Singapore also has a political framework that could force self-driving car adoption if the powers that be want it. Things that could be mandated such as self-driving only car lanes and motorways could be possible given the small size of Singapore and the powerful nature of the government. It would be unthinkable at present in America to force all cars on certain roads to be self-driving only. Singapore has the sort of culture that values "safety" over freedom.
In America, for better or worse, it would take a long time to mandate these things if they were even possible (considering that the same laws would need to be passed at least 52 times for states and territories). America has checks and balances--everybody gets their say.
>It would be unthinkable in America to force all cars on certain roads to be self-driving only.
There are private cities (Celebration, FL), private housing communities (1000s of them), private large corporate campuses, private shipping yards, private military complexes, private airports, industrial hangars - lots of places where you could safely force all cars to be self-driving without any governmental interference.
But I think for it to have meaningful impact we need interstate public roads to have self-driving. I haven't looked at the statistics, but most deadly accidents I read of late happen on interstate roads.
Edit: It's unthinkable at present but I believe if a city-state like Singapore undertook the experiment and it was successful it could eventually spread around the world like HD adoption. (That's why I said it would take a long time to happen in America.)
Not only is that most meaningful, it olso has the most bang for the buck (saves people the most time), AND is the easiest technically.
I wrote a long post about it before, but basically I expect self driving trucks on dedicated interstate lanes to be the start. Then self driving cars on the same lanes.
Over time all interstates will be self driving. City streets will never be unless we get AI with general intelligence.
> City streets will never be unless we get AI with general intelligence.
Driving is likely a much easier problem than artificial general intelligence. Even when it involves evading rampaging toddlers. Sure, if you want perfect safety, you'll need the car to anticipate a great deal —far more than humans currently do, like looking at the walkways as well as the drive lanes.
If we merely want something that's safer than a human driver that's probably nothing more than a (huge) engineering feat.
> Driving is likely a much easier problem than artificial general intelligence.
That isn't actually true. Scenario: Narrow one way road, you get to the bottom of it and there is an ambulance blocking the road. Behind you is a bunch of cars, ahead of you is a few cars.
You need to coordinate with other cars to move out of the way, you need to make use of driveways to turn around, you need to know to override the law and got the wrong way on a one way.
And that's just one scenario. You need general AI to handle all the different things that happen on city streets.
Highways is easier - it's controlled access. And if they block the road it's typically by officials who would have some way to remotely control the self driving cars and tell them what to do (not that that is not a car of worms in itself).
> If we merely want something that's safer than a human driver that's probably nothing more than a (huge) engineering feat.
Not even close. You would need to do better than 99.99999% perfect to do better than a human. That number is correct BTW - that's how rarely humans make a mistake that leads to a fatality. (Injury rate is not quite that good: 99.999% but that's still quite hard for a computer.)
We don't have that level of reliability in a phone, never mind a car. There are some appliances that are that good, but they tend to be very simple.
Such scenarios are rare enough that they could be handled either remotely, or not at all. The ambulance won't get through and its patient will die. Tough luck. But think of the many accidents those automated cars avoided in the first place. It sucks, but it's still worth it.
Also, cars don't have to be fully automated. They can still be remotely controlled, or otherwise warn their company that they should call a tow truck. My favourite example is automated trucks: you'll most likely have a central dispatch per region, with a few operators that manage 50 trucks each. Once a truck has a problem the computer can't handle, it just calls the relevant dispatch, where humans take over.
Also, automated cars don't have to do better than humans in every dimensions. Safety is paramount, but resilience in the face of unusual situations is not. It is okay for the computer to get you to your destination a bit more slowly, or even not at all from time to time. Even public transportation has way less than 99% reliability for my commute to work, and I still take it.
It's ALL edge cases. Everything. The non edge case are the easy part. It's the edge cases that is hard.
Saying it's fine "it's just edge cases" completely ignores the difficulty.
> But think of the many accidents those automated cars avoided in the first place. It sucks, but it's still worth it.
Did you read the rest of my post? Do you understand just how good the computer would have to be to have any impact at all? I have a feeling you don't.
