Waymo CEO Dismisses Tesla Self-Driving Plan: 'This is Not How It Works' (arstechnica.com) 185
An anonymous reader shares a report: Many Tesla fans view the electric carmaker as a world leader in self-driving technology. CEO Elon Musk himself has repeatedly claimed that the company is less than two years away from perfecting fully self-driving technology. But in an interview with Germany's Manager magazine, Waymo CEO John Krafcik dismissed Tesla as a Waymo competitor and argued that Tesla's current strategy was unlikely to ever produce a fully self-driving system. "For us, Tesla is not a competitor at all," Krafcik said. "We manufacture a completely autonomous driving system. Tesla is an automaker that is developing a really good driver assistance system."
For Musk, these two technologies exist along a continuum. His plan is to gradually make Tesla's Autopilot software better until it's good enough to work with no human supervision. But Krafcik argues that's not realistic. "It is a misconception that you can just keep developing a driver assistance system until one day you can magically leap to a fully autonomous driving system," Krafcik said. "In terms of robustness and accuracy, for example, our sensors are orders of magnitude better than what we see on the road from other manufacturers."
For Musk, these two technologies exist along a continuum. His plan is to gradually make Tesla's Autopilot software better until it's good enough to work with no human supervision. But Krafcik argues that's not realistic. "It is a misconception that you can just keep developing a driver assistance system until one day you can magically leap to a fully autonomous driving system," Krafcik said. "In terms of robustness and accuracy, for example, our sensors are orders of magnitude better than what we see on the road from other manufacturers."
Sour grapes. (Score:4, Interesting)
Re:Sour grapes. (Score:5, Funny)
Tesla should check their navigation systems. Running circles around is not the best way to destination.
Re: (Score:2)
Re: (Score:2)
Running circles around is not the best way to destination.
Actually orthodromes *are* the best way to your destination, and they *are* circles.
Re: Sour grapes. (Score:5, Funny)
The baseless defensiveness of Tesla fanboys is hilarious.
If they ever get in a cockfight with Apple fans, even the Abrahamic religions will tell them to take it down a notch. ;)
Re: (Score:3)
These days it's way too easy to write off things you don't like people saying because "they're just fans" or some equally bad excuse to discount valid arguments.
I'm not a Tesla fan (I live in a city and don't own a car) and definitely not an Elon Musk fan (he's basically a rich version of annoying teenager pulling stunts on TikTok), but I'm someone who loves technology and has spent years following it. And over time I've learned is to be much more exc
Re: Sour grapes. (Score:4, Interesting)
Vaporware? They have passengers in cars without drivers. Dismiss that at your peril but that's an incredible achievement.
Vaporware is Autopilot's FSD which doesn't *even* operate within a limited region under limited weather.
Waymo has a very clear path to a global rollout. It has a defined cost and blueprint to expand. Tesla has no clear path yet to a localized rollout.
It's like the difference between building an electric car that runs but costs $200,000 per vehicle and someone building a fusion powered car. Sure the fusion powered car would be amazing, but there is still fundamental... basic science that needs to be resolved before you can even say if it's feasible this decade or not.
Even Elon Musk admits that in 2019 the Autopilot approach he thought would be feature complete by the end of the year was a dead end and they had to scrap it and rewrite. That's the problem with still being in the basic science phase of engineering. You don't know if a promising lead will be the breakthrough, an incremental step or a blind corner leading to a dead end.
Waymo started from a "cost be damned, whatever it takes to prove feasibility" approach. Tesla started from a "well we can't spend $200,000 per car so we better find a solution that's cheap." approach. If Tesla can do it for $5k that will be great. But the $200k Waymo car is now already an $80k Waymo car. And in a couple years it'll be an $8k car and a few years after that a $1k car. Their hardware is dropping in price faster than Tesla is making software breakthroughs.
People have spent their entire lives pursuing "just 10 year away" fusion power. Meanwhile solar panels and wind has actually changed the world. Is fusion a better solution? Maybe. Can Tesla deliver FSD with the current hardware? Maybe, but they could get there a lot faster with Lidar and more robust maps and location specific hints in the maps.
It frustrates me that Tesla seems to have doubled down on both delivering an incredibly difficult to achieve solution... while also refusing to use any assistance. They did the exact same thing with automatic windshield wipers. They removed a $3 piece of hardware because "Vision can do it" but even with a very straightforward deep learning problem "Detect when windshield is wet" they're years in still far below the performance of a $3 component. The appeal of a generalized AI that solves every driving problem purely in software is causing pain for customers that could be solved with additional hardware.
Re: (Score:3)
LIDAR was a non-starter for Tesla in the beginning because a decent unit would cost as much as the entirety of the rest of the vehicle. Now Velodyne is wholesaling them for something like $100 each, at which point the only reason that I can see for Tesla to not use them is inertia.
