Hacker News Reader

Comment by juujian | Hacker News Reader

Incredible that this is too expense for a company like Tesla.

What is the HN opinion on Tesla skipping lidar? Having spent some time with computer vision in university I think it's insane to skip it - sure stereo reconstruction is powerful but lighting conditions have such an impact on results that having some robust depth data feels like a no-brainer and skipping it feels like malignant neglect.

As someone who's done a lot of computer vision, it is insane to skip it. And it's sad because what everyone missed from that viral Mark Rober video [0] was not the Looney Toons wall hit but the fucking kid in the smoke. Add all the cameras and AI you want, you ain't changing the laws of physics: visible light doesn't penetrate smoke. But radar does. Every (traditional) engineer knows that safe systems have redundancy. That safe systems have redundancy through differing modalities. Use cameras, but also use radar, lidar, and even millimeter wave. Using just cameras isn't just tying one hand behind your back, it's shooting yourself in the kneecap afterwards

[0] https://www.youtube.com/watch?v=IQJL3htsDyQ

The argument is that humans manage the task without lidar, and automation doesn't have to be perfect it just has to be better than humans, to be a net positive. It seems to me, you might as well use lidar if it's cheap enough, but the argument that computer systems can outcompete human drivers, without using lidar, is at least reasonable, although not yet proven.

Extending this line of thought I wonder why tesla didn't make cars on two legs and insisted on using wheels?

(Just wanted to make sure - this is not a stab at you, I'm well aware that the original argument is from tesla)

We could extend the argument more. Why build a self driving vehicle at all? Build a humanoid robot to drive the car for you! The argument that computer systems can outcompete human drivers, without using lidar, is at least reasonable, although not yet proven

(I didn't just want to just make sure - this is a stab)

It's a dumb argument on multiple accounts. While it's a routine argument in software engineering it's an argument that often will get people fired and sued in testimonial.

First off, humans don't do it "just by vision". Sure, we don't have lidar but we have hearing, we have touch, we have tons of experience. We can create world models for Christ's sake and that means modeling physics. I'm sure you've seen papers that claim world models but I'm a ML researcher who also has a physics degree and I'm not afraid to tell you that's bullshit. It's as honest as Altman calling GPT PhD level intelligence. A PhD has very little to do with the ability to recall information.

Second off, it doesn't matter much how humans do it. It matters how the car can. Why limit yourself. There's tons of cars with radar and lidar. They're not more expensive and they can see an object in fog or poor light conditions. It can do something humans can't do! Why in the world would you decide not to do that. You can make an argument about price but that argument changes when that thing becomes cheaper. When that happens you're now just someone adding danger for no reason. You can't argue that only cameras will be safer. It categorically isn't. The physics is in your way.

But that is the argument made when Tesla first said they were going to use only cameras. Because everyone knew lidar would come down with scale and that's why many other manufacturers went in that direction. Which is mutually beneficial, so Tesla would benefit from joining.

  > can outcompete human drivers

Third, be careful with those claims. I'm more willing to believe 3rd party reports like from NHSTA than directly from Tesla [0]

[0] https://www.forbes.com/sites/stevebanker/2025/02/11/tesla-ag...

Human suck at driving. Almost 2 million people are killed by vehicles on roads every year. While not all of this is due to limitations of vision, it's certainly a contributing factor. Tesla tries to handwave this saying that their systems don't text and drive, get distracted, etc..., but this strikes me as more "AI will become superintelligent real soon now and render humans obsolete" when actual experience with AI is that it is dumb, forgetful, and prone to hallucinations.

> visible light doesn't penetrate smoke. But radar does.

This is the key insight that frustrates me to no end about the whole thing. We have sensors that are better than human eyes, but we should limit ourselves to that because what? I don't use a calculator because it's slightly better at math, I use a calculator because it's fucking awesome at multiplying numbers in a way that my human brain can't remotely compete with. I want to be able to see where Elon is coming from but lately I can't.

Is Tesla getting into legal mess if they need to add sensors to make self driving work when they already sold that feature to car owners? Would that imply that they need to retrofit already sold cars with upgraded sensor packages?

Yes, and this is already turning out to be a problem for them. They've acknowledged that HW3 is not sufficient, and will be on the hook for those who bought the FSD package with those cars.

