You highlight the big challenge with computer vision and other computerized identification of things, regardless of the underlying technology. There is no "mental model" of what makes sense against which to check observations. There can be coded rules, but if something passes them, there is no extra sanity check like a person would do. Put another way, models have no common sense, and we don't really have a meaningful way to give it to them. This is going to remain a problem for self driving cars and is a big part of the reason why they have not lived up to expectations.
Humans use stereo vision to overcome things like this. I think some kind of stereoscopic strategy would be a feasible strategy, no? Why revert to the age-old argument of "common sense"?
Humans have more than just stereo vision sensors. Like GP said, we also have robust schemata, seeded by biology and improved over our entire lives, that allow us to make sense of the things we see. These heuristics are trivial for us to access, but relatively difficult to describe and implement in a software system. That's what "common sense" means in this context.
Stereo vision is overrated. One of my buddies is a pilot with 1 eye (since birth) and around 11 thousand landings. Top of his game. There is more to depth perception than stereo vision for sure
Funny, when I saw that I thought how did they miss this one? It's the moon, how is it possible no one has driven a Tesla at the right time in the right direction to have picked up the moon? The moon is out quite often. It doesn't seem like that much of an edge case to me, though I could be wrong.
The exact situation with the moons coloration changed by atmospheric conditions and being in just the right location over a long straight road to trigger this on a long stretch of straight road is probably quite rare. The software eventually does reject it as not a yellow light, only to change it’s mind back to the moon. I am wondering if this relates to the image classification system starting over from scratch regularly for some reason.
A friend has a 2018 Honda with auto-dipping headlights, it mistakes a near-full Moon low to the horizon as an oncoming car and switches the headlights to low-beam.
The real world is full of 'edge cases' that humans can deal with through their intuition and understanding of context. Humans are unlikely to mistake a yellow-ish moon for a yellow traffic light unless very intoxicated.
I'd like to see a Tesla on full autonomous autopilot take a drive in Calgary or Edmonton a few days after a heavy snow, when it's been mostly plowed but the lines are missing from the roads, all the lines at the side of the road are covered in snow, etc.
At level 3 they just reject things back to the driver if they can’t see road markings. Longer term GPS makes that an easier problem for self driving cars than bad weather. The post blizzard random trails through deep snow seem like a more difficult challenge, largely because it’s so rare.
Lanes in heavy snow covered roads are more of a suggestion. You can either drive on the bare 'wagon tracks' of road, or on the proper lane (that you can't even see) with 6-12 inches of snow on them.
Often when driving like this, you see a glimpse of the middle-yellow line inside one of the wagon tracks, and you realize that you're actually driving in part of the other lane. Fortunately the opposing traffic's wagon tracks are also shifted so that they're driving partially on the shoulder.
A GPS guided car would make a mess of the wagon track pattern that all the humans decided to just go with.
I am not talking 6-12 inches, when it’s 4+ feet of snow even pickup trucks need to stick to the valley’s which rarely have much to do with traffic lanes.
I think it is more fruitful to realize that humans also make mistakes. It may also be also beneficial to compare Human + FSD driving vs. Human driving (instead of just FSD vs Human).
As a general rule of thumb, if someone has a large, comprehensive data set about something, has claimed they've used it to back a conclusion that is very favorable to them or some cause they're supporting, and refuses to let anyone else see the data set, it's a reasonable bet that they've been "creative" with at least some of the interpretations and analysis.
It doesn't mean they're directly lying, in the sense of "You can't reasonable get these numbers out of the dataset." But it probably means that there are some games being played such that the number you're getting isn't the number it's strongly implied (or, occasionally, outright stated) to be. Or, if it is, it doesn't mean what they'd like you to think it means.
So, if you go back a few years to the first waves of numbers Tesla trotted out to "prove" how amazing Autopilot was, you roughly had:
- The fatality rate per mile of all drivers, in all vehicles, in all conditions
being compared with:
- The fatality rate per mile of expensive luxury cars, driven by older, well off drivers, mostly on limited access divided highways, in "easy" driving conditions.
You can certainly do the math and get a comparison, but it's not a particularly reasonable comparison, because:
- In general, six figure luxury cars are really, really safe. They're big, heavy, loaded with airbags and safety features, and really don't kill people very often at all. You can go look at a breakdown of death by vehicle type and class, and it turns out that luxury cars are very much under-represented in fatalities.