> They can still be remotely controlled
No they can't. Wireless internet service is not even remotely close to good enough to make that possible.
> Also, automated cars ... every dimensions. Safety is paramount
This is why I say highways only. That covers that majority of the use cases, and the most dangerous part, and the easiest part. It's a triple win. The use case for city roads is basically non existent with the exception of self parking.
> has way less than 99% reliability
That 1% doesn't mean death. It's means slowness. That's not what I mean. The self driving car has to be basically perfect or death is the result.
People very much underestimate just how good people are at driving.
Let's arbitrarily decide that 99.9% of driving time is made up of the "common" cases, and the rest (0.1%) is the "edge" cases. I'm pretty sure there are many more edge cases than common ones.
Now let's further assume that whenever an "edge" case occurs (that would be a couple times per hours), the cars just slows down until it stops or stop being in that edge case. It's not very convenient, but I'm quite sure it is rather safe.
> Do you understand just how good the computer would have to be to have any impact at all?
From what I have gathered, computers are already good enough to work on sunny roads with few surprises. Including towns. Or maybe the google cars don't work as well as I though they did? I wonder how many times the human had to push the emergency shutdown button in their tests.
> No they can't. Wireless internet service is not even remotely close to good enough to make that possible.
Just use a regular cell phone, and communicate through the speakers. 56kb/s is more than enough bandwidth (space stuff like the Mars rover use way less bandwidth than that, and suffer from way more latency).
> The self driving car has to be basically perfect or death is the result.
We don't need computers to be perfectly safe. We need them to be safer than humans. If automated cars kill on average half as many people as humans do, that's already a huge win. Not enough to stop there of course, but more than enough to switch.
Besides, I believe that abiding the traffic code and slamming the brakes whenever something goes wrong is not hard, and already safer than human driving (which take chances, speeds, rides too close to other cars…).
Bugs on the other hand may prove most problematic (see Toyota's unintended accelerations). We may need Nasa like processes to ship that code on the road.
Singapore may be small, but within that space (a bit over 700 square km) it has everything larger countries have. Towns, expressways, nature reserves, military bases, airports, etc.
Thus, it is an excellent place for testing such things as self driving cars.
That's only one lane that is enforced to a varying degree (hardly in CA), and subject to cheating. I think in order for cars to become safer we need to force the human out of the driver's seat. Even if we had a hybrid system, self-driving is still subject to other human drivers. Try taking away the keys of Americans. I ride a motorcycle and I would miss riding if it came to forcing people to adopt self-driving, but that's how we reduce highway deaths.
Edit: now that I think about it, cheating could be enforced by camera. Assuming the road had a computer network (like a much more reliable Bluetooth), I am sure cars could identify an unauthorized manually-driven vehicle and alert authorities.
I suspect there'd be a strong selection bias where the only people who'd continue to drive manually would be the better (or at least more attentive) drivers.
To enforce that, you could make the licensing regime more like an FAA private pilot rating and less like a token ritual for access to your God-given right to drive no matter how incompetent.
+1 I think you have it right, the political/social frameworks of Singapore could 'make this happen.'
I worked in Singapore in April and I loved the place. I felt very safe walking around at night for exercise, and the whole place was well run. I have read some criticisms of the government but people there seem happy, seemed to be employed, etc. A nice place and I would love to go back there!
From a product development and deployment perspective, it also makes a lot sense to start in Singapore. Unless a product is completely risk-free, it makes sense to validate outside of the largest and most valuable market.
> Singapore makes a particularly good testing ground for automated vehicles. Its manageable size (it’s about three-and-a-half Districts of Columbia), flat terrain, warm weather, and well-kept roads provide about as simple of an urban landscape as one could ask for. And its government is supportive of such technology, having formed an Autonomous Vehicle Initiative to oversee research in 2014.
I'd argue we already have too much testing under "ideal conditions." I want to see self-driving vehicles tested in snow, in rain, in strong winds, in fog, and on mountain roads.
Don't get me wrong, you absolutely want to start developing it under best conditions, but if they legitimately want to get this live by 2020 then we need to expand the scope to less idealised situations.
