8 ms·

The truth is we're really not that close to fully-autonomous cars. Even Elon Musk's most ambitious projections for "autopilot" are a car that can drive itself

by danhak 11y ago

The truth is we're really not that close to fully-autonomous cars. Even Elon Musk's most ambitious projections for "autopilot" are a car that can drive itself 90 - 95% of the time but will require an alert, licensed driver to take over if necessary. Secondly, I don't feel that the possible rise of truly autonomous vehicles (if that even occurs within the next ten years) affects the main point of this article. In fact, the advantages of battery-electric over gasoline become even more pronounced in a world of self-driving cars. The only drawback (speed to refuel) becomes moot as fully-charged vehicles can be dispatched while others direct themselves to charging or battery-swap stations.

toomuchtodo 11y ago
> The truth is we're really not that close to fully-autonomous cars. Google cars are pretty damn close to level 4 autonomous vehicles. They've driven more than a million miles fully autonomously, and Google runs their self-driving vehicle algorithms against simulated virtual worlds constantly (with data previously collected from their self driving cars). We are so much closer to fully autonomous cars than people think.
- jtokoph 11y ago
  I've always thought that the Google cars have driven 95%+ of their miles in Mountain View, CA where every single aspect of the roads have been digitized, tested and tweaked a million times over. If you stuck one of these cars in a brand new city, it wouldn't work nearly as well.
  - toomuchtodo 11y ago
    How quickly could a fleet of Google vehicles ingest a major metros geographic data? Assume a similar project as Street View, with minimum wage drivers piloting the data collection vehicles without autonomous mode turned on. Days? Weeks? Under 6 months for sure. EDIT: Current street view coverage: https://support.google.com/maps/answer/68384?hl=en https://support.google.com/maps/answer/68384?hl=en
  - thaumasiotes 11y ago
    Mapping the roads isn't challenging. How do you think humans perform at driving through brand new cities?
    - CountSessine 11y ago
      Humans don't need cm-scale precise maps in cities to avoid jumping curbs and running into streetlamps. But that's still a requirement for all autonomous cars that have been demonstrated.
      - thaumasiotes 11y ago
        I don't see why this matters, though. cm-scale maps aren't hard to produce.
        CountSessine 11y ago
        Aren't hard to produce everywhere?
        swombat 11y ago
        Not much harder to produce than the current street view maps that are almost everywhere...
      - peterfirefly 11y ago
        Not all autonomous cars require that, either. Nor is there any reason whatsoever to believe it will be a requirement in the future.
        CountSessine 11y ago
        No, but researchers agree that self-mapping or map-unassisted autonomous cars are much further off into the future than are perfect-map cars, like the ones Google is demonstrating. And the equipment (e.g. LIDAR) on the car is super-expensive. And has lots of moving parts. And is easily foiled. Etc.
  - mbreese 11y ago
    And the weather is nice 99% of the time. I'd like to see a Google self driving car work in the rain or snow.
    - bsder 11y ago
      You say this like people are good at driving in rain or snow. They are not. There is a reason for all of those pileup pictures whenever there is severe weather in the US. Self-driving cars are going to be WAY better in bad weather than any human could ever be. They will be able to "see" in conditions that humans couldn't dream of. They will be able to control steering far better on slick surfaces. They will be able to respond to emergencies far faster. I have said this before but it bears repeating, once we get self-driving cars, the insurance companies won't actually let you drive anymore.
      - danieltillett 11y ago
        I tend to agree on the insurance front, but I suspect that it won't be a ban, just a higher rate if you choose to drive manually.
      - CountSessine 11y ago
        You say this like people are good at driving in rain or snow. They are not. I've driven in 42 winters on icy roads with thick snow and limited visibility, without the aid of GPS which has a funny way of disappearing in inclement weather. Google's cars have been demonstrated driving in exactly 0 winters under those conditions. Because they know it would be a disaster. There is a reason for all of those pileup pictures whenever there is severe weather in the US. They're driving too fast. But they knew that. They just decided to do it anyways. But they didn't drive that fast thinking that the snow was actually falling leaves or that the windshield was dirty or that the car was suddenly moving backwards, or any of the other ridiculous things that an AI will begin thinking when it falls out of a switch statement, encounters a case it wasn't trained on, or loses its GPS signal. Self-driving cars are going to be WAY better in bad weather than any human could ever be I have said this before but it bears repeating, once we get self-driving cars, the insurance companies won't actually let you drive anymore. I completely agree. One day they will be completely superior to human drivers. But that isn't today and won't be for years. Many years. Like decades.