- Expensive cars tend to be driven by older, more experienced drivers. There's a bathtub curve of fatality rates by age (https://aaafoundation.org/rates-motor-vehicle-crashes-injuri...) - younger drivers without much driving experience die more, and once you get to 70+, your fatality rate goes up, but there's a sharp drop going to the 30-39 group, continuing to decline up to the 60-69 group. Guess what age range is going to be mostly buying Teslas?
- Divided highways tend to be fairly safe places to drive - especially in lower speed situations where a lot of people are going to be using Autopilot. You're exceedingly unlikely to die doing 20mph in rush hour traffic. Even running at speed, limited access highways are fairly safe places.
- And Autopilot simply doesn't work in bad enough weather. Again, a lot of traffic deaths are in pretty sketchy conditions - rain, snow, sleet... things that older Autopilot, especially, simply won't work in.
So, yes, a straight up comparison of the two numbers ("fatalities per million miles in Teslas with Autopilot engaged" vs "fatalities per million miles of all vehicles in all conditions") is at least somewhat deceptive.
That press release is not evidence that Tesla's Autopilot is saving lives, for at least three major reasons.
For one thing, Autopilot is more likely to be engaged for some types of driving than others. Comparing statistics with Autopilot engaged to the total of all miles driven is comparing apples and oranges.
Similarly, it's comparing statistics about the people who buy and drive Teslas with the entire driving public. You don't know how many accidents Tesla owners would have gotten into if they were in cars that didn't have self-driving features.
Finally, you can't draw any conclusions about whether lives were saved by just looking at the number of accidents and ignoring their severity. Hypothetically, if Autopilot caused fewer accidents at low speeds but more catastrophic collisions on highways, it could be easily be causing more deaths than it prevents.
This statistics simply proves that Tesla drivers know when it's safe to activate autonomous driving.
Make it self driving 100% of the time under all lighting and weather conditions and then compare the numbers.
The idea of AI and mathematics as a whole is to compute and do what is humanly impossible. It's expected that self driving means not just superhuman but almost perfect.
I disagree. None of the things AI or mathematics do are humanly impossible.
In fact most things are better done by humans, because they understand context, which most AI/ML currently doesn't.
AI is great for tasks that can be done without much context and can be incredibly efficient and fast compared to manual work. Everything else, including "self-driving" currently needs supervision or produces terrible results.
And the need is there. Tesla's previous CTO, JB Straubel, recently lost his wife to a driver who hit her while she was cycling [1]. Everyone on HN shits on Autopilot (and I'll be the first to admit it has flaws and isn't perfect), but humans are just terrible (tens of thousands of deaths per year, hundreds of thousands to millions of injuries). You don't have to be perfect, just better than humans.
The software will improve faster than humans will, as arguably, we've already hit our limit. Software doesn't get tired and it doesn't drive impaired. It makes mistakes, but so do we, and it's mistakes can be improved upon at scale. This is not blind faith in technology or software, but an acceptance of the limited capabilities of your median human (as well as the disadvantaged outliers).
Humans are no where near the limit to being attentive on the road. When you consider how many people are either distracted or tuned out to their surroundings, there is a ton of room for improvement, and not theoretical improvement, there are clearly some drivers that are consistently able to make correct decisions, whether there is risk or not.
This is not to say that software and hardware will not quickly overtake us, but while humans are on the road and in control, improvement will be possible.
I really don't think he referring to some physiological limit of individuals. It is the aggregate of society, which includes the prevalence of distracted driving.
We don't face the consequences, at least in my experience seeing the justice system play out. Drunken driver? Affluenza? Old person turning in front of a motorcycle killing the rider? You're getting off relatively easy for taking a life or maiming someone. I've followed the cases for each example given, the last being a close friend. The retiree who pulled out, in their 80s, still has their license and lives comfortably in their retirement community and my friend is dead.
You are right, the software should absolutely be held to a high standard (otherwise we're socializing losses and privatizing profits, also bad!), but please don't perpetuate a lie that existing drivers are held responsible to the degree they should for their mistakes (or even intentional acts) that cost lives. You'll find examples where someone is made an example of, but that is rare.
So disagree, please (checks and balances are important, as is the dialog we're having), but having seen what humans have to offer in this context, I will be the first, on my own time and resources, in front of regulators to argue for aggressive driver assist system development and implementation. Because the alternative is hot garbage we get by with because there are limited alternatives (ie no public transportation most everywhere and the inability to deploy it cost effectively), and we do a disservice to those who die or suffer because we could do better.
