Self Driving Cars?

[pj.mask]

This makes no sense. Of course I have avoided many accidents.

But that's just an "if". I can throw out any scenario I want, that doesn't mean it reflects reality. The question is, are they improving enough to reduce accidents overall.

We need to put failures into context - accidents per miles driven with a comparable benchmark (same profile of cars/drivers/conditions).

I already did that: Tesla with safety features vs Tesla in FSD shows:
46% safer 2019Q4
61% safer 2019Q3
49% safer 2019Q2
Etc

That lacks context, and you take it out of context. The higher accident rate w/o Autopilot engaged could be because Autopilot can't be engaged in that scenario. We need more info.

From Tesla's site:

So only on highways? IIRC, highways are ~ 4x safer than other roads. This all needs context and bench-marking to be meaningful.

I’m not aware of any limitation to when one can engage autopilot.

We are talking about “Autopilot”. Navigate on Autopilot is an additional feature, not sure the relevance here.

You’re not going to get a better statistic than Tesla with AP on vs AP off. Same drivers, same cars, same locations.

You might have some situational variation (maybe some do only use it on highways) but out of my antidotal sample size of 6, that is not the case.

If you can define "relative complete highway driving in good conditions", we can set our calendars to revisit this in 5 years. While we certainly will see improvements, I remain very skeptical that a sub $100,000 (un-subsidized) car could be relied on to be fully autonomous, even limited to a highway, and even limited to good conditions. Maybe on special segments of roads dedicated to testing the technology, but not just for any given road trip, where the driver can just read, take a nap, whatever, and let the car take over for hours. I could be proven to be wrong, but I highly doubt it.

There is a heck of a lot more to it than avoiding stopped cars.
-ERD50

You can always find more details to argue over and nitpick. Doesn’t change anything I said.

 

[JoeWras]

Then, do we not talk about the progress they are making along the way? Do we not talk about the deficiencies that still have to be corrected?

This is why the #1 rated cable program for engineers is "Engineering Catastrophes". We learn by watching the past screw ups.

 

[NW-Bound]

Somewhere, I read that car, insurance, and investment salesmen hate selling to engineers.

They ask too much, and want to know how it works.

 

[ERD50]

^^^ Musk is famous for overpromising on timing, but is that what’s most important? ...

And you can't ignore it either, not if we are talking about reality. Promises and predictions will not save a single life, only implementation of a superior system will. So yes, it's important. Implementation is everything.

And let's look at Musk's overpromising, to put some things in perspective. Musk said on Jan 10, 2016 (my emphasis, and my notes in [brackets] ):

https://www.theverge.com/2016/1/10/10746020/elon-musk-tesla-autonomous-driving-predictions-summon

"I think that within two years [ so by JAN 2018 ] , you’ll be able to summon your car from across the country," citing, for example, a Tesla owner beckoning their vehicle to drive solo from New York to meet him in LA.

""I think that within two years, you’ll be able to summon your car from across the country.""

"It will meet you wherever your phone is … and it will just automatically charge itself along the entire journey." ....

Musk later admitted that his predictions "might be slightly optimistic" and he clarified that sort of autonomous driving would require more than a simple software update to existing Model S and Model X vehicles. ...

So we are over two years past a two year prediction, and a Tesla isn't anywhere near this goal. You can barely summon a Tesla across a parking lot today. In this youtube video (and there are worse ones), http://bit.ly/2SZ3lkt the guy says that you need to continuously keep your eye on the car, and be ready to release the 'button if you see any problem. He says "if you are going to rely on the sensors to do everything, you are absolutely wrong, and you shouldn't do that!". And this is in a frickin' PARKING LOT at ~ 2 mph, and it drives like a 10 year old is behind the wheel!

... Or are the accomplishments themselves what are of most value whenever they come to pass? If Tesla achieves FSD in 2025, is that negated by not doing it by 2020? ...

You keep putting these "ifs" and "whens" in there. It makes it all rather meaningless. It's a little like me saying that "if" I cure cancer by 2030, it would be a good thing, right? Of course, and I wouldn't be criticized for not getting it done by 2025, but where is the evidence that I have any shot at that at all? Based on the progress we've seen, why would you expect Tesla to have FSD in 2025?

