Could it? There's a lot of signals pointing in different directions but it's pretty hard to miss both the one directly in front or the one right next to the car.
Plus, if you see both red and green lights, your first instinct as a driver (human or not) should be to stop and re-examine more closely, not run through it.
I suspect the driver, based on their commentary in the video, was trying to see if this was a one-off fluke or something more consistent (they mention the car already blew through that light once before).
I would hope a driver using FSD or any other system like it would be alert/aware to intervene...but for this video, it looks like they were trying to see how the car would react.
I'm insinuating they they knew they were approaching an intersection that they knew FSD had trouble with, purposefully did not slow down the cruise control to attempt a safe approach, and purposefully did not tap the brakes even when it was obvious the car was not stopping or slowing down in a safe manner, and then reacted on camera insinuating it was entirely surprising to them.
If none of this is true, than they are incredibly shitty drivers that were not paying attention to what was happening and let the car cruise through a red light. The exact same scenario would have occurred with a standard 1990 era cruise control system.
Well duh they're shitty drivers with no respect for a public road, they bought Tesla "full self driving" """beta""" and are testing it against unwilling other drivers!
"They're a bad person so this must be fake", or maybe they're a bad person which is exactly why they thought doing something like this is okay
I agree with your first paragraph. They tested the red light issue after experiencing it to make sure it was not a fluke. Their yelling was unnecessary.
...yes? It's pretty clear from the video this was done as a "test."
The article mentions the drivers experience working on FSD and they have a YouTube page dedicated to stuff like this linked in the article.
So the test involved letting the car go at 30mph through a live intersection, not taking any emergency steps such as checks notes stepping on the brake when approaching the red light at 25mph, and then cursing about it afterwards and STILL not making any attempts to correct it?
Unfortunately, yeah, seems that way...just like you said "they chose to do nothing" and luckily no one was in/approaching the intersection.
Look I'm all for knowing this information, that something like this can happen. But to be clear, in no way do I condone how it was done. This seems reckless for the sake of clicks.
There is a green light literally right next to it. I was also initially confused as to which lane either was pointing at. They are not offset angle wise very much.
In general I agree, but yesterday I saw a car blow through a stop sign at an intersection and then honk at the cars going perpendicularly through the intersection (they had no stop signs) :/ people are pretty happy to be confidently wrong.
And that's fine, but if you had a button to sell that exact person's driving to everyone as "full self driving", and you pressed it, I hope you understand most people would consider that a problem.
Just watched the video. Any driver who cannot correctly tell on an oblique intersection which light is theirs is a driver who should be fined until they can..or have their license taken away. Any programmer can clearly tell why FSD got it wrong (It saw the glow of the other road's green light on the angle) but nonetheless it was wrong and legally wrong.
Whoever is responsible for the intersection should also be fined. Our whole safety culture is based on being as clear as possible with redundancies from the consumer and the provider (i.e., assume the worst, provide the best).
Absolutely, but I think what most humans would do is slow down - especially at night - and not go through the intersection at 27 when they were approaching at 30. FSD appears to just make a binary "go/stop" decision and commit fully rather than slowing down when there is decreased confidence as a person would/should.
Definitely poor intersection design though as well. I've seen lights with extensions around the bulb to prevent them from being seen from obtuse angles, I wonder why these lights don't have those? Especially at night it would completely block your vision of the light(s) that don't apply to your road.
If the Tesla's self driving software had to be limited in speed based on it's confidence, it wouldn't be able to driver faster than 15 mph. Tesla's self driving is ENTIRELY based around "Fuck it, I'll probably be fine"
Everyone always gets upset at me when I bring this up. The infamous fatal "autopiloted" Model X crash in the Bay Area happened because the road lanes were painted to drive the car into the narrow side of a jersey barrier. The next day a near fatal accident happened in the exact same spot, likely because Caltrans as an organization has little to no liability and generally doesn't stay on top of freeway conditions.
There is an intersection in Houston that is one of the most dangerous in the USA that kills a dozen people a year due to poor design and planning from TxDOT. There are probably hundreds of intersections like this around the country that will need to be fixed.
If I go paint road lines into a wall painted like a tunnel, are the drivers who hit the tunnel wall responsible, or is it me who painted the road lines?
I don't buy any of the excuses, honestly. If the software can be that easily fooled by slightly broken white lines, it's not ready for release. That area of 101 simply isn't that difficult to parse (for humans). If it can't deal with that, how can it driven in places where the roads are truly awful? Lots of places in the USA alone with much worse road conditions than anything CalTrans is responsible for.
Also, there's a huge yellow and black chevron block visible in that picture that any human driver paying attention would have seen and easily avoided. Crosshatching is unnecessary if you're paying the least bit of attention to what's in front of you. What was the software's issue here?
