The dream of driverless cars is dying
spectator.co.uk
spectator.co.uk
The time it takes to get from say a car that just drives about randomly to a car that drives pretty well 75% of the time is about a day's worth of work with today's technology. Going from 75-80% is a week or two. 80-85% is months. Getting to the 90% is years, and who knows what we need for 99+%.
I did a self-driving car in GTA V project that streamed on Twitch 24/7. If the car wasn't improving noticeably day by day, people were getting angry and frustrated, as if the car was meant to be perfectly driving within months, surely!
There's definitely a major disconnect between the hype and reality of what the challenge of self-driving cars is. The bubble is just simply bursting at the moment, but the dream itself is not dying amongst actual engineers. It's just dying for the people who never understood how absolutely challenging the problem actually is.
I suspect the hype's been fueled by--in addition to the usual suspects--the growth of young professionals in urban areas who have this vision of never having to own a car and being driven around everywhere. For that to be a reality, self-driving has to happen right now--not incrementally over the coming decades.
Those on demand trip services (Uber, Lyft, etc) are already cheaper than owning a car. What you describe already exists.
Could have gotten a second car but I hate driving in peak hour traffic so I take the bus. Lyft is definitely convenient after hours since my last bus is at 6pm. The convinience comes at a steep price though.
Only until the VC dollars dry up.
How is uber cheaper?
Presumably for the "young, urban professionals" the GP referred to, which I'd expect are, at most, a family of 2 (in the GP commenter's eyes).
Even given this assumption, even that may be debatable, since it assumes the Uber/Lyft model is sustainable without self-driving cars.
I can't wait for self-driving tech, I'll probably buy it when it hits used car lots, though. I'm hoping by 2030 it'll be real. I don't think uber/lyft can sustain more than 4-5 more years on investors alone, without charging a LOT more for rides if they're going to continue using real drivers.
The real disruption will be when car companies instead of selling cars, simply sell memberships to a car sharing plan... pay $200/family member/month to Ford, get unlimited self-driving wherever anyone in the family is going. No need for insurance/etc... gas included.
That's an easy one.. for the same reason they assume everyone is a young, urban, technical professional living in a major tech hub.
I recall this being a source of criticism (though admittedly not recently) about the problems startups are solving being skewed toward the problems the above people (the same ones doing the solving) tend to have.
> (and always will be)
I'm not sure that's quite true, at least not in the sense that of individuals.
However, they may assume it of the population within a certain location, such as that urban tech hub in which they live, if the individuals who do form >=3 families move away to the suburbs, and are replaced by single individuals. This seems to happen in US cities.
> when car companies instead of selling cars, simply sell memberships to a car sharing plan
What's their incentive? Car-share already exists for non-self-driving cars, but through third parties. It's quite expensive.
> get unlimited self-driving
That seems a particularly unattractive option for the provide, as well as being political dynamite.
The future of cars, makes sense not to own... A car just dropped johnny (a kid I have no relation to) off at school a mile from my house 5 minutes before I need to be to work, this car then becomes mine till I get to work, from there it takes someone else to a brunch meeting... and so forth.
Using logistics it can free up roads by sending the closest available car to pick people up, and possibly re-arranging traffic patterns via networking to make roads clearer...etc.
Edit: one more attractive thing for car companies: Subscriptions. Someone paying 800 for life, is MUCH better than someone paying 20k once. They can know with better certainty their monthly recurring revenue when charting growth/etc..
Still, $200 seems far-fetched in light of the average American 1200 miles monthly. At under $.17/mi (not counting deadheading), that might not even cover depreciation/capital cost.
People are more apt to buy into an 'unlimited' plan. Road trips would cost extra (because you're gaining unlimited full-multi-day access to a vehicle), maybe something like $30/day + gas.
That's part of my point, though, that what the number is (or at least its order of magnitude), actually does matter.
For something like a per-person charge, if that number is high enough, for large enough families, people are going to start looking at whether they could be saving by using the per-vehicle pricing of buying versus renting. I'm also saying that this number must be high enough to be attractive for the car company.