Musk's above failed prediction is now almost 4 years old, and not even in the ball park. It's like first baby steps towards FSD on open roads. 2025 is just five years out. I just don't think it is reasonable to expect to go from first baby steps to FSD on open roads in five years, if we actually are observing the rate of progress to date.

... All the “experts” here tell us how much more difficult SDC is than anyone realizes, but then you want to judge them by when they reach level 4 or 5?

You don't have to be an expert to observe the rate of progress and come to some informed conclusions. You like to quote that "All truths go through.." nugget, but you know, all product development goes through stages too. As one of my bosses often said, nine women can't make a baby in 1 month.

I'm looking forward to improved safety systems in cars. I honestly believe that over the next 10 years, the most bang for the buck will be having the car systems work with the driver. Just making sure the driver is paying attention would go a long way to avoiding many accidents. It's like having an active passenger in the car - once in a while they alert you to something that maybe you would have missed, or maybe seen a little late. It doesn't need to be all/nothing for the driver, and I don't think that's the best approach (for now). The systems need to get far more sophisticated, and in the meantime, we could save lives with a more driver-interactive system.

-ERD50

 

[NW-Bound]

I'm looking forward to improved safety systems in cars. I honestly believe that over the next 10 years, the most bang for the buck will be having the car systems work with the driver. Just making sure the driver is paying attention would go a long way to avoiding many accidents. It's like having an active passenger in the car - once in a while they alert you to something that maybe you would have missed, or maybe seen a little late. It doesn't need to be all/nothing for the driver, and I don't think that's the best approach (for now). The systems need to get far more sophisticated, and in the meantime, we could save lives with a more driver-interactive system.

-ERD50

Enthusiasts say Tesla cars are safer.

Maybe they are right, but that's the car being active, and the driver watching it. It's the reverse of what you said above.

As I mentioned earlier,

Man+Machine > Man and also Man+Machine > Machine.

I love to get to Machine > Man, but that has to be proven first. Not by arm waving. I want real tests.

We did not get to commercial jetliners without going through biplanes, and propeller planes. Hey, biplanes and prop planes were very useful interims, were they not? But you cannot pretend they are jets!

 

[NW-Bound]

It doesn't need to be all/nothing for the driver, and I don't think that's the best approach (for now)...

Do people here remember Chris Urmson? He earned his Ph.D. in Robotics at CMU, and was one of the principal engineers at Google before he left to form his own company.

When representing Google, he made a famous TEDTalk about SDC, and said that his goal was to have it by the time his son turned 16, so that his son would not need to learn to drive. I forgot how young his son was then, but thought that was way, way, too optimistic.

However, I said to myself that perhaps they could have it, but it would still be too expensive for the masses, and the complexities would require a lot of maintenance and would not be commercially viable. In a way, Google (Waymo) has it in that form now, and their cars are driving around my town.

Well, I ran across an article that mentioned Urmson again. He admitted that his son has turned 16, and he was still working on it.

Sorry, but I digress...

It was Urmson and Google's position that a SDC has to be Level 4 or 5. An SDC with a lesser capability that requires the driver to take over may lead to abuse (drivers going to sleep or texting, etc...) and would not be safe. An SDC has got to be totally autonomous to be truly safe.

And that's why I have been watching Tesla to see if their users are as stupid as feared.

 

[Chuckanut]

PS. I occasionally watch some Youtube videos, in order to learn about the real-life performance of these systems. I usually do not read the comments, because they are made by ignoramuses. But one time, I found something that was hilarious. The commenter said Tesla autopilot was a "sentient" thing.

+1

The Dunning-Kruger Effect strikes again!

 

[ERD50]

I already did that: Tesla with safety features vs Tesla in FSD shows:
46% safer 2019Q4
61% safer 2019Q3
49% safer 2019Q2
Etc


I’m not aware of any limitation to when one can engage autopilot.

We are talking about “Autopilot”. Navigate on Autopilot is an additional feature, not sure the relevance here.

You’re not going to get a better statistic than Tesla with AP on vs AP off. Same drivers, same cars, same locations.

You might have some situational variation (maybe some do only use it on highways) but out of my antidotal sample size of 6, that is not the case. ....

I'm not following this. Why the heck would anyone ever turn of "Autopilot" if it as much better as you say? It sure sounds like Autopilot is engaged under certain cases, and it's likely that those cases are areas that are safer in general. So it isn't apples-apples at all.