It is absolutely a problem with the Tesla software that it wasn't able to parse that area properly. Doesn't necessarily mean the human driver wasn't at fault.
There are a dozen news articles that say it was missing, and even your picture looks like it's just the mount point, not the crash arresting system itself.
Page 2 (which is actually the 22nd page of the pdf) shows an image of the pre-crash state of the attenuator. It is present, but non-operational due to a previous crash. While that crash affected it's usefulness, it's visibility is only marginally reduced.
Sure, if you take a picture cutting out the context that makes it clear it's not a lane, then maybe it looks like a lane. I drove by that every day for years and never thought there was a lane there... Because it's absolutely obvious from the painted lines diverging from a point that it's not a lane. Now, can we make things even more obvious for distracted drivers and shitty AI, sure, and I'm all for that.
Thing is, a lot of exit lanes have white lines you have to cross, which I've wondered about as a driver. Usually you have to cross a white line to get into the hov lane, like an cloverleaf/onramp usually around a meter.
It’s an extremely common pattern for the lights for one direction of an intersection to be visible from the other (usually at an angle). Very few human drivers would confuse the two. At the very least they would slow down until they were able to confirm rather than driving full speed ahead like this car seemed to do.
So the road was pretty shittily marked, I was having a hard time figuring out wtf was going on (within the limits of the mediocre video), so its something that needs to be worked on.
What I don't get is why the driver/passenger are freaking out on camera. I'm going to guess its clearly for the social media outrage. They were already saying it was a red light before they passed it. They could have pressed the brakes at any time, but they decided to spend the time cursing. The car was already clearly approaching way too fast for the intersection (~25 mph) and they chose to do nothing.
I think they could see that no cars were coming from their vantage, so breaking wasn't necessary from a safety point. You are kinda underselling the whole "something that needs to be worked on." since you know... its in a bunch of cars currently on the road with other people who might not like to be part of the experiment.
The experiment that demonstrably makes cars safer than a human driver? Why would people not want to be a part of that? I sure do, I want to be safer when driving.
I think we just saw that these cars will blow through this red light… maybe even 100% of the time. That doesn’t align with your “demonstrably makes cars safer than a human driver.”
You want to pretend like we jumped to the engineering end and it’s a solved problem. It’s not. I know if my family drove through this intersection daily, I wouldn’t find the risk as acceptable.
There are all kinds of spots around the country where FSD will make mistakes with a high likelihood. Tesla has already released stats that plainly state that while autopilot or FSD are engaged, accidents are way less likely to happen than if they weren’t.
It’s part of the contract for beta users that you have to pay full attention to the road ahead and be ready to react at all times if the system makes a mistake. In the case of the clip in OP, they deliberately didn’t react and should be kicked off the program by Tesla and fined by the police.
>Tesla has already released stats that plainly state that while autopilot or FSD are engaged, accidents are way less likely to happen than if they weren’t
Because they are way more likely to be engaged on the easy miles and they are still bad enough that the driver pays attention
And if the accident happens right after the driver panic brakes, thereby taking control of the car... Does that count as an FSD accident or a driver-in-control accident?
I would be willing to bet, with Tesla's fraudulent ways, they are counting a brake before an accident a driver-in-control accident. They should be required to publish the time-since-FSD of every single accident in a Tesla vehicle.
There is no independently verifiable data in a great many publications of various large companies for obvious reasons. That’s why they’re legally forbidden from releasing bogus data.
What you should and shouldn’t trust is entirely subjective, I have no reason not to trust the stats Tesla has released. If Autopilot wasn’t good enough to warrant releasing those stats, they simply wouldn’t.
You’ll have to do some deeper statistical analysis to discredit what they’re saying.
> You’ll have to do some deeper statistical analysis to discredit what they’re saying.
I don’t need to discredit it, my default is to not believe it unless there has been verification by an independent body. Over time you learn to look at the wording of claims from a company and read between the lines. They may have cherry picked some stat under certain conditions but it won’t be representative - the small print will read stuff like “highway miles, during daylight hours when compared to similar vehicles” which gives them a bunch of wiggle room.
Many companies claim things in their marketing material, I’ve never know anyone to take it at face value but todays the day I guess.
if there were cars entering the intersection crossing the car would have detected that and hit breaks also, it detects cross traffic as well, which overrides a "green" light intersection from a confused tilt of the light
Even more strange, it was the second time it had done this. The person filming went back and did it again. At that point I feel like the correct response is "Yep, see, it just did it again" but instead it was played up for the video. That said, I'm not pro-Tesla nor do I think FSD is ready, just that this is a bad example.
