I was repeating this to people (typically I was downvoted, the same way I will be now), but the hype is just too huge. People need superheroes, in comics and in real life, they need to depend on some preacher, that will show them illusion of better life. And the marketing is using that, again we are in same scheme, instead of listening to technical people, everyone is relying on marketing which will say everything to boost up the stocks. <slowly clapping> Bravo! (I never had a driving license, I do all my "moving" with public transport or bike and self driving car would be usefull, but not yet - or sooner than in 10 years)
There's already a lawsuit from 4/2017 covering "Enhanced" Autopilot...
https://www.truthinadvertising.org/wp-content/uploads/2017/0...
Tesla’s Marketing and Sale of Vaporware
21. As Dictionary.com states, vaporware is “[c]omputer software that is advertised but
still nonexistent.” Tesla advertised vaporware to consumers, knowing full well that this particular
come-on would particularly excite its target market of high-tech aficionados.
Now we need one for self-driving for those who were duped into buying a Tesla just for that.
Not to mention all the AP1 owners who were misled with this tweet.
https://twitter.com/elonmusk/status/686279251293777920
In ~2 years, summon should work anywhere connected by land & not blocked by borders, eg you're in LA and the car is in NY
10 Jan 2016
AP1 owners who purchased Tesla in Q1/Q2'16 are SOL as in Q3'16 they released AP2 since AP1 didn't have full self-driving h/w despite his Jan'16 tweet.
Elon stated as early as the AP1 launch event that AP1 hardware was insufficient for full autonomy. The claim of a coming cross-country summon feature does not imply that it will be retroactively made possible on every Tesla vehicle ever produced.
If there is ever a successful lawsuit brought by AP1 owners, it is much more likely to focus on the promises that were made to them but never delivered, e.g. "meeting you at the curb" from your garage on private property.
The big leap in deep neural nets, gave us the illusion that we will solve many of our problems overnight. The fact is that we probably need at least one more decade to get things right
The leap was that now we have a plan for how to tackle these problems, we have tools that will actually work. Seemingly overnight those tools appeared and we realized their potential. However, we still need to learn how to best use these new tools and perfect them. Just because we can do something doesn't mean we can do it good enough, yet.
Do we know the tools that will work? Waymo uses expensive LIDAR's, and we dont know the price of the whole hardware set, since Waymo does not care about the price, they work on scientific problem.
Tesla claims that cameras are enough and they can put it into mass-produced cars without changing the cost significantly.
And you have to assume that optimizing this particular toolset gets you 100% (or at least approaching 100% in the limit) to solving the problem that you need to solve. Deep learning/neural nets popped for a variety of factors relative to a number of other techniques that had also been hanging around for ages. But there's no guarantee that past performance will be predictive of future results.
Yes definitively but if Waymo is a decade away then how far away from truly self driving cars is Tesla? 15 years? 20 years?
Tesla claims that LIDAR isn't neccessary but yet they constantly hit barriers and stopped trucks. If they can't even get the basics right how can Musk be sure that his strategy is the right one?
Neural nets are only a small part of the solution. Waymo gets this, but too many of the others overuse neural nets. The safe solutions are about 80% geometry and 20% neural net.
LIDAR and geometry give you a 3D map of the environment, at least the part of it in range of the sensors. Anywhere the vehicle is going to drive should be verified as flat. That's the first step. Waymo does this. Tesla doesn't. Whether Uber does will come out when the NTSB dissects their vehicle that ran over a pedestrian.
Those top-mounted LIDAR units aren't primarily looking for obstacles. They're looking for flat road. A pothole, a cliff, a car, or a pedestrian does not look like flat road. The software doesn't have to know why it's not flat to not go there.
Trying to classify stuff that isn't flat road is mostly for dealing with moving objects.
"What is that?" and "What will it do next" isn't a geometry problem. That's where neural nets come in. An acceptable result is "unidentified", which feeds back into the planner as "stay away from there".
Some people have gotten the idea that they just hook up a camera to a machine learning system, train it from a dashcam, and they have a quick path to self-driving. That will work most of the time. Until it kills someone.
> Tesla charges $5,000 for Autopilot's lane-keeping and advanced cruise control features. On top of that, customers can pay $3,000 for what Tesla describes as "Full Self-Driving Capability."
Is $5,000 what owners of other cars pay for similar driver-assist features?
It's a little high, compared to the last couple of cars I purchased with those features. Want to say that $3k was more in line with what I paid for a full safety/tech package, and that included a speaker/nav bump, etc.