Singapore certainly gets rain, of the solid-cube-of-water, when-you-hear-it-a-block-away-don't-keep-walking-to-your-destination-just-find-immediate-shelter, oh-so-that's-why-all-the-drainage-canals-are-so-deep variety.
It makes perfect sense to walk before running. I don't really see a "mountain road" use case for this tech anyway.
>I'd argue we already have too much testing under "ideal conditions." I want to see self-driving vehicles tested in snow, in rain, in strong winds, in fog, and on mountain roads.
Unfortunately, the vehicles don't yet work perfectly under "ideal conditions". I think this ought to be the first step. There's no need to have a "blanket" rollout of autonomous vehicles - start in ideal cities and then spread further.
Granted, I DO think that they'll find driving in snow/rain/winds to be fundamentally different enough from ideal urban driving that the exact methods the vehicle AI has "learned" may not directly apply. So, it may merit testing in parallel, because as you've said we need to stop ignoring these cases.
I guess my point is a long-winded way of saying "Yes we need to test those cases, but in addition to ideal urban conditions, not instead of."
Singapore is a city, and therefore already quite challenging. I've never been there, but I suspect there will be streets with playing children, scooters, etc.
Compared with most cities, there really aren't many playing children or other people in the road (you'd set autonomous vehicles to avoid Little India :-) ), the roads are impeccably maintained and there's virtually no traffic.
Living in Singapore, I disagree with you. Of course there are not many children playing on the main roads, but drive into a residential estate and things will be very different.
When I drive, I wish that your last statement was true. However, you're welcome to look at the traffic view of Google Maps and you can see that's wishful thinking. It's not as bad as places like Bangkok or Manila, but about as bad as your average European city.
I'm a European who's lived in Singapore. The traffic really doesn't compare with your average European city: there are a lot more cars per capita and no "off peak" plates here, and generally inferior capacity road networks and worse public transport alternatives too.
Pretty sure Google Maps' traffic maps are scaled based on averages for the road/area and not some absolute scale for easy comparison too (though this being HN there'll probably be someone that worked on it along to correct me in a moment)
Well, (human) drivers in Singapore are pretty bad (based on a few taxi rides), so there is that. Apparently lane markers are something your are supposed to drive on top of...
The Singapore driving style doesn't make sense if you're used to driving on western roads in western traffic. It makes perfect sense if you think of the lanes as aisles in a packed supermarket, and each car as a trolley. Cars don't so much 'navigate' as 'flock'.
Seems like we're far from going "all the way" on autonomous vehicles by tomorrow or next year, but to me the progress on this tech is pretty hard to predict 4 years into the future.
This is why I feel that Telsa's telemetry expertise will be the key to their success.
It's been covered a few times that they're constantly running AP algorithms in parallel with the driver when AP is not enabled and filtering the results back home. You're going to get a ton more real world scenarios that way, unknown unknowns and all that.
The very fact that these manufacturers are being extraordinarily cagey about the circumstances under which their automation works and where it doesn't should tell everyone everything they need to know about how 'around the corner' this tech is.
Think of it this way: non-optimal driving conditions is nowhere near the last hurdle for full autonomy. We're a long ways off.
Weather is still an issue AFAIK. Perhaps a 3d visioning system can fix this at some point, but I don't think we're there yet as those solutions are too slow for RTC.
The failsafe is to do what responsible drivers do when they can't cope with the conditions present: stop driving until conditions improve. The hard part is to reliably tell when the situation is no longer safe before something happens.
The times I've seen it rain in Singapore or kuala Lumpur it's all or nothing. Fine weather -> 2 foot of water on the road -> Fine weather.
I somehow don't think that scenario is in the test plan.
Do taxis actually disappear - which seems counter-intuitive; where would they go, and why would the drivers suddenly not want to take rides when it rains? - or is it just that all free taxis are grabbed immediately when it rains, thus there are no free taxis?
There we can answer that the seeming complete disappearance ox taxis when it rains will persist, because the ratio of customers per taxi shoots up when it starts to rain.