  - roel_v 11y ago
    That other company of which there was an article here a few months ago drove a car autonomously from LA to NYC. Volvo drives cars around Stockholm. BMW has self driving cars on the road. There are self driving trucks right now at work on oil fields in Alaska and in mines in Australia. It's true that we're not really there yet. And yes there are hurdles like regulation, what about heavy rain etc. But the pace at which things are going...
    - toomuchtodo 11y ago
      Freightliner has an autonomous semi-truck (the "Inspiration") working on Nevada roads currently: http://spectrum.ieee.org/cars-that-think/transportation/self-driving/freightliner-unveils-first-autonomous-semitruck-licensed-to-drive-itself-on-highways http://spectrum.ieee.org/cars-that-think/transportation/self... It (for now) requires a human driver, but operates at autonomous level 3.
    - CountSessine 11y ago
      No carmaker has a city-driving autonomous vehicle that doesn't require super-precise, cm-scale, unrealistically always-up-to-date maps for the AI. The autonomous super-size dump trucks drive along carefully pre-programmed routes using GPS. We're still a long ways away from practical and safe self-driving cars. AI and computer vision just isn't there yet.
      - swombat 11y ago
        "A long way away" - only true in terms of linear extrapolations of current progress. Think in terms of exponential progress, and we're about 5 years from the majority of trips being done by self-driving car, IMHO. The benefits of self-driving cars are so powerful that they will take over the world even faster than the electric vehicles described in this article...
      - CountSessine 11y ago
        Think in terms of exponential progress, and we're about 5 years from the majority of trips being done by self-driving car, IMHO 5 years is a fantasy. Continued exponential increases in device computing power is by no means guaranteed now that chips are at the 14nm feature size. And the recent progress in AI can only be partially attributed to improved computing power - much of it was from improved approaches to training neural networks. We could stumble into another AI winter and have to wait another 30 years for the next round of big algorithmic AI breakthroughs. I guarantee you that self-driving cars will not be the mainstream even in 15 years. Probably more.
  - ahmacleod 11y ago
    There was a time when you could say the same about street view. Now they're working on the bottom of the ocean.
  - krschultz 11y ago
    That's the whole innovation. They reframed the problem in order to make it easier. Everyone used to think about how to get the car to comprehend everything in the world. Google just provides the car a highly detailed map and then the car has fewer things to focus on. They replace some of the complexity by using huge amounts of data. It's a very 'Google' solution really. This is why Uber and others are scrambling to buy mapping companies. In the future the software goes hand in hand with a hyper detailed map.
  - nulltype 11y ago
    I watched one of these a few days ago get stuck at a 4 way stop near Castro St. I think the human driver had to take over.
- danhak 11y ago
  > They've driven more than a million miles fully autonomously Zero of those miles were at a speeds > 25 mph. And the cars have averaged > 1 accident per 100,000 miles.
  - toomuchtodo 11y ago
    Audi's TTS autonomous vehicle sped up Pike's Peak successfully 5 years ago: http://www.wired.com/2010/11/audis-robotic-car-climbs-pikes-peak/ http://www.wired.com/2010/11/audis-robotic-car-climbs-pikes-... "The car made five runs during a week of testing, reaching the 14,100-foot summit in 27 minutes. The very fastest drivers — in cars putting down more than 900 horsepower — do it in 10 minutes and change, so clearly the car isn’t terribly fast. But the point isn’t the speed, it’s the fact there was no one at the wheel. According to Audi, race officials said an expert driver in car like the TTS would complete the course in around 17 minutes." Note that none of the accidents Google cars have been involved in were while the vehicles were in autonomous mode. http://www.google.com/selfdrivingcar/reports/ http://www.google.com/selfdrivingcar/reports/
    - mod 11y ago
      None of them while a car was moving in autonomous mode* As I recall there were collisions where the autonomous car was sitting still. Others while in manual mode. Zero the fault of autonomous drivers, so far.
      - toomuchtodo 11y ago
        Regardless, autonomous vehicles will improve at a staggering pace due to computing power available to them, and sensors rapidly developing. ~33K [1] deaths a year from car accidents currently due to human drivers. I doubt autonomous vehicles can do worse than that. [1] http://www.cdc.gov/nchs/fastats/accidental-injury.htm http://www.cdc.gov/nchs/fastats/accidental-injury.htm
  - mikeyouse 11y ago
    I don't think this is accurate.. It was pretty common to pass the Lexus SUVs that Google was using for their autonomous platforms on the highway at speeds of 65+ MPH. I'm not sure if they're still around since my commute changed, but they were previously all over the Bay Area. I personally saw them on HWYs 101, 280, 880, and 92..