How is that road shittily marked? I could see, even from the crap video, exactly where to drive. Perfectly fine double lines in the middle, perfectly visible edge markings on the right side. Perfectly visible red light. Yeah there's a crappy slight bend right before the light and before you turn that bend that puts you head on with the red light telling you to stop, you're basically going straight towards a green light that's meant for the road coming from the right. I bet it got confused by that.
Which says something about his ability at this whole thing I guess, if he can't get it right even after having been through it. So, I guess you are saying it's not just shittily marked (which it's not) but also the lights are confusing, I re-watched it to check that.
The red light comes into view first actually and only then does the green one come into play, but fair enough that this doesn't say much. As soon as you're past the big bend and going straight towards the light (i.e. before the final bend) I can clearly see one red light that the car and driver are facing head on and another green light, which is slightly turned on an angle from the car/driver. As we get closer, we can even see that they did this on purpose, i.e. the red light is purposefully not right above the road and facing the road head on (which would only be the tiny part of if after the slight right bend) but is clearly meant for the people coming around the big bend, so they can see from afar that there's a red light. Then as you turn the slight right bend before turning left onto the main road, you have the regular lights as you would see at any t-intersection and they're clearly red.
Also any driver that is unclear of what the lights mean should simply slow down so they have more time to survey the situation, which if they interpret all the perfectly visible lights correctly at a late point only, they are slow enough that breaking would stop them where they should stop.
Instead it seems that people are siding with the auto-pilot, which thinks that he very clearly has a green light and takes the intersection at speed. It barely slowed down from 30 to 27 before taking the actual left turn of the intersection so it seems to have been very sure of what it's seeing. So if people are saying that this is an unclearly lit intersection, the auto-pilot disagrees and it thinks that it very clearly is that green light that's meant for the road its on.
I saw a massive car accident happen at night at an intersection. A drunk driver was coming down a hill rated at 45. He was doing 60 probably. Tboned someone turning on their green arrow across the intersection. Massive accident I stayed to help with for a bit.
A few days later I was going down the hill at night just like the drunk driver. For a second I saw the upcoming light and then one soon after it overlap because of the gradient on the hill and the distance between the lights.
I could definitely see AI and people not paying attention getting overlapping lights like this confused.
Pretty confusing intersection with the light not being above the lane. He himself wasn't sure at the start of the video which light was his. He was a "advanced driver assistance systems test operator" and didn't know that disengaging is how you teach it to do the right thing and that you should disengage if you are about to run a red? Posting it on social media is just the icing on the cake for him being terminated.
It's worth reiterating that even if FSD in all cars resulted in car crash deaths that numbered 1% of current annual deaths — 45,000, approximately 125/DAY in the U.S. — those 450 ANNUAL deaths would be considered a FAIL of FSD.
What’s with all the comments stating that the intersection could fool some human drivers too?
First, looking at it on 2D video shot at night is very different than actually being there. But more importantly, the acceptable threshold for autonomous vehicles is not and never will be “Better than the worst drivers.”
Everyone on here complaining about the road markings in the video is missing the point. Roads are frequently poorly maintained, and the markings are not consistent even from city-to-city (Or within cities). To be a reasonably full-self-driving vehicle, you need to be prepared for the imperfection which is roadways in the US (or anywhere).
we often make up for this in human judgment, but the machines are unable to make up for that, and that's a major gap.
78 comments
[ 4.1 ms ] story [ 150 ms ] threadPlus, if you see both red and green lights, your first instinct as a driver (human or not) should be to stop and re-examine more closely, not run through it.
I would hope a driver using FSD or any other system like it would be alert/aware to intervene...but for this video, it looks like they were trying to see how the car would react.
This entirely feels like a manufactured scenario since they simply did not act like a informed driver, and decided to freak out.
If none of this is true, than they are incredibly shitty drivers that were not paying attention to what was happening and let the car cruise through a red light. The exact same scenario would have occurred with a standard 1990 era cruise control system.
"They're a bad person so this must be fake", or maybe they're a bad person which is exactly why they thought doing something like this is okay
Look I'm all for knowing this information, that something like this can happen. But to be clear, in no way do I condone how it was done. This seems reckless for the sake of clicks.
But the serious issue with the FSD doesn't become moot because of this.
Definitely poor intersection design though as well. I've seen lights with extensions around the bulb to prevent them from being seen from obtuse angles, I wonder why these lights don't have those? Especially at night it would completely block your vision of the light(s) that don't apply to your road.
There is an intersection in Houston that is one of the most dangerous in the USA that kills a dozen people a year due to poor design and planning from TxDOT. There are probably hundreds of intersections like this around the country that will need to be fixed.
If I go paint road lines into a wall painted like a tunnel, are the drivers who hit the tunnel wall responsible, or is it me who painted the road lines?
The white lines of the two adjacent lanes made it look like there was a well-defined lane there, with the barrier right in the center.
