▲ 43 points
back
94 comments
Lidar will give you XYZ (z = distance) on a single datapoint, cameras give you XY and you need to calculate the Z from multiple cameras.
Lidar is ugly and expensive, but requires way fewer cycles to calculate distances. And it works through fog and even some obstacles, nor does it care about glare or get "scared" of shadows.
If we had infinite compute in the car, cameras would be the obvious solution ...but we don't. You can get "Pretty good" with cameras, but still people find specific spots where the car is 100% sure a shadow under a bridge is a car or something stupid like that.
> you need to calculate the Z from multiple cameras.
You can get 3D depth estimation from a single 2D image but I don’t know how good it is for automotive.
"depth estimation" precision matters! and it's still prone to optical illusions and other issues (many solutions struggle with road shadows, for example. sound familiar?)
I think the natural solution is a 360 lidar as a source of truth, followed by optical for everything else
$60,000 for car with sensors? I'm delighted to hear Waymo may be anticipating such low price points:
"We are hearing that Waymo is close to having its entire package (car + Waymo sensor suite and driver) at around $60,000"
The 100% tariffs on Chinese cars have been the biggest obstacle from not getting to that cost point sooner.
I've been a Electrek reader for almost a decade. It's interesting to see its editor Fred Lambert from super pro-Telsa/Musk when fighting EV vs. ICE cars to anti-Telsa/Musk when Musk got into politics.
Of course Musk knew multi-modality would win true Level 4 autonomous driving. It was a calculated tradeoff to work on vision first when LiDAR (components cost 10s of thousands and unreliable for continuous functioning) and Radar (components cost thousands) when Telsa was battling production hell and has real threat of not making it.
Millions of Tesla on the road and 100s of thousands supercharging stations later, LiDAR becomes more reliable and affordable (just like how Tesla made the lithium batteries affordable).
Could Tesla re-incorporate LiDAR + Radar into their sensing stack in the future. I think so! After all, just a few short years ago, Telsa threw away of their 300k+ lines of C++ initial AI software stack based of MobileEye OEM in opting in their own Deep Learning based neural network model. Now, they might still there are still room for camera/vision only assisted driving technology.
Waymo's modified Jaguar model costs $150k-$200k apiece fully loaded with 3 modalities. Tesla Cybercab production cost is below $30k. They are on DIFFERENT segment. Comparing them is like comparing Toyota Corolla vs. Feferrari for their engine output, absurd!
So, this article feels like another emotional biased hit piece of Lambert jabbing at Musk which is daily occurrence now at Electrek.
The general approach to pioneering work like self-driving is:
1) Make it work
2) Make it right
3) Make it cheap
For no good reason, Elon tried to solve step 3 before the other two. This is extra strange since SpaceX has done much better at pioneering work.
Lambert became anti-Tesla when Tesla told him that they weren't going to give him the 2 Roadsters he earned through referral fees.
Waymo's Zeekr and Hyundai based next gen platforms cost less than twice as much as the Cybercab and hold more than twice as many people. And have Lidar.
Tesla could add Lidar to their models, but it would cost them an extra couple of years. They're already well behind Waymo, in two years time Waymo will probably be available to the majority of the US population...
Tesla did not make Lidar affordable. They don't even use lidar.
They also did not make lithium batteries affordable. Panasonic and CATL did that. Tesla was simply the first to exploit that bew affordability.
I'm hoping Tesla reincorporate Lidar (and Radar) because it will mean that they are able to put reality ahead of politics.
Musk being who he is, however, I'm not holding my breath for that outcome.
Personal experience, but my Tesla Model 3 drives most trips from start to finish. If I have to intervene, it's usually because other drivers aren't letting it into a lane and I have to get aggressive. Friends I know in Dallas and Austin (simpler, friendlier streets than where I live) say they don't drive at all anymore. Maybe it's not good enough for a robotaxi service yet, but the rate of progress since my first Tesla in 2022 has been crazy to watch, especially in the last year.
Conversely, I've tried Waymo in Austin and SF many times, and while I enjoyed the experience overall (much better than Uber) it seemed to make a number of mistakes my Tesla wouldn't. In Austin they've struggled to pull up to the right pickup spot and taken very odd backroads (apparently b/c they struggle with certain intersections), and in SF a remote driver had to take over driving because my Waymo (along with another one behind me) got stuck behind a car that was double parked.
Any actual Tesla drivers want to weigh in here?
Because Waymo is largely unsupervised, they take a much more conservative route than Tesla does. Their route planner includes avoiding high-incident rate stretches of road.
If there is a stretch of surface road with a complex intersection and aggressive drivers, you are there to step in. The Waymo Driver instead will try to avoid that area.
Also, Tesla's hardware is not nearly complex or powerful enough to approach the low level of intervention or incident rate that Waymo has - but HW5 will be a massive step forward for them.