- ninkendo 11y ago
  The number of miles driven autonomously doesn't really matter if drivers have to intervene even for a second to resolve an issue. If the car is driving by itself 99% of the time, that's not good enough for the "share an autonomous car" use case. It's no good at 99.9% either. Or 99.99%. It's a lot more interesting if google's cars actually had no steering wheel or way of intervening. Because that's what we're talking about when we talk about the price reduction of a fleet of autonomous cars: the car driving to your house, completely on its own, with nobody in the car. 99.9% isn't good enough, and we're not even close to 99.9%. The difficulty of the last remaining difficult problems will only increase as we approach 100%, too. I think completely driverless cars are a problem on a similar scale as AI or nuclear fusion, but with the timescales decreased... we're probably going to be always perpetually "five years out" from having them (as opposed to fusion's perpetual "30 years"), even though we're making improvements all the time. We just don't understand the scale of the problems we haven't solved yet.
  - waterlesscloud 11y ago
    "It's no good at 99.9% either. Or 99.99%." That's not completely right. If that 99.9% covers 100% of certain use cases, they will be used for those cases. It's like much else in tech, it's not an instant change, it's going to be all about expanding the uses cases over time.
    - hvidgaard 11y ago
      That is the thing with driving - it's amongst other drivers, pedestrians, especially children. There are so few use cases where 99.99% is enough. I can only think of one, highway from A to B - how you get to and from the A and B respectively is not by selfdriving cars.
    - ninkendo 11y ago
      The thing with autonomous ride sharing though, is that the car has to make it from its last stop to your home on its own. And there's any number of scenarios that can happen on its way... a traffic light could lose power and turn into a 4-way stop, for instance. Or a big event at a large venue where hundreds or thousands of pedestrians are crossing the street and you kinda have to inch forward and wait for some of them to wave you on to go forward (basically everybody jaywalks at that point.) There's a million different 0.01% scenarios that could happen, and if you ramp the usage of a really really good autonomous car up to scale, you'll see them every day. Unless it can handle all of them without intervention, you simply can't ditch the driver, or you'd face huge PR nightmares when you either hit a pedestrian or you block an intersection and cause huge traffic in a major metropolitan area ("Mayor bans autonomous cars due to one too many gridlocks caused by them", etc.) I really think it's down to the PR issues and the fact that one too many mistakes can really kill the idea, that companies will be cautious to just not need a driver altogether. And for autonomous ride sharing, you must be able to do nearly half your miles without a driver. Hence why that last .01% is still crucial.
- Qwertious 11y ago
  Do they drive properly in rain/fog yet? IIRC they can't drive in rain, which is a hell of a limitation.

by danhak 11y ago

toomuchtodo 11y ago

> The truth is we're really not that close to fully-autonomous cars. Google cars are pretty damn close to level 4 autonomous vehicles. They've driven more than a million miles fully autonomously, and Google runs their self-driving vehicle algorithms against simulated virtual worlds constantly (with data previously collected from their self driving cars). We are so much closer to fully autonomous cars than people think.

jtokoph 11y ago
I've always thought that the Google cars have driven 95%+ of their miles in Mountain View, CA where every single aspect of the roads have been digitized, tested and tweaked a million times over. If you stuck one of these cars in a brand new city, it wouldn't work nearly as well.
- toomuchtodo 11y ago
  How quickly could a fleet of Google vehicles ingest a major metros geographic data? Assume a similar project as Street View, with minimum wage drivers piloting the data collection vehicles without autonomous mode turned on. Days? Weeks? Under 6 months for sure. EDIT: Current street view coverage: https://support.google.com/maps/answer/68384?hl=en https://support.google.com/maps/answer/68384?hl=en
- thaumasiotes 11y ago
  Mapping the roads isn't challenging. How do you think humans perform at driving through brand new cities?
  - CountSessine 11y ago
    Humans don't need cm-scale precise maps in cities to avoid jumping curbs and running into streetlamps. But that's still a requirement for all autonomous cars that have been demonstrated.
    - thaumasiotes 11y ago
      I don't see why this matters, though. cm-scale maps aren't hard to produce.
      - CountSessine 11y ago
        Aren't hard to produce everywhere?
        swombat 11y ago
        Not much harder to produce than the current street view maps that are almost everywhere...
    - peterfirefly 11y ago
      Not all autonomous cars require that, either. Nor is there any reason whatsoever to believe it will be a requirement in the future.