Here's a pretty good picture of what it looked like then:
https://www.autopilotreview.com/wp-content/uploads/2019/03/m...
You can see there is no hatching to make it obvious it was a non-drivable area.
I would have expected something like this might have made a difference:
https://driveractive.com/Assets/photos/rm25.png
(that said, why didn't radar help?)
Also, there's a huge yellow and black chevron block visible in that picture that any human driver paying attention would have seen and easily avoided. Crosshatching is unnecessary if you're paying the least bit of attention to what's in front of you. What was the software's issue here?
It is absolutely a problem with the Tesla software that it wasn't able to parse that area properly. Doesn't necessarily mean the human driver wasn't at fault.
I do think that it had already stopped at least one, maybe multiple crashes since it has been last replaced.
The discussion here though is about visibility, and the attenuator would have been quite obvious there.
Page 2 (which is actually the 22nd page of the pdf) shows an image of the pre-crash state of the attenuator. It is present, but non-operational due to a previous crash. While that crash affected it's usefulness, it's visibility is only marginally reduced.
What I don't get is why the driver/passenger are freaking out on camera. I'm going to guess its clearly for the social media outrage. They were already saying it was a red light before they passed it. They could have pressed the brakes at any time, but they decided to spend the time cursing. The car was already clearly approaching way too fast for the intersection (~25 mph) and they chose to do nothing.
You want to pretend like we jumped to the engineering end and it’s a solved problem. It’s not. I know if my family drove through this intersection daily, I wouldn’t find the risk as acceptable.
It’s part of the contract for beta users that you have to pay full attention to the road ahead and be ready to react at all times if the system makes a mistake. In the case of the clip in OP, they deliberately didn’t react and should be kicked off the program by Tesla and fined by the police.
Because they are way more likely to be engaged on the easy miles and they are still bad enough that the driver pays attention
I would be willing to bet, with Tesla's fraudulent ways, they are counting a brake before an accident a driver-in-control accident. They should be required to publish the time-since-FSD of every single accident in a Tesla vehicle.
There is no independently verifiable data that supports this.
You’ll have to do some deeper statistical analysis to discredit what they’re saying.
I don’t need to discredit it, my default is to not believe it unless there has been verification by an independent body. Over time you learn to look at the wording of claims from a company and read between the lines. They may have cherry picked some stat under certain conditions but it won’t be representative - the small print will read stuff like “highway miles, during daylight hours when compared to similar vehicles” which gives them a bunch of wiggle room.
Many companies claim things in their marketing material, I’ve never know anyone to take it at face value but todays the day I guess.
The red light comes into view first actually and only then does the green one come into play, but fair enough that this doesn't say much. As soon as you're past the big bend and going straight towards the light (i.e. before the final bend) I can clearly see one red light that the car and driver are facing head on and another green light, which is slightly turned on an angle from the car/driver. As we get closer, we can even see that they did this on purpose, i.e. the red light is purposefully not right above the road and facing the road head on (which would only be the tiny part of if after the slight right bend) but is clearly meant for the people coming around the big bend, so they can see from afar that there's a red light. Then as you turn the slight right bend before turning left onto the main road, you have the regular lights as you would see at any t-intersection and they're clearly red.
Also any driver that is unclear of what the lights mean should simply slow down so they have more time to survey the situation, which if they interpret all the perfectly visible lights correctly at a late point only, they are slow enough that breaking would stop them where they should stop.
Instead it seems that people are siding with the auto-pilot, which thinks that he very clearly has a green light and takes the intersection at speed. It barely slowed down from 30 to 27 before taking the actual left turn of the intersection so it seems to have been very sure of what it's seeing. So if people are saying that this is an unclearly lit intersection, the auto-pilot disagrees and it thinks that it very clearly is that green light that's meant for the road its on.
Some folks also seem to find it reasonable to let others know when something is unsafe.
Seen in that light, I don’t really see an issue and don’t think it’s cool to diminish the video by suggesting it was merely to produce outrage.
I, for one, don’t want to run into a wayward Tesla while on the highway.
A few days later I was going down the hill at night just like the drunk driver. For a second I saw the upcoming light and then one soon after it overlap because of the gradient on the hill and the distance between the lights.
I could definitely see AI and people not paying attention getting overlapping lights like this confused.
In those instances, the (temporary) light box was tilted enough to make it seem like the light could have been for the parallel road and not us.
I didn't bring it up because Beta detects and reacts to pretty much every other light like you'd expect.
First, looking at it on 2D video shot at night is very different than actually being there. But more importantly, the acceptable threshold for autonomous vehicles is not and never will be “Better than the worst drivers.”
we often make up for this in human judgment, but the machines are unable to make up for that, and that's a major gap.