This is maybe the best reason to allow autodrive. We need to let it learn because it will get better. Whatever we pay now in terms of lives lost is made back in the future, in lives saved.
All depends on our faith in the system getting better of course. There's always going to be dissatisfying things surrounding it.
Gotta be up front about the state of the art though, can't let people sell a thing like it's the finished item.
I absolutely agree with your entire comment. The system must be taught, but everyone must participate to the best of their ability. Can't mess it up, there are people we haven't even met yet who are counting on us, collectively, to do better.
Considering the resources spent on Tesla's “self-driving” systems… if you used that money on public transport, how many lives could you (improve and) save by reducing the number of cars on the road? (My guess: you could build a massive amount of public transport on a local scale, but not much on a global scale, so you'd probably only save a several dozen lives and significantly improve a few tens of thousands.)
Factoring in that bad drivers start being really negligent when there cars are “self-driving”, I'm not certain this will save more lives than the opportunity cost. But if it's mostly used by good drivers, and bad drivers get off the road, it could be huge.
Public transportation in the US will never replace the automobile. There is simply too much infrastructure to replace and not enough GDP. It's a fine solution for SF, Chicago, and NYC, but not the rest of the country.
Illinois used to have electric rail connecting every podunk town for a nickel. It was called the "interurban" and was scrapped for the war effort in the 1940s. The reason half the towns (that now require cars) even exist is because the train had to stop for fuel and water. Not enough GDP is a pretty sad state of affairs for the richest nation on earth.
US urban sprawl is costing so much GDP it is not even funny. And i don't understand why they don't invest in septic and would rather install a full sewer system for less than 200 people.
Tesla has spent about $8.5B on R&D in the past decade. Of course, not all of that was spent on FSD, but let's assume it was.
Let's also assume our hypothetical public transport is built in San Francisco. San Francisco recently spent $1.5B on a 1.7 mile subway line. It will also take a decade from when work begun to its opening next year.
That works out to be a 9.7 mile subway that opens in 2032.
You can't conflate overall R&D development with manufacturing and construction costs, that's a fallacy. I don't know what the exact numbers are but Tesla has definitely invested well over $100b in their supercharger networks, manufacturing facilities, and even cost per car and it will still be far less of a public good affecting far less less people directly and indirectly than a reasonable public transport network.
Its going to be a tradeoff that current USA safety first culture is ill prepared to deal with appropriately. ~30,000 people die and massive injuries and property damage happen each year in the USA. How many people's deaths should we accept as a society to reduce those numbers by, say, two orders of magnitudes? 1,10,100,1000,10000? I hope we don't choose not to develop this tech just because automatic cars will kill 100's of people per year but save ten of thousands also.
A million people are killed on the roads yearly world wide. How many deaths would you accept to reduce that to 10,000/year?
If a human driver kills somebody in an accident, he is responsible.
Who is if the drives autonomous?
I guess still the driver but he wouldn't feel responsible or he would never active autonomous driving.
Incorrect. I've had the chance to use it. I'd take responsibility, in the same way I must whenever I operate a vehicle. They are good systems and frankly already handle a lot of tasks better than people - like driving at 50+MPH on busy highways, because they don't get bored, tired, or angry.
What "USA safety first culture"? If you kill someone with your car today, mostly nothing happens. The system is supremely setup to excuse even the most negligent, reckless behavior as long as it happens in a car. (Any car! Bought a huge, pointless truck with impossible sight lines? No problem!)
As it stands, being able to blame Autopilot has a great chance of improving the system. What do you think happens when the Tesla Autopilot has one of its predictable jerks and maims a pedestrian? YouTube is full with documentation of just these kinda problems, and suing Tesla in civil court promises a lot more cash than some random deadbeat.
Driving regulations have always been inconsistent. Just the fact motorcycles are legal despite their 24x mortality rate should tell you it has much more to do with politics than safety.
It will be all boring company tunnels on mars, no surface roads. So no moons.
A lot of autonomous car problems would go away if companies and governments went for enhanced roads instead of trying to create a generic solution on the first try. No need to detect traffic signs or lights if that information was made readily available by the road you are driving on.
I'd go farther and say that it's not possible to drive correctly in some cases on current roads. Everyone who's driven east/west with the arc of the sun parallel to the road knows this.
Sounds an awful lot like Positive Train Control, which has a hard enough time keeping communication up with just one vehicle per track section. You're not going to get very many nines of uptime when dealing with electronics exposed to the environment.