      - CountSessine 11y ago
        No, but researchers agree that self-mapping or map-unassisted autonomous cars are much further off into the future than are perfect-map cars, like the ones Google is demonstrating. And the equipment (e.g. LIDAR) on the car is super-expensive. And has lots of moving parts. And is easily foiled. Etc.
- mbreese 11y ago
  And the weather is nice 99% of the time. I'd like to see a Google self driving car work in the rain or snow.
  - bsder 11y ago
    You say this like people are good at driving in rain or snow. They are not. There is a reason for all of those pileup pictures whenever there is severe weather in the US. Self-driving cars are going to be WAY better in bad weather than any human could ever be. They will be able to "see" in conditions that humans couldn't dream of. They will be able to control steering far better on slick surfaces. They will be able to respond to emergencies far faster. I have said this before but it bears repeating, once we get self-driving cars, the insurance companies won't actually let you drive anymore.
    - danieltillett 11y ago
      I tend to agree on the insurance front, but I suspect that it won't be a ban, just a higher rate if you choose to drive manually.
    - CountSessine 11y ago
      You say this like people are good at driving in rain or snow. They are not. I've driven in 42 winters on icy roads with thick snow and limited visibility, without the aid of GPS which has a funny way of disappearing in inclement weather. Google's cars have been demonstrated driving in exactly 0 winters under those conditions. Because they know it would be a disaster. There is a reason for all of those pileup pictures whenever there is severe weather in the US. They're driving too fast. But they knew that. They just decided to do it anyways. But they didn't drive that fast thinking that the snow was actually falling leaves or that the windshield was dirty or that the car was suddenly moving backwards, or any of the other ridiculous things that an AI will begin thinking when it falls out of a switch statement, encounters a case it wasn't trained on, or loses its GPS signal. Self-driving cars are going to be WAY better in bad weather than any human could ever be I have said this before but it bears repeating, once we get self-driving cars, the insurance companies won't actually let you drive anymore. I completely agree. One day they will be completely superior to human drivers. But that isn't today and won't be for years. Many years. Like decades.
- roel_v 11y ago
  That other company of which there was an article here a few months ago drove a car autonomously from LA to NYC. Volvo drives cars around Stockholm. BMW has self driving cars on the road. There are self driving trucks right now at work on oil fields in Alaska and in mines in Australia. It's true that we're not really there yet. And yes there are hurdles like regulation, what about heavy rain etc. But the pace at which things are going...
  - toomuchtodo 11y ago
    Freightliner has an autonomous semi-truck (the "Inspiration") working on Nevada roads currently: http://spectrum.ieee.org/cars-that-think/transportation/self-driving/freightliner-unveils-first-autonomous-semitruck-licensed-to-drive-itself-on-highways http://spectrum.ieee.org/cars-that-think/transportation/self... It (for now) requires a human driver, but operates at autonomous level 3.
  - CountSessine 11y ago
    No carmaker has a city-driving autonomous vehicle that doesn't require super-precise, cm-scale, unrealistically always-up-to-date maps for the AI. The autonomous super-size dump trucks drive along carefully pre-programmed routes using GPS. We're still a long ways away from practical and safe self-driving cars. AI and computer vision just isn't there yet.
    - swombat 11y ago
      "A long way away" - only true in terms of linear extrapolations of current progress. Think in terms of exponential progress, and we're about 5 years from the majority of trips being done by self-driving car, IMHO. The benefits of self-driving cars are so powerful that they will take over the world even faster than the electric vehicles described in this article...
    - CountSessine 11y ago
      Think in terms of exponential progress, and we're about 5 years from the majority of trips being done by self-driving car, IMHO 5 years is a fantasy. Continued exponential increases in device computing power is by no means guaranteed now that chips are at the 14nm feature size. And the recent progress in AI can only be partially attributed to improved computing power - much of it was from improved approaches to training neural networks. We could stumble into another AI winter and have to wait another 30 years for the next round of big algorithmic AI breakthroughs. I guarantee you that self-driving cars will not be the mainstream even in 15 years. Probably more.
- ahmacleod 11y ago
  There was a time when you could say the same about street view. Now they're working on the bottom of the ocean.
- krschultz 11y ago
  That's the whole innovation. They reframed the problem in order to make it easier. Everyone used to think about how to get the car to comprehend everything in the world. Google just provides the car a highly detailed map and then the car has fewer things to focus on. They replace some of the complexity by using huge amounts of data. It's a very 'Google' solution really. This is why Uber and others are scrambling to buy mapping companies. In the future the software goes hand in hand with a hyper detailed map.
