Tesla Full Self Driving (FSD) info - interesting stuff!

[Rick Law]

The situation in Denmark is a little special though, they have crazy high car sales taxes there (something like 180%) so any discount really matters!   I recall seeing a Mercedes in Copenhagen with a personalised number plate:  "PAID 3X"

Yeah, but the same thing happened in Hong Kong - Tesla sales went from 2000 in 2016 down to 32 in 2017^[1].

China, Denmark, anywhere else, Tesla as a maker of only EV cars needs to keenly aware that they are not yet price-competitive to non-EV cars.  They have to live in the two corners - (1) richer folks and (2) government assistance be it by regulation or financial assistance.

Corner number two (subsidies) is getting smaller.  So for EV-only makers, they will be increasingly depend on corner 1 - the richer folks.  For the richer folks where money matters less, quality (or the aura of it) is important.

Unfortunately, Model 3 is not a good thing for corner-1.  Can you imagine an "entry level Rolls Royce" for the average folks?  Mercedes tried that with the 190's in the 1980s.  It didn't go over too well.  I think that is the reason why the BMW 1 series is now history in the USA.  Toyota/Honda (and others) understood that, they started Acura and Lexus.  Instead of diluting the brand,  Toyotas are for the penny-counters and Lexus are well, for those with (or think they have) more money to spend.  Those in corner 1 really would prefer to be separated from the average folks - like first class in a plane with a curtain separating it from economy.

I think, Tesla needs to keep their Teslas with an aura around it - all the ooh's and aah's.  People will pay extra for that.  Start another brand for model 3 and sell it in a separate facility.  But they need money to carry that out, and money isn't what they have in excess...  Elon and his team better figure this out and get themselves on better footing...

Oh, regarding China - given China is a new rich country, that separation would be particularly important.  My life-experience tell me, new rich are more status conscious and would want the distinction more.  I could be wrong, and China's new rich would not be as status conscious as I anticipate - but if I am right and they try to sell Teslas as primium car but also Model 3 (or less) that is cheap (cheaper) in the same shop, that would be a disaster in China.

Reference:
[1] Financial Times magazine, 2018, February: "Tesla sales drop in Hong Kong after tax breaks removed"
"Touted as the city with the highest number of Teslas per capita by Mr Musk in early 2016, just 32 of the company’s vehicles were registered in Hong Kong between April and December, figures from the Transport Department showed. That is down from nearly 2,000 in the same period in 2016."
^{https://www.ft.com/content/2b8eb480-0a45-11e8-839d-41ca06376bf2}

[nctnico]

In the Netherlands the tax breaks on EVs are partly removed (no tax breaks on EVs over 50k euro). Sales of the Tesla Model-S and Model-X have dropped to insignificant numbers. The only model selling in reasonable numbers is the much cheaper Model 3.

[Rick Law]

In the Netherlands the tax breaks on EVs are partly removed (no tax breaks on EVs over 50k euro). Sales of the Tesla Model-S and Model-X have dropped to insignificant numbers. The only model selling in reasonable numbers is the much cheaper Model 3.

Looks like the are loosing the "premium image"...    Dinner plate please!  Their goose is cooked...

[maginnovision]

It's a premium image that people enjoyed getting a break on. Now it's an expensive vehicle with smaller breaks which keep shrinking and with every one sold makes it less exclusive. Time will tell where this goes.

[extide]

Mercedes tried that with the 190's in the 1980s.  It didn't go over too well. 

FWIW, The 190 series became the C class, which is not only the most popular model that Mercedes sells in the US, it's the most popular model out of all luxury brands.

[TerraHertz]

I find the topic of self driving cars fascinating. Not so much for the technology, but what it reveals about people's not-so-rational dreams and expectations.  Much the same as the apparent yearning for 'AI servants and industrial production', but with the added dimension of staking your life on the software controlling a high speed vehicle.

A few points:
* Any highly automated system with some kind of always-on net connectivity will be susceptible to remote hijacking. Don't even bother to talk about computer security, in a world where (for eg) hardware level backdoors are built into all contemporary Intel CPUs, all network routers, all cell phones, etc. Ask Michael Hastings about the pros and cons of remote vehicle operation.

* All complex software systems are prone to obscure bugs and design mistakes. Some are more stupid than others (eg Boeing's 'please fly us into the ground right now' brilliant idea.)