That isn't the end of the world, but it'd turn into a much bigger problem if they also had to add additional sensors and body modifications to support those sensors.

Solution: just never implement FSD.

For a long time, I honestly thought their solution might be something like this. Either that or they could ship more advanced hardware for 3-5 years before updating the software, so that most vehicles would have the new hardware.

My understanding is that they went the opposite direction - their cars used to have lidar, but don’t anymore.

Worse, they turned them off for the older vehicles with a software update.

They never had lidar. They had a very low resolution radar that was used for AP, and some pretty terrible ultrasonic sensors with massive blind spots.

IIRC Musk specifically said that the cars had sufficient hardware for FSD mode and advertised them as such. Tesla would have to retrofit the LIDAR sensor or pay money back to their customers if they rolled out FSD with LIDAR.

> What is the HN opinion on Tesla skipping lidar?

Short-sighted and egotistical.

There likely have been deaths and injuries that would have been prevented by lidar, and there will likely be more in the future.

It strikes me as foolhardy that the U.S. allows self-driving vehicles to use the public roads without having to pass safety tests. I'd blame the lack of government control of public spaces as well as Tesla's engineering.

Musk is allowed to test in production with 1 ton metal machines racing at 100 km/h without it entailing legal responsibility. Amazing the influence that a few million believers on social media can buy you these days

interesting claim i read in another thread a couple weeks ago:

>Tesla Vision is, currently, legally below minimum human vision requirements and has historically been sold despite being nearly legally blind.

https://news.ycombinator.com/item?id=43605034

My opinion is that skipping lidar is nonsense. Check Mark Rober's video, published one month ago, "Can You Fool A Self Driving Car?" where he compares a Tesla with other Lidar-equipped cars: https://www.youtube.com/watch?v=IQJL3htsDyQ

> the HN opinion

I'm not sure why you'd think HN has a monolithic opinion, this is a site with myriad different users.

Maybe they're more asking for the whole breadth of opinions available from the HN community?

True, but you will find that some subjects align people more closely than most others.

There will be time in the very near future (read five years time) people will not buy vehicle (car, bike, etc) without lidar as the price become insignificant as car reverse camera, and it become commonplace.

Personally now I'll not buy any vehicle without assisted camera parking and apparently many people will agree with this important feature including Marques Brownlee [1].

[1] Reviewing my First Car: Toyota Camry Hybrid! [video]:

https://youtu.be/Az6nemkRB1Y

A similar situation to Jobs reciting research on how optimal the one button mouse is. A thought bubble.

Why, we will perhaps never know. But likely they were early and it was deemed too expensive back then, or didn't find a supplier they could work with. Now there's too much prestige in it and they can never back down which would be admitting to a mistake.

It would be one thing if it was a one time event but then they repeated that playbook with the lack of a rain sensor.

I think it was a valid decision that turned out to be incorrect and is staying put as a result of stubbornness. People really like criticizing decisions in hindsight especially here where the armchair engineer with the benefit of hindsight is too common.

People have been criticizing this decision from the get go. It may have upgraded from engineers to the general public but let's be honest, the latter doesn't matter in a topic like this anyways

They should have had the benefit of hindsight as well thanks to testing.

I can only speak for myself, but I work on this stuff in this industry: Tesla’s choice is asinine at this point. It’s one thing to claim cameras only and find that won’t work and pivot, but they are so dug in that they can’t admit they were wrong and won’t do so. So it’s asinine now.

>It’s one thing to claim cameras only and find that won’t work and pivot, but they are so dug in that they can’t admit they were wrong and won’t do so. So it’s asinine now.

Didn't they start with lidar or radar or similar and then go back to only using vision based technologies?

What do yes call a person that has only visual sensory? Disabled.

Humans have sight, touch, taste, sound, smell, and vascular sensory. Only a portion of systems used in self drive automation.

Humans also have a vestibular system[1] and proprioception[2], which I think are very important in making judgments about safe following distance, car handling, and road conditions, in particular in adverse conditions like strong side winds and slippery or icy roads, which may not always be visually obvious.

While some of this can be handled by an IMU, I think humans still have a strong advantage in fusing their various sensory inputs, thanks to millions of years of evolution.