Re: (Score:2)
Tesla is selling cars right now, in the place I'm in, that can, you know, on the roads I might drive, actually perform the basic mechanics of driving. That's just a whole lot more exciting.
As long as your roads never include something like a stopped fire engine.
Tesla is very far from full self-driving. Their marketing, OTOH, claims it's here today.
Re: (Score:2)
Re: (Score:2)
I suspect stereo can be a perfectly good replacement for LIDAR, but it's not like Tesla has stereo either. Their multi-angle cameras are all right next to each other, useless for stereo.
Re: (Score:3)
I suspect stereo can be a perfectly good replacement for LIDAR, but it's not like Tesla has stereo either. Their multi-angle cameras are all right next to each other, useless for stereo.
The outer two cameras are about as far apart as your eyes are, so no, no so much. You really don't want them to be too far apart, because if they were, then all of the things you'd see in one camera but not in the other would make generating a depth map harder.
Re: (Score:2)
Well, Waymo didn't go the wrong way, and Tesla isn't running circles around them.
Waymo is aiming at the Taxi market. If the car costs an extra $30K because of the Lidar units on it, it'll just get amortized over the life of the car - if that's 200,000 miles, then Lidar adds $0.15 / mile to the cost of a ride. If Waymo can start running autonomous Taxis 5 years earlier and with no competitors because they chose the more expensive route, they'll still make a lot of money. And, today, they have autonomous t
Re: (Score:2)
Waymo runs fully autonomous taxis in Phoenix Arizona today https://waymo.com/waymo-one/ [waymo.com] If you're into this sort of thing there's a fairly interesting interview with their CTO by Lex Fridman at https://lexfridman.com/dmitri-... [lexfridman.com] (he's also made several interviews with Musk and is an AI researcher himself).
Re: (Score:2)
All taxi businesses in the US combined had an estimated revenue of about $31 billion [bizfluent.com] in 2019. Tesla's revenue by itself in 2019 was roughly $24.5 billion [macrotrends.net].
Re: (Score:2)
>>> we'll see that Tesla's path is the only one that can work
Why do you say "can"? Waymo is doing the same thing as Tesla is doing, with the same sensors, with the exception that they replace an extremely low resolution radar for distance measuring with a high resolution LIDAR. At worst, you could say that the two systems end up with equivalent performance but Tesla's is cheaper, but there really isn't an argument that Tesla's system is fundamentally better than Waymo's, and is the only one that
Re: (Score:2)
Why do you say "can"? Waymo is doing the same thing as Tesla is doing, with the same sensors ...
The GP said Tesla's competitive advantage was having over a million cars being driven tens of billions of miles per year collecting data which can be used to train automated driving AIs. Waymo does not have tens of billions of miles of video and radar to collect to train their autonomous units.
Who knows if the GP is correct that it requires this much data to train these systems, but that was the reason he claimed Tesla was the only company with a viable path to creating autonomous cars.
Re: (Score:2)
Waymo is multiple orders of magnitude below Tesla in terms of available data. They - literally - can never catch up on that front. Ultimately, I believe (and this is one of my few actual opinions that I express) that huge amounts of data are the only thing that can crack this particular nut.
Re: (Score:2)
I believe that huge amounts of data are the only thing that can crack this particular nut.
That's quite an interesting claim. While remembering (via neural networks) every single situation that could possibly occur is certainly one way to handle said situations, there are very good human drivers out there who haven't driven billions of miles. Either those drivers are actually terrible (how do we measure that if they never get into an accident?), or there are alternative ways of learning that doesn't involve dumping mountains of raw data onto the learner.
Moreover, a self-driving car shouldn't nece
Openpilot Comma.ai (Score:2)
It has nothing to do with what sensors are better. Tesla has over a million cars on the road now and all of them continuously watch the road. It's also believed that they watch how humans drive and compare it to what the computer thinks it should be doing.
And thus the sensors do actually have something to do with it: by dropping expensive sensors, Tesla's bet is to make a (relatively) cheaper car (at least compared to robo taxis) that can actually be bought en masse, and thus making it possible for there to be millions of cars out-there gathering the bigdata.
A LIDAR system which alone costs half the price of an entry-level EV would have been a hurdle in developing such a large fleet.
The point is that this is a "big data" problem, and Tesla has enough cars on the road to get that big data. Nobody else does.
Another player who is also taking the "big data" approach real serious and t
Re: (Score:2)
Velodyne is now selling LIDAR units for around $100, there's no reason not to include them now since Tesla's cameras each cost more than that.