* The nature of the problem prevents 'perfection.' Road driving is a highly complex task, with infinite numbers of obscure marginal and tricky cases. Even alert, skilled humans get fooled sometimes and end up in accidents. Some of which are high speed and fatal. No level of AI is going to completely eliminate all situational f*ck-ups.  Self-driving cars ultimately involve a statistical risk evaluation: What added probability of death or maiming do you accept for the 'convenience' of not being in control of the vehicle yourself?

* What's the actual benefit of not being in control? What kind of person could actually nap, or concentrate on some productive task, while being carried at speed by an AI system? I definitely never could.  Idle conversation, maybe. But that's not something I'd choose to even want, let alone try.   I can relax in trains, because that infrastructure is very good at achieving an extremely low risk. Planes, because again quite low risk, and skilled humans are in charge (unless a new Boeing...) Cars on ordinary roads - never!

* At a political level, there are too many ideologues trying to impose restrictions on individual travel, and also track such travel. Taxing travel at a rate per Km, requires logging of travel. Automated cars can be legislated to support that. Tracking for social control, ditto. Schemes like China's 'social credit', also would find automated cars useful.  How about governments mandating things like rationed travel, selected days when you are allowed to travel, zones you are not allowed to enter, and so on? Automated cars enable all that kind of bullshit.

* Privacy. Given that Google, Amazon, etc 'home assistants' are being revealed to upload a lot of information about conversations in the home (and commonly making comical interpretation mistakes), do you want that in cars as well?

* Legal liability. I can't wait till this issue becomes prominent through increasing numbers of accidents in which no human adult was responsible.  I wonder if legal principles for dealing with AI in general, will be founded in road accident case precedent? That will go well, not.

Then there are the fundamental philosophical issues with increasingly general purpose AI systems (which is where self-driving cars will be forced to go, by virtue of the complexity of the problem.) There are many questions here. For instance, how wide is the window of workability, in which the car is smart enough to get the job done, but not smart enough to get in a snit about something you said to it, and refuse to do the job? (Or decide to suicide, with you as passenger.)

Meanwhile, I'm keeping my old, pre-engine-management-computer car.  Self-driving car technology is good for one thing - entertainment, in watching a fairly predictable developing social insanity.

[apis]

but with the added dimension of staking your life on the software controlling a high speed vehicle.

Why trust a car at all, it's a pile of complicated software and hardware. Most people stake their life on incredible complex hardware and software every day.

Even alert, skilled humans get fooled sometimes and end up in accidents.

No one thinks self driving cars will be perfect, but some people realise that neither are humans. In fact humans are **** poor drivers and that's the main reason why replacing them with a computer is not such a bad idea. Even the Tesla "autopilot" cruise control is claimed to already have much better accident statistics than human drivers.

It's safe to be a naysayer regarding this because it won't become widespread overnight, might take decades, but self driving cars are no doubt the future, mainly because it will be cheaper than human drivers and money rules.

I understand that some people might enjoy driving and not like the idea of a future where the average man might not have the option to drive, but other than that I don't really understand why anyone would be against robotic cars or other forms of automation. Just seems like ludditis to me.

[apis]

Waymo disengagement rate for 2018 was 1 disengage per 11,017 miles out of 1.28 million self-driven miles in CA 2018. That corresponds to about one disengage per year for an average driver in the US.

"By the end of 2018 we’d driven another six million miles, which means our self-driving cars have now covered 10+ million real-world road miles. The connection between our real-world miles coupled with our 7+ billion miles in our simulation is key to our improvement rate."

https://medium.com/waymo/an-update-on-waymo-disengagements-in-california-d671fd31c3e2



"At one disengagement per 11,017 miles, Waymo’s closest competitor is GM-owned Cruise Automation at 5,205 miles. Apple’s effort is last at a disengagement every 1.1 miles, though the company disagrees with the metric."

 

[nctnico]

* The nature of the problem prevents 'perfection.' Road driving is a highly complex task, with infinite numbers of obscure marginal and tricky cases. Even alert, skilled humans get fooled sometimes and end up in accidents. Some of which are high speed and fatal. No level of AI is going to completely eliminate all situational f*ck-ups.  Self-driving cars ultimately involve a statistical risk evaluation: What added probability of death or maiming do you accept for the 'convenience' of not being in control of the vehicle yourself?

Now you are assuming self driving cars will always be worse than human drivers. Automatic anti-skid systems (which exist for a long time) are already proving you wrong.

