James Bond, sitting in the night train, is having dinner with Vesper Lynd. "Wait a minute, you're saying self-driving cars are not safer than human drivers?". Bemused, James takes a sip of Bordeaux, and quips "You've noticed?".
I am a signalling engineer, with experience in design and delivery of autonomous trains (really autonomous, no driver, no nothing). Seriously, the conclusion makes me want to scream:
> Shouldn’t ADS developers be required to prove an ADS is at least as safe as a human driver for a specific ODD before allowing it out onto public roads?
To deliver an ADS thou shall:
1/ Identify the standards you are going to comply to (and I am talking about real standards, like EN 50126, SMS compliant with AS9100, not a Waymo brochure)
2/ Define your safety objectives (hint: for an ADS, it's going to be way way waaaaaay higher than human error rate). You don't self-define these. You need to talk to the safety authorities of the area were you intend to circulate.
3/ Build a safety case explaining how you achieve credibly these objectives (spoiler: testing will not be enough)
4/ Have your case examined by an independant safety assessor, accredited by a relevant national or trans-national authority
5/ Put your ego on a shelf.
Otherwise, thou shall kill people.
Simple. But it seems the automotive industry is ready to go to any length to avoid going through the concept of an independant safety assessor.
It's also not good enough for them to be better than average human drivers. There are many bad drivers and if you are a good driver, you wouldn't want to switch to anything less safe than your driving
Although a driver is not at fault in every accident they are involved in. Therefore it is possible that the good driver would be safer with an increased self-driving environment even if that individual's odds of causing an accident increase as long as the odds of other people causing an accident decrease.
I have nothing but a hunch to back it up, but I have a feeling getting the worst drivers off the road would provide a greater benefit to public safety that would outweigh the damage caused by downgrading the skills of the best drivers.
Some drivers are “better than others” but there are no “good drivers”.
Everyone gets tired, everyone’s gets older, everyone gets distracted, everyone drives in a new location (potentially with a different driving style).
I don’t know when it tips to be better off with the computer, but I’ll feel comfortable as soon as the crash/fatalities rates are equal or better than Uber/Lyft.
So if there are no 'good drivers', everyone must be a 'bad driver', so if self driving cars are to be 'good', they need to be better than at least 100% of humans. High bar!
> It's also not good enough for them to be better than average human drivers
Of course it is, because it means there will be less deaths per million miles driven, which surely is the goal.
yes, eventually we want them to be better than the average driver, then better then the best, then (hypothetically) perfect. But the reality is the day they're better than the average, we'll be saving lives if we all switch over.
If you are talking about making it mandatory for everyone to switch, so human drivers are banned from the roads, I doubt you will be able to sell mere equality of performance by autonomous systems to the voting public. You can make as many greater good statistical arguments as you like, but you are still saying you are going to make many individual drivers and their passengers less safe.
This kind of equivalence is also ignoring the possibility of a catastrophic system failure that simply can't happen with human drivers.
no, it would only mean less deaths per million miles driven if those worse than self driving cars switch to it. If the best drivers switch to self driving cars, there will be more deaths per million miles. Duh
Have you ever accelerated into the back of a fire truck, which was stationary on the freeway, with bright flashing lights on it, in dry conditions, with ideal visibility and lighting, while travelling at a mere 31 miles per hour? [1]
I mean, I acknowledge that some accidents will always be unavoidable. Urban driving where someone steps out from between vehicles. Challenging conditions like fog and dust clouds. Sensor malfunctions. A driver coming in the opposite direction swerving into your lane. I'm not saying a self-driving car should avoid 100% of accidents.
But avoiding large clearly visible stationary objects? I'd like to think most drivers on the road can manage that.
There are things you can easily check:
Number of prior crashes.
Do you drive alone?
Do you listen to radio/music/podcast while driving?
Are you texting, making phone calls or talking to passengers while driving?
Do you drive late/when you are tired?
Do you always drive the same routes?
Any health issues, age related or otherwise?