[1] https://en.wikipedia.org/wiki/Vestibular_system

[2] https://en.wikipedia.org/wiki/Proprioception

Well I haven't used taste nor smell in my driving yet. Touch only as far as vibration and steering wheel torque (both not difficult to sense with electronics).

That narrows it down a bit.

Never once smelled or tasted something funny when driving and realize you’d better pull over… like… right now? I have.

It was a valid experiment and pushed computer vision but it clearly failed a long time ago. The fact that Teslas are not only still sold without lidar but “autopilot” is pushed as safe is disgusting.

People will die (and have already died) horrifically because of this decision. It’s morally bankrupt.

I assume the same would apply to any car not using LIDAR? Or just Tesla because they decided on a different tech stack?

Radar technology offers a range of applications, including the ability to detect objects around corners, behind obstacles such as brick walls, and even penetrate human bodies at specific frequencies. However, when multiple sensors yield similar results, it becomes challenging and costly to discriminate which results are valid.

Operating radar at a specific frequency, such as 2.45 GHz (a microwave frequency often used due to its affordability), can be ineffective in environments rich in water droplets (e.g., rain), as these can dominate the radar signals. Higher frequencies enable the detection of smaller water droplets, but switching between frequencies can be expensive. Additionally, varying the radar's detection range to identify objects of different sizes complicates the calculations, involving factors such as minimum and maximum range, power, and time on target.

Cameras typically detect non-moving objects by comparing successive images. In contrast, radar can identify both stationary and moving objects and determine their direction relative to the sensor by emitting a frequency and analyzing the reflected pulses. Lidar, on the other hand, uses light to measure the distance to objects in its path, employing a photoreceptor to capture the reflected light.

That'll likely be why LIDAR is used rather than RADAR.

I was responding to Tesla skipping Lidar but apparently it’s a sore topic. As pointed out by someone else Tesla used to use ultrasonic radar sensors but stopped.

“ Until this month's change, all Model S and Model X EVs intended for North America were equipped with radar sensors but the company has been building new Model 3 and Model Y vehicles without any front radar sensors since May 2021. That's when Tesla announced a change for those models away from radar to Tesla Vision” 0

I don’t think many people understand how these systems work but we’re not on a radar or engineering forum.

0. https://www.caranddriver.com/news/a39250157/tesla-no-radar-s...

This isn't too expensive for Tesla, it's just nowhere near the level needed for an AV. Automotive lidars are 10-20 scans/second, rated for dust/rain/etc, and need a range of at least 50 meters, but 100-200 is more ideal. Not a fan of Tesla's approach, but I wanted to clarify that it's not like they can just use a lidar like this and call it a day. The specs are completely different and that really drives up cost!

Maybe something like this: https://linuxgizmos.com/sony-introduces-as-dt1-described-as-...

> With a horizontal field of view of 30° or more and 576 ranging points (24 x 24), the sensor supports a frame rate of 30 fps, with a reduced 15 fps mode for maximum distance operation.

The requirements for electronics in a car are pretty extreme (temp, durability), not that I disagree, but it's not apples to oranges.

> requirements for electronics in a car are pretty extreme

+ the salaries of everyone working on that stuff, not just assembly but also writing the code to support it

Not that I disagree, either: at the volumes that a modest car company puts out, I'd assume it's easily worth the, say, 3% cost premium on the car's total price to have something that can actually see things you don't see and thus makes a safer system. It might even reduce costs by having lower requirements for the vision hardware and software, but that's not something I can know. There's a lot of unknowns here that I think mean we can't really do a good comparison indeed

That hasn’t stopped Tesla before. They have a track record of treating automotive-grade quality standards as optional when doing electronics sourcing[1].

As the article notes, Tesla conveniently “fixed” their thermals and durability issue that caused by inventing a feature called cabin overheat protection and marketing it as for people/animals overheating and not for the non-automotive-spec electronics in the cabin.

If you can’t bring auto quality electronics to the car, just change the car so it avoids standard auto thermal conditions ¯\_(ツ)_/¯

https://www.thedrive.com/tech/27989/teslas-screen-saga-shows...

Don’t they constantly get tested as the safest car in the world? I saw it years ago in some American news and the first google result is from New Zealand last year https://www.drivencarguide.co.nz/news/tesla-model-y-is-the-s...