I think you wouldn't have to go that far, even things like identifying traffic lights would be a lot easier if you had an up to date list of their locations. Worst case I can think of are car accidents where the people involved have to place warning signs on the road and construction/cleanup crews that are slowly moving along the road, but that would still be a limited amount of exceptions to handle instead of trying (and failing) to handle everything.
The choice to remove radar has definitely reduced the chances that I will buy a tesla. It would be one thing if they lowered the price, but the "full self driving" upgrade continues to be $10k usd.
Plus the cost of the latest FSD computer upgrade, even if Elon Musk promised your old model "has all the required hardware for Level 5 autonomy". Your fault for believing him, pay up.
And even then - all the features are in perpetual Beta and don't work.
I've had a Model 3 for 3 years with FSD. My experience:
Autopark - trying to get it to show up is almost impossible. When it does show up, the chance of curb rash is between 95-99%.
Summon - for the most part works if it can connect. still does some sketchy stuff some times.
Smart Summon - literal piece of shit that will always do the wrong thing and then just stop in the middle of the road.
Navigate on Autopilot - kind of works, but it makes the auto-lane change feature finicky. Without NOA, auto lane change works fine, with it enabled, fails 80% of the time with the car jerking back into the original lane.
I saw that. I do like that I would be able to try it out for a month. I would be much more amenable to it if they took the monthly fee off the flat purchase price.
And yet, most of the time, humans don’t seem to require radar to distinguish the moon (or a big yellow light far away) from a traffic light.
I think stereo vision won’t help humans much at the distances we’re talking about (might be different for a car with cameras over a meter apart, but that will suffer from bad resolution compared to the human fovea).
Parallax probably is a clue we use. You want to brake gently to optimize comfort for the passengers, so let’s say you want to start braking 10 seconds away from the traffic light. Since sin(x) ≈ x, the image of the traffic light will move up about 10% every second at that moment.
On a camera, that could be less than 100 pixels, but I think that should be detectable.
⇒ you don’t have to discriminate between the moon and a tragic light, you just have to find the nearby yellowish blobs.
Is it publicly known what Tesla’s self-driving does? Given the above, I think I would try to extract a depth field from the known speed of the car and the camera images, discard or deprioritize whatever is more than 20 seconds out, and then let some ML algorithm loose on detecting points of interest such as traffic lights, cars, etc. (As opposed to “feed a ML algorithm lots of video, and hope it will learn to deprioritize stuff that’s far away)
Neural nets esp. those than can run several times a second on a car (not in datacenters) are still very far from human level of perception.
A human disambiguates by looking around, lots of prior knowledge about how the world works, etc. That doesn't seem to be the case with the current Teslas.
The Tesla keeps seeing it as a traffic light then rejecting it as not a yellow light over and over in the UI. Just the right atmospheric conditions plus just the right angle probably don’t show up a lot in it’s sample sets.
GPS gives accurate timing and location which should make filtering out the moon straightforward, but it might just be overly cautious as a level 3 system.
It's not really rejecting it, it's just getting confused why the yellow light is still there when stationary objects typically fly by at about the same speed you're moving forward.
Whatever stereo-depth perception of the camera they're using apparently can't tell the difference between where a traffic light should be and a celestial body. OTOH, I have mistaken a Burger King light-up sign for a full moon when it was coming up behind some trees, so I can't fault them too badly.
The full clip shows the false yellow light is absent from the UI for long periods of time.
Watch a Tesla UI for a while and you will regularly show objects being reclassified. What’s breaking that loop is the fact the moons relative position is fixed so after rejecting it as not a traffic light the software then sees that same input and starts treating it like a new object after a while. I don’t know if the slowing behavior is related to the misclassified moon or a response to the model being confused.
My significantly dumber car keeps picking up speed limit signs for side roads and supermarkets, sometimes as low as 5mph. I’m glad it cannot slam on the brake when it sees them!
92 comments
[ 2.7 ms ] story [ 173 ms ] threadI'd like to see a Tesla on full autonomous autopilot take a drive in Calgary or Edmonton a few days after a heavy snow, when it's been mostly plowed but the lines are missing from the roads, all the lines at the side of the road are covered in snow, etc.
Often when driving like this, you see a glimpse of the middle-yellow line inside one of the wagon tracks, and you realize that you're actually driving in part of the other lane. Fortunately the opposing traffic's wagon tracks are also shifted so that they're driving partially on the shoulder.
A GPS guided car would make a mess of the wagon track pattern that all the humans decided to just go with.
