Mark Rober just set up one of the most interesting self-driving tests of 2025, and he did it by imitating Looney Tunes. The former NASA engineer and current YouTube mad scientist recreated the classic gag where Wile E. Coyote paints a tunnel onto a wall to fool the Road Runner.
Only this time, the test subject wasn’t a cartoon bird… it was a self-driving Tesla Model Y.
The result? A full-speed, 40 MPH impact straight into the wall. Watch the video and tell us what you think!
Show someone footage of 9/11 and they‘ll think of 2001. Show someone footage of a burning or crashing Tesla 20 years from now and they‘ll think of 2025.
Who was the idiot that removed LiDar to cut costs?
/s
Elon removed the radar. Tesla cars never had lidar. What an idiot Musk.
He did say lidar was “useless” though.
He’s had humans don’t use LiDAR so his cars shouldn’t have to. Of course humans have a brain, and he’s cars don’t, but you can’t tell him anything.
Human drivers also make automobiles one of the most dangerous ways to travel.
Human drivers also make automobiles one of the most dangerous ways to travel.
Same goes for Teslas. Just that the human that’s making them a dangerous way to travel is Elon Musk.
I think it’s also reasonable to say a human dying because of their own actions is different than a human dying because a big corp cut costs on safety features in an entirely autonomous car where the human has no ability to stop what’s happening. (You can control them in current teslas, but they’re working on cars without human controls as well)
Weymo have them beat then. They already have cars that essentially act like taxis and the only control the passenger has is an emergency stop button. But they don’t control the speed they don’t control the steering or anything.
But those systems absolutely do have lidar because in order for them to be widely accepted they don’t need to be a bit safer than humans they need to be a lot safer than humans.
He also said the government doesn’t use sql.
Bahaha, what kind of a bizarre statement is that?
Was he trying to imply the government only uses spreadsheets and nosql DBs?
Or did he think it was necessary to point out that your average government employee isn’t writing their own SQL to grab data they need?
Even then, that’s not really correct. People grab data through sql queries all the time. Mostly because all the front ends are trash.
Someone said something he didn’t like so he blurted out the first ignorant thing that he thought of, as usual.
Tesla had camera+radar+sonar, and that wasn’t their own tech - they used mobileye EyeQ back then. When they switched to in house tech they gradually ditched the radar and sonar which made no sense to me. But at the time I saw their lead say in an interview that this is superior and I believed. not anymore.
they said doing so cut costs but obviously lidar/radar/sonar only gets cheaper over time, let alone the extra r&d costs because a vision only system is much more difficult to develop.
It was removed because it was giving false positives. They should have upgraded it with lidar but decided to just remove it.
It was removed because of supply chain issues.
That would make sense if they put it back after, but they didn’t
Yeah, it might drive straight into a wall but at least it isn’t returning false positives!
They are so expensive too! /s
Who would have known electronics gets cheaper all the time?? /j
Except GPUs 😥
deleted by creator
I’m so glad I wasn’t the only person who immediately thought “This is some Wile E. Coyote shit.”
I mean, it is also referenced in the article and even in the summary from OP.
And extensively in the video too.
The actual wall is way more convincing though.
still, this should be something the car ought to take into account. What if there’s a glass in the way?
That might have been an even „simpler“ test.
Yes, but Styrofoam probably damages the car less than shards of glass.
Glass is far more likely to cause injuries to the driver or the people around the set, just from being heavier material than styrofoam.
Yes, I think a human driver who isn’t half asleep would notice that something is weird, and would at least slow down.
Glass would be very interesting, might actually confuse lidar also.
As much as i want to hate on tesla, seeing this, it hardly seems like a fair test.
From the perspective of the car, it’s almost perfectly lined up with the background. it’s a very realistic painting, and any AI that is trained on image data would obviously struggle with this. AI doesn’t have that human component that allows us to infer information based on context. We can see the boarders and know that they dont fit. They shouldn’t be there, so even if the painting is perfectly lines up and looks photo realistic, we can know something is up because its got edges and a frame holding it up.