> maybe not 100% electric, but maybe 100 mpg with electric assist.
I'd say that's misguided, then, because fuel cost alone is too low a fraction of total operating cost.
> Road trips would cost extra
So like today's cellphone "unlimited" data? In that case, sure, an actually-limited subscription plan would easily be attractive to a car company, since they could write in any limits they want and change them any time, as well as vary the pricing at any time.
How many consumers would opt for that instead of outright buying is another matter.
> It's just dying for the people who never understood how absolutely challenging the problem actually is.
You're laying a lot of blame at the feet of the excited people, but in their defense, much of their excitement is thanks to hucksters (cough Elon cough) who've been insisting that fully autonomous cars are just a year or two away. It's not like these people can be reasonably expected to understand how difficult the problem domain is, especially when they've seen recent AI successes at other problems that were popularly considered impossible (e.g. AlphaGo).
And, as you say, a lot of people assume the mostly unanticipated advances in deep learning in a short period of time carry over to everything even vaguely related to AI and cognitive science more broadly.
Here's the playlist with all of those videos if anyone else is interested: https://www.youtube.com/playlist?list=PLQVvvaa0QuDeETZEOy4Vd...
> How about electronic publishing? Try reading a book on disc. At best, it's an unpleasant chore: the myopic glow of a clunky computer replaces the friendly pages of a book. And you can't tote that laptop to the beach. Yet Nicholas Negroponte, director of the MIT Media Lab, predicts that we'll soon buy books and newspapers straight over the Intenet. Uh, sure.
Takeaway: people expect way too much out of next five years of the future, and typically it takes twenty years for those expectations to become fully realized. (I believe there is an Andy Grove quote to the same effect, but I couldn't find it in time.)
But the proof of concept of auto-cars generally works. The kinks don't seem insurmountable, just a series of incremental improvements and economies of scale.
Flying cars on the other hand have some really big problems to tackle, such as noise pollution. And robots with "common sense" still seem a long ways off. There is no proof of concept that's even close.
Why does no one consider a helicopter to be a "flying car"? What can a flying car do that a helicopter can't? The only difference I see is that a flying car would be cheaper and owned by more people, which a helicopter could easily be if people actually wanted a flying car. Helicopters are expensive because they're very niche devices.
The fact that almost no one owns a helicopter proves that people don't actually want flying cars.
To your second point, people with enough money and training do commonly buy them for personal use. (I know this anecdotally from a relative involved in sales of Airbus helicopters.) If you addressed the cost and safety issues no doubt more people would buy them.
So did cars when they were first introduced. Since they solved a need, there was a great incentive to make them easier to use and easier to repair. That incentive doesn't exist for helicopters because most people don't own helicopters.
>To your second point, people with enough money and training do commonly buy them for personal use.
I'm glad you pointed this out, I didn't make that argument because it would ruin the surprise, but yeah. People do use flying cars, and they're called helicopters. They're just used by very rich people or people with enough training to operate them. But think in cities like NYC, helicopters are used just like cars would be, because using a car can be unpractical. Functionally there is no difference between Michael Bloomberg traveling to Newark via a limo versus via a helicopter: both park at his building and can transport him from his home to his destination. It's a flying car.
>If you addressed the cost and safety issues no doubt more people would buy them.
The incentive to improve cost and safety only exists if demand exists. Cars only got cheaper and safer when the public began buying them. Flying cars (helicopters) will only become safer and cheaper if people start demanding them as well. The fact that they're not, IMO, proves people don't actually want flying cars.
No one in NYC, not even Bloomberg, is landing, let alone parking, his helicopter at his office. It is completely impractical in terms of safety, space and noise, even if it were just one person doing it, and absurdly infeasible to think it could be scaled up to, say, just the number of people using the Lincoln tunnel in any given half-hour. There are a few helipads around the edges of the city (a number that is likely to decrease on account of noise), that cater to some of the transportation needs of a number of people that can be counted in the hundreds per day, not hundreds of thousands.
