Tesla: Failure of the FSD's degradation detection system [pdf]
static.nhtsa.gov
static.nhtsa.gov
Does it not detect them at all, or fail to deal with detected sensor degradation adequately? Does "Full Self Driving (assisted)" slow down under conditions of poor visibility?
Does Tesla even look for the road surface? One big advantage of those up-top LIDAR units is that you have a good scan of the pavement ahead. If you're not sensing flat pavement ahead, don't go there. That's basic. Vision-only systems, going back to Mobileye, have been overly dependent on looking for known kinds of obstacles. Original Mobileye could only detect car rear ends.
Under conditions of poor camera visibility?
Humans drivers can see under conditions that cameras cannot, and people will otherwise misinterpret “visibility” as referring unpredictably (and with personal biases) towards either human sight and/or camera processing and/or lidar processing.
And yet randos on the web keep asserting he's not an engineer. Is there any factual basis for this? Is it just that he doesn't have a degree with that word in the title?
I suppose most people don't know that physics degrees are largely accepted as engineering degrees.
He continually says dumb things that aren't true or reasonable and has never worked in the field he's a rich boy who bought things with daddy's aparteid money.
Obvious things first, cameras have way worse contrast and low light sensitivity than human eyes.
Humans have much more evolved logical thinking capacity, even the stupid ones can figure stuff out that modern AI struggles with.
Humans have other sensors, too that they use to plausibility check the picture they see. I.e. one of the best sensor fusion systems on the planet.
When in doubt humans can figure out whether it's a lens occlusion or a some other artifact in their vision by virtue of moving their head around.
There's probably other things I'm not thinking of. In any case to make full self driving work we should first start by using all available tech to make it safe. When you have safe tech you can slowly start removing individual sensors while verifying that safety remains high. As the experience and system evolves there will be optimization potential.
And until we have that low light thing and high contrast figured out, camera alone doesn't cut it.
I personally feel like that isn't really true any more.
But try asking your favorite LLM what happens if you're holding a pen with two hands (one at each end) and let go of one end.
Seems fine to me?
https://chatgpt.com/share/69bcd01a-a750-800d-95f5-3b840b9ee2...
https://gemini.google.com/share/edc223bb6291 (the try again gave a woman, oops)
Even Midjourney couldn't do it.
Researchers built the Winnograd Schema Challenge more than a decade ago to assess common sense reasoning, and LLMs beat that challenge task around GPT 4.
If you ask them in isolation they may write a script to solve it "properly", but I guess this is because they added enough of these to the training set. But this workaround doesn't scale.
As soon as I give the LLM a proper problem and a small part of it requires numeric reasoning, it almost always hallucinates something and doesn't solve it with a script.
If the logic/math is part of a larger problem the miss rate is near 100%.
LLMs have massive amounts of knowledge, encoded in verbal intelligence, but their logic intelligence is well below even average human intelligence.
If you look at how they work (tokenization and embeddings) it's clear that transformers will not solve the issue. The escape hatches only work very unreliably.
I have been broadly quite happy with gpt 5.4 xhigh's reasoning on things like performance engineering tasks.
They're just not that good - nowhere near human vision performance. And a human in a car has a surprisingly good view of the road and a very fast pan tilt system to look around.
Tesla's do not actually have 360 degree full binocular vision coverage - nor the ability for a camera to lean left or right to improve an ambiguous sensor picture.
So while I fully believe that vision only self driving could work, within the constraints of automobile platforms and particularly the Tesla and it's current camera deployments, it is not remotely similar enough to human visual fidelity for that to solve a valid argument.
Humans are hard to compete with! I'd want LIDAR & RADAR just to give me an edge.
Tesla’s actually have zero binocular vision coverage because the cameras have different focal lengths and are too close even if they did have the same focal lengths.
They are also below minimum vision requirements for driving in many states.
And this is an amazingly stupid statement. Humans drive with most of their senses, not just vision. In fact our proprioception plays an important role in driving.
Even Tesla's use of cameras is poor because they're monocular and fixed in place. Most humans have binocular vision and those visual sensors have multiple degrees of freedom and the ability to adjust focus.