Re:Sour grapes. (Score:5, Interesting)
He’s just pissed that he went the wrong way (LIDAR), and that Tesla is running circles around them.
You can take a fully autonomous waymo taxi today in phoenix az, no backup driver no remote: https://waymo.com/waymo-one/ [waymo.com] afaik tesla doesn't have anything similar.
Re: (Score:2)
He’s just pissed that he went the wrong way (LIDAR), and that Tesla is running circles around them.
Sounds very unlikely. I had pretty much the same thing explained to me by an expert in self-driving tech a long time ago. Driver assist is missing fundamental capabilities and parts and you cannot retrofit them. This two tasks look the same but they have fundamental differences.
Re: (Score:2)
Lidar was the right choice. It's now become affordable and compact - Volvo is fitting it to cars this year - and Tesla is still nowhere near having a fully self driving vehicle or even matching Waymo's L4 tech.
Re:Sour grapes. (Score:5, Interesting)
> but my experience has been that the car does not understand closing speed or physics very well
Well, I watch the display every time I'm in the passenger seat (gotta give the wife some time before she gets her own...) and I see problems that are far worse than that.
1) it sees perhaps 10% of crossing traffic, and maybe 15% of crossing people
2) it sees perhaps 75% of traffic going the same direction, if that. Cars in right-turn lanes often do not appear
3) it often sees cars in the opposite-direction lane as passing, as the car enters and leaves the blind spots between the sensors
4) lots of things are just wrong - absolutely anything yellow or orange on the side of the road is a traffic cone, including sharp-turn chevrons, fire plugs, even in one case a construction hat on top of a sign.
5) it slams on the brakes for random reasons every 150 km or so on average. Some times it does it and never does it again at that location, including while driving along a practically empty 401 westbound on a weekday afternoon where the nearest car was at least 30 lengths away and in a different lane (thank god no one was behind me, it SLAMMED them). In other cases, you can see what's causing it but no idea why it thinks it has to stop, like a road curve with wooden crash-barriers but it's not clear why it thinks it has to stop. Perhaps the pole spacing is at one of the fundamentals of the radar carrier or something?
It's years away. This is why I refuse to buy "full self-driving". Well, that at the fact that this car is so freakishly fun to drive I can't understand why anyone would buy one and then not drive it! Even on long trips, which I do all the time, steering is simply not a chore enough that I'd want to avoid it.
Now that said, Waymo's comment is totally bogus. They're claiming that you can't start simple and then expand out, but that's exactly what they're doing too! Different definitions of expansion, granted, but the same process in the end.
Re: (Score:2)
Yeah... we put an emotional support traffic cone in the garage to keep our car happy.
Much of what you describe I attribute to a poorly developed physics engine, which is where LIDAR might be able to augment information. In ~3,000 miles though, I have only had a handful of times with truly phantom braking... but it interprets the warning sign with a blinking light as a traffic signal and tries to stop every time on me; I have learned to reflex accelerate or disengage autopilot in most of the situations.
The n
Re: (Score:3)
Bear in mind that the FSD beta code is a major rewrite that uses all of the cameras in a 3D-fused fashion, rather than the 2D fusion that non-FSD beta folks experience. So unless you're part of the beta program, what you're seeing is a long way from what the beta folks are seeing.
Go check out some of the FSD videos, and I think you'll find that they're a lot further along than you think (except for funny/mostly harmless stuff, like identifying dogs as people once in a while). I'm not convinced it will be
Re: (Score:3)
except for funny/mostly harmless stuff, like identifying dogs as people once in a while
I thought dogs were the best people?
Re: Sour grapes. (Score:2, Insightful)
Re: Sour grapes. (Score:2)
Solving for anomalies is just a trailing end of edge cases.
You missed your calling.
Probably even true (Score:4, Insightful)
Probably true. But is it relevant?
A general self-driving is of course much more work than a specialized self-driving car. But most miles and hours are logged in rather special, and largely closed environments: Interstates, Autobahn, or what the equivalent is called in your country. Start there. You don't need a car that can handle each and every broken traffic light in rural Alabama. If you start there, it's OK if the system needs an additional road-side guidance assistance. (Could be as simple as making sure that making sure that road markings are in good condition)
In general 80:20 principle.
Re: (Score:3, Insightful)
In general 80:20 principle.
So, would you bet your life with those odds?
Re: (Score:3)
In general 80:20 principle.
So, would you bet your life with those odds?
I believe the poster was referring to the Pareto principle. [wikipedia.org]
Re: (Score:2)
In general 80:20 principle.
So, would you bet your life with those odds?
The Tesla owner is betting the lives of others as well. Probably a good thing then that Teslas see ghosts [youtube.com].
Re: (Score:2)
That's not a 80% chance of having an accident, but if it works 100% on the 80% of my miles spent driving on the Autobahn - yes.