The idea that if there is a wish, a solution would have been found by now, does not stand up to the most cursory examination - for one thing, if it were so, not many people would be dying.
I agree it's a genuinely difficult problem, but we might agree for different reasons. I think it's a genuinely difficult problem because there is no way to make a flying car as safe as a car that is stuck to the ground. There's just not. Normal cars break down, they run out of gas, they stop working all the time. You coast to a stop and call a tow truck. Flying cars break down and you plummet to your death. Normal cars crash into each other in just 2 dimensions. Flying cars crash into each other in three dimensions, and then you plummet to your death. Normal cars get stuck in traffic and sit with their engines at idle, producing small amounts of pollution, noise, wind and heat, yet still contribute substantially to global warming. Flying cars sitting at idle in a traffic jam still have to hover above the ground, which means they're running their engines almost as hard as if they were actually moving, producing tons more waste byproduct than even the most gas-guzzling SUV.
Of course, my argument isn't "flying cars are a bad idea", so all of those points might seem offtopic. But it links back to the idea that we already have flying cars and we've already realized that it's a terrible idea. Short of true anti-gravity, a flying car is always going to be less safe, more polluting, more complicated mechanically, and is always a bad idea. If people actually wanted flying cars, we would deal with the downsides long enough for smart people to overcome the reliability and price issues, but overcoming "plummeting to your death" is a genuinely difficult problem that only has one solution: true anti-grav that works even when the power is disconnected. Since that's a far-future sci-fi plot, the closest thing we have to flying cars is a helicopter.
It just turns out that using them as reliable daily individual transport is a really bad idea.
I think you may be misinterpreting my last sentence. What I mean is that, just because medical science has not delivered immortality, it doesn't mean that there aren't lots of people who would prefer not to die. You could substitute specific causes of death, such as cancer or alzheimer's, into that argument, if you think that most people have come to terms with their mortality.
Let's beg the question and assume people want to go to the mall with their flying car, that's their primary use case. They can't do this right now because there are no parking spaces at the mall for their helicopter. There are no parking spaces for their helicopter because no one is flying to the mall with their helicopter. If people did, the mall would build parking for them. This is shown by the fact that places where people regularly fly helicopters, there is dedicated helicopter parking. It's just as valid for me to say "the fact that helicopters exist but are uncommon proves people don't want flying cars" as it is for you to say "helicopters are not flying cars".
I'll ask the question again: why do people not consider helicopters to be flying cars? They do exactly what people want a flying car to do: they take off and land vertically, they can fit in a reasonably sized parking space, they can easily maneuver in three dimensions, and they hold passengers and cargo. I argue that helicopters are flying cars, and they're proof that most people don't actually want flying cars.
However they also get it wrong. It's hard to predict. Others in the thread have pointed out Asimov's stories that featured FTL warp drives and computers that fit in only one room.
One thing I've noted from observing 20th century progress is that progress is much, much faster for things that are less capital intensive per iteration. It's one reason I've been predicting that solar power will eat the world. Nuclear power is capital intensive per iteration (per power plant or research reactor), so even if it's superior in some ways it has a longer iteration time and so it will lose the race. I think this neatly explains why space tech stagnated and computer tech zoomed way ahead. So today we still have chemical rockets and still get most of our electricity globally from burning coal and gas (100+ year old tech), but we have supercomputers in our pockets.
Bill Gates:
> We always overestimate the change that will occur in the next two years and underestimate the change that will occur in the next ten.
Roy Amara:
> We tend to overestimate the effect of a technology in the short run and underestimate the effect in the long run.
Kurzweil has called it ”the intuitive linear view” vs ”the historical exponential view”:
> Most technology forecasts ignore altogether this “historical exponential view” of technological progress. That is why people tend to overestimate what can be achieved in the short term (because we tend to leave out necessary details), but underestimate what can be achieved in the long term (because the exponential growth is ignored).
