guess it was not a self driving car.
FWIW, I see -- here in Phoenix -- at least one Waymo car on the streets every day. Never saw so many before this spring (and I drive a lot).
Deflect much, Kevin?
golack is exactly right: drivers for long-distance trucks are THE biggest problem our supply-chain fuster-cluck faces right now. If a driver-team could take ten 54-footers from LA to Tucson, another to Midland, a third to Jackson and a fourth to Atlanta or Charlotte that would get TEN times the stuff to Spartanburg in the same amount of time that a single driver team takes today.
No, it wouldn't solve the GHG issue -- only trains can do that, but they aren't fast and reliable enough for a lot of shipments. They are getting better but the rails know that those truck-trains are coming very soon, so they don't want to invest the money to triple-track every mile between LA and Sierra Blanca (UP) or Clovis (BNSF) and double track the rest of the routes to Atlanta. That's what it would take to APPROACH the speed and reliability of trucks.
As several mentioned above humans are pretty good at dealing with ambiguity and unique situations that computers just can't deal with. The other day I was reminded of this seeing a delivery van deal with an unusual overhanging metal ramp in a parking area. Most of the ceiling of the place had plenty of height, but this one metal edge was hanging down and would have acted like a can opener for any tall enough vehicle that drove close to it on the edge of the driving lane. It wasn't marked at all. Sure, in an ideal world this thing would have not been there or covered with bright colored warnings or something, but it wasn't. The van driver was able to drive past it with maybe 6 inches to spare. Would a self-driving van have been able to do that? Or, the other day I had to slow up and swerve around some construction worker on the edge of the road who had extended a long, barely visible level out across the driving lane. What caught my attention from a distance was this guy just standing near the road in a place where people don't normally stand, even though off the road. Would a computer recognize that? These daily driving problems are just endless.
My sf movie group just watched and discussed Blade Runner. Back in 2019, we were supposed to have "replicant" humanoids who were indistinguishable from regular humans, only stronger, faster, and smarter (and which we cheerfully enslaved). We were supposed to have massive migrations off-planet to other worlds, made ready for us by those replicants. We were supposed to have flying cars!
But they didn't drive themselves.
Oh, and they didn't have cell phones in "Blade Runner", either. It's funny what writers will think is "futuristic" at any given time.
Who could’ve guessed that self driving cars would be so difficult? I mean, except all the academics who’ve studied the subject for the last half century of course.
I wonder if this unable-to-predict-every-possible-encounterable-problem problem is why we still have pilots in planes…? 🤔
Precisely. It has always struck me as odd that so much effort was put into automating driving, instead of first tackling much easier problems like trains and ships. (Automated takeoffs and landings of aircraft in a range of conditions is already solved, but, as you say, the pilots are there to deal with the unusual). But of course it was because the automobile market is so large; nobody was going to get obscenely rich selling automation for shipping or rail.
As you often find, it's solving that final 10% or so of a problem that makes engineering challenges difficult. Unlike Atrios, I do think we'll get there, but I think it will be a decade or two further out than you, Kevin. I also think it will also need to come with some civil re-engineering to make roads more self-driving-car friendly.
Have fun in Paris!
Bourgeois competition is glorious.
Being an engineering psychologist, I have long had questions about how self-driving cars will be integrated into the automotive-transportation system. For example: Other than traffic jams/road work, when were you last on a road where the average speed of traffic was at or below the posted limit? Will self-driving cars be allowed to exceed limits? To match the average speed of other traffic? The 75th-percentile speed? What speed will be allowed when there is no other traffic nearby? For another example: how aggressively will self-driving cars be allowed to drive? Will drivers have control over that? Will aggressive drivers continue to drive their cars, leaving already-careful drivers as the market for automation?
And how will human drivers behave when 50% of the cars on the road are autonomous vehicles? Will people habitually cut off autonomous vehicles, knowing that an autonomous vehicle will always break? Will teenage knuckleheads start playing "chicken" with autonomous vehicles just for kicks?
+!
You will have to have some roads where you have to drive strictly by the laws, where self-driving cars can go, and some roads where you can drive like you drive now, where self-driving cars will not go.
Considering the amount of detectors that self-driving cars must have, anybody that drives badly on the strictly-legal roads will be caught immediately.
The non-strictly-legal roads will probably become crazies only in short time.
My other example is from my daily commute, on semi-rural roads, typically one lane each direction with a solid double yellow line down the middle. On trash day you find yourself behind a truck stopping every few feet. In practice this is less of a problem than one might think. You edge the nose of your car around to see what is on the other side, then zip around the truck. The trash guys typically will signal if there is oncoming traffic you can't see, but the path of wisdom is to wait until the truck has crested the hill or gotten through the curve. Regardless, this is all technically illegal, as it requires crossing that double line. What would a self-driving car do? I have asked this question many times and never gotten a straight answer.
This kind of good natured ribbing is what I love most about the blogosphere.
Just say it. "I was wrong and Atrios was right." YOU CAN DO IT!
"It turns out to be a hard problem!" I wrote a blog piece back in either 2015 or 2016 asking a series of awkward questions about how self-driving cars would work. My questions weren't answered, but a lot of people were eager to assure me that I was a luddite. Most of the questions are still unanswered. It was obvious at the time that the enthusiasts were reporting that 95% of the problem had been solved, and that solving the remaining 5% was obviously a trivial matter.
I have had a Tesla with the FSD suite since 2019.
Oh, its solvable, and will be solved. I mean, as to when, I have not clue. From the improvements I have seen over the past three years none of them have been predictable in the sense of "if today the car can do X, then three months from now or three years from now it will be able to do Y."
The problem is not the "driving part" - my car already drives on FSD better than I do, its the recognition part.
There is currently the tech argument going on between cameras, radar, and lidar, all plus GPS. I am not qualified to comment.
However, the huge problem of converting camera input to recognizable objects is proceeding nicely. Once all the objects are recognized having the car respond appropriately is not going to be a problem.
I used to wonder whether human "anticipation" of driving events was going to be unsolvable, as FSD mostly relies on "reacting" rather than anticipating. But seeing how well a computer, who never gets tired or distracted, can react has caused me to not wonder about this any more
Self driving cars won't need to anticipate, basically. Humans need to anticipate to be a good driver because our reactive abilities are limited.