Keep in mind that most cars (including my two examples) only really have lane centering and adaptive cruise, though, as far as driver-assistance features go. Tesla's ability to change lanes and do basic highway nav is above and beyond the norm. I'd probably pay another $2k for that. Also, the way most of these packages work, you often can't get them a la carte--they're on top of another package, so the effective cost might actually be $5k depending on how that's structured.
Paying $3k for vapor that won't arrive before I sell the car/return the lease would piss me off, though.
"Is $5,000 what owners of other cars pay for similar driver-assist features?"
Although at its core, that's what Autopilot is (nothing more than Lane-Keep Assist and Adaptive Cruise-Control), to leave it at that is not giving it enough credit. Tesla's Autopilot is far beyond any other manufacturer's drive-assist features that I have personally driven (and I tested quite a few before purchasing my Model S).
In fact, a lot of people are saying that how well the assist features work is part of the reason why we saw the unfortunate fatality where the Model X ran right into a median. It is so good that people (wrongly) begin to place too much trust into it, and don't pay attention when they need to. When I first got my car (Feb 2017), it was shortly after the big MobileEye departure, and TeslaVision was still in its infancy. Never in a million years would I have taken my hands off the wheel or stopped paying attention when using this early version of TeslaVision AutoPilot. That thing tried to kill me more than once. The current version works very, very well and I'm totally happy with having paid and additional $5,000 for it.
In order to sincerely claim a car has the hardware needed to support a full self-driving capability the car would need several LIDARs, several good quality cameras, a high-precision GPS receiver, and a respectable Inertial Measurement Unit.
If you get all that hardware, $5,000 is a great price; back in 2005, any one of those components would have set you back that much.
Of course, if Tesla doesn't have that full complement of sensors - and from the price, they might not have all those sensors - the promised self-driving capabilities may never materialise.
> In order to sincerely claim a car has the hardware needed to support a full self-driving capability the car would need several LIDARs, several good quality cameras, a high-precision GPS receiver, and a respectable Inertial Measurement Unit.
No, it needs decent cameras with coverage in all directions. That's all what a human needs so that's enough hardware.
Of the vehicles that completed the DARPA Grand Challenge and the DARPA Urban Challenge, how many do you think used LIDAR on their vehicles?
I'll tell you: 100%. Out of 11 teams finishing, every single one used LIDAR.
And I'll tell you why too: With LIDAR you need much simpler algorithms, and that directly translates into fewer bugs, meaning fewer instances of your vehicle crashing into stationary concrete barriers on clear days.
We have brains (at least, I like to think most of us do) which are capable of many, many times the amount of processing power which currently exists in AP2 and AP2.5 cars. Elon himself has mentioned this delta in processing power on at least one occasion, which is why AP2.5 is even a thing.
Simply stating that "humans use eyes to drive, so car can use cameras to drive!" is foolish.
(For those who don't know - AP2.5 is the current hardware suite on Tesla's which, among some other smaller upgrades, doubled the processing power of AP2 cars. AP2.5 has two NVIDIA PX2 units as opposed to AP2's one)
> We have brains (at least, I like to think most of us do) which are capable of many, many times the amount of processing power which currently exists in AP2 and AP2.5 cars.
I think it's just a matter of writing the correct software. Our brain seems really underpowered to me. And considering what even insects can do with highly limited brainpower, I think the hardware is good enough. All the gains need to come from software now.
I think the latest Honda Civic offers it for $1000 in US (gor the lower end trims that do not include it by default).
There in Europe I think all trims have this (ACC, LKAS, CMBS).
Well, the point of the article is that the "full self-driving capability" is not actually available yet, and there is no firm delivery date (AFAIK, that's so for the whole industry, not just Tesla, but I think Tesla is the only manufacturer selling it.)
Sure, but the suite of driver-assist features (e.g. lane assist) for $5,000 is available for Tesla and its competitors, though most reviews say that Tesla is currently leading the pack.
Why not sell full autonomy for certain roads only? Like, put sensors every three feet on the interstate, and have a sensor in the car read the proximity sensor. Perfect full auto driving.
There's no reason to think you need sensors. Although it's not autonomous, look at GM's Cadillac system and you have a pretty good indication that we're probably fairly close (<10 years) from being able to have systems that are completely autonomous on freeways. May need some sort of beacon to flag construction zones and the like but it seems on the horizon.
Which would actually be a big win for safety and being able to generally deal with long, boring drives. That doesn't buy you robo-Uber or cars-on-demand but those are likely much harder problems.