ADDED: One thing Stoll did get quite wrong was missing the utility of mass amounts of often uncurated information. Though, in this regard, he has a lot of good company among science fiction authors who mostly envisioned centralized encyclopedias created by experts, not Wikipedia or the Web generally.
Somewhere there's a great review, written somewhat recently, on all the things he had wrong. I can't seem to find the one I'm thinking of. I guess he was correct about how things on the web are never permanent.
The author right, in 2003, I doubt you saw many people reading books on the beach with a clunky laptop or whatever version of kindle existed back then. So, from the mid 90s to mid 2000s, the author was decidedly right for 10 whole years, which is a long time.
Now, if he continued to say that not many people would read ebooks on the beach from 2008-2018, he would have been wrong about that bet. But I doubt he would have felt that way, because by 2008, it was much clearer where the internet was headed compared to in 1998
Oh, yes, sitting on a wet and windy beach trying to read a newspaper sounds superior to reading it on an e-reader in every conceivable way.
It seems almost a certainty that we will gradually ease into the concept. Some cars will be able to do some things autonomously, a little, for some people, some of the time. Eventually the cost/benefit levels will improve, more features will roll out, and sure over a long time scale things might look revolutionary.
In other words this big complicated category of tech will grow like most other technology has. This isn't an individual app like Instagram or WhatsApp it's a whole category of technology, like the concept of automated factories, or container based shipping, or the internet itself. It could be decades, or longer. Or not.
But as a casual observer it's pretty easy to look around and notice that there really aren't any practical applications of autonomy with mass adoption, and conclude that it's going to be a minute, and that we'll definitely have ample time to notice as we slowly get closer.
https://news.ycombinator.com/item?id=16353541
Here's my negative scenario: self-driving is "AI-complete"; you can't really hit all the edge cases without solving AI in general, which is more than 30 years away (Kurzweil is the wild optimist and predicts 2045).
edit: Although since writing that I have become aware that Kurzweil thinks the Turing test will be passed in 2029; the "singularity" is the 2045 date.
https://futurism.com/kurzweil-claims-that-the-singularity-wi...
In any case, I think it's realistic that level 4 self-driving cars won't be widely deployed for end-to-end trips until 2029-2045. This doesn't seem to be what many people think right now. They think it's going to be an "instant roll-out" like the iPhone -- either it works or it doesn't.
I remember reading about this exact thing in the 90s and thinking that it was completely incredible, but I didn't even notice it when it happened. We first got computers that could sort-of translate, but not impressively so, then computers that could sort-of speak, but not impressively so, then computers that could sort-of listen, but not impressively so. All these technologies improved, little by little, until they got to a point where they are pretty impressive, but we're used to them.
Cars are already going the same way, my 5-year-old car can already follow/match speed with the car in front of it and warn me about signs on the road. Little by little, it'll do more things on its own, until I won't be needed any more.
For arbitrary language pairs, it doesn't. Try translating between Greek and Chinese or Japanese at a level sufficient to get around town and do everyday, tourist-y things.
Machine translation works well between pairs of languages that are very similar and for which there are many parallel texts. It's all downhill from there (e.g. you have to go via an intermediary language, typically English, to get to the target language etc, which starts to seriously degrade performance).
Auto-cars are not different. In ideal (i.e. laboratory) conditions they may work sufficiently well. Diverge from those conditions and you're gambling with the passengers' lives.
This has been the case for a very long time. I learned something recently that surprised me. To phrase it as a quiz: when do you think was the first coast-to-coast (completely) driverless journey? Answer below.
______________
When are we going to get the first car, hands off the steering wheel, feet off the pedals, drive coast to coast in the US? Yeah, well, it actually happened in 1995 with the Navlab Project from Carnegie Mellon University.
Courtesy of Rodney Brooks:
https://cdn.arstechnica.net/wp-content/uploads/2018/06/2nd-P...