I don't think he's talking Pareto (Score:2)
More like the 90% of the work is 10% of the cost type. Crossed with 80% self-driving is good enough in most cases.
Re: (Score:2)
So, would you bet your life with those odds?
They are pretty good odds compared to what you bet with when you get in your car every day, which is that close to 80% of all people currently on the road have been in an accident at some point.
And as for mortality... Well if you're going to be in an accident by all accounts you're best of being in a Tesla when you are given their safety ratings exceed even those of Volvos.
If Telsa screws up bad enough (Score:5, Insightful)
Re: (Score:3)
Self driving cars are nothing like nuclear power. One nuclear accident can have very wide ranging effects and costs hundreds of billions of Euros.
A single self driving car can't do any more damage than a single person with a car.
Re: (Score:2)
Re: (Score:2)
You're also not saving on number of autos. Automakers are still focused on the old school idea of one or more cars per person. We need to get to a model where we have better mass transit and not assuming that we replace mass transit with "ride share" or autonomous vehicles. What if the computer world was like that and you solved congestion and performance issues by just buying more ram or a faster internet because optimization is too hard? Oh wait...
Re: Probably even true (Score:2)
Except we'll forever be stuck with a half-assed crutch because nobody has the resources or a chance to enter the market anymore, because of something that juuust about kills the energy to actually do something about it, but is not one bit further away from total shit than absolutely necessary. So maximum stable grinding-down infuriation with no way out due to how brains fundamentally work.
Let's please to it properly from the start.
Because the most long-lasting solution in the world is the temporary crutch.
Re: (Score:2)
I would really like self driving to come to complete fruition, but I think this last 20% bit is going to take a lot longer to crack, and probably parallels some advances that need to occur in machine intelligence. My gold standard would be putting a kid in a car and sending them to the grandparents house 10 miles away. Or trusting the car with the grandparents when they are unable to drive safely. In those instances the fear is less about some horrific accident with a mismarked center divider at 65 mile
Re: (Score:2)
That's a good test. But my first goal would be, I drive to the Interstate as normal, "log in" and take over an hour later when i'm on the exit ramp.
Re: (Score:2)
Yes, the grandparent standard.
I should have 30 or 40 years before the kids take my keys. I'm particularly interested in something being available by then, and having it before the horrible moment.
In the meantime, it would be *nice* if I had something that would let me decide to have an extra round at the sports bar during an exciting game--but I wouldn't be likely to shell out for such a thing unless it could also be my permanent geezermobile . . .
hawk
Re: (Score:2)
I think self driving is not really that important. A better solution to the commute is to just hop on the bus, light rail, subway, etc. If those don't work well, then improve the mass transit or figure out ways to eliminate the need for transit in the first place. Crowded freeways filled with self driving cars with all the passengers working on their laptops for more billable hours seems like a dystopia. Self driving for trucks seems fine but for the commuter it feels like just another gadget type thing
Re: (Score:2)
>>> better solution to the commute is to just hop on the bus, light rail, subway
OK, if it worked that way. Instead it works as:
1. Walk to the stop, which smells like urine and probably has someone sleeping on the bench.
2. Stand around with a bunch of people in the cold, or heat, waiting for the bus or light rail car to show up. Some of the people you're standing next to make you feel...nervous.
3. Get on the transit with all those people you've been waiting with.
4. Take 5 times as long to get to
Re: (Score:2)
Well, walking is great. Fit a half hour of exercise into your 23 hour workday. No need to drive to 24 Hour Fitness. And if the freeways are standing still the mass transit may be faster in many cases.
If that's all too radical, then maybe work from home. Or run for transportation board to improve the facilities. And hope that don't bulldozer your house to make way for a bypass freeway.
Re: (Score:2)
For my regular short trips I definitely prefer mass transit, and would in an unfamiliar city as well if it were possible. For those situations where "you can't get there from here", such as driving to our cottage, a vehicle that I don't have to drive would be wonderful. I hate driving, and my wife can yell at Autopilot as much as she wants as long as she's not complaining about my driving I don't care.
Re: (Score:2)
Mass transit in America is typically awful. So much was designed around the car and mostly because the cities were still newish and had room to expand. Europe felt very different, so mass transit because more popular. I spent several weeks taking the train in Helsink to work out of town, even though I was allowed to take a taxi and have it reimbursed, and there was always a line of taxis waiting outside the hotel or the office, and they all spoke great English and they had memorized all the addresses and
Re: (Score:2)
I start with a product that actually does what their CEO, PR department says it can do. Elon has been claiming fully autonomous capability since 2016, and I'm here to tell you that those capabilities weren't there then, and they aren't there now.
