Why Tesla’s cars keep crashing
theguardian.com
theguardian.com
Welp, he's wrong. He won't admit it. More people will have to die and/or Tesla will have to face bankruptcy before they fire him and start adding lidar (etc) back in.
Real sad because by then they probably won't have the cash to pay for the insane upfront investment that Google has been plowing into this for 16 years now.
Technofetishists and low-level tech/IT folks love to lap up the crap coming out of elon's mouth.
The issue came when he promised every car would become a robotaxi. This means he either has to retrofit them all with lidar, or solve it with the current sensor set. It might be ego as well, but adding lidar will also expose them to class action suits.
The promise that contributed to the soaring valuation, now looks like a curse that stops him from changing anything. It feels a bit poetic.
A LIDAR can get near-exact distances between objects with error margins of something like 0.2%, even 100m away. It takes an absolute expert human to accurately judge distance between themselves and an object even 5 meters away. You can see this in the youtube movies of the "Tesla beep". It used to be the case that if the Tesla autopilot judged a collision between 2 objects inevitable, it had a characteristic beep.
The result was that this beep would go off ... the humans in the car know it means a crash is imminent, but can't tell what's going on, where the crash is going to happen, then 2 seconds "nothing" happens, and then cars crash, usually 20-30 meters in front of the Tesla car. Usually the car then safely stops. Humans report that this is somewhere between creepy and a horror-like situation.
But worse yet is when the reverse happens. Distance judgement is the strength of LIDARs. But they have weaknesses that humans don't have. Angular resolution, especially in 3D. Unlike human eyes, a LIDAR sees nothing in between it's pixels, and because the 3d world is so big even 2 meters away the distance between pixels is already in the multiple cm range. Think of a lidar as a ball with laser beams, infinitely thin, coming out of it. The pixels give you the distance until that laser hits something. Because of how waves work, that means any object that is IN ONE PLANE smaller than 5 centimers is totally invisible to lidar at 2 meters distance. At 10 meters it's already up to over 25 cm. You know what object is smaller than 25 cm in one plane? A human standing up, or walking. Never mind a child. If you look at the sensor data you see them appear and disappear, exactly the way you'd expect sensor noise to act.
You can disguise this limitation by purposefully putting your lidar at an angle, but that angle can't be very big.
The net effect of this limitation is that a LIDAR doesn't miss a small dog at 20 meters distance, but fails to see a child (or anything of roughly a pole shape, like a traffic sign) at 3 to 5 meters distance. The same for things composed of beams without a big reflective surface somewhere ... like a bike. A bike at 5 meters is totally invisible for a LIDAR. Oh and perhaps even worse, a LIDAR just doesn't see cliffs. It doesn't see staircases going down, or that the surface you're on ends somewhere in front of you. It's strange. A LIDAR that can perfectly track every bird, even at a kilometer distance, cannot see a child at 5 meters. Or, when it's about walking robots, LIDAR robots have a very peculiar behavior: they walk into ... an open door, rather than through it 10% of the time. Makes perfect sense if you look at the LIDAR data they see, but very weird when you see it happen.
Worse yet is how humans respond to this. We all know this, but: how does a human react when they're in a queue and the person in front of them (or car in front of their car) stops ... and they cannot tell why it stops? We all know what follows is an immediate and very aggressive reaction. Well, you cannot predict what a lidar sees, so robots with lidars constantly get into that situation. Or, if it's a lidar robot attempting to go through a door, you predict it'll avoid running into anything. Then the robot hits the wood ... and you hit the robot ... and the person behind you hits you.
Humans and lidars don't work well together.
Huh? The most basic skill of any driver is the ability to see if you're at a collision course with any other vehicle. I can accurately judge this at distances of at least 50 meters, and I'm likely vastly underestimating the distance. It is very apparent when this is the case. I can't tell if the distance between us is 45 vs 51 meters, but that is information with 0 relevance to anything.
> The result was that this beep would go off ... the humans in the car know it means a crash is imminent, but can't tell what's going on, where the crash is going to happen, then 2 seconds "nothing" happens, and then cars crash, usually 20-30 meters in front of the Tesla car. Usually the car then safely stops. Humans report that this is somewhere between creepy and a horror-like situation.
This is a non-issue and certainly not horror-like. All one's got to do is train themselves to slow down / brake when they hear the beep. And you're trying to paint this extremely useful safety feature as something bad?
> Worse yet is how humans respond to this. We all know this, but: how does a human react when they're in a queue and the person in front of them (or car in front of their car) stops ... and they cannot tell why it stops? We all know what follows is an immediate and very aggressive reaction.