- nulltype 11y ago
  I watched one of these a few days ago get stuck at a 4 way stop near Castro St. I think the human driver had to take over.
danhak 11y ago
> They've driven more than a million miles fully autonomously Zero of those miles were at a speeds > 25 mph. And the cars have averaged > 1 accident per 100,000 miles.
- toomuchtodo 11y ago
  Audi's TTS autonomous vehicle sped up Pike's Peak successfully 5 years ago: http://www.wired.com/2010/11/audis-robotic-car-climbs-pikes-peak/ http://www.wired.com/2010/11/audis-robotic-car-climbs-pikes-... "The car made five runs during a week of testing, reaching the 14,100-foot summit in 27 minutes. The very fastest drivers — in cars putting down more than 900 horsepower — do it in 10 minutes and change, so clearly the car isn’t terribly fast. But the point isn’t the speed, it’s the fact there was no one at the wheel. According to Audi, race officials said an expert driver in car like the TTS would complete the course in around 17 minutes." Note that none of the accidents Google cars have been involved in were while the vehicles were in autonomous mode. http://www.google.com/selfdrivingcar/reports/ http://www.google.com/selfdrivingcar/reports/
  - mod 11y ago
    None of them while a car was moving in autonomous mode* As I recall there were collisions where the autonomous car was sitting still. Others while in manual mode. Zero the fault of autonomous drivers, so far.
    - toomuchtodo 11y ago
      Regardless, autonomous vehicles will improve at a staggering pace due to computing power available to them, and sensors rapidly developing. ~33K [1] deaths a year from car accidents currently due to human drivers. I doubt autonomous vehicles can do worse than that. [1] http://www.cdc.gov/nchs/fastats/accidental-injury.htm http://www.cdc.gov/nchs/fastats/accidental-injury.htm
- mikeyouse 11y ago
  I don't think this is accurate.. It was pretty common to pass the Lexus SUVs that Google was using for their autonomous platforms on the highway at speeds of 65+ MPH. I'm not sure if they're still around since my commute changed, but they were previously all over the Bay Area. I personally saw them on HWYs 101, 280, 880, and 92..
ninkendo 11y ago
The number of miles driven autonomously doesn't really matter if drivers have to intervene even for a second to resolve an issue. If the car is driving by itself 99% of the time, that's not good enough for the "share an autonomous car" use case. It's no good at 99.9% either. Or 99.99%. It's a lot more interesting if google's cars actually had no steering wheel or way of intervening. Because that's what we're talking about when we talk about the price reduction of a fleet of autonomous cars: the car driving to your house, completely on its own, with nobody in the car. 99.9% isn't good enough, and we're not even close to 99.9%. The difficulty of the last remaining difficult problems will only increase as we approach 100%, too. I think completely driverless cars are a problem on a similar scale as AI or nuclear fusion, but with the timescales decreased... we're probably going to be always perpetually "five years out" from having them (as opposed to fusion's perpetual "30 years"), even though we're making improvements all the time. We just don't understand the scale of the problems we haven't solved yet.
- waterlesscloud 11y ago
  "It's no good at 99.9% either. Or 99.99%." That's not completely right. If that 99.9% covers 100% of certain use cases, they will be used for those cases. It's like much else in tech, it's not an instant change, it's going to be all about expanding the uses cases over time.
  - hvidgaard 11y ago
    That is the thing with driving - it's amongst other drivers, pedestrians, especially children. There are so few use cases where 99.99% is enough. I can only think of one, highway from A to B - how you get to and from the A and B respectively is not by selfdriving cars.
  - ninkendo 11y ago
    The thing with autonomous ride sharing though, is that the car has to make it from its last stop to your home on its own. And there's any number of scenarios that can happen on its way... a traffic light could lose power and turn into a 4-way stop, for instance. Or a big event at a large venue where hundreds or thousands of pedestrians are crossing the street and you kinda have to inch forward and wait for some of them to wave you on to go forward (basically everybody jaywalks at that point.) There's a million different 0.01% scenarios that could happen, and if you ramp the usage of a really really good autonomous car up to scale, you'll see them every day. Unless it can handle all of them without intervention, you simply can't ditch the driver, or you'd face huge PR nightmares when you either hit a pedestrian or you block an intersection and cause huge traffic in a major metropolitan area ("Mayor bans autonomous cars due to one too many gridlocks caused by them", etc.) I really think it's down to the PR issues and the fact that one too many mistakes can really kill the idea, that companies will be cautious to just not need a driver altogether. And for autonomous ride sharing, you must be able to do nearly half your miles without a driver. Hence why that last .01% is still crucial.
Qwertious 11y ago
Do they drive properly in rain/fog yet? IIRC they can't drive in rain, which is a hell of a limitation.