This test, in the context of the title of this article, relies on a fairly dumb pretense that:
- Computers think like humans
- This is a realistic situation that a human driver would find themselves in (or that realistic paintings of very specific roads exist in nature)
- There is no chance this could be trained out of them. (If it mattered enough to do so)
This doesnt just affect teslas. This affects any car that uses AI assistance for driving.
Having said all that… fuck elon musk and fuck his stupid cars.
I am fairly dumb. Like, I am both dumb and I am fair-handed.
But, I am not pretentious!
So, let’s talk about your points and the title. You said I had fairly dumb pretenses, let’s talk through those.
- The title of the article… there is no obvious reason to think that I think computers think like humans, certainly not from that headline. Why do you think that?
- There are absolutely realistic situations exactly like this, not a pretense. Don’t think Loony Tunes. Think 18 wheeler with a realistic photo of a highway depicted on the side, or a billboard with the same. The academic article where 3 PhD holding engineering types discuss the issue at length, which is linked in my article. This is accepted by peer-reviewed science and has been for years.
- Yes, I agree. That’s not a pretense, that’s just… a factually correct observation. You can’t train an AI to avoid optical illusions if its only sensor input is optical. That’s why the Tesla choice to skip LiDAR and remove radar is a terminal case of the stupids. They’ve invested in a dead-end sensor suite, as evidenced by their earning the title of Most Lethal Car Brand on the Road.
This does just impact Teslas, because they do not use LiDAR. To my knowledge, they are the only popular ADAS in the American market that would be fooled by a test like this.
Near as I can tell, you’re basically wrong point by point here.
Excuse me.
-
Did you write the article? I genuinely wasn’t aiming my comment at you. It was merely commentary on the context that is inferred by the title. I just watched a clip of the car hitting the board. I didn’t read the article, so i specified that i was referring to the article title. Not the author, not the article itself. Because it’s the title that i was commenting on.
-
That wasn’t an 18 wheeler, it was a ground level board with a photorealistic picture that matched the background it was set up against. It wasnt a mural on a wall, or some other illusion with completely different properties. So no, i think this extremely specific set up for this test is unrealistic and is not comparable to actual scientific research, which i dont dispute. I dont dispute the fact that the lack of LiDAR is why teslas have this issue and that an autonomous driving system with only one type of sensor is a bad one. Again. I said i hate elon and tesla. Always have.
All i was saying is that this test, which is designed in a very specific way and produces a very specific result, is pointless. Its like me getting a bucket with a hole in and hypothesising that if i pour in waterz it will leak out of the hole, and then proving that and saying look! A bucket with a hole in leaks water…
Y’all excused, don’t sweat it! I sure did write the article you did not read. No worries, reading bores me sometimes, too.
Your take is one of the sillier opinions that I’ve come across in a minute. I won’t waste any more time explaining it to you than that. The test does not strike informed individuals as pointless.
-
I agree that this just isn’t a realistic problem, and that there are way more problems with Tesla’s that are much more realistic.
Tell that to the guy who lost his head when his Tesla thought a reflective semi truck was the sky
Well that seems like a realistic problem. Not picture of tunnel
This doesnt just affect teslas. This affects any car that uses AI assistance for driving.
Except for, you know… cars that don’t solely rely on optical input and have LiDAR for example
Fair point. But it doesn’t address the other things i said, really.
But i suppose,based on already getting downvoted, that I’ve got a bad take, either that or people who are downvoting me dont understand i can hate tesla and elon, think their cars are shit and still see that tests like this can be nuanced. The attitude that paints with a broad brush is the type of attitude that got trump elected…
based on already getting downvoted
In this case, yes, but in general, downvotes just mean your take is unpopular. The downvotes could be from people who don’t like Tesla and see any defense of Tesla as worthy of downvotes.
So good on you for making the point that you believe in. It’s good to try to understand why something you wrote was downvoted instead of just knee-jerk assuming that it’s because it’s a “bad take.”