They wanted to sell “self driving ready” packages 10 years ago, when LiDAR actually was expensive. So at the time, they had to make big deal about LiDAR being unnecessary.

But now it has come down in price reportedly by more than a factor of ten so at some point a logical person would revisit that decision

it's not only about logic; his ego is now involved in it which virtually guarantees it will never be revisited.

Would this not also be the case had Tesla embraced the tech and installed thousands into their cars?

The lidars used on self-driving vehicles are far more capable and far more expensive.

Not by that much current generation hardware for cars is $500-700. And some of the oem expect to bring it price down below $200 with the next generation equipment. Now that BYD put self driving in almost every car it will supercharge adoption and lidar prices might drop even a lot faster with economies of scale.

My (tenuous) understanding is that the challenge with lidar isn't necessarily the cost of the sensor(s) but the bandwidth and compute required to meaningfully process the point cloud the sensors produce, at a rate/latency acceptable for driving. So the sensors themselves can be a few hundred bucks but what other parts of the system also need to be more expensive?

That seems very unlikely to me. Automotive applications are already doing things like depth reconstruction based on multiple camera angles and ML inference in real time. Why should processing a depth point cloud be significantly more difficult than those things?

The basis for my understanding is a convo with a Google engineer who was working on self-driving stuff around 10-15 years ago -- not sure exactly when, and things have probably changed since then.

At the time they used just a single roof-mounted lidar unit. I remember him saying the one they were using produced point cloud data on the order of Tbps, and they needed custom hardware to process it. So I guess the point cloud data isn't necessarily harder to process than video, but if the sensor's angular resolution and sample rate are high enough, it's just the volume of data that makes it challenging.

Maybe at that time 10-15 years later we have graphic cards doing actual ray tracing lidar computing is way less complex. Anyway the $200 I is for the whole system not just sensors so that would include signal processing

Makes sense. Maybe doing self driving well just requires a ridiculously high bandwidth regardless of data source. Related, the human visual system consumes a surprisingly large quantity of resources from metabolic to brain real estate.

The whole point of lidar is to massively increase the amount of ranged data you have to work with.

This doesn’t seem to stop Teslas competition in self-driving cars from implementing it; and succeeding far more in safety and functionality while doing so.

What is the cost of a human life worth?

edit: seriously, a $4,000 sensor and an extra, say, $3,000 for an upgraded computer module so your car can drive itself is just too much too afford?

Valuation of a statistical life is $5-10M, depending on who you ask[0].

So it’s too much to afford, or at least not singularly justifiable, unless more than 1 out of every 2000 cars kills someone in a way that would be prevented by LIDAR.

0: https://www.sciencedirect.com/science/article/pii/S109830152...

At this point having "something" would probably even beat having nothing.

I guess it's simply a big numbers thing. If you sell lots of cars, shaving a couple of hundred dollars of each car adds up.

Karpathy addressed this question at the time:

https://news.ycombinator.com/item?id=33397093

Of course he was working for Tesla back then. His opinions might be different today given that Elon is no longer signing his paycheck.

What a weird argument by Karpathy. He has a degree in physics. How does this dude not know that radar can see things not possible through camera vision. That argument there doesn't make any sense. That there's supply chains and things break and this makes it unsafe? Well that's true for every single bolt, every nut, ever little thing. I understand a drive to simplicity but you can't just throw fancy words in there like entropy while ignoring the literal physics that says camera + radar is less entropy than camera no radar. There is literally more (unique!) information available to you!

His opinions aren’t much different in interviews I’ve heard since, although of course that doesn’t mean he’s completely unbiased now.

Money better spent on marketing. Like that song about "him having a plan".

After all car sales don't drive the stock market. Public opinion does.

I’ll bet a lot of Tesla investors are wishing neither of those applied these days.

Do we really need LiDAR in a Tesla? I own a Chevy Trax and it has LKAS and ADAS. Not even using LiDAR just sensor fusion with camera and radar. It’s a cheap car too. It’s car assisted driving.

I have driven a Tesla once but not with the added feature.

This isn't really something you'd ship in a car though. It's cool that we have such a rich ecosystem of devices that this can be made "off-the-shelf" - but for production use in a car? Not really practical.

Reply on news.ycombinator.com