Phantom braking so bad I want to return my car

[sleepydoc]

And this is the point. What you are asking for is a system with low false positives and reliable true positives (i.e. low false negatives). And that's fine, we all want that. But all these other brands you hold up as examples only have (to your knowledge) low false positives .. you have no way to know what the rate of false negatives is (that is, the car fails to brake when it should). So they fail to meet your own criteria for "good" examples of AEB.

If all you want is a car that has very low false positives for AEB, that's easy .. get a car wtihout AEB entirely (or turn it off). Problem solved!

You have multiple flaws in your reasoning.

First, you conveniently assume that 'all the examples' have a high false negative rate and therefore fail and then by implication use that to justify Tesla's high false positive rate. Since you have no data on Tesla's false negative rate the conclusion that Tesla's system is better than (or is even as good as) other cars' systems is nothing but speculation. Second, you provide no evidence that other cars' systems are failing.

Then you say, well, if all the false positives are a problem, just turn it off! Really? Your solution to a system that's so poor that it drives people crazy is to turn it off? You're basically admitting it's a failure.

We have more than ample evidence that Tesla's system has a high false positive rate. People give all manner of excuses: "it saw a shadow," "there was a truck," "it was confused by a snow drift," or, the best one of all "it must have seen something, you just missed it." How about the obvious reason: it's just a crappy system. Ultimately, in all of these cases the car is braking when it doesn't need to which is the definition of a false positive.

AEB systems are by definition emergency backup systems for rare circumstances. If you set a goal of catching all cases with any system that is less than perfect you end up with a high false negative rate to catch a few rare events, some of which would have been caught anyway but do so at a cost - cars brake unnecessarily, potentially getting rear ended or causing rear endings further down the line, aggravation, poor driving experience, etc. Then they turn the system off as you suggest and we're actually worse off then we would have been if they had set the sensitivity properly in the first place. Ah, but you say "if someone rear ends you it's their fault for not paying attention." Well, if the system has a false negative and you hit something, it's your fault for not paying attention. See how that works? it's not that tough.

Finally, you're assuming that all of these events are due to the AEB system. If they are then the AEB system activation in Teslas is probably a thousand fold higher than pretty much every other car I've driven (no, I'm not exaggerating there, I'm underestimating if anything.) If they are not AEB events then we not only have an issue with the AEB system but evidence that the other systems have problems too. Either way, Tesla is failing here.

 

[sleepydoc]

Have they really?

Federal Regulators Investigate Honda Unexpected Braking Complaints - Consumer Reports

Consumer Reports shares details of NHTSA's investigating consumer complaints that the automatic emergency braking system on some Honda vehicles unexpectedly activates.

www.consumerreports.org

well, yes. Like everyone else here you are either intentionally or (hopefully) unintentionally mixing AEB systems with adaptive cruise systems. I haven't used or talked to anyone who's used Honda's adaptive cruise system but false AEB activations are a separate issue (which Tesla also has, by the way)

 

[sleepydoc]

Well, it clearly wasn't a phantom. I would call it a calibration issue. Is it your car, or is it generalized?
I am sorry if you already said, but have you done a recalibration on your cameras? I know that when we got our car we just drove home. Across Boise, then up and down and up the canyons. It calibrated somewhere along the way. But that was certainly not the 20 miles in the center right hand lane of a 6 lane freeway Tesla recommends. If I start having problems like you describe, that's the first thing I would do (first chance I got. Not convenient).

Yes - I actually tried recalibrating it about 6 weeks ago to see if it would help the problem. Unfortunately it didn't.

 

[sleepydoc]

So you're making an excuse for Tesla? LOL LOL LOL. Come on. Don't be such a fanboi. Every other car manufacturer has figured this out already. Sheesh.

It's amazing the amount of mental gymnastics some people go through to explain, justify and just plain deny that Tesla may actually have a system that doesn't work well.

One of the frequent excuses I see is "TACC was braking for a truck in the oncoming lane." Ok. let's take this at face value - and say it truly thinks a truck is going to hit you. Aside from being wrong, there's the fact that it doesn't brake for cars in the same lane. Why? The car is smaller, but if I'm traveling 45 MPH, hitting a car at 90 MPH really isn't any better than hitting a truck at 90 MPH so why is the car not an issue but the truck is? And what about all the other trucks that it doesn't brake for? Either it's failing in the first case or failing in the second case. Either way it has an unacceptably high failure rate.

 

[TheLex]

Have they really?