How good is your eye sight and your hearing?
I am sure there are still more things you can check objectively.
1. You can only really go on metrics of crashes with your fault, currently.
2. You can't really be certain
3. It would be good to be able to collect data of driver faults without crashes. But how do we do it without giving insurance companies more reasons to hike prices?
and it seems to me from a bit of lookings for statistics just now that if you've been driving for a few years and haven't crashed, you're better than average
> Levels 3-5 are considered Automated Driving Systems (ADSs) in which the driver does not need to pay attention to the road.
I'm going to say that's patently wrong.
as the article points out at level 3, the control system is likely to drop out and punt hard decisions over to the driver at any time. This is exceptionally dangerous because there is almost never enough time to react properly (unless the driver was actively paying attention)
at level 3 you really still need to pay attention.
I think Lvl3 is one of the most dangerous levels a car can be at, based on the experience Waymo (back then Chauffeur) made: Humans are even worse at monitoring than they are at driving. It's incredibly boring. Their test-drivers (who got extensive training for what to do and what not to do) started to do anything but pay attention, including taking a nap, before they pulled the plug on the incremental approach.
And we see this partially with Tesla now. Some drivers start to trust the system so much that they start to do anything but pay attention - partially even bypassing the pesky systems that remind them that they are required to take over at any moment.
I think you're being a little selective with your quoting there, as the article goes on to say
> One big issue for Level 3 vehicles is that a crash might occur in the 10 seconds the driver spends taking over, so Level 3 vehicles will probably need to include an ODD [operational design domain] where 10 seconds is reasonably safe (e.g. low-speed highway traffic jams).
Without a nationwide Vehicle to Vehicle (V2V) and Infrastructure to Vehicle (I2V) communication system the answer is probably "No" and that's the elephant in the room no one is really talking about yet.
That's because the I2V would be hugely expensive to implement and maintain. Car companies are not going to pay for it, and it'd be a tough sell to taxpayers.
More expensive cars and roads are not a great move right now. When we need is more efficient and less expensive cars and better driver training programs.
>When we need is more efficient and less expensive cars and better driver training programs.
I bet human drivers are already mostly safe - I think most accidents are broadly in 3 categories, (1) impairment (drinking, age, sleep) or (2) distraction (cell phone, makeup) or (3) weather. The first two wont be solved by better training programs and the last would be really hard to make a requirement (how do you train people to drive in the snow in Los Angeles)?. If you want safer roads, it would be more wise to spend that money on public transportation. On a per capita basis it would be cheaper and safer.
That said I don't think the rush to autonomous driving is as altruistic as the developers claim. For $250B you could massively overhaul the transportation systems for major cities. However there is little upside to capture by doing do; while autonomous driving has the potential to create a last-mile autonomous transportation system that would massively reduce the costs of transporting goods and people (by removing labor costs). Capturing that upside is far more lucrative.
I think the question is less "we will only see AV commonplace when it's unilaterally safer than humans", and more "we will see AV commonplace when the value generated exceeds the potential costs of fatal accidents". It's after AV is commonplace will the value captured from human drivers be used to reduce costs in the failures of AV. For example, if UPS were to convert their 50% fleet to AV, then they would use the money saved to then invest in V2V or I2V infrastructure in order to accelerate the transition of the rest of their fleet.
I see you're looking at this mostly from a business value perspective and I guess in the end that's what it all comes down to, but I feel there is so much more value added than just "less drivers needed."
The cost of I2V for example can be offset by less waste in both space used to park cars, drive cars and maintain cars. You can utilize roads much better by having a lot of tiny single passenger vehicles that park themselves outside of the city when traffic is low. Also, travel time with public transportation is always longer than point-to-point, so people save time over that. There's too much for a HN comment that I can think of as benefits of AV.
Since you're allowed to drive in the snow when you have a driver's license it should be part of the training. You can get skid training courses where a slippery road is 'simulated' on a wet plate which should be possible in LA ;)
Also, aren't steep fines a good "training program" for distractions like holding your phone and messaging when driving?
