Car crashes killed 37,133 people in the US in 2017(arstechnica.com)
arstechnica.com
Car crashes killed 37,133 people in the US in 2017
https://arstechnica.com/cars/2018/10/car-crashes-killed-37133-people-in-the-us-in-2017/
23 comments
>Why don’t you develop a face and eye tracking system to detect if the test driver is dozing off or otherwise not paying enough attention and act accordingly?
At least in Waymo's case, because their explicit goal is to go directly to level 5. Unlike some companies who think intermediate steps are desirable, for Waymo test drivers are specifically that, for testing. The accident (which was just on HN) if anything supports the hypothesis that there should either be Level 1, maybe a little 2, or 5, and that's it.
So spending time and effort on developing more complex technology like that purely for the special limited time case of R&D is probably not worth it vs just plain having better and more professional driver practices. Waymo can and should be blamed for having shifts of more then say 4 hours, insufficiently rigorous selection/training, not having a 2 person requirement, online real time observation by a control center, etc. But it's still a one-off situation. Where technology comes into play is scaling and general usage by non-professionals, but for the specific, low volume case of testing it is absolutely possible to have low-tech or no-tech handled well.
For companies who claim level 3/4 are valuable stepping stones rather then a valley, yes driver attention is integral to their proposal in general deployment and thus requiring that also be part of their technology makes sense. But it's not just over engineered for a level 5 shift, it's to some extent making excuses for what is a direct and simple management failure. If managers overwork and badly train a fork lift operator or any other equipment operator and an accident results, we don't say the fork lift needs eyeball trackers, we say management failed at following best practices and is at fault.
At least in Waymo's case, because their explicit goal is to go directly to level 5. Unlike some companies who think intermediate steps are desirable, for Waymo test drivers are specifically that, for testing. The accident (which was just on HN) if anything supports the hypothesis that there should either be Level 1, maybe a little 2, or 5, and that's it.
So spending time and effort on developing more complex technology like that purely for the special limited time case of R&D is probably not worth it vs just plain having better and more professional driver practices. Waymo can and should be blamed for having shifts of more then say 4 hours, insufficiently rigorous selection/training, not having a 2 person requirement, online real time observation by a control center, etc. But it's still a one-off situation. Where technology comes into play is scaling and general usage by non-professionals, but for the specific, low volume case of testing it is absolutely possible to have low-tech or no-tech handled well.
For companies who claim level 3/4 are valuable stepping stones rather then a valley, yes driver attention is integral to their proposal in general deployment and thus requiring that also be part of their technology makes sense. But it's not just over engineered for a level 5 shift, it's to some extent making excuses for what is a direct and simple management failure. If managers overwork and badly train a fork lift operator or any other equipment operator and an accident results, we don't say the fork lift needs eyeball trackers, we say management failed at following best practices and is at fault.
Even IF a level 5 self-driving becomes practical soon (I doubt it based on the fact that AI still lacks commonsense), many human riders would still want an option of manual override.
In the linked article of my post above, the manual override went wrong because the driver fell asleep and that could occur to any rider as well.
Obviously, one could make it harder to override the autonomous system but it would result in slower activation which could result in an accident for any car that still lacks commonsense.
In fact, taking into account the commander/rider’s state of attention should be part of commonsense needed for any level 5 self-driving system.
In the linked article of my post above, the manual override went wrong because the driver fell asleep and that could occur to any rider as well.
Obviously, one could make it harder to override the autonomous system but it would result in slower activation which could result in an accident for any car that still lacks commonsense.
In fact, taking into account the commander/rider’s state of attention should be part of commonsense needed for any level 5 self-driving system.
>Even IF a level 5 self-driving becomes practical soon (I doubt it based on the fact that AI still lacks commonsense), many human riders would still want an option of manual override.
And they should be accommodated why exactly? If they want manual they can just buy a standard car with l1/l2.
>In the linked article of my post above, the manual override went wrong because the driver fell asleep and that could occur to any rider as well.
No, not just Waymo but even GM have explicit plans for final cars to not even have so much as a steering wheel at all, and pushing for Congress to allow an exception in that case from full FMVSS-compliance. The manual override and functions are just for the development process, the whole point of level 5 is that passenger involvement is eliminated. Anything less then that isn't level 5 and gets back into the problem of almost not needing any attention, except when you suddenly do. I'm not commenting on how close they are, just on what the goal is and why developing drive attention technology purely for testing purposes makes no sense vs just having better test procedures (and paying for it).