Limiting the use to highways appears to meet this criterion, but it does not yet do so, on account of problems, like stationary-obstacle detection and adverse weather, that are relevant to highway driving.
It doesn't fire up the people who want robo-Ubers but it would have a lot of benefits, including safety, and seems like a much more boundable problem. I could actually imagine something like this being approved for consumer use decades before the more general case.
Because there is a significant difference in kind between human driving with driver assistance and self-driving, because having both on the road creates problems that's neither alone faces, and because once self-driving becomes viable for general use, regulatory and insurance factors will probably drive a rapid transition to a state where it's the only thing most people can do on most publicly accessible roads.
> or container based shipping
Intermodal container-based shipping was a fairly rapid global phase change, going quickly from introduction to international standards to global dominance of shipping. So it's kind of an odd example to use to counter the idea of a rapid phase change.
I think such a system can be built. It's called a train.
Here in Berlin, two children were killed by these monsters on one day last month, with the children doing nothing wrong.
I was walking down a street the other day, early morning, and it stank with food scraps, garbage - it was one of those maintenance streets you find at the back or sometimes even the front of fast food outlets, delis and cafes. The backend of that industry looks unappealing and smells as bad even though the businesses and city are doing everything they can to keep it organized.
With pneumatic tube rubbish disposal - which already exists - it would improve the quality of a large portion of the city, not to mention those children wouldn't have had to suffer.
More generally as I see it we are in the Dark Ages of underground tunnel making but if it could be made 1000x affordable lots of problems start to go away for our infrastructure. I've seen The Boring Company's proposals of course but I suspect it's on the wrong track. What you want is a creation capable of creating a criss-cross of tunnels - a whole network (similar to dark fiber, just waiting for land development so it can be turned on) which is deep beneath any other infrastructure and is linked to the surface through vertical channels e.g. a connection to somebody's house goes up, not across as we assume pipes are supposed to. I think the cost of extra pump power would be as nothing to getting rid of every legal and utility availability headache!
Seems like a recurring theme. It's near a half century for the moon landings. Should revisit the concept of batteries too. I'm reading a book called This Victorian Life and am constantly impressed by lots of little innovations now forgotten. Two concepts I think we'll have to revisit: heliostats for indoor natural lighting and terrariums aka self contained biospheres for space/underground.
Given our litigious society, I think that driverless systems for big trucks need to be far far safer than the human-guided equivalent. When accidents happen, they have the potential to do far more damage.
For OEMs, I think it's less about a pot of gold and more of a defensive move. They're willing to throw money at trying to mitigate a huge business risk.
These companies might not know (or even have an opinion on) how close the leaders like Waymo are to making it work. But they know that, if it does work, a new era will begin when nobody will be able to sell a car without it.
And it's almost moot whether they believe it will work. They're protected as long as they're putting sufficient resources into it that they stay roughly as far along as others.
Certainly, the car needs to be able to recognize conditions where it needs to turn over control with, say, one minute warning. And we're not at that point today. But one can imagine things like radio beacons for construction zones for example. It's not easy but it seems much easier than the more general case and it's also possible to get too hung up on truly weird corner cases like traffic going down the freeway in the wrong direction.
A general enough system would have no edge cases. It is IMO the required solution, and a driver always on the defensive is what we all are, at lease here in South America.
Just following the rules would have killed me years ago. Other drivers are reckless.
Just looking at the first problem, a four lane highway is a four lane highway, regardless of how vehicles comport themselves. It seems more productive to solve congestion by getting at least some of the vehicles off the ground using emerging passenger drone technology. [0] That incidentally looks like an easier place to attack automation as there are fewer corner cases. Commercial aircraft achieved high levels of automated flight decades ago.
[0] https://www.dezeen.com/2017/04/28/kitty-hawk-prototype-perso...
As for that Kittyhawk prototype - one prop failure at altitude and the pilot is dead. There's a reason they're testing over water at very low altitudes. Quadcopters have a very bad failure mode: they always crash. You want at least six motors, if not more, to avoid a crash every time a single motor fails. But, of course, more motors requires more battery power, which raises the weight, which makes it less maneuverable, and so forth.