Even if you wanted to only use vision for navigation it's irresponsible to not use binocular configurations that get more reliable depth sensing.
That's enough for vastly more depth perception than any human eyes.
I have 20 toes. Therefore I should fly 10 times better.
Amazing how you lost your critical sense just because you want Tesla to succeed and drink everything Musk says.
At this point I truly don't understand why anyone cares what that liar says.
It might be dishonest (if he doesn't believe it is possible), but I don't think he's saying that the current systems have reached the mark.
I also own a Tesla, and there is no indication shown to the user that FSD's vision is degraded. They need to add this in.
For example, numerous times I have been driving my Tesla with FSD activated with ostensibly a clean and clear windshield when suddenly the car will do the "clean the windshield in front of the camera routine" without any indication that the car's camera is degraded. If people haven't seen this "clean the windshield routine", the wiper fluid is dispensed and the wiper will vigorously wipe in front of the camera only -- the rest of the windshield only gets a cursory wipe.
This indicates to me that the camera has poor visibility and I am not informed or aware of this as a driver, which is concerning. I am often curious if there is a thin occluding film on the windshield in the camera box in front of the camera, or something that has degraded FSD's vision, but they do not give you the ability to view the camera feed, nor do they notify you that the vision is degraded. I think a "thin occluding film" may be in the camera box because my normal windshield outside of the camera box started to show a thin chemical film after a couple of months, which apparently (according to a Google search) happens when a new car off-gasses, adding a thin film of chemical byproduct to the windshield. This is my first new car so I've no idea if this is normal or not.
As with all things FSD, it does sometimes and not others. I've driving my parents' Tesla with FSD engaged and it did complain when the windshield got dirty but didn't say anything when it drove into fog. (I took over manually.)
In any case, it seems reasonable to me that the human should be making the decisions once conditions become adverse. It’s a simple liability issue for the car company but also I’d rather trust my own judgment if it’s only 80% certain it’s not driving me off a cliff.
Systems built from cameras that are only nearly as capable as human eyes and software that is only nearly as capable as the human brain will fall short overall. To match or surpass human performance, the individual components need to exceed human abilities where possible--and that's where LiDAR provides an advantage.
If the cameras are a little less sharp in some sense is a minor rounding error in comparision.
Comparing human and camera acuity is difficult. But saying Teslas have cameras that are a little less capable than human eyes is unfounded.
You can say that about the original IBM PC from 1981. That doesn't make the IBM PC better at driving.
"Number of miles driven in situations where [quality of conditions is greater than some threshold] versus all conditions."
"If you don't count the games we'd definitely have lost, our winning percentage is so much higher!"
This must be one of the most stupid takes that gets repeated non stop by Tesla fans.
I just don't get it. Humans also have emotions and other biological senses that Computers don't have. You just cannot compare both. What makes human so good at driving is that they can react relatively well to unknown new events. Teslas cannot do that, and with the current hardware never will.
That seems to be better than that try to continue to run the vehicle. What would you expect it to do?
If it is foggy I just don't drive, anybody that expects me to drive when conditions are bad can go and drive themselves.
If the condition is a little fog and little rain and little snow/sleet I hate to break it to you but those are very permitting. In most of the continental US the number of days where driving conditions for an (below)average human and such that it is wiser not to get on the road is very small. If the "robo"taxi technology you posses cannot match that of a (below)average human you got nothing but vaporware you've been pitching as "done deal" for more than a decade.
:-)
That argument is dumber and dumber any time I think about it. And we haven't even gotten into the fact that human eyes and its partner in crime the brain work much different than a camera.
But yes I agree we should hold self driving cars to a higher standard.
That's not good enough. Too many accidents at manual takeover. The new standard, which Mercedes has demonstrated and China is mandating, is that the system must be able to pull the vehicle over and stop safely when there are problems.
But it's that 0.1% of situations where the results will be catastrophic. Sure, you can detect vehicles, traffic cones, bikes (both bicycles and motorcycles), people, mopeds, traffic lights, lane markings, everything you'd expect on a road.