Elon, stop with the distortion field and put-up or shut-up.
Re: (Score:2)
There are (at least) two fundamental problems with this approach:
A system which is good enough to handle the general case of driving (clear roads, moderate tra
LIDAR not "orders of magnitude better" (Score:5, Interesting)
In terms of robustness and accuracy, for example, our sensors are orders of magnitude better
LIDAR is not "orders of magnitudes better" than cameras, if for no other reason than they cannot sense color - used heavily in road signage and invaluable in distinguishing some kinds of objects.
Granted LIDAR can sense shapes pretty well and knows where objects are to a pretty good degree. While that is helpful close up in order to have a really robust general self driving system you really need to be able to sense color...
The reason Waymo works where it does, is not because the sensors they have are so awesome but because they have pre-mapped out the road so heavily the car is easily able to tell exactly where it is via the combination of GPS and LIDAR data. Anywhere outside of the mapped area and it is toast.
Tesla is building a truly general self driving system you can use on any road, anywhere. And it's not like they are gradually adding features and expecting FSD to pop out - instead they spent a LONG time gathering reaction data from the fleet of Tesla cars so they could train a driving model that could handle unexpected events, and currently they have FSD in alpha where people can turn on FSD and more directly train it by taking over when the self driving aspect fails, adding further refinement to everyday driving.
Re:LIDAR not "orders of magnitude better" (Score:5, Insightful)
Fundamentally, LIDAR gives you a series of timing measurements, which you can process into a point cloud. A camera gives you a series of intensity measurements; if you've got two of them, you can compute displacements, and process those into a point cloud. Everything from there on is processing.
One of the problems with this debate is that people usually talk about both systems as black boxes. Waymo says their LIDAR gives them a "top down view of the world." It doesn't, of course, it gives them a series of timing measurements, which they can process into a top down view of the world, just like Tesla (or Waymo) can do with cameras.
Waymo's strategy is to put *all* the sensors, and the very best of those, on their cars. They've got LADAR, Doppler RADAR and cameras. Each of those has some strengths that make up for weaknesses in the other systems. The drawback is extreme cost. Tesla is taking the other approach, building something that's practical and then seeing what they can do with it.
It would be surprising if Waymo didn't have a fully self driving car first. It would also be surprising if you couldn't buy one for normal car prices from Tesla first.
Re: (Score:2)
Tesla has to derive depth from highly complex and fragile multi-frame methods which take into account the vehicle speed ... it's a fucking shitshow compared to LIDAR. Adding colour from a normal camera to a LIDAR image is trivial on the other hand, which Waymo almost certainly already does.
If Tesla at least had stereo it would make deriving depth far more robust, but they don't have that either.
Re: LIDAR not "orders of magnitude better" (Score:4, Insightful)
Re: (Score:2)
Because of the distance we are also forced highly complex and fragile multi-frame methods ... but we are much better at interpretation, when we're paying attention.
Biology agrees with me that the more robust stereo method is superior though when feasible, we got two eyes after all. For a car stereo vision is feasible, just putting a camera on the far corners of a car gives more than enough stereo separation for accurate depth at several hundred meter distance.
Re: (Score:2)
As it turns out you have two eyes, which is stereo.
Re: (Score:2)
As it turns out you have two eyes, which is stereo.
2 eyes only give stereo in close conditions. With driving, 2 eyes really doesn't add much, since distance vision is more important. Maybe for close parking, but not driving.
Re: (Score:3)
As it turns out you have two eyes, which is stereo.
2 eyes only give stereo in close conditions. With driving, 2 eyes really doesn't add much, since distance vision is more important. Maybe for close parking, but not driving.
Indeed. Also this seems to be generally unknown. Humans do distance estimates for distant objects by size. Which requires reliably recognizing these objects, something machines cannot do today and will remain unable to do for the quite a while yet.
Re:LIDAR not "orders of magnitude better" (Score:5, Insightful)
It's painfully clear you have no applied experience working with sensors in a practical setting, or performing sensor fusion.
LIDAR and cameras sense fundamentally different things. When gauging distance to other objects, LIDAR is certainly orders of magnitudes better than cameras in a few ways that are extremely important to self-driving. First and foremost, LIDAR provides a raw measurement of distance to another object at an extremely high rate, with relatively little post processing (as compared to multiple image fusion). Tesla self-driving has at least one fatality that has made headlines which was due to a visual/color as primary sensor approach.
Adding color to LIDAR is not that hard. It's not like Waymo has never heard of a camera or cameras are some arcane technology that's hard for folks to use... in fact, they show that they do use cameras in conjunction with LIDAR: https://blog.waymo.com/2020/03... [waymo.com]
You can get rough distance data from 2+ good camera feeds, but the 1) distance data is going to be fairly rough and 2) getting two good camera feeds is harder than you think it is. The human visual system has a broad range of visual adaptation techniques that cameras don't have. Glare, dirt, not enough light, too much light, vibration, orientation change due to temperature shifts, so on and so forth.