>In fact, taking into account the commander/rider’s state of attention should be part of commonsense needed for any level 5 self-driving system.
Are you sure you actually understand what the "levels" are here? It's certainly widely regarded as an imperfect classification system, particularly at 2, 3, and 4, but 1 and 5 are clear enough and 5 is "start to finish without even a human in the vehicle at all, or everyone sleeping".
And they should be accommodated why exactly? If they want manual they can just buy a standard car with l1/l2.
>In the linked article of my post above, the manual override went wrong because the driver fell asleep and that could occur to any rider as well.
No, not just Waymo but even GM have explicit plans for final cars to not even have so much as a steering wheel at all, and pushing for Congress to allow an exception in that case from full FMVSS-compliance. The manual override and functions are just for the development process, the whole point of level 5 is that passenger involvement is eliminated. Anything less then that isn't level 5 and gets back into the problem of almost not needing any attention, except when you suddenly do. I'm not commenting on how close they are, just on what the goal is and why developing drive attention technology purely for testing purposes makes no sense vs just having better test procedures (and paying for it).
>In fact, taking into account the commander/rider’s state of attention should be part of commonsense needed for any level 5 self-driving system.
Are you sure you actually understand what the "levels" are here? It's certainly widely regarded as an imperfect classification system, particularly at 2, 3, and 4, but 1 and 5 are clear enough and 5 is "start to finish without even a human in the vehicle at all, or everyone sleeping".
> Are you sure you actually understand what the "levels" are here? It's certainly widely regarded as an imperfect classification system, particularly at 2, 3, and 4, but 1 and 5 are clear enough and 5 is "start to finish without even a human in the vehicle at all, or everyone sleeping".
I understand ‘Level 5’ and I am an active AI researcher. Do you understand what kind of technology is required to get to Level 5?
Do you think level 5 can be achieved without understanding how humans behave in a vehicle? Commonsense is needed for Level 5.
For an example: What if someone sleepwalking or rather sleep kicking and break the windows open then leave a leg or an arm hanging out of it? Do you think a Level 5 car should not monitor how the passengers behave at all?
Do you have real information/references rather than speculation that Waymo and GM have plans to launch only after achieving Level 5 rather than Level 4 which most observers agree on?
Do you have a cost/benefit calculation of why the driver detection system is “not worth it” since it is better to wait for Level 5 before launching anyway? By the way, the driver detection system is something a team of competent AI researchers can do today (perhaps imperfectly but good enough) while no one even knows for sure how to achieve Level 5.
When do you expect Level 5 can be achieved? If your answer is less than 10 years, you appear to know more than most of the top AI researchers I have talked with or read their prognostications.
I understand ‘Level 5’ and I am an active AI researcher. Do you understand what kind of technology is required to get to Level 5?
Do you think level 5 can be achieved without understanding how humans behave in a vehicle? Commonsense is needed for Level 5.
For an example: What if someone sleepwalking or rather sleep kicking and break the windows open then leave a leg or an arm hanging out of it? Do you think a Level 5 car should not monitor how the passengers behave at all?
Do you have real information/references rather than speculation that Waymo and GM have plans to launch only after achieving Level 5 rather than Level 4 which most observers agree on?
Do you have a cost/benefit calculation of why the driver detection system is “not worth it” since it is better to wait for Level 5 before launching anyway? By the way, the driver detection system is something a team of competent AI researchers can do today (perhaps imperfectly but good enough) while no one even knows for sure how to achieve Level 5.
When do you expect Level 5 can be achieved? If your answer is less than 10 years, you appear to know more than most of the top AI researchers I have talked with or read their prognostications.
>I understand ‘Level 5’ and I am an active AI researcher.
Then frankly your posts make even less sense, unless you've gotten so deep into AI research that you've lost sight of what humans can accomplish without technology.
>Do you understand what kind of technology is required to get to Level 5?
Utterly irrelevant to the discussion at hand, beyond that by definition L5 means zero human involvement is required.
>Do you think level 5 can be achieved without understanding how humans behave in a vehicle?