That's going to be the make or break for me, personally. If Tesla can't make their solution work well on HW5, then they have serious structural issues in their program.
Tried FSD once. It was very confused about a traffic circle, cruised up to it & slammed on the brakes. Got rear ended (at low speed).
The incomplete, miscategorizing and glitching display of other vehicles and obstacles does not give me confidence that the car is able to correctly identify what’s around it, let alone drive accordingly.
Even the lane departure warnings are 90% false positives (but I’d rather be warned than not).
Just driving normally is much less stressful and more fun than supervising the car’s attempt at driving like a teenager.
How long ago was it that you tried FSD? My car (HW4) had a lot difficulty like this in the early days of FSD 11 but over the past thirty months it's gone from "good at highway driving, not very good at navigating city streets and it might try to kill me" to "nearly perfect at everything". Over the past six months, most of my interventions have been to avoid it automatically parking in handicap spaces and minor navigation issues when I know there's a quicker/better route. Since FSD 14.2, I've done 23k miles with 94% of that on Full Self Driving. Longest streak was 930 miles where it took me on a road trip out of state and then back home and I never touched the wheel; it handled parking at chargers and diners, navigating strange destinations (small towns, back roads), avoiding two deer that I almost certainly would have hit, and managed to keep 85mph while navigating traffic during rush-hour in Chicago. I've never owned a product, much less a vehicle, that got so much better over time.
Will the Tesla handle getting stuck behind a double parked car without your intervention?
I thought the pitch for just-cameras was that the hardware was cheaper and that it's what humans are "using", evidence that it should be sufficient. Just-cameras was to take more engineering to build upnfront but pay off in lower costs in the end. That argument disagrees with the graph that's presented.
The argument made more sense 5+ years ago, but the hardware costs of sensors such as LIDAR has plummeted.
In addition, a certain CEO of a certain car company has made very aggressive sales pitches about how LIDAR isn't necessary and even going so far as to sell unfinished self driving solutions as being completely capable of doing so.
So hypothetically, if that car company were to adopt LIDAR, they would have been selling their solution under false claims in a sales capacity.
Sure, cameras (eyes) are mostly all that humans are using, but two points:
1) the vehicles aren't quite matched with the same experience and intelligence (right now) as exists behind a human's two "cameras"
2) we hold self-driving robots to a much higher standard. meatbags are getting smushed in large numbers daily by other humans, but a single robot smushing a human is headline news for days. giving the robot vastly more information about the world around it, in much higher detail will eventually allow it to drive much better than us.
[obviously we should hold them to higher standards, we should create something much better at driving than the best humans]
> 94% fewer serious-injury crashes than human drivers
That is a remarkable number. You can avoid 50% of multi-vehicle crashes by being a good driver yourself, but the other 50% are caused by the other driver. To get above 50% you have to drive defensively and do a good job modelling the actions of other drivers.
From my experience of potentially serious injury crashes, they are mostly from going too fast and crashing into something. The ones where I've stopped and some other bad driver has gone into me are mostly too slow to cause injury.
i know this doesnt take away from the point of the article, but the whole thing is written by ai which just immediately makes me distrust it
Also worth watching the source presentation which is really interesting and informative beyond the Waymo specifics: https://www.youtube.com/watch?v=Gp4zrV3-6N8
> “Humans of course can drive with just eyes, so there’s that proof of existence,” he said. “If the goal were to just approximately match human performance or to build an assist product, that’s a very reasonable way to go.”
It feels like there's a lot of assumptions baked in here. For example, do humans get in accidents because eyes are insufficient driving sensors? Or because human drivers get tired, distracted, make mistakes, etc.
Also, you can put 10 cameras on a car looking in all directions at the same time, which feels materially different from a human that only has 2 eyes which can only look in a single direction at a time.
It's possible that massively superhuman driving performance can't be achieved with just cameras, but I'm not seeing a strong case for that being made here.
Humans also have prior world domain knowledge. A human driver seeing a guy running across the road behind an obstacle can infer that the human will probably pop out on the other side (and prepare accordingly). A human can see the silhouette of an object with 1 eye and still infer that it has depth because they know what objects are, which is why one-eyed people can drive legally everywhere in the US. This inference extends to everything, not just shapes but behaviors of anything in the road, animate and inanimate objects.
Also humans can move their heads around which gives extra parralax information and thus better depth perception than two static cameras would imply.
Humans of course generally have human level intelligence and Teslas don't yet which means with similar inputs they'll probably do worse. Also I'm not sure Tesla's cameras provide as good an input as sitting in the driving seat? Certainly I can't drive nearly as well looking at the reversing screen as looking through the window.