Maybe a Tesla owner can clarify this for us?

RE: SDC in 5 years in good conditions -

...
You can always find more details to argue over and nitpick. Doesn’t change anything I said.

You said:

... I think we should have relative complete highway driving in good conditions within 5 years. (Good meaning no snow, heavy rain, with decent road markings or traffic) once they have enough examples of how to avoid stopped cars.

My comments are not nitpicks. It's just defining the conditions for SDC. Are you saying by 2025 the driver can take a nap under clear weather, clear road, normal traffic? Certainly stopped cars aren't the only thing standing between then and now.

-ERD50

 

[NW-Bound]

... It's a little like me saying that "if" I cure cancer by 2030, it would be a good thing, right? Of course, and I wouldn't be criticized for not getting it done by 2025, but where is the evidence that I have any shot at that at all?

Well, promises sell stocks! Hence, they are useful.

Elizabeth Holmes made a lot of promises and sold a lot of Theranos shares to well-heeled people (it was a private company and was never public). She made herself the youngest billionaire woman ever.

She would have gotten away with it, if she did not deliver phony products that endangered people's life. The charges against her are "nine counts of wire fraud and two counts of conspiracy to commit wire fraud, for distributing blood tests with falsified results to consumers".

There was no charge for BS'ing. It is not a prosecutable offsense. And we would not have a jail big enough for that anyway.

 

[ERD50]

Enthusiasts say Tesla cars are safer.

Maybe they are right, but that's the car being active, and the driver watching it. It's the reverse of what you said above.

As I mentioned earlier,

Man+Machine > Man and also Man+Machine > Machine.

I love to get to Machine > Man, but that has to be proven first. Not by arm waving. I want real tests. ...

....

It was Urmson and Google's position that a SDC has to be Level 4 or 5. An SDC with a lesser capability that requires the driver to take over may lead to abuse (drivers going to sleep or texting, etc...) and would not be safe. An SDC has got to be totally autonomous to be truly safe. ...

What I'm envisioning as saving the most lives in the near-term is different than Tesla's current approach. Man + machine as you say, but instead of the car doing much of the controlling and telling the driver "Hey, you pay attention too", I'm talking about actively involving the driver, and using the machine to assist with this.

And I think the Cadillac system does some of this. What I'm envisioning is a system that monitors the driver, and makes sure they are paying attention, and aware of potential dangers.

For example, it could monitor head movement, and see that the driver has been looking left/right, and checking the mirrors at some expected rate. If the car sense another car/pedestrian to the right, and driver is looking left, maybe it's time to alert the driver?

That sort of tech is far easier to implement than taking over for the driver, and could still be developed along the same path as full SDC, it's not either/or.

For those who keep pointing out how bad the average driver is, let's think about that. We can look up some data, but I bet almost all of these driver mistakes that result in accidents are due to inattention/distraction. Drivers rarely just make a driving mistake. How often does someone say, run a red light because they made a mistake and thought "red means GO"? Or end up the wrong way on a one-way street, because they made a mistake about what "ONE WAY - DO NOT ENTER" means? Or rear-end a stopped car ahead of them because they made a mistake, and didn't think you had to stop for a stopped car?

No, they run red lights and enter one-way streets or rear-end a stopped car because they were not paying attention. So while we are waiting for cars to get good enough to be autonomous, why not engage the driver, and use both resources to provide safety?

If the driver refuses to pay attention (sleepy, drunk, medical issue), the car could pull over. How many lives would that save?

Some people will say drivers won't want this nagging from their car. But then you are saying it's OK to keep killing people while we wait for fully SDC. I'd much prefer the good now, rather than wait for the perfect later.

-ERD50

 

[RonBoyd]

https://www.morningbrew.com/daily/s...hlights-tesla-driverless-tech-vulnerabilities

Researchers at McAfee (the same software security firm your dad single-handedly kept in business in the ’90s) managed to trick Tesla vehicles into speeding using...electrical tape.

How they did it: Technicians put a strip of tape on the “3” on a 35 mph speed limit sign, causing a 2016 Tesla Model X and Model S to read the sign as 85 mph and accelerate cruise control accordingly.

 

[Chuckanut]

OK, that's good. No texting or surfin' on the smartphone then. How long does it let you take your hands off the wheel?