When it's fully rolled out, it's going to kill a loooot of people.
Do you have any data to back this up?
Because there is definitely evidence that shows that it is saving lives: https://twitter.com/Model3Owners/status/1383545728488808453
I think it is more fruitful to realize that humans also make mistakes. It may also be also beneficial to compare Human + FSD driving vs. Human driving (instead of just FSD vs Human).
What about when Tesla's marketing says FSD can do[2] things it can't do?
[1] https://www.thedrive.com/tech/39647/tesla-admits-current-ful...
[2] https://old.reddit.com/r/TeslaLounge/comments/oey4g1/for_tho...
It doesn't mean they're directly lying, in the sense of "You can't reasonable get these numbers out of the dataset." But it probably means that there are some games being played such that the number you're getting isn't the number it's strongly implied (or, occasionally, outright stated) to be. Or, if it is, it doesn't mean what they'd like you to think it means.
So, if you go back a few years to the first waves of numbers Tesla trotted out to "prove" how amazing Autopilot was, you roughly had:
- The fatality rate per mile of all drivers, in all vehicles, in all conditions
being compared with:
- The fatality rate per mile of expensive luxury cars, driven by older, well off drivers, mostly on limited access divided highways, in "easy" driving conditions.
You can certainly do the math and get a comparison, but it's not a particularly reasonable comparison, because:
- In general, six figure luxury cars are really, really safe. They're big, heavy, loaded with airbags and safety features, and really don't kill people very often at all. You can go look at a breakdown of death by vehicle type and class, and it turns out that luxury cars are very much under-represented in fatalities.
- Expensive cars tend to be driven by older, more experienced drivers. There's a bathtub curve of fatality rates by age (https://aaafoundation.org/rates-motor-vehicle-crashes-injuri...) - younger drivers without much driving experience die more, and once you get to 70+, your fatality rate goes up, but there's a sharp drop going to the 30-39 group, continuing to decline up to the 60-69 group. Guess what age range is going to be mostly buying Teslas?
- Divided highways tend to be fairly safe places to drive - especially in lower speed situations where a lot of people are going to be using Autopilot. You're exceedingly unlikely to die doing 20mph in rush hour traffic. Even running at speed, limited access highways are fairly safe places.
- And Autopilot simply doesn't work in bad enough weather. Again, a lot of traffic deaths are in pretty sketchy conditions - rain, snow, sleet... things that older Autopilot, especially, simply won't work in.
So, yes, a straight up comparison of the two numbers ("fatalities per million miles in Teslas with Autopilot engaged" vs "fatalities per million miles of all vehicles in all conditions") is at least somewhat deceptive.
For one thing, Autopilot is more likely to be engaged for some types of driving than others. Comparing statistics with Autopilot engaged to the total of all miles driven is comparing apples and oranges.
Similarly, it's comparing statistics about the people who buy and drive Teslas with the entire driving public. You don't know how many accidents Tesla owners would have gotten into if they were in cars that didn't have self-driving features.
Finally, you can't draw any conclusions about whether lives were saved by just looking at the number of accidents and ignoring their severity. Hypothetically, if Autopilot caused fewer accidents at low speeds but more catastrophic collisions on highways, it could be easily be causing more deaths than it prevents.
In fact most things are better done by humans, because they understand context, which most AI/ML currently doesn't.
AI is great for tasks that can be done without much context and can be incredibly efficient and fast compared to manual work. Everything else, including "self-driving" currently needs supervision or produces terrible results.
The software will improve faster than humans will, as arguably, we've already hit our limit. Software doesn't get tired and it doesn't drive impaired. It makes mistakes, but so do we, and it's mistakes can be improved upon at scale. This is not blind faith in technology or software, but an acceptance of the limited capabilities of your median human (as well as the disadvantaged outliers).
[1] https://www.nytimes.com/2021/07/01/business/boryana-straubel...
This is not to say that software and hardware will not quickly overtake us, but while humans are on the road and in control, improvement will be possible.
You are right, the software should absolutely be held to a high standard (otherwise we're socializing losses and privatizing profits, also bad!), but please don't perpetuate a lie that existing drivers are held responsible to the degree they should for their mistakes (or even intentional acts) that cost lives. You'll find examples where someone is made an example of, but that is rare.