What are you trying to say here? If the car in front of me brakes I brake too. I do not need to know the reason it braked, I simply brake too, because I have to. It works out fine every time because I have to drive in such a way to be able to stop in time in case the car in front of me applies 100% braking at any time. Basic driving.
Generally, what you're describing as predicting is more accurately called assuming. Assuming that things will go how one wants them to go. I call that sort of driving optimistic: optimistically assuming that the car in front of me will continue going forward and that there is nothing behind that huge truck that's blocking my view of the upcoming intersection, so I can freely gas it through.
That mindset is of course wrong; we must drive pessimistically, assuming that any car may apply max braking at any time and that if any part of our line of sight is obstructed, the worst case scenario is happening behind it - there is a high speed object coming towards us at a collision course that will reveal itself from behind the obstruction at the last second. Therefore, we must slow down when coming around a line of sight obstruction.
That's probably because for things moving in straight lines at constant velocity you don't need to be able to measure distance at all accurately to figure out if they are on a collision course. You just need to be able to tell if the distance is decreasing.
First, you just have to note if their angular position is changing. If it is then they are not on a collision course.
If the angular position is not changing, then you have to check if the distance is decreasing. If it is they are on a collision course. If it is not then they aren't.
If you take advantage of the fact that cars generally have distinctly different front ends and back ends and that most of the time cars are traveling forward you don't even have to estimate distance. If the angular position is not changing just note if the direction the car is pointing has its front closer to you than its back or not. If its front is closer than its back then it is on a collision course. Otherwise not.
You will need to make some adjustments due to cars having volume. A near miss for point cars could still be a collision for cars with volume, but this should be fairly easy to deal with.
Can you tell me the distance between 2 objects, each 50 meters away from you, down to 1 cm? That's the superhuman part. Even the distance between you and an object 10 meters away down to a few millimeters is impossible for a human.
It can also tell you 4-5 seconds before it happens which objects are going to collide. Not just which object YOU are going to collide with, but any collision between any 2 objects if they are within the range of the LIDAR.
But they see fundamentally different things than humans. So humans will never work together nicely with LIDAR guided robots.
But radar and ultrasound did not cost a lot and he got rid of those too, suggesting it was more than cost that made him go vision only.
Heck, they even use vision for rain sensing instead of the cheap and more effective sensor everyone else uses (which is just some infrared LEDs and photodiodes that measures the change in internal reflection at the outer side of the windshield when the critical angle changes when the windshield gets wet).
They did get rid of the radar at a moment when there was a shortage of parts. They had the choice, ship now without the part, or wait and ship less cars.
Maybe that was always the plan, and the shortage only accelerated the decision.
Might?
And yet here we are where the testing grounds are our public roadways and we, the public, are the guinea pigs.
https://thevictoriancyclist.wordpress.com/2015/06/21/cycling...
To me (keen cyclist and non-driver), it seems like the newspapers were pushing back against the freedom that cycles were giving to women. One of my favourite pro-cycling quotes is from the suffragette Susan B. Anthony (1896):
> “Let me tell you what I think of bicycling. I think it has done more to emancipate women than anything else in the world. It gives women a feeling of freedom and self-reliance. I stand and rejoice every time I see a woman ride by on a wheel…the picture of free, untrammeled womanhood.”
It would be on-brand for the newspapers to demonise cycling if it was allowing women to escape their restrictions.
Nowadays, there's doesn't seem to be much negative reporting about other car-shaped EVs, but just Teslas which pre-dates the anti-Musk viewpoints. Also, the reporting isn't just about autonomous crashes, so it would seem to me that Teslas do have an issue with quality. (Here in the UK, I only know one person with a Tesla and he's had several minor issues with it).
However, I do see parallels with the early anti-cycling reporting and current anti e-scooter/e-bike/e-motorbike reporting here in the UK, though I suspect that some of that is pushed by the motor lobby, although we do have a lot of illegal e-motorbikes being ridden around our cities.
I never will understand this horrible decision. It isn't good design if it kills people. I wonder why this isn't regulated. They could at least implement a "push to pop up" functionality that works without battery power or have a narrow slot under the handle.
This is probably core to their legal strategy. No matter how much data the cars collect they can always safely destroy most because this allows them to pretend the autonomous driving systems weren’t involved in the crash.
At this point it’s beyond me why people still trust the brand and the system. Musk really only disrupted the “fake it” part of “fake it till you make it”.