No, it’s just a bad take. Every other manufacturer of self driving vehicles (even partial self driving, like automatic braking) uses LiDAR because it solves a whole host of problems like this. Only Tesla doesn’t, because Elon thinks he’s a big brain genius. There have been plenty of real world accidents with less cartoonish circumstances involving Teslas that also would have been avoided if they just had LiDAR sensors. Mark just chose an especially flashy way to illustrate the problem. Sometimes flashy is the best way to get a point across.
I agree the wall is convincing and that it’s not surprising that the Tesla didn’t detect it, but I think where your comment rubs the wrong way is that you seen to be letting Tesla off the hook for making a choice to use the wrong technology.
I think you and the article/video agree on the point that any car based only on images will struggle with this but the conclusion you drew is that it’s an unfair test while the conclusion should be that NO car should rely only on images.
Is this situation likely to happen in the real world? No. But that doesn’t make the test unfair to Tesla. This was an intentional choice they made and it’s absolutely fair to call them on dangers of that choice.
That’s fair.
I didn’t intend to give tesla a pass. I hoped that qualifying what i said with a “fuck tesla and fuck elon” would show that.
But i didn’t think about it that way.
In my defense my point was more about saying “what did you expect” the car to do in a test designed to show how a system that is not designed to perform a specific function cant perform that specific function.
We know that self driving is bullshit, especially the tesla brand of it. So what is Mark’s test and video really doing?
But on reflection, i guess there are still a lot of people out there that dont know this stuff, so at the very least, a popular channel like his will go a longway to raising awareness of this sort of flaw.
A camera will show it as being more convincing than it is. It would be way more obvious in real life when seen with two eyes. These kinds of murals are only convincing from one specific point.
That’s true, but it’s still way more understandable that a car without lidar would be fooled by it. And there is no way you would ever come into such a situation, whereas the image in the thumbnail, could actually happen. That’s why it’s so misleading, can people not see that?
I absolutely hate Elon Musk and support boycott of Tesla and Starlink, but this is a bit too misleading even with that in mind.So, your comment got me thinking… surely, in a big country like the US of A, this mural must actually exist already, right?
Of course it does. It is an art piece in Columbia, S.C: https://img.atlasobscura.com/90srIbBi-XX-H9u6i_RykKIinRXlpclCHtk-QPSHixk/rt:fit/w:1200/q:80/sm:1/scp:1/ar:1/aHR0cHM6Ly9hdGxh/cy1kZXYuczMuYW1h/em9uYXdzLmNvbS91/cGxvYWRzL3BsYWNl/X2ltYWdlcy85ZTUw/M2ZkZDAxZjVhN2Rm/NmVfOTIyNjQ4NjQ0/OF80YWVhNzFkZjY0/X3ouanBn.webp
A full article about it: https://www.atlasobscura.com/places/tunnelvision
How would Tesla FSD react to Tunnelvision, I wonder? How would Tesla FSD react to an overturned semi truck with a realistic depiction of a highway on it? JK, Tesla FSD crashes directly into overturned semis even without the image depiction issue.
I don’t think the test is misleading. It’s puffed up for entertainment purposes, but in being puffed up, it draws attention to an important drawback of optical-only self-driving cars, which is otherwise a difficult and arcane topic to draw everyday people’s attention to.
Good find, I must say I’m surprised that’s legal, but it’s probably more obvious in reality, and it has the sun which is probably also pretty obvious to a human.
But it might fool the Tesla?Regarding the semi video: WTF?
But I’ve said for years that Tesla cars aren’t safe for roads. And that’s not just the FSD, they are inherently unsafe in many really really stupid ways.
Blinker buttons on the steering wheel. Hidden emergency door handles, emergency breaking for no reason. Distracting screen interface. In Denmark 30% of Tesla 3 fail their first 4 year safety check.
There have been stats publicized that claim they aren’t worse than other cars, when in fact “other cars” were an average of 10 year older. So the newer cars obviously ought to be safer because they should be in better conditions.
I remember elon saying something along the lines of his camera system being just as good and they thusly don’t need to employ things like LIDAR.
And that’s what you get for cheaping out on tech and going with cameras over lidar. Not only that, but Tesla removed all the radar technology that literally every car uses for collision detection about a year ago.