My totally unscientific test indicates about 10-12 seconds before it buzzes at me and brings up a warning on the dashboard.

Thankfully, no pedestrians were run over and no automobile was damaged during these tests.

 

[Midpack]

Some people will say drivers won't want this nagging from their car. But then you are saying it's OK to keep killing people while we wait for fully SDC. I'd much prefer the good now, rather than wait for the perfect later.

Who’s waiting? No one has to. More and more brands/models, including Tesla, are offering more and better driver assistance features every day. There have been quite a few posts about it here as well while you’ve been repeating your I’d rather have driver + assist features POV over and over for years.

 

[NW-Bound]

My totally unscientific test indicates about 10-12 seconds before it buzzes at me and brings up a warning on the dashboard.

Thankfully, no pedestrians were run over and no automobile was damaged during these tests.

Thanks. I should search the Web for the answer, but was lazy. When I am ready to buy a new car, I will research these safety features carefully.

We just came back from downtown, an unusual outing for us at this hour. I was telling my wife about the simple car-following/lane-keeping for stop/go rush-hour driving, and she immediately said that would be very helpful.

I am very much for full-blown SDC, seeing that my 86-year-old mother driving skill has become quite poor, yet she insisted on driving. And I will be needing it too in the future. And so, I have been watching this technology to see if anyone has it.

Waymo's cars look to be the best, and if they sell it for $150K, it is still worth it to me. When I am old enough, I will pay even more. Something that valuable would be part of the inheritance to my kids. It's not wasted.

 

[NW-Bound]

https://www.morningbrew.com/daily/s...hlights-tesla-driverless-tech-vulnerabilities

Researchers at McAfee (the same software security firm your dad single-handedly kept in business in the ’90s) managed to trick Tesla vehicles into speeding using...electrical tape.

How they did it: Technicians put a strip of tape on the “3” on a 35 mph speed limit sign, causing a 2016 Tesla Model X and Model S to read the sign as 85 mph and accelerate cruise control accordingly.

It's worse than that. They can cause a car to drive off the road, or to brake for a phantom pedestrian.

How? By using a drone to project lane markings onto the road, or a picture of a pedestrian. They can also flash the image of a fake speed limit sign to fool the car.

 

[ERD50]

Who’s waiting? No one has to. More and more brands/models, including Tesla, are offering more and better driver assistance features every day. There have been quite a few posts about it here as well while you’ve been repeating your I’d rather have driver + assist features POV over and over for years.

I'm not aware that Tesla (as one example) does anything at all to assure the driver is paying attention, other than a simple check that they have done something with the steering wheel.

If you've been reading my posts for years and years, the difference should be obvious. Do I need to say it again? I'd like to see the system actively monitoring the driver, their head movements, and are they paying attention to potential problems.

Take the earlier example of the test of cars hitting those mock-pedestrians. A failure of the systems results in injury/death.

But the approach I advocate would be continuously monitoring the driver for situational awareness. An alert driver probably wouldn't hit that pedestrian, whether the car system saw it and alerted the driver or not.

As data is collected, we can see at what point the system alone is so good as to make the driver superfluous. Then we have arrived at full SDC!

-ERD50

 

[NW-Bound]

The after-market add-on Level 2 AP $1000 kit I wrote about earlier monitors the driver's eyes. Impressive, eh?

Tesla 3 owners discovered a cockpit camera that Tesla did not tell them about. When asked what it was for, here's what Musk said:

"It’s there for when we start competing with Uber/Lyft & people allow their car to earn money for them as part of the Tesla shared autonomy fleet. In case someone messes up your car, you can check the video."

Tesla could use the interior camera to monitor the driver, but it chose not to.

Tesla seems to have a more permissive attitude. Because the owners can push the operating envelope beyond what other car makers would allow their owners for safety reasons, naive people believe Tesla cars are more advanced. Another marketing advantage!

 

[ERD50]

It's worse than that. They can cause a car to drive off the road, or to brake for a phantom pedestrian.

How? By using a drone to project lane markings onto the road, or a picture of a pedestrian. They can also flash the image of a fake speed limit sign to fool the car. ...

To be fair, if I approached that projected image on a dark street, I'd hit the brakes too. Something's not right, even though I'd likely realize it isn't a person, I'd wonder what it is and slow down.