(Where I live (north Europe), you practically don't flag taxis on the street but order it on the phone; when I was living in Beijing, I think the rain phenomenon was simply that all taxis were taken as the first drops came down.)
It's worse than that: taxi drivers count the amount of money they made in a day and stop when it's enough. Then if it rains they make their day quickly and go home, while they will cruise much longer during low demand days.
This is a classic example of short term view and the invisible hand going bad, found in Thinking fast and slow.
Citation needed. I would think most taxi drivers are smart enough to figure out that it is better to work a few more hours more on rainy days and sit in the park on sunny days than to work a few hours more on sunny days and sit in the park on rainy days.
You got one. 'Thinking Fast and Slow' by Daniel Kahneman. Sometimes people don't optimize long term and just prefer to get 'daily' rate and go back to their families. This might turn bad if sth happens and they can't ride for some time and could use the 'rain' money. Book is really good, highly recommended.
In NYC, I've heard that quite a few drivers have an agreement with the taxi owner that makes it barely-profitable for them to do more than a given number of rides - and with a rain, that number is achieved much earlier in the shift.
That would explain it, but I've heard this complaint about taxis disappearing when it rains in very many cities - it would be odd if they all had similarly broken incentives for the drivers.
During heavy rain, road conditions are not good and visibility is poor. Hence, the probability of getting into accidents is higher. (Think car and health insurance.) However, the fare price does not increase to compensate. In Singapore, where the demand of taxis regularly exceeds the supply, it makes more sense to go take your meal and wait out the downpour (typically over in one or two hours).
combination of (1) more riders (2) fewer taxis because (a) taxi companies cover damage on a limited basis i think (b) driver pays for repairs out of pocket and then goes through a lengthy and painful assessment and reimbursement process (c) out of work while the car is in the workshop
Perhaps they’ll also help with the Singaporean phenomenon of the seeming complete disappearance of taxis when it rains—an almost daily occurrence in the tropical clime.
My question is whether the disappearance is real, or just perceived, as when it starts to rain, the available taxis are flagged to use immediately.
I'm pretty sure it's the latter - when it rains, people take all the taxis. This is what Singaporeans say, and locals in nearby Kuala Lumpur have said the same.
> Are self-driving cars able to handle rainy conditions?
Seriously? I pretty sure all current vehicles need to go through extreme weather testing before they can be brought to market. For example my dad worked on a project testing Ford vehicles in extreme cold and snow conditions in northern MN.
I can foresee several phases in the development of "robocars".
Phase 1 - dedicated lanes for robocars, to mitigate problems with the still-dominant human-piloted cars. Robocars capabilities at primitive level, will likely have human-usable controls.
Phase 2 - once their safety and compatibility have been established, robocars will freely mix with human-piloted cars. Robocars will accommodate errors committed by human-piloted cars, and will no longer need human control.
Phase 3 - robocars dominate the roads, while laws and economic structures (e.g. insurance) will favor robocars. Robocars will have localized coordination capability.
Phase 4 - robocars will act in concert over large (metropolitan) distances to smooth traffic flows. Human-piloted cars will become as horses are today - a hobby for the rich.
Admittedly, I am mixing adoption phases with technological ones.
I think the important bit is the government buy in and support for this. As far as autonomous driving is concerned, legislation might end up being the bottle neck in most nations.
I agree with your view on the matter, and the article did suggest that this was the case, however, they didn't show it.
In the very same paragraph that the article made note of the government's support, it also conceded that the Delphi already had pilot programs in the US.
I would expect to see some concrete examples of laws being passed.
I am not, by the way, saying that Singapore is not auto-car friendly or that they won't have them earlier or anything like that. Just that this article doesn't argue that fact.
(Edit: Not to be coy with my view: I think all the rich countries will get them at approximately the same time.)
Do you have a source for that claim? I might have fallen victim to Google's PR here, but I was under the impression that they are reliable already - at least in controlled conditions (low speed, good weather, well mapped area).
Depends on the threshold for reliability.