Federal Regulators Investigate Honda Unexpected Braking Complaints - Consumer Reports

Consumer Reports shares details of NHTSA's investigating consumer complaints that the automatic emergency braking system on some Honda vehicles unexpectedly activates.

www.consumerreports.org

OK so Honda isn't so great at this. But Honda is not a self-proclaimed high tech company that's claiming to be the market leader in advanced AI. As for my own vehicles, neither has suffered from phantom braking and I've been all over the Western US and Western Canada, including towing a big fifth wheel while using ACC.

 

[Fourdoor]

Friend of mine with a Model Y that's about a year old has put on 40,000 miles, largely on 2 lane highway going north and south here in Idaho. He uses TACC and FSD (which, afaik doesn't use radar anymore) except in towns and just told me he recalls maybe 2 episodes of significant slowing in all of that time. This is on curves and hills and grades and trestles and bridges and narrow roads with log trucks as well as everything else you might imagine.

How can his experience be so different from the daily "slam on the brakes" episodes that others are experiencing? Same software and same basic hardware, presumably. Maybe recent hardware is different, or there are just some lemons out there in this regard.

I absolutely believe some people are experiencing this.

I have, in the 1 week of ownership and 300 miles driving canyon/mountain roads, had a couple of slowings. More like a hesitant let off the gas than actual braking and certainly not slam on the brakes. Both times it felt like the software was being cautious (tight curve ahead, large truck close to the centerline) so I don't even know if that counts as PB.

I hope I don't experience the drastic stuff others are, but I also understand that it's absolutely a judgement call in the software programming and it's a choice between more false positives vs more false negatives. For head-ons, I'll take the false positives if they are what I have and my friend has experienced.

Is he in the FSD Beta program? If so, then yes his radar has been disabled... but it is my understanding that "regular" FSD still uses the full suite of sensors that the car came with.

Keith

PS: As much as I bitch and moan, even I acknowledge that the system is getting better over time. It is now almost as good as it was when I purchased the car 8 months ago!

 

[Fourdoor]

And this is the point. What you are asking for is a system with low false positives and reliable true positives (i.e. low false negatives). And that's fine, we all want that. But all these other brands you hold up as examples only have (to your knowledge) low false positives .. you have no way to know what the rate of false negatives is (that is, the car fails to brake when it should). So they fail to meet your own criteria for "good" examples of AEB.

If all you want is a car that has very low false positives for AEB, that's easy .. get a car wtihout AEB entirely (or turn it off). Problem solved!

Pretty sure if they crash into an oncoming car head on and the download of system data shows that the car didn't brake it would be proof of a false negative. All cars have data logging now and having a bad false negative bias would show, and it would affect their safety rating. If one of the cars mentioned by other participants in this thread has a combination of a reputation for never having PB AND it has a poor safety rating I would suspect that it has bad biasing in it's software.

Keith

 

[Fourdoor]

Yes - I actually tried recalibrating it about 6 weeks ago to see if it would help the problem. Unfortunately it didn't.

Another point on the recalibration idea. If you bring up PB to a service center they do not tell you to recalibrate the cameras, they tell you that it is normal behavior that will be solved through system updates.

When service centers will not take an appointment on an issue and give a generic "that's normal" response to a complaint that means it is wide spread, and that it is a known issue.

Keith

 

[drtimhill]

Pretty sure if they crash into an oncoming car head on and the download of system data shows that the car didn't brake it would be proof of a false negative. All cars have data logging now and having a bad false negative bias would show, and it would affect their safety rating. If one of the cars mentioned by other participants in this thread has a combination of a reputation for never having PB AND it has a poor safety rating I would suspect that it has bad biasing in it's software.

Keith

Ideally yes, someone would be tracking this, but do you know if that is happening? First, someone would have to track all such collisions, THEN track if the car had some form of AEB, and THEN obtain data from the car maker, including if the AEB responded and THEN correlate this data. I admire your optimism, but I'm rather skeptical.

 

[DrGriz]

Is he in the FSD Beta program? If so, then yes his radar has been disabled... but it is my understanding that "regular" FSD still uses the full suite of sensors that the car came with.

Keith

PS: As much as I bitch and moan, even I acknowledge that the system is getting better over time. It is now almost as good as it was when I purchased the car 8 months ago!

I am not sure if he is using beta or not. Kind of think he's not the sort to do that. I can ask, though. People were complaining about PB well before vision only, however. @sleepydoc has a radar car, AFAIK.

 

[sleepydoc]

I am not sure if he is using beta or not. Kind of think he's not the sort to do that. I can ask, though. People were complaining about PB well before vision only, however. @sleepydoc has a radar car, AFAIK.

I have a 2020 Model Y, delivered 7/31/2020. I don't think it has radar but I'm in the FSD beta program so from my understanding radar should be turned off even if I do have it.

Just adding another variable - It's not clear to me how much FSD affect TACC and how much my experiences will differ from others' because of that.