It's something to keep in mind but it's not an issue itself.
> The YouTuber Mark Rober, a former engineer at Nasa, replicated this behaviour in an experiment on 15 March 2025. He simulated a range of hazardous situations, in which the Model Y performed significantly worse than a competing vehicle. The Tesla repeatedly ran over a crash-test dummy without braking. The video went viral, amassing more than 14m views within a few days.
> The real surprise came after the experiment. Fred Lambert, who writes for the blog Electrek, pointed out the same autopilot disengagement that the NHTSA had documented. “Autopilot appears to automatically disengage a fraction of a second before the impact as the crash becomes inevitable,” Lambert noted.
In my previous comment I was wondering why would anyone still trust Tesla’s claims and not realistically assume the worst. It’s because plenty of people will only worry about it when it happens to them. It’s not an issue in itself until after your burned to a crisp in your car.
That quote isn't playing games about whether it was engaged or not. If that's a lie it's equally easy to make the lie whether the system disengages or stays engaged.
I'm taking issue with a very specific scenario, not claiming tesla is honest in general.
Tesla bragged about the cars giving a ton of data, and showed it when this suited the company and it was good for the image. But every time it was controversial like an unexplainable accident potentially caused by the car itself the data was somehow not transmitted, or lost, or irrelevant.
I’m not sure why you have such a hard time understanding the issue, or insist on what you’re “pretty sure” about when all evidence (they cite the NHTSA and experiments conducted privately by a NASA engineer, as well as the string of coincidental data unavailability for controversial accidents) points to the contrary. The article provides evidence and discussion on all these points. Nonetheless you ignore all that and stick to your “I’m pretty sure” with fanboy abandon. Sets a really low bar for future conversations.
In theory. Maybe.
Have they ever done that?
You're citing entirely different bad behavior. That's not evidence for my question. The article has claims of stonewalling and claiming no data at all and one case where they said the software wasn't even installed, but those are not the scenario I asked about.
Calling me a tesla fanboy for wanting evidence for the correct claim instead of a completely different claim is pretty ridiculous. I'm not being pro tesla here.
And the reason I said "pretty sure" is that people bring up that scenario over and over and over, but nobody has ever shown an example of it being real, despite having tons of examples of other tesla problems.
> This is probably core to their legal strategy. No matter how much data the cars collect they can always safely destroy most because this allows them to pretend the autonomous driving systems weren’t involved in the crash.
Like, what judge would get fooled by that? It's dumb software engineer thinking.
> This is probably core to their legal strategy. No matter how much data the cars collect they can always safely destroy most because this allows them to pretend the autonomous driving systems weren’t involved in the crash.
I didn't have a reply to this until a couple days ago, when it hit me: it's a deliberate safety choice. They never concealed that AP/FSD disengages at the last second, it's always been stated: "because there's nothing else it can do". While I agree with that, the choice tells more about the possible consequences if AP/FSD remained active.
What would happen if AP/FSD were not disengaged before the crash? Suppose the cameras/sensors, the self-driving computer or the actuators got damaged to some degree on collision, not enough to make some or all of them inoperable, but enough to provoke uncontrolled movements after the crash. You may have a vehicle trying to accelerate without control, steering at random, repeatedly crashing into a wall, ... making it impossible for rescuers to reach you.
During an imminent unavoidable collision the last second of travel is basically watching the accident unfold, so if AP/FSD were at fault it would be apparent regardless of the disengagement.
However, I don't personally believe that as even topics which are more to do with the products rather than the man seem to get flagged extremely quickly.
“The whole Tesla fleet operates as a network. When one car learns something, they all learn it. That is beyond what other car companies are doing.” Every Tesla driver, he explained, becomes a kind of “expert trainer for how the autopilot should work”.
Good grief.
I’m not afraid of Tesla drivers. I’m concerned about uncommanded action such as accelerating without driver input. Over reliance on an unreliable autopilot can be considered reckless. My own car has a number of features, from lane keeping to adaptive cruise control that make long and boring trips easier, but it’ll complain loudly if I drive with a single hand for more than a couple seconds.
Even the bad design of the door handles, which should open when airbags are deployed, the door loses power, or the vehicle comes to a stop, is a red flag for me.
TBH I see this more as a “firefighters aren’t being given the right tools” issue, as this is far from unique to Tesla, and the tools have existed since laminated side glass became a requirement - but don’t seem to yet be part of standard issue or training.
https://www.firehouse.com/rescue/vehicle-extrication/product...