As of the end of 2015 in the 1.3 million miles Google cars had driven on public roads in California, human drivers had been required to prevent 13 incidents where Google assessed a collision would have happened without human intervention, a further 56 where the driver intervened for safety reasons and 272 times due to more minor sensor issues (using quite conservative thresholds)
That might not be bad for new technology and was showing trend improvement, but it doesn't compare favourably with accident rates for human drivers
You likely under estimate the number of accidents a human has in that time frame it ~108 years worth of driving. 1 accident per 8 years is hardly terrible. Further there rate of problems has likely been dropping so the most recent 120,000 miles are the most important.
What is in your opinion the acceptable number of traffic death through self-driving cars?
In addition, impaired drivers causing accidents will pay a fine or go to jail. Who takes the responsibility of impaired sensors in self-driving vehicles?
But once they have accident rates that compare favorably with accident rates for human drivers, they'll have a saleable product. Self-driving cars don't have to be perfect, they just have to be better than people.
To be allowed on regular public roads without humans behind the wheel in most jurisdictions, they're going to have to be better than people in virtually all conceivable circumstances rather than just on average. (This isn't even necessarily a bias in favour of the status quo or irrational fear of machines though those exist too; the analogy would be a person with unusually quick reaction times who drives 100k miles every year without incident still being liable to lose their licence if they're proven to have a bug in their system that makes them occasionally inclined to drive under the influence or faster than speed limits...)
And, as a practical matter, not to fail in ways that average people (aka the voters) won't see as crazy and incompetent--even if the occasional tired or under the influence human driver might have done something equally bad.
Also, let's say the accident rate was slightly better than that of an average human's.
Among people, there are some good drivers and some poor drivers. The average driving ability --- including fast reflexes for dealing with accident situations, for example --- is somewhere in between.
Is it fair to possibly force self driving cars on the good drivers?
(they probably shouldn't jump on the self-driving car bandwagon early anyway, but just a thought.)
People (especially males I've noticed) seem to equate good driving with great reflexes.
I have pretty good reflexes, but that doesn't make me a good driver.
I equate good driving with common sense, not speeding over the limit or taking unnecessary risks, not tail gating people, and in general not being a tool.
I.e. when you're driving a 2-tonne metal tank going at 80 km/h, good == safe.
In this sense, autonomous cars have already far surpassed us.
I apologize if I made it seem like I meant great reflexes meant good driving. Instead, I included it because great reflexes might help people get out of accident situations (say, like properly pulling the handbrake when the car starts spinning out of control, or turning away from a deer in the right direction last second, etc.)
Again though, I was discussing the situation in which cars were only slightly above average, meaning there were likely situations where some people would do better than it and might be putting themselves in more risk going into a self-driving car.
(if your point was more to make sure people don't take the wrong message away from my post, sorry for this response then!)
This assumes self driving without human intervention. If your in the car you can keep an eye out for danger and still be far more relaxed. Open road on a sunny day, no cars, sit and chill. Creeping along at 5mph in a traffic jam, relax it's all good. Driving in a more complex situation ok, manual time.
People suck at paying attention to boring situations. So, the easiest part of the trip is likely the best thing to automate.
Remember, the actual accident rate is very low. Further cars and humans are likely to make different mistakes so if you supervise and the car does sub second reactions and 24/7 360 degree vision the combined rate is likely to be great.
Force? I said they had a salable product. Even if every car sold today was self-driving, the average car on the road is eleven years old. It's going to be awhile before anyone is forced into anything. At least in the US.
You don't need a source for common sense. It will take years for self driving cars to become reliable. There are many adverse conditions the car will need to be able to deal with in the real world. Weather and all other chaotic systems are not to be dismissed casually if we want self driving cars to become a reality in our lifetime. Just because something will take a long time to reliably engineer doesn't mean it's not a worthy problem to solve.
I guess it depends on what you mean by "reliable". If you mean it won't crash more than a human, then we can pretty much do that under certain conditions. If you mean it will accept a trip 24/7 no matter the weather or road, it will take a while.
Do you have a source for this? Everything I can find is not apples-to-apples - they compare the reliability of a self-driving car under extremely careful and limited conditions, specifically by people who are developing, testing or evaluating the car... To the reliability of cars driving in snow, rain, glare, and operated by people with limited cognitive abilities and reflexes.