From a pure "worst case failure mode" ground traffic is a better first automation target for moving small quantities of people.
Amazon's scheme for drone delivery seemed kind of crazy at first glance but it does have some sound economics behind it. Here's a reference that seems to have a reasonably balanced view: http://thedronegirl.com/2017/05/07/drone-delivery-economics-...
Capacity is a function of maximum density at speed, which both consistency of behaviour and speed and accuracy of detection and reaction have a role in setting. Automated driving — if it is exclusive — addresses congestion (provided added “deadhead” runs don't offset max density gains), and that's always been part of the self-driving vision, even back when it was about “smart roads” and not smart cars on dumb roads.
> (b) extremely inefficient energy use in transportation, especially personal vehicles.
Self-driving might help address that, too, as it allows lower friction exchanges between local personal transport and higher capacity—and higher efficiency—shared long-range transport.
Lanes are as wide as they are and following distances are as long as they are because people are terrible at driving. A mostly-driverless 4-lane highway could more than double capacity just by reducing following distances when computers rather than human beings are making the decision when to brake and how hard to brake.
Humans also cause traffic jams by causing accidents or rubbernecking at accidents - that'll get better too.
On the other hand if you could distribute traffic flow vertically the potential for scaling goes up considerably. It's a hard problem to solve (read: expensive) but building roads is not exactly cheap either.
BTW I'm not against autonomous vehicles. It would be great if they worked. But I think we're more likely to see near-term investment payoffs using it for incremental safety improvements, something that already seems to be happening.
Example: I currently drive to work 20km. As my wife needs occasionally as well, we have two. With an autonomous car we could forgo one and I send it back after I arrive at work. In the evening it'll pick me up again.
Self-driving cars can park far more densely without fear of being blocked in and can park inconveniently far away while still being reasonably available. Thus cities will need less on-street or near-street parking and can reclaim a lot of that space for traffic or transit or bikes or people or commerce.
If your wife only occasionally needs a car, couldn't she call a local on-demand one - an automated uber? That seems cheaper than paying to send an empty car on two daily 20km trips...
Take driving in the snow, for instance. It might be completely impossible to do it safely for the next twenty years by having a computer manage the drive path, but designs might create the ability for cars to form mini-trains where the lead car is driven by a human and the cars fall in line together. Then you could still get in a self-driving car, it could still get you from Toronto to Ottawa. It could even get you right to your front door! They'd just have to pay some train drivers (conductors?) when the weather is too harsh for to put it in full auto mode.
We're going to get there. The benefits are too obvious and we can work around the persistent problems.
Sure we will, no doubt about that.
The question is when?
In 10 to 20 years from now level 4 sounds reasonable, in 2 to 5 years from now it sounds not, notwithstanding the (daily) announces about this or that new autonomous driving system or related "breakthrough".
It's the social problems that are hard for AI, like how to handle a construction zone that looks like you need to break the law to get around. Those don't always have easily generalizable answers.
For these cases, the car could have a manual mode (?)
Regardless of how you feel about the future of driverless cars, this is a safe article to skip - there's no content.
Since the hype and exuberance was entirely self created, how can anyone else be blamed? This is like blaming others for taking you seriously.
It would be nice if it would happen faster, we could cut the number of car accidents in half, car related fatalities by a quarter, and zero speeding tickets...
“Shall hold manufacturer, software developer, etc not liable for any injury or death from the result of using car...”
That's how you can tell if it's driverless.
(Also they don't really work reliably: try using lane assist with roadworks markings and watch your car veer off into the barricades. Or just too-unclear-for-machine—but-blatantly-obvious-for-human markings, like the dead Tesla driver from a few weeks ago encountered. Even a fourteen year old doesn't make that sort of mistake.)
Not sure how to address that except through more regulation which doesn't exactly help on the innovation front...