But what about the mattress that fell out of someone's truck? If the car doesn't know what a mattress is and what it looks like, it can't really adequately determine its size based on the monocular vision that Tesla has. Sure, maybe it could use motion vectors between video frames to make a guess, but I'm not convinced that's going to work well, especially relative to LIDAR.
Steering back to the subject at hand...
> "In the crashes that ODI has reviewed, the system did not detect common roadway conditions that impaired camera visibility and/or provide alerts when camera performance had deteriorated until immediately before the crash occurred."
I don't think I've ever had my Tesla disable Autopilot based on road conditions, though maybe it's because when conditions are bad, I've just taken manual control preemptively. I've let it go through construction areas where cones are guiding traffic outside the painted lines, and surprisingly, it's handled it fine, though I've only done this at low speeds (~20 mph).
Camera visibility is another story. In heavy rain at night, I've had it not allow me to enable AP, though I've never had it disable AP and tell me to take control. However, it HAS limited the cruise speed based on visibility.
All this to say...
...anybody buying Tesla's FSD is being swindled, as far as I'm concerned. "FSD (Supervised)" is a scam. If you have to supervise it, it's not self-driving. It's just a party trick that you have to watch to make sure nothing goes wrong.
FSD is a better driver than me 99% of the time.
99.9% of driving of sea-level, non-rainy, near-the-coast California/Austin weather, maybe. I would guess it's a no-go in the inland foggy conditions in CA, for example, or freezing rain in TX.
In terms of ALL conditions? 60-70% maybe.
Human vision only? Sure. Cameras only? I'm sceptical of the quality of these cameras. Particularly given, unlike the human eye, they don't do the rapid involuntary movements that make human eyes ridiculously robust.
Also, I guarantee you FSD would recognize a mattress as an obstruction.
Which is probably why you still can’t take a Robotaxi anywhere, but Waymo does millions of paid rides every month.
No software update is going to give working FSD to all those older models on the road. I don't think they can even have HW4 retrofitted either. Ideally FSD would be disabled on those cars because it's confirmed never going to work properly but people paid for lifetime FSD on them (newer cars are subscription only).
It's really unfortunate that FSD is judged as a singular feature when it is vastly different depending on the car. HW3 cars are getting cut down versions of FSD software, which is useless garbage because it drives like an idiot. HW2 and earlier are probably not getting FSD updates at all anymore. The majority of Teslas on the road are HW3 or earlier, run significantly worse (unsafe) FSD versions, and yet Tesla still allows it to be turned on.
As for road surface, yeah, that is how it should be (as you describe). Higher end Audi model/trims do road surface scanning and will adjust ride height, vehicle camber, even (to the greatest extent possible) help you avoid potholes while maintaining lane positioning. This is - or should be - table stakes.
>Does Tesla even look for the road surface?
It is a shameful engineering design to leave out LIDAR and it has cost human lives.
Let's hope Musk does not leave out something important for the moon landing. His proposal for it is absolutely ridiculous, it looks like a children's book fantasy and many smaller top-heavy craft have already toppled on the moon!
One giant leap?
The gravity is weaker so just jump down /s
(ChatGPT answered, I was too lazy to computer 40m with gravity being 1.62m/s on the lunar surface.)
Velocity on impact: ~11.4 m/s km/h: ~41 km/h mph: ~25.5 mph
If Tesla FSD is better than the average driver using it, isn't it still a net win, even if it might crash in different scenarios than a human? Especially because a human has a window to override FSD, but FSD doesn't really get a chance to override a human, except in limited scenarios like automatic emergency braking. And it gets more people using it by providing FSD at a lower cost?
https://www.thedrive.com/news/tesla-autopilot-fails-wile-e-c...
> If Tesla FSD is better than the average driver using it, isn't it still a net win, even if it might crash in different scenarios than a human?
That was the analysis when the industry was in its infancy. I think a lot more work has to go into that argument for people to accept it now that the driverless car industry has been operating for a decade+, it's not really clear that this pans out.