Frankly, trying to do a self-driving system without LIDAR seems like driving with blinders on. You can do it, but almost anyone would be significantly better at it with their full visual field.
Re: (Score:2, Informative)
It's painfully clear you have no applied experience working with sensors in a practical setting, or performing sensor fusion.
I have done experimental work in that area, including combining that data for use in neural networks.
LIDAR and cameras sense fundamentally different things.
Yeah that's kind of my whole point. Thanks for playing though!
LIDAR provides a raw measurement of distance to another object at an extremely high rate, with relatively little post processing
Yes, and??? it takes more post-processi
Re: (Score:2)
While I agree with everything you said your analogy has to be the most self-defeating thing I've ever seen. On account of the fact that cameras on a Tesla give the car a complete 360 degree visual field, it's actually the complete opposite to driving with blinders.
Stick to what really happens: You're driving with one eye closed. Or you're on a motorbike with no visor and it's dusty. Those a far better analogies of the problems facing cameras.
Re: (Score:2)
100% fair.
Re: (Score:2)
... sensors orders of magnitude better than what we see on the road from other manufacturers.
as a bragging right is moronic. Sensor accuracy is already far, far better than needed with clever and innovative algorithms, your average human manages easily. I’m looking forward to the clever new approaches that quantum computing will enable by removing computational complexity - that would be brag worthy. If you could slap on a gigapixel camera and call it solved we’d already be riding in them.
Re: (Score:2)
Not sure why you say that; there are plenty of LIDAR systems with full color capability.
The disadvantage of LIDAR is that you need a lot more processing power to convert the point cloud into useful information for driving, and you end up with a lot of non-material information for driving decisions.
My experience with autopilot on a Tesla is that it has a number of limitations that might not be addressable with cameras alone. (Which is why they upgraded their forward-facing RADAR platform.) I do not think the
Re: LIDAR not "orders of magnitude better" (Score:2)
You are speaking like they are mutually exclusive.
I had assumed Waymo, like everone sane, would use both. And Tesla tried to just do a hack job with only cameras for no reason other than flashiness and attention being the only thing Musk actually cared about. Aka being the equivalent of a Chinese import product with the main goal of tricking you.
Re: (Score:2)
Actually Tesla went all-camera because at the time any decent LIDAR cost more than their car was supposed to. That's no longer the case, I wouldn't be at all surprised to see the radar replaced with a lidar in the next year or two.
Re: (Score:2)
You don't need color to distinguish road signs, they're designed with color blind people in mind. Not only that, I'm sure it would be no problem for people to drivie a car with only B/W vision.
Re: (Score:2)
LIDAR is a complement to cameras, not a replacement.
LIDAR can't "sense shapes" because in current implementations the point cloud isn't dense enough, but it does know precisely where an obstruction is ahead of it. You can't use LIDAR to detect a pedestrian 50 feet in front of a car, but you can use LIDAR to detect that there is something between 5 and 6 feet tall and between 1 and 3 feet wide precisely 49.5 feet in front of the car. If you combine that with a vision system that can classify shapes, and a
Re: (Score:2)
This is probably what the dismissive comments about Tesla were really about... Waymo's never going to be able to make big bucks on robo-taxis and they don't seem interested in manufacturing cars, so their plan seems to
The answer is "All of the above" (Score:4, Informative)
It's best to use all three technologies together: LIDAR, image recognition, and pre-mapping of routes. Triangulating various clues and experience together is how humans drive successfully.
Self-driving is a tricky task to do well, and the computer needs as much info as it can get. For example, do you slam on the breaks if an object suddenly hops into your path? If it's a plastic trash bag in the wind, it's probably not worth slamming on the breaks if there are cars behind you, but rather apply a gentle deceleration. If it's a person crossing the road, then hit the breaks on full. LIDAR probably can't make that distinction alone.
If auto-cars get a reputation for being break-happy, causing rear-end collisions, there will be a backlash. It might not even be fair, as such cars may do other things better than humans, but perception matters in society, and skepticism abounds.
And I'm a better driver on my way too and from work because I know the routes. For example, I know spots where other drivers tend to get confused due to screwy lines or signs, and thus I'm slower and more careful at those spots and formed contingency plans, such as seeing if there are oncoming car in the opposite lane so that I can swerve into it if necessary. Experience matters. The older parts of towns have really screwed up streets, as they were expanded in ad-hoc ways as population increased.