What we “need” is hard to say, but what we will get is what will sell more cars. Maybe those two align, but there might not beer a strong connection between them.
For example ambient lighting inside the car is largely a gimmick, but a fun one. Clearly not something benefiting society, but it seems to sell cars.
I have my doubts about V2V and I2V communications. They'd simplify the task for automated systems, but come with a rat tail of problems.
I2V is incredibly hard to scale, and is a duplicated implementation of a pre-existing system (road markings & signs). Any system with duplicated implementations will grow inconsistencies. In those situations, the automated driver will be driving based on a differently perceived environment than the human driver - and it'll be hard to know. The automated system would also need to read lane markings & signs to recognize these inconsistencies and act accordingly, at which point the I2V communication is just a crutch to help the automated vehicle, reducing it's utility and hence the cost-value ratio (which is already insanely high).
V2V communication is unreliable by the fact that older vehicles won't support it, so automated vehicles won't be able to communicate with some (initially: most) cars. They'll need to be able to drive safely without communicating with nearby vehicles, so V2V is again just a "bonus" system a vehicle must not rely on.
Automated drivers will be safer than human drivers with time. It's an incredibly hard problem to solve, but it will be solved. Technology is improving continuously, and different companies are trying different approaches to solve the problem - which makes me optimistic in terms of automated drivers. Becoming safer than human drivers is mostly a matter of being able to drive with an automated system based on infrastructure optimized for humans. Sources of accidents will shift: From driver impairment, distraction and ignorance of rules, to misinterpretation of the environment. Additionally, and I think this is the key: If an accident happens, at best the involved human drivers improve their reaction to the specific situation. Automated drivers can share this new knowledge across the entire fleet.
Do you think it's (near-)impossible for computers to at least equal an awake human driver without that? If SDCs can "simply" achieve near-human capabilities, then they've already exceeded human safety simply by virtue of not being subject to the same distractions and impairments as humans.
Mind you, this is a hypothetical argument. If any of the major players thought they could make a strong statistical safety case, they'd already be publishing press releases and scaling.
> More expensive cars and roads are not a great move right now. When we need is more efficient and less expensive cars and better driver training programs.
You’re making a good faith argument.
Would increasing minimums for car insurance make us more or less safe? I believe it would make us more safe, intuitively, and the effect would be reinforced by taking people off the road who can’t really afford to drive.
The elephant in the room is public transportation motivating bad drivers to stay away from the car. "Oh but they won't," and maybe, but I'm pretty sure there are a lot of people who would like to have someone else drive, not to mention bypassing the tickets and accidents (and repairs) that bad drivers experience.
Well, the biggest elephant in the room is ...there are no self driving cars out there in the wild.
Any near future self driving car will not be safer than a sobar, alert human driver any time soon, unless they have a directive to halt the vehicle at any difficult scenario.
We will not know before such a car exists and is tested under the exact same circumstances vs a human driver.
The edge cases are problem not only because they actually are a technical problem (which is undeniable) but because edge cases are an optics problem.
When self-driving goes wrong due to mishandled edge cases, people will be willing to ignore massive amounts of statistical evidence that self-driving cars are safer than humans on average, and obsessively focus on the the failed edge cases.
"I don't care if self-driving cars have N times fewer accidents per kilometer, what about how that one one time when the Wile E. Coyote painted a picture of a tunnel on a cliff, and then a self-driving car tried to go into it and blew up in a giant ball of fire? It was all over the news man; self-driving is a joke."
Higher safety can be achieved by them being simply “as good” or even worse to some extent but the removal of the driver from multi passenger rides reduces casualties sufficiently to make them safer on average.
Readit News
I am a signalling engineer, with experience in design and delivery of autonomous trains (really autonomous, no driver, no nothing). Seriously, the conclusion makes me want to scream:
> Shouldn’t ADS developers be required to prove an ADS is at least as safe as a human driver for a specific ODD before allowing it out onto public roads?