Yeah, absolutely. If how humans behave in the vehicle mattered it wouldn't be L5. Active sabotage is out of scope since there is no difference there then a human driver. If you are a passenger and suddenly start hitting your driver in the head on the interstate what could possibly happen? Oh right, accident and everyone is badly injured or dies.
>What if someone sleepwalking or rather sleep kicking and break the windows open then leave a leg or an arm hanging out of it?
Then they'll have shattered glass and bad cuts all over their leg or arm which would likely wake them up and also hurt a lot. They should probably tell their L5 car to go to the ER and also are you fucking serious? It's not as if humans sleeping in vehicles including ones with glass windows is some new thing and somehow an epidemic of people kicking through windows has never come to my attention. Feel free to share your stats on that one.
>Do you think a Level 5 car should not monitor how the passengers behave at all?
No. On the contrary, self driving cars monitoring their passengers at all times sounds terrifyingly dystopian. Exactly what country/agency do you do AI research for?
>Do you have real information/references rather than speculation that Waymo and GM have plans to launch only after achieving Level 5 rather than Level 4 which most observers agree on?
Amongst others: "GM Says Car With No Steering Wheel Or Pedals Ready For Streets In 2019", https://www.npr.org/sections/thetwo-way/2018/01/12/577688125.... A lot of mass publication stuff doesn't use "level 5" specifically because it's not terminology everyone will understand, but if it has zero manual controls and human involvement beyond high level orders ("Go here", "stop") that's what it is.
>Do you have a cost/benefit calculation of why the driver detection system is “not worth it” since it is better to wait for Level 5 before launching anyway?
Why would I need to? You're the one making the assertion that this is something Waymo should do, so presumably you have a cost/benefit analysis right? Waymo apparently doesn't though, and I'm just observing why that would be the case. There are plenty of actual for real jobs being done all the time worldwide that are of the "99% boredom, 1% HOLY SHIT WE'RE ABOUT TO DIE" sort, it's a tough but well known problem that can be worked on in a restricted setting with training, checklists, scheduling and support, no brand new tech required. You're asserting that they should develop a whole new untested complex technology, even if less complex, that will immediately become worthless (or even a negative for those worried about mass surveillance) in deployment. Why do you think this is better then just following decent professional best practices for R&D testing?
>When do you expect Level 5 can be achieved? If your answer is less than 10 years, you appear to know more than most of the top AI researchers I have talked with or read their prognostications.
Irrelevant, and you should really rethink your logic here if you think otherwise. Whether it's 1 year, 5 years, 10 years, or 30 years it's not the public's problem, it's Waymo's and their investors. They can't deploy L5 until it's ready, the timeline for "ready" is just their issue. If they're satisfied that with how their development is going just with human testers and training and the results are what the public wants then how they get from A to B is up to them.
Then frankly your posts make even less sense, unless you've gotten so deep into AI research that you've lost sight of what humans can accomplish without technology.
>Do you understand what kind of technology is required to get to Level 5?
Utterly irrelevant to the discussion at hand, beyond that by definition L5 means zero human involvement is required.
>Do you think level 5 can be achieved without understanding how humans behave in a vehicle?
Yeah, absolutely. If how humans behave in the vehicle mattered it wouldn't be L5. Active sabotage is out of scope since there is no difference there then a human driver. If you are a passenger and suddenly start hitting your driver in the head on the interstate what could possibly happen? Oh right, accident and everyone is badly injured or dies.
>What if someone sleepwalking or rather sleep kicking and break the windows open then leave a leg or an arm hanging out of it?
Then they'll have shattered glass and bad cuts all over their leg or arm which would likely wake them up and also hurt a lot. They should probably tell their L5 car to go to the ER and also are you fucking serious? It's not as if humans sleeping in vehicles including ones with glass windows is some new thing and somehow an epidemic of people kicking through windows has never come to my attention. Feel free to share your stats on that one.
>Do you think a Level 5 car should not monitor how the passengers behave at all?
No. On the contrary, self driving cars monitoring their passengers at all times sounds terrifyingly dystopian. Exactly what country/agency do you do AI research for?
>Do you have real information/references rather than speculation that Waymo and GM have plans to launch only after achieving Level 5 rather than Level 4 which most observers agree on?