There's lots of tricks people could play on human drivers too. I'm more concerned about the normal everyday challenges.

-ERD50

 

[Midpack]

I'm not aware that Tesla (as one example) does anything at all to assure the driver is paying attention, other than a simple check that they have done something with the steering wheel.

If you've been reading my posts for years and years, the difference should be obvious. Do I need to say it again? I'd like to see the system actively monitoring the driver, their head movements, and are they paying attention to potential problems.

Take the earlier example of the test of cars hitting those mock-pedestrians. A failure of the systems results in injury/death.

But the approach I advocate would be continuously monitoring the driver for situational awareness. An alert driver probably wouldn't hit that pedestrian, whether the car system saw it and alerted the driver or not.

As data is collected, we can see at what point the system alone is so good as to make the driver superfluous. Then we have arrived at full SDC!

-ERD50

Some Toyota/Lexus, Cadillac and BMW’s have/had driver monitoring systems. Driver monitoring could make a significant difference in some cases (basic inattention), but it won’t do as much for drivers e.g. under the influence, exhausted, road rage, speeders/reckless drivers, night driving/poor weather disorientation, potholes, wrong way drivers, animal crossings, construction, blown tires at speed. SDC will do much more.

If there’s a demand for driver monitoring (or its legislated in), automakers will fall all over themselves to provide it. And then we’ll see the impact of DMS on the way to SDC.

 

[NW-Bound]

To be fair, if I approached that projected image on a dark street, I'd hit the brakes too. Something's not right, even though I'd likely realize it isn't a person, I'd wonder what it is and slow down.

There's lots of tricks people could play on human drivers too. I'm more concerned about the normal everyday challenges.

-ERD50

Yes, I would also brake for a projected human figure until I am sure it's fake.

On the other hand, swerving off the road is not likely. People do not follow lines blindly, particularly bright lighted lines. And they also recognize other visual cues, such as the curbs, the road shoulders, the trees along the road, etc...

Humans also recognize the flashed traffic signs.


PS. This kind of hacking is why the vulnerabilities of each particular SDC implementation have to be studied and tested. Imagine when you give the rider no way of overriding the computer. Anybody can be turned into dead meat by a personal enemy, or just a mass murderer.

 

[ERD50]

Some Toyota/Lexus, Cadillac and BMW’s have/had driver monitoring systems. Driver monitoring could make a significant difference in some cases (basic inattention), but it won’t do as much for drivers e.g. under the influence, exhausted, road rage, speeders/reckless drivers, night driving/poor weather disorientation, potholes, wrong way drivers, animal crossings, construction, blown tires at speed. SDC will do much more.

....

I disagree. I don't think you are applying enough creativity to this option. Maybe you are missing my point that this would interact with SDC technology as it develops, but as a driver enhancement, rather than taking over driving (until it has proven itself capable - a long ways off, IMO)?

under the influence, exhausted, - The system I describe absolutely could help. It would see that they weren't paying attention (eyes not scanning the road, not reacting to things the car system sees) and after a few warnings, it would pull over, maybe call 911.

road rage, speeders/reckless drivers, Why couldn't a system like this detect speeding, or reckless driving (crossing lanes, abrupt motions, etc). If the road rage is affecting driving, it would be detected, heck maybe even detecting stress through analyzing pupils?

night driving/poor weather disorientation, potholes, wrong way drivers, animal crossings, construction, blown tires at speed. - Again, making sure the driver is attentive will help just about everything.

As I've said before, much of the monitoring of the driver is relatively simple to achieve today. While we are waiting for SDC to be fully capable, why not combine it with tech that keeps the driver engaged, and reap the benefits of both systems (two 'heads' are better than one)? I'm not sure why you seem to have any opposition to this. It would seem to help advance SDC tech, because we can engage an imperfect SDC system with an active human as part of the system, and learn and improve the SDC faster, as there would be more hours of use on average.

And because the human is an active part of the system, the SDC can be more 'relaxed' and would have fewer false alarms, as it isn't as critical if it misses something. Since the human is engaged, if the SDC misses something, it's just didn't add value in that instance - still better than today as it has been keeping the driver engaged. But it didn't blindly go ahead and run over a pedestrian if it missed it, it just didn't alert the driver.

-ERD50

 

[NW-Bound]

But, but, but ERD my friend, the system you describe will not sell.