The radar module on my truck costs $70.
The richest man on earth doesn’t think the lives of your vehicle’s passengers are worth $17.50 a pop.
And that’s to a knuckledragger like me, buying a single radar unit online. I’m sure the manufacturer gets insane quantity discounts.
I seem to recall that fElon prevented the self driving team from utilizing LIDAR for any part of the system, instead demanding that everything run off of optical input. Does anyone else remember the same?
What’s cool is, Tesla’s used to have radar sensors at least, but Elon removed them from production to save money. Even if you have a car from back then, the software no longer uses them and they’ll just physically unplug it them the next time you have the car serviced, as it’s just a drain on the battery at this point 🙃
meanwhile our subaru has lidar for adaptive cruise control and emergency braking
I didn’t realize EyeSight had different versions, on the Solterra it looks like it is indeed LIDAR.
My Crosstrek has the older dual camera setup for depth perception, it would not be fooled by a picture of a road on a wall… I’m surprised the Teslas are.
they’ll just physically unplug them the next time you have the car serviced
So, (looks at watch), in an hour?
So did they unplug it
Yikes.
Came here to actually write this. Everyone remembers that. He made Tesler the hated shit it is today.
As a space nut I seriously hope that he never gets a chance to do anything similar with SpaceX. Thankfully he’s mostly been kept away from important things thus far.
Don’t get me wrong, I know SpaceX’s closet is overflowing with skeletons. But since Congress has been so kind as to continuously cut NASA’s budget for the last few decades, I have to rely on SpaceX and other private companies to keep our space endeavors going.
I’m (was) huge SpaceX nerd, but last year or so I’m less and less. He always was dumb narcissist asshole, but now I can’t take it anymore. Also the idea that we’ve fucked up this planet and need to move somewhere else, by doing thousands of launches finishing this planet always made me sick. If someone would take him out, I probably would come back to liking the company.
To the famously already fucked planet Mars, no less.
Yes, I recall at the time experts saying it was a terrible mistake and Elon saying Machine learning will bridge the gap.
The real reason was to increase margins.
Tesla never had LIDAR. That’s the little spinny thing you see on Waymo cars. They had RADAR, and yes it was removed in 2021 due to supply shortages and just…never reinstalled.
It was removed due to supply chain, but Musk did seem to legitimately think optical only was better.
Yes. He took too much inspiration from Stanford University’s “Stanley” winning the DARPA Grand Challenge in 2005. This was an early completion to build viable autonomous vehicles. Most of them looked like tanks covered in radar dishes but Stanford would up taking home the gold with just an SUV with cameras on it.
It was an impressive achievement in computer vision, and the LiDAR-encrusted vehicles wound up looking like over-complex dinosaurs. There’s a great documentary about it narrated by John Lithgow (who, throughout it, pronounces the word robot as “ro-butt”). Elon watched it, made up his mind, and like a moron, hasn’t changed it in 20 years. I’m almost Musk’s age so I know how the years speed up as we go on. He probably thinks about the Stanford win as something that happened relatively recently. Especially with his mind on - ahem - other things, he’s not keeping up with recent developments out in the real world.
Rober just made Musk look like the absolute tool he is. And I’m a little worried that we may see people out there staging real world versions of this somehow with actual dangerous obstacles, not a cartoonish foam wall.
I did low-key get the squiggles before writing the article. I thought, from an ethical hacking disclosure-type perspective, that this info might cause folks to… well, ya know, paint tunnels on walls.
Then I looked, the cat was already out of the bag, the video had something like 5 million views on it in the 4 hours it took me to draft the article. So I shared it, but I definitely did have that thought cross my mind. I am also a little worried on that score.
Iirc they were using a combination of lidar and radar, but Elmo wanted to cut costs.
Ah okay. I was genuinely curious if I was remembering correctly because I definitely know it’s been awhile since I’d read anything on the subject.