Obviously slowing to a stop is preferable to slamming into a barrier. But we shouldn't allow cars on the road that routinely just come to a stop on highways (nor ones that slam themselves into barriers).
Just as we shouldnt currently be allowing cars onto the road that operate without a users hands on the controls and their full attention on the road ahead. I see your point but if Tesla has fine print that says you must keep your hands on the controls, and knows when you are not they have the ability to do things to prevent that from happening. Not just leaning on the goodwill and promise of the driver.
it has been suggested by many that the ideal start would be to take over existing HOV lanes and turn set them aside for self driving cars. They have unique markings already and from that a system could be developed to mark any road sufficiently and economically.
better yet it can probably be paid for by people who want to use it and manufacturers who want to sell their solutions.
the elephant in the room though is that is has been marketed as a safety system and it cannot operate in the conditions we need it most. just live BEVs there are too many players going in too many directions that regulation and standards have not caught up.
Their marketing claim of "Full Self-Driving Hardware" is evil-genius - it's very similar to the Halting problem[0] - Turing proofed that we can not develop an algorithm that can predict whether any program will eventually halt.
It's very similar with this marketing claim. We can never show that the hardware is not capable of being self-driving. Maybe, someone, at some point in the future, could pull it off. Even if the rest of the industry uses Lidar, beefier chips etc., it is not proof that it can not be done with that hardware. Tesla can keep playing that game, until virtually no one owns the current generation of cars anymore.
Eh. The phrase "technically correct" comes to mind. Sure, it won't be formally disproven, but if Tesla doesn't deliver FSD while these cars are on the road, that will definitely (and very rightfully) be seen as an inability to deliver.
Also, I think the conversation about Lidar in this article (and perhaps among some readers here) misses a key point. Elon doesn't think merely that one can build a Level 4 car without Lidar; he thinks it's vastly preferable for a Level 4 car not to have any dependency on Lidar -- not only for cost reasons, but also for performance in degraded atmospheric conditions (rain, snow, fog, etc.). It's a difference in strategy, not merely an attempt to do more with less.
Visibility range and quality of pictures in degraded atmospheric conditions also goes down significantly. Not sure if it can be an argument for Lidar vs Cameras discussion.
I think it's very different from the halting problem because there is no perfect solution (like there is with "either the problem halts or it does not"). A self-driving car just needs to be good enough, where "good enough" means X% more reliable than the average human driver.
I personally am terrified of human drivers and will not walk close to the edge of a road because I don't trust them to be paying attention. As shown recently, self-driving cars appear to not yet be at that level (though I don't know the numbers - perhaps they are).
When self-driving cars reliably have same-or-few accidents per year (by volume) then I think they can claim "Full Self-Driving".
Humans write the software the power those self driving cars... was it an error that always repeated it self or did only happen every billion iterations? I’m all for self driving but it’s still horrifying to think of the edge cases with software touching the physical world...
It's moreso the cases where the ML model(s) fail(s) to perform the correct action. Essentially the inputs after being passed through the weighted neural network fail to sum up to the correct number == rip pedestrian/passengers.
You used the work Marketing. The moment it is called Autopilot, the general population will believe it is so.
A thousand disclaimers and alerts will not be remembered 10 days later, but the word Autopilot will stick. People hear auto-pilot and they think airplanes flying 100% autonomously (which of course is not the case).
As long as that name stands, people will be doing stupid and dangerous things like watching Harry Potter, or anything-but-looking-at-the-road.
I would call it "enhanced cruise control", but hey, that doesn't sell as much as "Autopilot", does it?
Not having to constantly make minor adjustments to throttle position to maintain the correct speed is a big bonus for me, both in traffic and out of it. I find I can focus much more on watching my surroundings, and whats happening 500m down the road, rather than constantly looking at the speedometer. This also makes a big difference to how tired I feel after long trips. The same applies to the Lateral Control part of Autopilot, although I use that less.
(To be fair, many other cars provide adaptive cruise control too, but Tesla's is the best I've tried so far).
I'm interested in trying out one of the current adaptive cruise control systems. I pretty much got out of the habit of using cruise control. Most of the roads I drive on have enough traffic that I find cruise control leads to me prioritizing maintaining a constant speed rather than sensibly staying in the flow of traffic. I.e. shifting lanes rather than slowing down a couple mph.
TL;DR: Less mental fatigue, more alert, and more aware of my surroundings. The difference between constantly making micro adjustments and not having to was much greater than I expected.