Side note: since the pandemic, drivers seem crazier than normal. I've had too many close calls and saw two accidents near me. The world's gone nuts. Carmageddon
Re: (Score:2)
One of the more spectacular Autopilot failures was when it ran the Tesla into the back of a stopped fire engine. In at least two different incidents.
You don't use LIDAR because it can make out shapes. You use LIDAR because it has a very precise measurement of range to everything around it. When you have that measurement, you stop before hitting the fire engine.
With cameras, your measurement is limited by your ability to do object recognition, which is what failed on the back end of those fire engines.
market niches (Score:5, Interesting)
Re: (Score:2)
https://waymo.com/waymo-one/ [waymo.com]
Re: market niches (Score:2)
You think there is no market for flying carpets?
Build one and if you can make it safe, you will see.
Because... (Score:4, Interesting)
incremental improvement of a system can never produce as good or better a product then one created in the full by intentional design. *warning joke* I assume he is also a new world creationist then ;) *LOL*
In any case , the augment that Telsa can't use scalability and incremental improvement to get to the same place really doesn't wash.
Beside that. I don't really want a self driving car. I'm not sure i want one that drives me around the city based on a GPS. GPS isn't always correct and the computer doesn't automatically 'see' that problem. I would be perfectly happy with a functioning 'autopilot' So I can read a book when I take 4 and 6 hour road trips to see family etc. rather then risk road hypnotism and fatigue. Sometimes consumers will happily use MP3's rather then demanding uncompressed audio, because there is a point at which 90% of the value is found in 70% of the functionality ( and 50% of the cost).
Re: Because... (Score:2)
Yes it does "wash".
A bicycle will never be able to get to the moon. No matter how many incremntal improvements you make. It must stop being a bicycle for that. No matter how long you drag it out before you give up on wheel-based propulsion.
Cameras without LIDAR suck. That is not a hard thing to get.
Yes, if you work hard enough, you can halfway make it work. But who's stupid go the hard route, *and* introduce an unnecessarily insanely much bigger amount of complexity and error-proneness out of pure bone-head
Re: (Score:2)
Re: (Score:2)
>>>Cameras without LIDAR suck.
So we shouldn't allow humans (who are only equipped with two cameras) to drive because they suck? Well, I'll give you that, humans do suck at driving, but not so much because they can't see and make sense of the world in front of them with their cameras.
If the goal is never to do autonomous driving until it's perfect, well, we'll never do it.
Re: (Score:2)
Re: (Score:2)
A bicycle will never be able to get to the moon. No matter how many incremntal improvements you make. It must stop being a bicycle for that. No matter how long you drag it out before you give up on wheel-based propulsion.
Cameras without LIDAR suck. [...]
Yes, if you work hard enough, you can halfway make [cameras] work
If not for the fact that humans have been relying solely on two—sometimes less, and oftentimes low fidelity—biological cameras to drive autonomously for over a hundred years, your argument that cameras are insufficient for full self-driving might actually make some sense. Unfortunately for you and Waymo's CEO, the argument makes absolutely no sense because we already know with certainty that cameras are more than sufficient. In fact, they're literally the only way that full self-driving has work
Re: (Score:3)
In any case , the augment that Telsa can't use scalability and incremental improvement to get to the same place really doesn't wash.
That's not the argument. The argument is Teslas only have cameras, which greatly limit their ability to do full self-driving. Sometimes image processing fails, and that means you run into the back of a stopped fire engine.
LIDAR will give you a precise measurement to everything around you, which will let you know there's something big and not moving in front of you, even if your image processing doesn't recognize it.
The problem for Tesla is they now have a lot of vehicles on the road that only have cameras
Re: (Score:2)
I'm not sure i want one that drives me around the city based on a GPS. GPS isn't always correct and the computer doesn't automatically 'see' that problem.
No one is proposing or building self-driving car systems that rely on GPS, other than for navigation. GPS is way too imprecise for positioning a car within a roadway and, as you say, it knows nothing about obstacles or temporary changes. Any functional system has to do all of the driving based on seeing what's around it.
And I'd prefer if Americans learned to drive. (Score:2, Insightful)
The process for getting a driver's license in America is a joke.
No wonder people are attracted to something that is still bad but maybe not as bad as them to do it for them, so they can become even worse at it.
We should build our roads and cars and train our drivers to be able to do 200 mph easily and without stress, even when they are tired. *While* still being in control of their own fortune, aka still being people. Not externalized blobs of meat with eyes, of the feudal corporation-state swarm queens.
3227 (Score:4, Insightful)
Germany has some of the most arduous driver training in the world. The number in the title reflects how many people died on your roads due to your excellent driving in 2019.
Per motor vehicle there are 9 other western countries with a lower death toll.