Funny thing is, the price of lidar is dropping like a stone; they are projected to be sub-$200 per unit soon. The technical consensus seems to be settling in on 2 or 3 lidars per car plus optical sensors, and Chinese EV brands are starting to provide self driving in baseline models, with lidars as part of the standard package.
Cameras and radar, I believe. Never lidar.
Did he want to cut costs or did he want a network of cameras at his control all over the world?
Yes.
I remember there being claims from him or his team about lidar being a dead end that would not scale as well as computer vision.
I believe he claimed that since humans use their vision to drive that computer vision was more than enough.
I don’t know about you, but I also rely on sounds & feel when I drive. I also know that the human eye has evolved to detect motion, filter out extraneous information, and send just the important bits to the brain so that it doesn’t get overloaded with everything the eye sees. Computer vision is the exact opposite from that, having to process every bit of every image the camera sees.
I also know of many times my vision fails. Driving into a sunrise for example
I don’t know about you, but I also rely on sounds & feel when I drive.
Of course. When I feel myself driving into a wall, I stop immediately.
You must connect with the road, every km or so stop and hug the asphalt.
since humans use their vision to drive that computer vision was more than enough
Surprised he didn’t swap out the wheels with legs while he was at it
o p t i m u s
Yep! That’s what I’m thinking of. It was Elmo. The real engineers objected.
Was just thinking this
A single LiDAR sensor prevents this kind of issue
Even RADAR prevents this and the cars had RADAR! They started disabling RADAR for the older cars since the new ones don’t even have the hardware installed.
I’m trying to find an article that covers what I remember but I know for sure that it’s been a good while since I saw the info I recall. Hopefully I can dig something up.
That might actually be the exact article I’m thinking about. Thanks!
Indeed there is a lidar car in the video and it works way better in many scenarios.
Is that just to cover his ass cuz he was promising backwards-compatible FSD for models that don’t have LIDAR?
Entire video is worth watching. He also snuck a chest mounted lidar into Disney and mapped some rides.
This would be hilarious if it weren’t for shitty cars causing deaths.
That said, I always wondered why we don’t find a system like RFID that could penetrate concrete and asphalt, and plant passive receivers in roads? We re-pave roads so damn often in this country (the U.S.) it seems like we could’ve knocked it out in the past couple of decades, minus our most rural areas.
I know RFID itself isn’t strong enough, but I imagine that would’ve been an easier problem than figuring our complete self driving. Not to mention making GPS a secondary system for U.S. road travel in most cases.
Maybe it’s just a dumb shower thought?
I have another idea: echolocation and laser measurement.
Echolocation is specifically audio based.
Lidar is a similar technique, but much more accurate and precise.
Project a grid of laser beam, read when the laser bounces back, you know the distance to that part of the grid.I don’t know the value of echolocation in this case, as I’m generally ignorant here, but it’s straight wild to me that they went purely on visuals.
Tesla used to also have radar (and maybe lidar?) but they removed it as a cost cutting measure. If you ever see older videos of a Tesla slowing down or stopping due to a potential collision a few cars ahead, that’s from before they switched to only relying on cameras. The collision avoidance was significantly better back then.
What you’re describing is just a higher level of autonomy. If I remember correctly, you’re describing level 3 whereas Tesla’s are level 2. I believe VW made a level 3 proof of concept mini bus back around 2020 but the legislation doesn’t allow for the sensors in the road yet because… Oh that’s right. A level 2 car manufacturer owns like half the world right now which means nobody is allowed to innovate or do better than him. Huh, that sucks.
Things that happen when you rely exclusively on optical sensors, i.e. cameras. But that’s just cheaper, more money for Nazi Elon.
Are we reeeeally sure optical sensors with fast image recognition software are cheaper than LiDAR?
The hardware is, which is the important part at scale: even if the code is 10x more expensive when you sell millions of the car it becomes pennies/car
Painted walls, the natural enemy of the Tesla.
Turns out having radar is rather important…
It’s dirt cheap, too. If this was a cost-cutting measure, it was a thoroughly idiotic one. Which feels like the mark… of a certain someone I can think of
A genius :-)
This is some amazing stuff! Buy $LAZR! 😁
