Waymo CEO explains why Tesla’s camera-only self-driving falls short
electrek.co
electrek.co
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 can get 3D depth estimation from a single 2D image but I don’t know how good it is for automotive.
I think the natural solution is a 360 lidar as a source of truth, followed by optical for everything else
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.
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.
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...
They also did not make lithium batteries affordable. Panasonic and CATL did that. Tesla was simply the first to exploit that bew affordability.
Musk being who he is, however, I'm not holding my breath for that outcome.
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?
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.
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.
I’m clearly not a typical driver. I love keeping the camera views all on screen for situational awareness. I don’t use cruise control and rarely use GPS navigation, even back when I used to drive cross country in previous cars. I guess I just got used to memorizing entire routes in my long distance biking days. Google Maps Street View really helps me understand exactly where I’ll be going.
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.
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]
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.
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.
https://www.reddit.com/r/TeslaFSD/comments/1u5eetx/hw4_fsd_a...
Tesla is making a huge bet on robotaxi, it’s quite possible it will come to fruition.
I can't purchase a Waymo at any price.
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"!
Recent updates broke parking in my specific parking spot, but I'm confident this regression will be fixed.
but you said "FSD" which stands for FULL Self Driving. You didn't say FSDS, or FSD Beta, or "Tesla's Level 2" self driving.
It is NOT a FULL self driving solution yet. They are lying and you are repeating their lies.
folks you can't make this stuff up
Obviously, our disagreement is with Tesla branding. I don't care what they call it (I liked Autopilot), it's incredibly useful.
I wish Elon wasn't so stubborn about admitting he was wrong.
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.
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.
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.
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.
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.”
An Ojai costs less than a mid tier luxury vehicle. Is that not cheap enough?
And the price is only coming down.
Seems like a bad way to legislate this, since everyone else on the roads pays the price for cut rate self driving
* with prerequisite that "safer than the avg human" is the minimum bar for any system.
But much more importantly, there is all sorts of post-processing happening that is so weird and complex and really a little decoupled from the raw visual information that it's futile to try and compare that to any digital operation you're thinking of in terms of cameras: If you were to look at a still captured from your eyes, you'd see an upside-down image that's only sharp in the very center of the image, blurry and desaturated pretty much everywhere else, grossly so in the corners, with several big black or blurry spots all over.
From that perspective, eyes are awful really. Yet still, your brain manages to turn the stereoscopic vision into a spatial representation, calculate the movement and keep track of other objects relative to you, fill in areas you can't actually see, detect things moving in your peripherals, highlight things of importance, and so much more.
edit: I also want to point out that the fanboys conflate documented evidence in wikipedia of things EM has said with vitriol. It's like when people get banned on social media for hate speech for quoting something the POTUS said.
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
Why? So long as they're insured and their owners can be held criminally liable for injuries just as a human driver would be, why does the bar have to be higher?
A huge portion of the risk of human drivers is due to confounding human factors like negligence, distraction, medical issues, fatigue, etc. Responsible drivers can take most of these factors into their own hands -- not driving while sleepy, not driving drunk, not using their cell phone. These people don't want a vehicle
A better comparison would be comparing the capability of a self-driving car to the skill of a human driver driving under ideal circumstances of health/responsibility/etc.
Even today, you cannot fully avoid this risk: when you get into a vehicle driven by a stranger, you don't know a priori whether the driver is excellent or average. Ideally customer feedback weeds out poor drivers and reduces accident rates, but I'm not privy to data that would prove or disprove whether that actually happens - especially because new drivers join the pool every day.
But besides that, the people who are in the socioeconomic demographic to buy an expensive self driving car overlap with the demographic that have lower driving risk. For them, the self-driving car is more dangerous than driving themselves.
For a personal standard, it is reasonable that an individual purchaser may use their own risk profile as that standard... as many of the risks involved are not random but can be controlled through choice or are inherently personal. For example, about half of all traffic deaths are due to drowsy or drunk driving.
Accident risk for any given driver is highly distributed. For someone making a marginal choice to purchase a car, the choice they are faced with is choosing between the safety profile of a self-driving car and their own personal risk profile.
2) Politics - loss of jobs and urge to ban them - there needs to be a counterargument
3) Government does mandate a lot of safety features. Just insure it wont fly as this is number one risk for everybody.
The bar should be safer than a human and as affordable as possible.
Of course not!
So, you may be surprised to find that Waymo's sensor package doesn't cost $200k. Whoops!
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.