[Rick Law]

Mercedes tried that with the 190's in the 1980s.  It didn't go over too well. 

FWIW, The 190 series became the C class, which is not only the most popular model that Mercedes sells in the US, it's the most popular model out of all luxury brands.

You have a point there.  I am thinking how snob-appeal may affect sales to the rich clientele.  I don't think anyone will dispute that it has an impact, but it would be merely one of the many variables that would affect the financials.  No doubt some of the S class buyers would rather not the C class folks clogging up the shop, but when you add up all the dollars the seller/manufacturer make, the case is not as clear whether it is more profitable to focus on high end.  Your point is absolutely right and I have to adjust my thinking here.

Walking myself back from snob-appeal and related, the issue of quality remains.

I still stand by my point that for Tesla FSD is not as important as interior trim falling off new cars or wind-shield cracking by itself.  That two were both cited by Consumers Report as quality issues.

When the car is falling apart by itself, mean while, Tesla is hoping I would trust my life to that car auto-driving me to somewhere...  That I think is silly.

[wilfred]

Waymo disengagement rate for 2018 was 1 disengage per 11,017 miles out of 1.28 million self-driven miles in CA 2018. That corresponds to about one disengage per year for an average driver in the US.

Anything above zero means you need a steering wheel and an attentive driver. No steering wheel or no automation. A hybrid mix is the worst of all possibilities. It means you can't trust the automation or you can't blame the driver.

The family of someone struck by an auto needs to seek a just outcome and lawyers at 20 paces with a large corporation is going to be an issue. You can't treat it the same a a plane crash due to a firmware error. You can't take the cars off the road you can't spend millions on a crash investigation.

The only solution I can see is to have each vehicle carry a designated driver who will accept the legal responsibility for the firmware choices. Perhaps they insert a card to load ethical parameters to customise the firmwares programming. Which obviously means you can't set the car to circle the block whilst you pick up your dry cleaning. Nor can you use it to freight children to school or Saturday morning sports. Nor can you get dropped off at the door and have the car go and find a parking spot.

If autos are the solution then I don't think the problem is well understood. Buses and trains will solve congestion better. Clustering people closer to work and community facilities will work better. Children will be better off walking to school close to home or socialising with friends on a bus or train.

[apis]

Waymo disengagement rate for 2018 was 1 disengage per 11,017 miles out of 1.28 million self-driven miles in CA 2018. That corresponds to about one disengage per year for an average driver in the US.

Anything above zero means you need a steering wheel and an attentive driver. No steering wheel or no automation. A hybrid mix is the worst of all possibilities. It means you can't trust the automation or you can't blame the driver.

When it comes to anything safety-critical it would have to be zero (it will never be exactly zero but close and what is most important: better than a human driver), no argument there.

For the most part safety means don't crash into something or drive off the road, which you can do deterministically quite well using good sensors (lidar), good maps and traditional control theory. That's what Waymo is doing. Tesla has taken a very different approach, they use only cheap sensors (cameras) and then use machine learning algorithms to convert the video feed into a 3D representation of the world. Such machine learning based computer vision systems have been making really amazing progress the last decades. But as we have seen, the Tesla system sometimes makes fatal mistakes. The old school, but more expensive, Waymo approach is much more robust. Waymo using lidar means they don't have to convert a 2d representation into 3d, they get a very accurate 3d representation directly from the lidar sensor (and on top of that they add sensor information from cameras and radar).

Tracking objects and steering the car to avoid objects (short term) is a solved problem as long as you have an accurate 3d representation of the world. The most difficult parts are things like trying to predict other drivers behaviour, pedestrians intentions and things like that.

However, there are many corner cases that aren't safety critical but can still be hard for a car to deal with, e.g. because getting out of the situation requires technically breaking the law (driving somewhere you're not allowed to driver normally when the normal road is blocked). Would it be legal to program the car to break the law? But It's not so critical if the car can't find the route because the road is blocked, it can just drive somewhere where it can stop safely and ask the passenger or a tele-operator for help. Maybe there is an obstacle on the road it's not sure how to deal with (there was an event with a mattress on the road the self driving cars stopped for while human drivers just drove over it.) The cars will be programmed to be overly cautious and it might be annoying if they stop for something they don't have to stop for, but it won't be a major problem as long as it doesn't happen too often. I imagine these kind of minor problems could be solved in 5-15 minutes and if it only happens once every year then it won't be so bad. And one would expect the number of disengagements to keep getting lower as they improve the software and sensors.