Sort of -- in the sense that on some occasions I'll use TACC by itself without Autosteer (most notably because TACC works great in the rain, but Autosteer doesn't). My experience is that TACC alone is still a huge improvement on regular driving. Autosteer is another big improvement on top of TACC. It's a hard thing to quantify but I'd say they are similarly helpful. Obviously the effect is increased when you put them together.
I test drove Tesla and specifically to test autopilot. I liked it less than Q7 automatic cruise control and line steering. In particular, to my taste Tesla was more/unnecessary aggressive in braking when cars changed lines in front of me.
I use Q7 ACC pretty much all the time nowdays. I do accelerate and sometimes brake manually of course. However I enjoy the fact that I don't need to be 100% focused on driving when it is enabled. It is similar to how I operate autopilot in the plane: monitor it does the right thing but don't bother making constant minor adjustments. I am still mentally in the game, just for a different reason/goal.
I've found it akin to being able to ask someone to grab the wheel for a moment, while I do some quick task. I'm still able to feel if the car veers, or suddenly decelerates, so I'm able to take control quickly, but I don't have to worry I'm going to plow into the guy in front of me who suddenly slammed on his brakes, because I took my eye off the road for 5 seconds.
On a lighter note, my dad learned to steer using his knee so he could eat while driving. I can't help but imagine how much easier his life would be with autopilot (probably safer too).
I wonder about the idea of not "full" self-driving, but with turns and lane changes being automated instead of a steering wheel for example for idiot-proof driving. A joystick may be available as a fallback for manual steering. Of course, there would also be collision detection for example too.
66 comments
[ 2.6 ms ] story [ 124 ms ] threadhttps://www.truthinadvertising.org/wp-content/uploads/2017/0... Tesla’s Marketing and Sale of Vaporware 21. As Dictionary.com states, vaporware is “[c]omputer software that is advertised but still nonexistent.” Tesla advertised vaporware to consumers, knowing full well that this particular come-on would particularly excite its target market of high-tech aficionados.
Now we need one for self-driving for those who were duped into buying a Tesla just for that.
AP1 owners who purchased Tesla in Q1/Q2'16 are SOL as in Q3'16 they released AP2 since AP1 didn't have full self-driving h/w despite his Jan'16 tweet.
If there is ever a successful lawsuit brought by AP1 owners, it is much more likely to focus on the promises that were made to them but never delivered, e.g. "meeting you at the curb" from your garage on private property.
Tesla claims that cameras are enough and they can put it into mass-produced cars without changing the cost significantly.
Pretty large gap, don't you think?
Tesla claims that LIDAR isn't neccessary but yet they constantly hit barriers and stopped trucks. If they can't even get the basics right how can Musk be sure that his strategy is the right one?
The strategy is right one as long as it sells cars.
Musk is a salesman in engineer's clothing.
LIDAR and geometry give you a 3D map of the environment, at least the part of it in range of the sensors. Anywhere the vehicle is going to drive should be verified as flat. That's the first step. Waymo does this. Tesla doesn't. Whether Uber does will come out when the NTSB dissects their vehicle that ran over a pedestrian.
Those top-mounted LIDAR units aren't primarily looking for obstacles. They're looking for flat road. A pothole, a cliff, a car, or a pedestrian does not look like flat road. The software doesn't have to know why it's not flat to not go there.
Trying to classify stuff that isn't flat road is mostly for dealing with moving objects. "What is that?" and "What will it do next" isn't a geometry problem. That's where neural nets come in. An acceptable result is "unidentified", which feeds back into the planner as "stay away from there".
Some people have gotten the idea that they just hook up a camera to a machine learning system, train it from a dashcam, and they have a quick path to self-driving. That will work most of the time. Until it kills someone.
Is $5,000 what owners of other cars pay for similar driver-assist features?
Keep in mind that most cars (including my two examples) only really have lane centering and adaptive cruise, though, as far as driver-assistance features go. Tesla's ability to change lanes and do basic highway nav is above and beyond the norm. I'd probably pay another $2k for that. Also, the way most of these packages work, you often can't get them a la carte--they're on top of another package, so the effective cost might actually be $5k depending on how that's structured.
Paying $3k for vapor that won't arrive before I sell the car/return the lease would piss me off, though.
Although at its core, that's what Autopilot is (nothing more than Lane-Keep Assist and Adaptive Cruise-Control), to leave it at that is not giving it enough credit. Tesla's Autopilot is far beyond any other manufacturer's drive-assist features that I have personally driven (and I tested quite a few before purchasing my Model S).