For example, today you can look at a car and predict how it's going to behave because you have a good model for how people drive. But let's say in the future driverless cars are much "safer" on paper than human drivers, but they behave very differently from them such that it's hard for people to predict their behaviors.
Now you've created a highly dynamic system where you don't have a good model for the all the actors because some of them behave one way but others behave a completely different way. Does this increase the overall safety of the system or decrease it, despite the new actors being statistically safer than the current ones?
I don't think you can with great confidence say what's going to happen just by looking at crash rates and comparing to the current system. You're going to change the whole system by introducing large numbers of actors who "crash in different scenarios than a human"
Yes!!! Thank you hopefully I will get credit for inventing this attack :)
>> IncreasePosts 39 minutes ago | parent | context | flag | on: Tesla: Failure of the FSD's degradation detection ...
What are wile e coyote attacks? Painting a tunnel entrance on a wall?
>> If Tesla FSD is better than the average driver using it, isn't it still a net win, even if it might crash in different scenarios than a human?
I don't think so because it is fooled by simple things that could easily be prevented and counting on a human to override is very risky because the human is simply not alert in the passive mode.
I think cameras are great, but there is no excuse not to also use LIDAR.
tesla cars killing people would be all over the news each time and nobody would care that similar or even marginally smaller amount of people would die anyway. People simply expect way more for giving up control.
Is it really that hard to see?
What's much worse is that Tesla is hiding the reasons for the crashes in the crash reports (the only company with FSD).
I'm still waiting for a jurisdiction to either create a liability safe harbor for self-driving systems with lidar or outright ban cameras-only systems on public roads.
Overall, yikes.
Oh good, Tesla vehicles apparently struggle with the task of "Hey, there's a car there" in degraded conditions.
Probably don't need to worry about that while driving though.
>Tesla also described internal data and labeling limitations that prevented a uniform identification and analysis of crash events with the subject system engaged. ODI believes this limitation could have led to under-reporting of subject crashes over portions of the defined time-period.
I thought Tesla was a "Software" company!
This report is insanely vague though. It's very preliminary, opened yesterday.
Yeah I think posting this here is premature without any details.
Maybe I'm misremembering things, but I feel like 4-5 years ago we didn't have these clickbait headlines that fed political discourse. It feels like reddit culture has permeated this place for a while.
Anytime one of Elon Musk's company has a misstep, the headlines violently shoots to the top of the front page.
This is an expected and understood result given the hardware and software involved.
You will not get past these issues without a RADICAL improvement in camera technology paired with specialized, dedicated processing hardware matched against several (and I mean several several) "common" environment profiles.
FSD is a scam. It's not safe. It is not technically sound.
The fact that there aren't many more accidents with the system is a by product of consistent and well thought out road standards, car standards, other safety systems present on cars, and driver education.
The report is not premature and it's not premature to comment on them.
Can you clearly and explicitly state why you feel like the report or the commentary is premature?
Eh, while I agree with you on the permeation of reddit culture on this board, this post is in no way clickbait or political in nature.
In fact, the title of this post is literally copy and pasted from the problem description.
The reason this stuff shoots to the top is because Elon Musk and his companies are a red alert menaces to society. People are sick of him and the damage he causes with his money, and wish he and his cars would just fuck off for good. From his cars slamming into people and property, his website spreading hate, his starships raining fiery debris, or he's personally taking an axe to government programs we rely on, everyone has cause to be absolutely done with his antics.
But since businesses can apparently unleash autonomous murderbots onto the public roadways with zero repercussions for 10+ years, I guess we'll have to settle for endless flamewars about Musk's campaign of destruction on HackerNews instead.
I'm not sure that's the case anymore. Each Tesla model has gotten more spartan over the years. And the interiors have never been all that "premium" when compared with other manufacturers. They should still offer the most comprehensive safety features, but whether or not thats because of "luxury" or not, I'm not sure.
Some people upvote everything slightly negative about the topic: "see how bad it is!!!"
Some people flag everything slightly negative about the topic: "we rather not let you see how bad it is"