The family of someone struck by an auto needs to seek a just outcome and lawyers at 20 paces with a large corporation is going to be an issue. You can't treat it the same a a plane crash due to a firmware error. You can't take the cars off the road you can't spend millions on a crash investigation.

The only solution I can see is to have each vehicle carry a designated driver who will accept the legal responsibility for the firmware choices. Perhaps they insert a card to load ethical parameters to customise the firmwares programming. Which obviously means you can't set the car to circle the block whilst you pick up your dry cleaning. Nor can you use it to freight children to school or Saturday morning sports. Nor can you get dropped off at the door and have the car go and find a parking spot.

I think we will mainly see taxi services for quite some time, not privately owned self driving cars. If one of the cars cause a problem it would be the taxi company who is responsible (and they in turn would hold the manufacturers of the cars responsible).

You never need to worry about parking, a car picks you up when you need it and drops you off at your destination. It will be like taxi, but without the driver.

If autos are the solution then I don't think the problem is well understood. Buses and trains will solve congestion better. Clustering people closer to work and community facilities will work better. Children will be better off walking to school close to home or socialising with friends on a bus or train.

Busses and trains can also be made self driving. Busses and trains are great as long as they are filled with people. But busses and trains also drive around almost empty for a large part of the day. Proponents of PRT (podcars) have always claimed that on demand services with smaller cars are more efficient than traditional mass transit systems. The problem with cars today is that everyone wants their own car, so we have one person driving around with 3 seats empty (or more) and most of the time the car is parked somewhere. We normally don't use taxi today because it's too expensive to pay a human chauffeur, but if you don't need a human driver that changes.

[apis]

For the most part safety means don't crash into something or drive off the road, which you can do deterministically quite well using good sensors (lidar), good maps and traditional control theory. That's what Waymo is doing. Tesla has taken a very different approach, they use only cheap sensors (cameras) and then use machine learning algorithms to convert the video feed into a 3D representation of the world. Such machine learning based computer vision systems have been making really amazing progress the last decades. But as we have seen, the Tesla system sometimes makes fatal mistakes. The old school, but more expensive, Waymo approach is much more robust. Waymo using lidar means they don't have to convert a 2d representation into 3d, they get a very accurate 3d representation directly from the lidar sensor (and on top of that they add sensor information from cameras and radar).

I might add that Waymo also uses machine learning, and Tesla has radar sensors as well as cameras (but no expensive LIDAR). But from what I know there is a large difference in their approach and Tesla rely much more on machine learning.

In the last question of this talk the Waymo engineer also mentions they use a hybrid approach, only using machine learning when it improves safety (even if they have to use the more expensive sensors).

"You want to be safe in the environment, so you don't want to make errors in perception, prediction and planning. And the state of machine learning is not at the point where it never makes errors."

https://youtu.be/Q0nGo2-y0xY?t=3774

[Marco]

The problem with radar is that for the moment it's just doppler for a huge FOV, not ToF distance based imaging... which makes it almost completely useless.

[maginnovision]

I'm pretty sure ti has mmwave that can give you size, speed, and direction of travel for detected objects. Seems ok to me. Of course all the inputs are fed into a filter to determine what's most likely happening you never accept everything at face value. Except for Tesla, they seem like they're ML or nothing.

[wraper]

For the most part safety means don't crash into something or drive off the road, which you can do deterministically quite well using good sensors (lidar), good maps and traditional control theory. That's what Waymo is doing. Tesla has taken a very different approach, they use only cheap sensors (cameras) and then use machine learning algorithms to convert the video feed into a 3D representation of the world. Such machine learning based computer vision systems have been making really amazing progress the last decades. But as we have seen, the Tesla system sometimes makes fatal mistakes. The old school, but more expensive, Waymo approach is much more robust. Waymo using lidar means they don't have to convert a 2d representation into 3d, they get a very accurate 3d representation directly from the lidar sensor (and on top of that they add sensor information from cameras and radar).

I might add that Waymo also uses machine learning, and Tesla has radar sensors as well as cameras (but no expensive LIDAR). But from what I know there is a large difference in their approach and Tesla rely much more on machine learning.

In the last question of this talk the Waymo engineer also mentions they use a hybrid approach, only using machine learning when it improves safety (even if they have to use the more expensive sensors).