To deliver an ADS thou shall:
1/ Identify the standards you are going to comply to (and I am talking about real standards, like EN 50126, SMS compliant with AS9100, not a Waymo brochure)
2/ Define your safety objectives (hint: for an ADS, it's going to be way way waaaaaay higher than human error rate). You don't self-define these. You need to talk to the safety authorities of the area were you intend to circulate.
3/ Build a safety case explaining how you achieve credibly these objectives (spoiler: testing will not be enough)
4/ Have your case examined by an independant safety assessor, accredited by a relevant national or trans-national authority
5/ Put your ego on a shelf.
Otherwise, thou shall kill people.
Simple. But it seems the automotive industry is ready to go to any length to avoid going through the concept of an independant safety assessor.
EDIT: Bullet points syntax
I have nothing but a hunch to back it up, but I have a feeling getting the worst drivers off the road would provide a greater benefit to public safety that would outweigh the damage caused by downgrading the skills of the best drivers.
Some drivers are “better than others” but there are no “good drivers”.
Everyone gets tired, everyone’s gets older, everyone gets distracted, everyone drives in a new location (potentially with a different driving style).
I don’t know when it tips to be better off with the computer, but I’ll feel comfortable as soon as the crash/fatalities rates are equal or better than Uber/Lyft.
Of course it is, because it means there will be less deaths per million miles driven, which surely is the goal.
yes, eventually we want them to be better than the average driver, then better then the best, then (hypothetically) perfect. But the reality is the day they're better than the average, we'll be saving lives if we all switch over.
This kind of equivalence is also ignoring the possibility of a catastrophic system failure that simply can't happen with human drivers.
How do you objectively judge if you are such a good driver that you won't benefit?
I mean, I acknowledge that some accidents will always be unavoidable. Urban driving where someone steps out from between vehicles. Challenging conditions like fog and dust clouds. Sensor malfunctions. A driver coming in the opposite direction swerving into your lane. I'm not saying a self-driving car should avoid 100% of accidents.
But avoiding large clearly visible stationary objects? I'd like to think most drivers on the road can manage that.
[1] https://www.ntsb.gov/investigations/AccidentReports/Pages/HA...
2. You can't really be certain
3. It would be good to be able to collect data of driver faults without crashes. But how do we do it without giving insurance companies more reasons to hike prices?
https://thanaverage.xyz/
I'm going to say that's patently wrong.
as the article points out at level 3, the control system is likely to drop out and punt hard decisions over to the driver at any time. This is exceptionally dangerous because there is almost never enough time to react properly (unless the driver was actively paying attention)
at level 3 you really still need to pay attention.
And we see this partially with Tesla now. Some drivers start to trust the system so much that they start to do anything but pay attention - partially even bypassing the pesky systems that remind them that they are required to take over at any moment.
> One big issue for Level 3 vehicles is that a crash might occur in the 10 seconds the driver spends taking over, so Level 3 vehicles will probably need to include an ODD [operational design domain] where 10 seconds is reasonably safe (e.g. low-speed highway traffic jams).
That's because the I2V would be hugely expensive to implement and maintain. Car companies are not going to pay for it, and it'd be a tough sell to taxpayers.
More expensive cars and roads are not a great move right now. When we need is more efficient and less expensive cars and better driver training programs.
I bet human drivers are already mostly safe - I think most accidents are broadly in 3 categories, (1) impairment (drinking, age, sleep) or (2) distraction (cell phone, makeup) or (3) weather. The first two wont be solved by better training programs and the last would be really hard to make a requirement (how do you train people to drive in the snow in Los Angeles)?. If you want safer roads, it would be more wise to spend that money on public transportation. On a per capita basis it would be cheaper and safer.
That said I don't think the rush to autonomous driving is as altruistic as the developers claim. For $250B you could massively overhaul the transportation systems for major cities. However there is little upside to capture by doing do; while autonomous driving has the potential to create a last-mile autonomous transportation system that would massively reduce the costs of transporting goods and people (by removing labor costs). Capturing that upside is far more lucrative.