No, I was just going by what ma2rten said. But I think it is a fair comparison in the context of "show we allow this?". People drive in the rain at night without anyone considering banning it, so if we can make good-weather autonomous cars safer than that then it's good enough to start.
Well, given the current safety level of new cars, which is coming down towards 1 fatality per one billion miles driven, and a fleet of 100 000 vehicles driving autonomously for 10 000 miles per year, you need to collect data for five to ten years just to show that your autonomous cars are safer than regular cars.
Tesla, with the current AutoPilot, will have to collect data for closer to 20 years in order to have good enough statistics to claim AutoPilot is safer. (Since AutoPilot only works in some situations, they wont get 10 000 miles per year.)
If they're only reliable in controlled conditions, they're not reliable. The really hard problems that would impact the feasibility of self-driving cars (as the popular imagination sees them) arise in unanticipated situations.
After all the work that was done with the DARPA Grand Challenge(s) in the past decade, it would be an embarrassment to Google if they couldn't get the cars to work in controlled conditions, rather than it being a staggering achievement in getting them to where things are at now.
I really want self-driving cars to happen, but I see the biggest impediment to that vision of the future being the general public's level of optimism and credulity WRT this stuff, to say nothing of the tech community's optimism and credulity. It's a domain that is a nearly infinite bucket of incredibly hard problems, problems that may ultimately prove insoluble as currently specified. If everything goes well, then maybe in 20 or 50 years a lot of transportation will take place in self-driving cars which operate with sets of known constraints in environments that are (to some degree) controlled.
If everything doesn't go well, then people will keep talking about how self-driving cars are inevitable, and how in a just a few years they're going to pick you up at your house and drive you to work using the exact same roads set up the exact same way as roads are now, doing the exact same commute that they might've been doing when driving themselves. This credulity will push the money and the technology forward, until too many disappointing setbacks occur, and then all the money dries up and nobody's talking about self-driving cars anymore.
If you tell somebody that something is inevitable and almost here and it's just a matter of throwing enough resources at it, it's much easier to get people to give you money to do those things, but as soon as anything happens that doesn't fit that script, they will assume you've been lying to them all along (or just aren't credible), and the R&D money goes away. Whereas if expectations are set appropriately, it's harder to get that money, but as long as there are achievable, realistic goals and people aren't basically pitching magic, then the funding is more likely to stick around during the rough patches.
Let's assume that we have all the algorithms discovered, the prototype works excellently. For example, all the problems with snow and weather have been solved, as well as the logistical problem of making an extremely accurate 3D map of the world. Google has their prototype working everywhere: it's ready to be implemented as production quality (it has to be better than Android).
Writing high quality, reliable code (for example, in airplane control software), including testing etc, tends to take 5-8 years. So from a project management perspective, that's the reasonable estimate.
Why would anyone believe a company's claims? I'm not trying to be a jerk, but why would we trust a source who has a vested interest in misleading us?
Maybe phrased less combatively: Do you have a source that's not Google (or a Google-slanted PR piece) that has hard data about what's possible today and what will be possible soon?
Singapore definitely has some of the worst driving standards of any major city I've lived in around the world, so if you can stand up against them with a self-driving car, that is definitely promising.
I'm comparing Singapore to other first-world big cities, but there are other places in Asia where you could make a similar claim.
One of the most surprising discoveries upon moving there was that, unlike nearly everywhere in the West, in Singapore a pedestrian does not have the right of way. In the U.S. this is ingrained into all driver education courses, but in Singapore they do not even slow down around pedestrians who are crossing a street.
The other observation that jumps out to anyone I know who has visited Singapore is how the buses are driven. A Western bus driver would likely lose his job for driving as erratically as they do in Singapore. In fact just a couple weeks ago one of these drivers killed a pedestrian [0, 1]. I'm surprised it doesn't happen more often.
Self-driving cars would be in for quite a test on the roads of Singapore.
In America, for better or worse, it would take a long time to mandate these things if they were even possible (considering that the same laws would need to be passed at least 52 times for states and territories). America has checks and balances--everybody gets their say.