Per billion km travelled there are 5 other western countries with a lower death toll *note that most countries don't publish this stat so it could be much higher.
Per person there's 7 other western nations with a lower death toll.
All of these countries don't train their drivers as well as Germans do. Hell one of them is Italy (fucking LOL).
If you think you can solve the problem through driver training then you're as ignorant as our most prolific ignorant poster BAReFO0t. ... Wait... You are BAReFO0t. Well you are consistent at least.
No one can rival Musk at media spin (Score:2)
The bigger question... (Score:5, Insightful)
Re: (Score:2)
On the flip side, if Waymo comes up with a workable, affordable system, they can sell it to all the other automakers who don't have such a system, and not have to make the cars themselves. I can see room for both models in the next few years.
Re: (Score:2)
As an owner of a Tesla Model 3, I'd say that the habit will be tough to break, but if someone makes a compelling Taxi service I'd give up my car.
But, it can't be just any taxi service. When I want to go somewhere, it has to be there even if I want to go NOW (OK, I'll give them 5 minutes to show up). It has to be priced reasonably, it has to have inviting vehicles, and it has to work for me - it should work just as well for popping down to the corner for a gallon of milk as driving 350 miles to go visit my
He picked the wrong thing to brag about (Score:2)
tldr but robustness and accuracy is exactly the kind of thing that you can evolve towards, just develop better sensors. Being able to drive safely with bad sensors would be something to brag about.
Re: (Score:2)
tldr but robustness and accuracy is exactly the kind of thing that you can evolve towards, just develop better sensors
Nope. The sensors are already in thousands of Teslas on the road today, and their owners paid thousands to have those sensors installed on the promise of full self-driving, so you can't tell them "you have to buy this package now".
"Just add better sensors" is not something Tesla can easily do at this point. Which is what this dude is saying.
As a Tesla owner who uses AutoPilot... (Score:5, Interesting)
As a Tesla owner who uses AutoPilot and has paid for FSD... I agree. Mostly.
I'm skeptical of Tesla's ability to reach level 5 with their optical-only system, and I think that Waymo is in a completely different category than Tesla, not just miles ahead but on an entirely different level. Waymo has already achieved level 4 autonomy in a moderately-demanding environment (Phoenix city streets), and their use of LIDAR in addition to cameras, is a huge differentiator. I agree that Tesla is not a competitor to Waymo, and that Elon's dreams of every Tesla being able to operate as a self-driving taxi are a long way off -- and probably not achievable with current hardware.
That said, I think Tesla is pretty close to achieving level 4 in a much more limited environment: Well-traveled freeways in good repair and in reasonably good weather. And Tesla has an enormous telemetry advantage, because they have hundreds of thousands of cars on the road. If you combine those things, I could see Tesla being very close to offering true full self-driving from entrance ramp to exit ramp. They just need to map out the sections of freeway that have extremely low disengagements, and aggressively monitor for road work, plus make sure their system can do a good job of coming to a safe stop when approaching an area the car can't handle.
Unfortunately, Tesla seems to be focusing on the hard problem (full autonomy in complex city environments), rather than the low-hanging fruit of freeway autonomy. I think they should focus entirely on making that work perfectly, enabling Tesla owners to use their vehicles as sleeper cars (especially if they can also fully automate recharging). If my car could do that, I would drive rather than fly when I have to go the home office. Of course, my wife would be annoyed at me taking her car away for a week at a time, but that's okay because the mileage reimbursement from my employer would come very close to covering the payment on a Cybertruck, so I'd buy one and use that.
Level 4 freeway capability will change medium-distance travel, when it comes. And I think Tesla is actually very close to it. I'm sure Waymo could do it even better, but changing short-distance travel is even more lucrative, so they're probably better off keeping their eyes on that prize.
None of it works, all of it is fake (Score:2)
Re: Bashing the competition... (Score:5, Insightful)
Seeing a basic rational observation as "bashing", aka a personal emotional thing, is generally a sign of a trigger, a cult and/or a mental illness.
Re: (Score:2)
Re: (Score:2)
That would depend on whether that was bashing or offering a basic rational.
Re: Tell that to Teslas running autopilot beta (Score:2)
You sound like the type of pointy-haired boss that picks the quickly cobbled together demo mockup facade built on a swamp over the guys that don't have any above ground walls up yet because they built a proper foundation and basement.
If you do it right, you usually take a long time for a good core, and then will easily quickly finish the shell. If you are a sleazy cheat, you will do the hollow shell first, then graft a shitty core in there, which will take longer in total, and then make a new shell that isn
Re: (Score:2)
I've seen some videos on youtube that seem to show Autopilot in full control through all types of drviving
It's always helpful when people post their crimes to YouTube.
Re: (Score:2)