Competitor explains why competition's product falls short - film at 11. I'd say the proof is in the pudding, which system shows the best results under different circumstances. I suspect the end result will be that a combination of cameras and some scanning apparatus - lidar, radar, sonar - gives the best results and that self-driving systems will converge on that combination.
We already have those results. It’s why Waymo is doing millions of paid rides per month while Tesla is LARPing a robotaxi service that’s basically doing zero fully autonomous rides, as far as anyone can tell.
I can purchase a self-driving Tesla at a reasonable price, in-line with other cars of the same class.
I can't purchase a Waymo at any price.
You can't purchase a level 3 Tesla at any price, either.
To me, true self-driving starts at level 3. Otherwise, it's supervised.
And Tesla selling their product as "Full Self Driving (Supervised)" is like selling "Fully Cooked Meal (Frozen)". No! You call it a "frozen meal"!
If Teslas could self drive you could sit in the back seat, or sleep, or send the car somewhere without being in it. Elon really wants them to, but Tesla's do not self drive.
You can't buy a self-driving Tesla today. Even Tesla tells you that in the fine print. Its self-driving (supervised). If it was actually self driving, they wouldn't be saying (supervised).
"FSD (Supervised)" is a marketing name... which they updated with the parenthetical almost certainly because of brewing legal issues with false advertising. It does not self-drive.
https://www.tesla.com/support/fsd#pay-attention
> FSD (Supervised) is an advanced driver assistance system that is intended to be used only with a fully attentive driver. It does not turn a Tesla into a fully autonomous vehicle.
"self-driving" in better than ideal conditions.
What does this have to do with anything at all?
Right. But are $200K worth of extra sensors actually worth it or are 10+ cameras already safe enough?
This article makes no attempt to answer the actual meat of the question.
Ultimately, people will vote with their wallet.
p.s. I wonder if you could prevent glare by placing an LCD-like device in front of the lens dynamically blocking out the sun.
Where did you get $200k worth of extra sensors?
Also, contrary to your statement, the article directly addresses the meat if the question:
> The disagreement is whether “good enough to help a human drive” ever becomes “good enough to remove the human,” and his answer is no, because the sensing tops out before you get the last few nines.
> Now, Tesla would point to its active Robotaxi service as an example of “removing the human”, but Dolgov also said that camera-only is also good for a demo, and that’s basically what Tesla’s Robotaxi service is right now.
LIDAR units are nowhere near that expensive.
Even ten years ago, the Velodyne top of car units only cost about $80,000.
The short range ones, found at vehicle corners and on the rear, are now around $200 - $1500. The long range car-top unit is still expensive, around $10,000, and that's dropping. The whole sensor suite on a Waymo is probably under US$20,000 right now. If that.
The future direction for automotive seems to be more car-top sensors with narrower fields of view that add up to a full circle. Only the one that looks ahead needs 300m range.
Article has a paragraph about how lidar is rapidly getting cheaper, so the camera comparison is only temporary
Dolgov also pushed back on the one knock that camera-only fans always reach for: cost. Waymo is on its sixth generation of hardware, and he said each generation has drastically cut the price. Betting against sensors like lidar on today’s prices, he warned, means betting on “a number that has a fairly short shelf life.”
"Ultimately, people will vote with their wallet"
Seems like a bad way to legislate this, since everyone else on the roads pays the price for cut rate self driving
> Ultimately, people will vote with their wallet.
Nope, you need a regulatory approval, insurance etc. Self driving cars need to be 10x the average driver. So the trillion dollar question is when Lidar will meet all requirements and when camera-only.
1) Both in 2028
2) lidar in 2030 and cameras in 2035
3) Gov will require lidar for full self driving
Personally I would rather vote to ban them if they aren't safe enough. If car makers want to cheap out at the expense of public safety they should go to prison for the people they kill.
You’re like 2-3 orders of magnitude off on your cost there.
Which is why we need strong regulation. People shouldn’t be able to vote with their wallet when other peoples lives are at stake.
> But are $200K worth of extra sensors actually worth it
Of course not!
So, you may be surprised to find that Waymo's sensor package doesn't cost $200k. Whoops!
> I wonder if you could prevent glare by placing an LCD-like device in front of the lens dynamically blocking out the sun.
Elon has already lied, on an investor call, about how such things are not necessary - he referenced how "Tesla cameras can count photons", as a reason why they were not affected by fog, rain, or snow.
Except, no they can't, and they never will be able to. Not the cameras in any Tesla now (and even then). It's a completely different type of camera (and it would do effectively nothing useful from a visualization perspective, let alone functional camera for human viewing). And even if they changed the camera, or added some, well, you can't count photos if you're not in an enclosed space (you know, like the open road).
This is where Elon gets really dangerous, in my mind, he knows enough to be dangerous, he knows what to say that convinces the layperson, or investor, passes the initial sniff test only to dramatically fall over when you investigate.