"You want to be safe in the environment, so you don't want to make errors in perception, prediction and planning. And the state of machine learning is not at the point where it never makes errors."

https://youtu.be/Q0nGo2-y0xY?t=3774

Waymo approach is garbage frankly. They may have a very small number of disengagements when driving in areas with their high detail maps. But when on a large scale, it's a dead end approach. You will never have precise high detail maps for everything, not to say without errors. Also changes happen all of the time, keeping it up to date is a huge challenge on it's own. It's simply unfeasible to make up to date high detail maps for every area. Not to say when it comes to bad weather, LIDAR becomes an erratic piece of garbage.

[G7PSK]

I am begining to think that all Tesla related material should be moved to the Dodgy technology page.

https://youtu.be/sAQlLu5ttOk

[wraper]

I am begining to think that all Tesla related material should be moved to the Dodgy technology page.

Not again  . Just look statistics on how many cars catch fire and think again if you should have posted this. Not to say it's off topic.
https://www.google.com/search?biw=2048&bih=1166&tbm=isch&sa=1&ei=HavKXPzgK4aFmwXs0pG4DQ&q=car+fire+garage&oq=car+fire+garage&gs_l=img.3..0i24.17775.19340..19545...0.0..0.58.341.7......1....1..gws-wiz-img.......0j0i67j0i5i30j0i8i30.DwDjUrYwVWE

[Marco]

I'm pretty sure ti has mmwave that can give you size, speed, and direction of travel for detected objects.

It's possible, it's just not currently on the cars ... or it wouldn't be running into fire trucks.

[janoc]

I will just post link to this here:

https://jalopnik.com/automation-transformed-how-pilots-fly-planes-now-the-s-1834176244

It specifically discusses Tesla's approach to this and why it is a huge problem. There are decades of research on the impact of human factors when dealing with automation (e.g. in aviation) and, frankly, I am shocked how clueless some people are about this, despite this being an engineering forum.

Talking about "disengagement rates" - hello, when the car drives you into wall with a fully engaged automation because it got confused, does that count? Or when it drops the mess in your lap with 3 seconds to react - you will be dead before you realize what is happening, unlike in a plane where you usually have minutes and kilometers of empty air ... Talking how it is safer than human drivers - with only data for this being the oft-quoted claim from Musk using data from Autopilot which is supposed to be used only on highways - where the least number of crashes occurs, etc.

Or someone making the argument that because antiskid systems are better at preventing loss of control than humans are, it somehow follows that automatic cars will be better too (never mind the several orders of magnitude difference in the complexity of the problem ...).

This debate is pretty much the same every time Tesla is mentioned. The fanboys come out of the woodwork and will defend every nonsense that caused people to die in the past only because now it comes from Tesla, so laws of physics and human psychology somehow don't apply anymore.

There is also this:
https://techcrunch.com/2019/05/01/tesla-sued-in-wrongful-death-lawsuit-that-alleges-autopilot-caused-crash/

Yes, that wasn't a fully autonomous car and yes, the driver will most likely be found at least partially at fault - but if Tesla loses this, it will be pretty much game over for both their Autopilot and most likely the fully self driving cars in their current iteration as well. Every lawyer will jump on every single minor scrape hoping to sue bejeesus out of these companies. And they will have only themselves to blame because, as the plaintiffs wrote in the suit, "Tesla is beta testing its Autopilot software on live drivers".

I have nothing against Musk (e.g. what he achieved with SpaceX is amazing!) or Tesla but if we keep the debate to this style where every criticism is brushed away with "I don't hear you, go away, Luddite!", sticking head in the sand and/or mindlessly repeating hype about machine learning as some sort of magic mantra that will fix everything just give it enough data, people will die - in completely avoidable and preventable accidents.

[madires]

Another issue most people are not aware of is that the driver assistance systems and systems for autonomous driving can be fooled easily. This was demonstrated several times by security researches at hacker cons. There are a lot problems which need to be addressed before we may think about trusting a self-driving car.

[apis]

Waymo approach is garbage frankly. They may have a very small number of disengagements when driving in areas with their high detail maps. But when on a large scale, it's a dead end approach. You will never have precise high detail maps for everything, not to say without errors. Also changes happen all of the time, keeping it up to date is a huge challenge on it's own. It's simply unfeasible to make up to date high detail maps for every area. Not to say when it comes to bad weather, LIDAR becomes an erratic piece of garbage.