Have you seen T-Al thread where he describes the video of a guy going 120mph with his Tesla? Do you think people spend money on a car to have you put restriction on how they drive? You will not let them overspeed, zigzag, tail gate, do brake check, sleep at the wheel, read books, play with their smartphone? Get outta here.

That kind of system would have to be mandated by law, so that all car makers are on equal footing. No way you can have such laws passed.

 

[Midpack]

^^ We disagree, but no need to rehash it all again. Driver monitoring will help in some circumstances. But from above - a badly impaired, reckless or angry driver isn’t in full control even if they’re fully aware. And in many of the other situations the driver is disoriented with split seconds to react, driver monitoring isn’t going to help a driver surprised by a blown tire, an animal coming out of nowhere, a pothole they didn’t see until they hit it, driving in rain/snow when you really can’t see and suddenly your off the road at speed.

 

[NW-Bound]

What I was saying in my post above is that while responsible drivers may welcome a monitoring system that helps them to be safe, there are people who want to drive in an aggressive manner. See video below.

They will not buy such a system unless it is mandated on all cars, because they will buy a more permissible car that lets them do what they want. It is also hard to define what the system should allow a driver to do. Unless you have a government design for all cars to use, the makers will all do different things. We already talk about how Tesla lets the users "play" with the AP in iffy scenarios, while other makers don't.

The best way is what Waymo describes. Unfortunately, we will not have it anytime soon.

 

[ERD50]

But, but, but ERD my friend, the system you describe will not sell.
...
That kind of system would have to be mandated by law, so that all car makers are on equal footing. No way you can have such laws passed.

Well, we have all sorts of rules mandated by law that restrict our freedoms. In this case, it would only be restricting bad behavior anyhow. I'm not so sure it would be rejected to the degree you say.

People willingly pay for some of these systems today. Blind spot monitoring, lane departure warnings, back-up cameras with object warnings - most as an extra cost option. Even if some are a bit 'nagging'.

I've had a few false alarms on my car, but outside of that, I've always been aware when it says I'm drifting out of the lane (I did it on purpose to avoid something), or that I was approaching a car ahead fast (I knew it was there, had my foot hovering over the brake, but was expecting the car to speed up a bit). So it 'nags' a little, but I happily accept that if in 1 case out 10,000, it alerts me to something I didn't see, and it keeps me out of an accident. Any future car I buy will have this (and more). I'm accepting it, and even looking for it and paying for it.

^^ We disagree, but no need to rehash it all again. Driver monitoring will help in some circumstances. But from above - a badly impaired, reckless or angry driver isn’t in full control even if they’re fully aware. ....

A badly impaired driver that is 'fully aware'? That makes no sense. If they are driving so normally that a somewhat capable SDC system couldn't detect it, they must not be badly impaired.

I don't think it's a re-hash - you never 'hashed' it in the first place! You seem to reject this, but like above, I see no logic in your dismissal of this.

... And in many of the other situations the driver is disoriented with split seconds to react, driver monitoring isn’t going to help a driver surprised by a blown tire, an animal coming out of nowhere, a pothole they didn’t see until they hit it, driving in rain/snow when you really can’t see and suddenly your off the road at speed.

Which would be the case today. And today, and for who knows how many years, an SDC system can't be depended upon to handle them w/o intervention anyhow. How can it not be better to have a system that makes sure the driver is actively engaged?

An actively engaged driver will be handle all of the above better than a driver who let his/her attention lapse.

To put this in the simple logic terms that NW-Bound has used:

An (imperfect, but engaged driver plus an imperfect SDC) > (an an imperfect SDC alone).

What the heck is your issue with that? I really don't understand.

Some more logic:

Let's say (imperfect, but engaged driver plus an imperfect SDC) gets to the point of reducing accidents/injuries/fatalities per mile to 20% of current rates. Great.

But lets say tests show that letting the driver disengage caused that number to go up to 30% of current rates. It would be unethical to eliminate driver engagement, that's still thousands of fatalities per year, etc.

Until SDC is so good that there is a statistically negligible advantage to retaining the driver, how can we advocate for a disengaged driver?

And I honestly feel that driver engagement would advance the rate of improvement of SDC, as we'd get more data.

It's pure logic, where's the flaw?

-ERD50