In fact, a lot of people are saying that how well the assist features work is part of the reason why we saw the unfortunate fatality where the Model X ran right into a median. It is so good that people (wrongly) begin to place too much trust into it, and don't pay attention when they need to. When I first got my car (Feb 2017), it was shortly after the big MobileEye departure, and TeslaVision was still in its infancy. Never in a million years would I have taken my hands off the wheel or stopped paying attention when using this early version of TeslaVision AutoPilot. That thing tried to kill me more than once. The current version works very, very well and I'm totally happy with having paid and additional $5,000 for it.
If you get all that hardware, $5,000 is a great price; back in 2005, any one of those components would have set you back that much.
Of course, if Tesla doesn't have that full complement of sensors - and from the price, they might not have all those sensors - the promised self-driving capabilities may never materialise.
No, it needs decent cameras with coverage in all directions. That's all what a human needs so that's enough hardware.
I'll tell you: 100%. Out of 11 teams finishing, every single one used LIDAR.
And I'll tell you why too: With LIDAR you need much simpler algorithms, and that directly translates into fewer bugs, meaning fewer instances of your vehicle crashing into stationary concrete barriers on clear days.
Am I defective?
We have brains (at least, I like to think most of us do) which are capable of many, many times the amount of processing power which currently exists in AP2 and AP2.5 cars. Elon himself has mentioned this delta in processing power on at least one occasion, which is why AP2.5 is even a thing.
Simply stating that "humans use eyes to drive, so car can use cameras to drive!" is foolish.
(For those who don't know - AP2.5 is the current hardware suite on Tesla's which, among some other smaller upgrades, doubled the processing power of AP2 cars. AP2.5 has two NVIDIA PX2 units as opposed to AP2's one)
I think it's just a matter of writing the correct software. Our brain seems really underpowered to me. And considering what even insects can do with highly limited brainpower, I think the hardware is good enough. All the gains need to come from software now.
Er, yes.
Tesla is really good at following painted lane lines. Very smooth steering. Sometimes right into solid obstacles.
Which would actually be a big win for safety and being able to generally deal with long, boring drives. That doesn't buy you robo-Uber or cars-on-demand but those are likely much harder problems.
better yet it can probably be paid for by people who want to use it and manufacturers who want to sell their solutions.
the elephant in the room though is that is has been marketed as a safety system and it cannot operate in the conditions we need it most. just live BEVs there are too many players going in too many directions that regulation and standards have not caught up.
It's very similar with this marketing claim. We can never show that the hardware is not capable of being self-driving. Maybe, someone, at some point in the future, could pull it off. Even if the rest of the industry uses Lidar, beefier chips etc., it is not proof that it can not be done with that hardware. Tesla can keep playing that game, until virtually no one owns the current generation of cars anymore.
[0] https://en.wikipedia.org/wiki/Halting_problem
Also, I think the conversation about Lidar in this article (and perhaps among some readers here) misses a key point. Elon doesn't think merely that one can build a Level 4 car without Lidar; he thinks it's vastly preferable for a Level 4 car not to have any dependency on Lidar -- not only for cost reasons, but also for performance in degraded atmospheric conditions (rain, snow, fog, etc.). It's a difference in strategy, not merely an attempt to do more with less.
I personally am terrified of human drivers and will not walk close to the edge of a road because I don't trust them to be paying attention. As shown recently, self-driving cars appear to not yet be at that level (though I don't know the numbers - perhaps they are).
When self-driving cars reliably have same-or-few accidents per year (by volume) then I think they can claim "Full Self-Driving".
One software bug can make every Tesla go into casual murder mode on a similar piece of road.
A thousand disclaimers and alerts will not be remembered 10 days later, but the word Autopilot will stick. People hear auto-pilot and they think airplanes flying 100% autonomously (which of course is not the case).
As long as that name stands, people will be doing stupid and dangerous things like watching Harry Potter, or anything-but-looking-at-the-road.
I would call it "enhanced cruise control", but hey, that doesn't sell as much as "Autopilot", does it?
(To be fair, many other cars provide adaptive cruise control too, but Tesla's is the best I've tried so far).
I wrote up my experience in slightly longer form here: https://news.ycombinator.com/item?id=16773903
I use Q7 ACC pretty much all the time nowdays. I do accelerate and sometimes brake manually of course. However I enjoy the fact that I don't need to be 100% focused on driving when it is enabled. It is similar to how I operate autopilot in the plane: monitor it does the right thing but don't bother making constant minor adjustments. I am still mentally in the game, just for a different reason/goal.
On a lighter note, my dad learned to steer using his knee so he could eat while driving. I can't help but imagine how much easier his life would be with autopilot (probably safer too).