Priority one is no accidents (Waymo has no fatalities, Tesla 3).
Priority two is comfort.
Priority three is that there no annoying disengagements that cause delays. Waymo has the lowest number of disengagements per mile by a large margin. Tesla is still on the "cruise control" stage.

Google has already demonstrated they are able to map the parts of the world where they would want to drive in the foreseeable future, so the need for maps is not much of a limitation. Using maps is safer so why not use them if you can? Clearly they are able to deal with some changes in the map environment or else they wouldn't be able to drive as much as they do.

All sensors have both advantages and disadvantages, but lidar is without a doubt the best type of sensor you can have. Waymo drives in light rain at the very least. If they can't use the lidar in some extreme conditions they will be in the same bad situation as Tesla, having to rely on only cameras and radar. It's not like the lidar is the only sensor. Lidar + camera + radar > camera + radar, no matter how you look at it.

The most difficult part for a self driving car is to take the sensor data and create an accurate representation of the surroundings. That is also one of the most safety critical tasks (you don't want to miss any pedestrians on the road like Uber did, or road dividers or trucks, etc, like Tesla did). So it should be obvious that using the best possible combination of sensors to minimise the risk of confusion is the way to go.

Musk has said the following about lidar apparently:
“In cars, it’s freaking stupid. It’s expensive and unnecessary. And as Andrej was saying, once you solve vision, it’s worthless. So you have expensive hardware that is worthless on the car.”
Basically he's admitting that computer vision isn't good enough yet and that he is betting on that they will solve vision before everyone else.

Tesla has convinced people to collect the road data for them and people even pay a premium to be their safety drivers. If something goes wrong Tesla just blames it on the driver not being alert enough: "it's only cruise control". It's clever marketing if nothing else, but it's killing people. They get a lot of data for free. Machine learning on it's own isn't good enough yet, but Tesla is betting that they will get to good enough before the competition, and by then they will be sitting on more data and hardware with lower marginal cost.

But there is no reason to think other companies like Waymo won't be able to remove the lidar whenever it's safe to do so. It's not like Waymo and others are unable to use more machine learning and less expensive sensors whenever they choose to.

[G7PSK]

Might well be the wrong place,probabley the wrong page as well. There are 272.1 million cars on the road in the US last year of which 171,500 caught fire Tesla has built 300,00 cars and from what I can find online there are at least 2 fires a week relating to them. I will leave the math.

[wraper]

Google has already demonstrated they are able to map the parts of the world where they would want to drive in the foreseeable future, so the need for maps is not much of a limitation. Using maps is safer so why not use them if you can? Clearly they are able to deal with some changes in the map environment or else they wouldn't be able to drive as much as they do.

Yow fresh they generally are? 1 year, 2 years? I've seen plenty of situations when map came out yesterday, but does not match actual road which was rebuilt differently a few months ago. Not to say, it's not just satellite view that is required in this case.

[apis]

I have nothing against Musk (e.g. what he achieved with SpaceX is amazing!) or Tesla but if we keep the debate to this style where every criticism is brushed away with "I don't hear you, go away, Luddite!", sticking head in the sand and/or mindlessly repeating hype about machine learning as some sort of magic mantra that will fix everything just give it enough data, people will die - in completely avoidable and preventable accidents.

The main problem I see is a lot of people who say it's impossible but they don't have an informed explanation of why they believe so or any data to back it up. It's just heated opinions and there seems to be very poor understanding of how these cars work. People on the internet say it's not possible and in the meantime these cars are driving tens of thousands of miles uninterrupted in the real world, apparently unaware it's not possible. Seems pretty disingenuous to me.

I agree there's reason to be critical but it's also annoying that people assume all self driving cars are equal. Tesla is way behind the others in this field. If you want to assess the technology you should look at the technology leader which is Waymo not Tesla.

Mostly the criticism is on the lever "it's impossible, there are too many nuances and edge cases". Well, obviously not, they are already at a level where they can drive completely uninterrupted in California for what corresponds to a year for an average person in the US. (With zero fatalities of course). They've made the cars overcautious to guarantee no accidents, but that means there will be false positives (disengagements), so the difficult part is to get those down to an acceptable level without increasing the risk of serious accidents.

Self driving cars won't take over by 2019 as some (Elon) says, but there's nothing to indicate they won't get there eventually imo.