I think the question is less "we will only see AV commonplace when it's unilaterally safer than humans", and more "we will see AV commonplace when the value generated exceeds the potential costs of fatal accidents". It's after AV is commonplace will the value captured from human drivers be used to reduce costs in the failures of AV. For example, if UPS were to convert their 50% fleet to AV, then they would use the money saved to then invest in V2V or I2V infrastructure in order to accelerate the transition of the rest of their fleet.
The cost of I2V for example can be offset by less waste in both space used to park cars, drive cars and maintain cars. You can utilize roads much better by having a lot of tiny single passenger vehicles that park themselves outside of the city when traffic is low. Also, travel time with public transportation is always longer than point-to-point, so people save time over that. There's too much for a HN comment that I can think of as benefits of AV.
The same way you train people to drive in the snow in Norway, Sweden, Finland, Denmark etc. the 8-11 month of the year there is no snow on the ground.
Also, aren't steep fines a good "training program" for distractions like holding your phone and messaging when driving?
For example ambient lighting inside the car is largely a gimmick, but a fun one. Clearly not something benefiting society, but it seems to sell cars.
iirc, AD rate is 1 in 1mil (ideal conditions).
I2V is incredibly hard to scale, and is a duplicated implementation of a pre-existing system (road markings & signs). Any system with duplicated implementations will grow inconsistencies. In those situations, the automated driver will be driving based on a differently perceived environment than the human driver - and it'll be hard to know. The automated system would also need to read lane markings & signs to recognize these inconsistencies and act accordingly, at which point the I2V communication is just a crutch to help the automated vehicle, reducing it's utility and hence the cost-value ratio (which is already insanely high).
V2V communication is unreliable by the fact that older vehicles won't support it, so automated vehicles won't be able to communicate with some (initially: most) cars. They'll need to be able to drive safely without communicating with nearby vehicles, so V2V is again just a "bonus" system a vehicle must not rely on.
Automated drivers will be safer than human drivers with time. It's an incredibly hard problem to solve, but it will be solved. Technology is improving continuously, and different companies are trying different approaches to solve the problem - which makes me optimistic in terms of automated drivers. Becoming safer than human drivers is mostly a matter of being able to drive with an automated system based on infrastructure optimized for humans. Sources of accidents will shift: From driver impairment, distraction and ignorance of rules, to misinterpretation of the environment. Additionally, and I think this is the key: If an accident happens, at best the involved human drivers improve their reaction to the specific situation. Automated drivers can share this new knowledge across the entire fleet.
Mind you, this is a hypothetical argument. If any of the major players thought they could make a strong statistical safety case, they'd already be publishing press releases and scaling.
You’re making a good faith argument.
Would increasing minimums for car insurance make us more or less safe? I believe it would make us more safe, intuitively, and the effect would be reinforced by taking people off the road who can’t really afford to drive.
But I don't agree about costs being the biggest issue atm. A lack of V2V/I2V standards is currently the largest issue.
Any near future self driving car will not be safer than a sobar, alert human driver any time soon, unless they have a directive to halt the vehicle at any difficult scenario.
We will not know before such a car exists and is tested under the exact same circumstances vs a human driver.
When self-driving goes wrong due to mishandled edge cases, people will be willing to ignore massive amounts of statistical evidence that self-driving cars are safer than humans on average, and obsessively focus on the the failed edge cases.
"I don't care if self-driving cars have N times fewer accidents per kilometer, what about how that one one time when the Wile E. Coyote painted a picture of a tunnel on a cliff, and then a self-driving car tried to go into it and blew up in a giant ball of fire? It was all over the news man; self-driving is a joke."
Higher safety can be achieved by them being simply “as good” or even worse to some extent but the removal of the driver from multi passenger rides reduces casualties sufficiently to make them safer on average.