So disagree, please (checks and balances are important, as is the dialog we're having), but having seen what humans have to offer in this context, I will be the first, on my own time and resources, in front of regulators to argue for aggressive driver assist system development and implementation. Because the alternative is hot garbage we get by with because there are limited alternatives (ie no public transportation most everywhere and the inability to deploy it cost effectively), and we do a disservice to those who die or suffer because we could do better.
All depends on our faith in the system getting better of course. There's always going to be dissatisfying things surrounding it.
Gotta be up front about the state of the art though, can't let people sell a thing like it's the finished item.
Factoring in that bad drivers start being really negligent when there cars are “self-driving”, I'm not certain this will save more lives than the opportunity cost. But if it's mostly used by good drivers, and bad drivers get off the road, it could be huge.
I found lots of photos from the 50s trying to find more info
https://www.cera-chicago.org/Blog/3315468
We will never get massive public transport, no matter how much money is spent. That is the harsh reality
Let's also assume our hypothetical public transport is built in San Francisco. San Francisco recently spent $1.5B on a 1.7 mile subway line. It will also take a decade from when work begun to its opening next year.
That works out to be a 9.7 mile subway that opens in 2032.
A million people are killed on the roads yearly world wide. How many deaths would you accept to reduce that to 10,000/year?
Zero.
As it stands, being able to blame Autopilot has a great chance of improving the system. What do you think happens when the Tesla Autopilot has one of its predictable jerks and maims a pedestrian? YouTube is full with documentation of just these kinda problems, and suing Tesla in civil court promises a lot more cash than some random deadbeat.
A lot of autonomous car problems would go away if companies and governments went for enhanced roads instead of trying to create a generic solution on the first try. No need to detect traffic signs or lights if that information was made readily available by the road you are driving on.
https://www.tesla.com/support/transitioning-tesla-vision
Plus the cost of the latest FSD computer upgrade, even if Elon Musk promised your old model "has all the required hardware for Level 5 autonomy". Your fault for believing him, pay up.
* Isn't actually Full Self Driving
I've had a Model 3 for 3 years with FSD. My experience:
Autopark - trying to get it to show up is almost impossible. When it does show up, the chance of curb rash is between 95-99%.
Summon - for the most part works if it can connect. still does some sketchy stuff some times.
Smart Summon - literal piece of shit that will always do the wrong thing and then just stop in the middle of the road.
Navigate on Autopilot - kind of works, but it makes the auto-lane change feature finicky. Without NOA, auto lane change works fine, with it enabled, fails 80% of the time with the car jerking back into the original lane.
Can't wait for FSD though /s
So if you’re in Europe/Asia/Oceania or in the market for an S/X - you’re in luck!
I think stereo vision won’t help humans much at the distances we’re talking about (might be different for a car with cameras over a meter apart, but that will suffer from bad resolution compared to the human fovea).
Parallax probably is a clue we use. You want to brake gently to optimize comfort for the passengers, so let’s say you want to start braking 10 seconds away from the traffic light. Since sin(x) ≈ x, the image of the traffic light will move up about 10% every second at that moment.
On a camera, that could be less than 100 pixels, but I think that should be detectable.
⇒ you don’t have to discriminate between the moon and a tragic light, you just have to find the nearby yellowish blobs.
Is it publicly known what Tesla’s self-driving does? Given the above, I think I would try to extract a depth field from the known speed of the car and the camera images, discard or deprioritize whatever is more than 20 seconds out, and then let some ML algorithm loose on detecting points of interest such as traffic lights, cars, etc. (As opposed to “feed a ML algorithm lots of video, and hope it will learn to deprioritize stuff that’s far away)
A human disambiguates by looking around, lots of prior knowledge about how the world works, etc. That doesn't seem to be the case with the current Teslas.
The Tesla keeps seeing it as a traffic light then rejecting it as not a yellow light over and over in the UI. Just the right atmospheric conditions plus just the right angle probably don’t show up a lot in it’s sample sets.
GPS gives accurate timing and location which should make filtering out the moon straightforward, but it might just be overly cautious as a level 3 system.
Whatever stereo-depth perception of the camera they're using apparently can't tell the difference between where a traffic light should be and a celestial body. OTOH, I have mistaken a Burger King light-up sign for a full moon when it was coming up behind some trees, so I can't fault them too badly.
Watch a Tesla UI for a while and you will regularly show objects being reclassified. What’s breaking that loop is the fact the moons relative position is fixed so after rejecting it as not a traffic light the software then sees that same input and starts treating it like a new object after a while. I don’t know if the slowing behavior is related to the misclassified moon or a response to the model being confused.