Amongst others: "GM Says Car With No Steering Wheel Or Pedals Ready For Streets In 2019", https://www.npr.org/sections/thetwo-way/2018/01/12/577688125.... A lot of mass publication stuff doesn't use "level 5" specifically because it's not terminology everyone will understand, but if it has zero manual controls and human involvement beyond high level orders ("Go here", "stop") that's what it is.
>Do you have a cost/benefit calculation of why the driver detection system is “not worth it” since it is better to wait for Level 5 before launching anyway?
Why would I need to? You're the one making the assertion that this is something Waymo should do, so presumably you have a cost/benefit analysis right? Waymo apparently doesn't though, and I'm just observing why that would be the case. There are plenty of actual for real jobs being done all the time worldwide that are of the "99% boredom, 1% HOLY SHIT WE'RE ABOUT TO DIE" sort, it's a tough but well known problem that can be worked on in a restricted setting with training, checklists, scheduling and support, no brand new tech required. You're asserting that they should develop a whole new untested complex technology, even if less complex, that will immediately become worthless (or even a negative for those worried about mass surveillance) in deployment. Why do you think this is better then just following decent professional best practices for R&D testing?
>When do you expect Level 5 can be achieved? If your answer is less than 10 years, you appear to know more than most of the top AI researchers I have talked with or read their prognostications.
Irrelevant, and you should really rethink your logic here if you think otherwise. Whether it's 1 year, 5 years, 10 years, or 30 years it's not the public's problem, it's Waymo's and their investors. They can't deploy L5 until it's ready, the timeline for "ready" is just their issue. If they're satisfied that with how their development is going just with human testers and training and the results are what the public wants then how they get from A to B is up to them.
My assertion is that they will deploy Level 4 cars first. In fact, Waymo is deploying Level 4 cars to the public soon.
If you are so sure that Waymo and GM will deploy Level 5 and skip Level 4, based on your level of understanding of the technology and the economics involved, please publish a blog post or an Ask HN and see how many experts agree/disagree with you.
Otherwise, are you willing to bet with the people who disagree?
By the way, no steering wheel does not mean Level 5.
https://www.techrepublic.com/article/autonomous-driving-leve...
If you are so sure that Waymo and GM will deploy Level 5 and skip Level 4, based on your level of understanding of the technology and the economics involved, please publish a blog post or an Ask HN and see how many experts agree/disagree with you.
Otherwise, are you willing to bet with the people who disagree?
By the way, no steering wheel does not mean Level 5.
https://www.techrepublic.com/article/autonomous-driving-leve...
This already exists and is being deployed as a consumer-grade feature.
I think that this will be a necessity for mass rollout of self-driving cars.
I think that this will be a necessity for mass rollout of self-driving cars.
I think many cars have fatigue detection systems already.
Driver monitoring is an area of active development.
I think 50 years from now we're going to look at this fatality rate much like we look at 1930s-1950s era cigarette advertising, or pre-antiseptic era medieval medical practices like bloodletting.
So, 10+ 9/11 equivalents (compared on number of deaths) in one year.
The autonomous driving companies are aiming to change this headline to "People driving cars killed 37,133 other people", this is similar to the argument that "Guns kill people", isn't it?
"Guns kill people" is a straw man. The actual argument is that guns make it easier to kill people (and yourself) on a whim.
Autonomous vehicles have nothing to do with gun control because no one is trying to teach handguns to operate themselves.
Autonomous driving companies can and should be judged on how well their software lowers rates of collisions compared to human drivers.
Autonomous vehicles have nothing to do with gun control because no one is trying to teach handguns to operate themselves.
Autonomous driving companies can and should be judged on how well their software lowers rates of collisions compared to human drivers.
How does this compare to elsewhere in the world, in some comparable metric (e.g deaths per driven distance or hour)?
It's very bad. The US is an outlier when it comes to developed countries. It's closer to third world levels.
Why don’t you develop/deploy a face and eye tracking system to detect if the test driver is dozing off or otherwise not paying enough attention and act accordingly? (Warn first, and if needed, take the car to a safe spot and stop.)
The system should be useful in real deployment with passenger/driver on board as well, since the self-driving car needs to know if it should obey human’s manual command, which should also depend on the person’s state of attention at the moment.
[1] Waymo’s self-driving car crashed because its human driver fell asleep at the wheel https://qz.com/1410928/waymos-self-driving-car-crashed-becau...
[2] In addition, we all heard about the case which the test driver was watching a movie and a pedestrian was killed.