 

[WhiteWi]

Vision only M3 and MY are ever since May of 2021.

 

[drtimhill]

I have a 2020 Model Y, delivered 7/31/2020. I don't think it has radar but I'm in the FSD beta program so from my understanding radar should be turned off even if I do have it.

Just adding another variable - It's not clear to me how much FSD affect TACC and how much my experiences will differ from others' because of that.

The only effect is that, since the FSD stack is built using the TeslaVision stack, cars with FSD beta will NOT use radar even if its installed. None of the FSD stack code has been back-ported to the TACC stack (yet).

 

[Fourdoor]

The only effect is that, since the FSD stack is built using the TeslaVision stack, cars with FSD beta will NOT use radar even if its installed. None of the FSD stack code has been back-ported to the TACC stack (yet).

Are you sure on that? My daytime TACC usefulness has improved by a huge amount in the last month, I had assumed this was due to porting of FSD Beta improvements over to the regular TACC system. Or were you talking about the TACC stack for the radar equipped non-FSD Beta cars?

Keith

 

[drtimhill]

Are you sure on that? My daytime TACC usefulness has improved by a huge amount in the last month, I had assumed this was due to porting of FSD Beta improvements over to the regular TACC system. Or were you talking about the TACC stack for the radar equipped non-FSD Beta cars?

Keith

Reasonably sure, but at some point the dividing lines between these stacks get fuzzy. Certainly TACC is not using the overall FSD BEV vision stack (since that integration is the whole V11 thing Elon talks about). It's quite possible that some of the lower NNs have been shared, but even that I doubt, since it would potentially cause regressions in TACC that would divert attention away from the FSD beta work.

It's quite possible Tesla have been fixing some TACC issues independently of FSD work of course (possible in preparation for the stack unifications).

 

[karuna]

I have a 2021 Model Y Performance. Phantom Braking was a THING until V11. Now its no longer an issue.

Phantom braking is back again. For a few days in between it seemed like it was solved. Not sure if the newer updates brought the problem back.

 

[stopcrazypp]

It's amazing the amount of mental gymnastics some people go through to explain, justify and just plain deny that Tesla may actually have a system that doesn't work well.

One of the frequent excuses I see is "TACC was braking for a truck in the oncoming lane." Ok. let's take this at face value - and say it truly thinks a truck is going to hit you. Aside from being wrong, there's the fact that it doesn't brake for cars in the same lane. Why? The car is smaller, but if I'm traveling 45 MPH, hitting a car at 90 MPH really isn't any better than hitting a truck at 90 MPH so why is the car not an issue but the truck is? And what about all the other trucks that it doesn't brake for? Either it's failing in the first case or failing in the second case. Either way it has an unacceptably high failure rate.

Yes, actually hitting a car is significantly better, given the acceleration you experience from a head on collision varies by the relative mass of the two vehicles, so hitting a truck (especially a semi truck) is far worse than hitting a car. It is very basic physics:
Truck Collision Example
At the extreme end, you can think of trains vs cars. As long as the train is not derailed, most of the time the train conductor is fine even though they don't typically wear seat belts or have airbags. However, the car tends to be completely destroyed.

This is even ignoring that trucks tend to be higher and thus make the crash structures in Teslas far less effective (or even useless in the situation of taking off the roof as described later below).

There is also the fact that there have been instances where people in Teslas were decapitated from crossing under a truck (first article describes two examples):
Tesla driver killed after smashing into truck had just enabled Autopilot – US crash watchdog
This one the driver/passenger was "lucky":
UPDATE: Tesla Crashes Into Another Semi: Federal Investigation Underway
Or killed after smashing into or being smashed by a truck:
A man dies after his Tesla Model S gets hit by a dump truck, likely the first death in the Model S under "regular traffic conditions"
Tesla driver killed in crash posted videos driving without his hands on the wheel
1 dead in Paradise Valley crash involving Tesla and truck

When they slam into a car, usually Teslas end up killing the people in the other car (while those in the Tesla do better), due to Teslas being heavier on average and having better crash structures:
Tesla head on collision with a Honda
A Tesla driver is charged in a crash involving Autopilot that killed 2 people

The car being tuned to be more "scared" of trucks makes sense in this context.

 

[stopcrazypp]

I have a 2020 Model Y, delivered 7/31/2020. I don't think it has radar but I'm in the FSD beta program so from my understanding radar should be turned off even if I do have it.

Just adding another variable - It's not clear to me how much FSD affect TACC and how much my experiences will differ from others' because of that.

Tesla didn't remove radar until May 2021. There is no way you don't have radar unless you mean 2021 instead of 2020.
Transitioning to Tesla Vision

If you are in FSD Beta however, it's turned off regardless.

 

[sleepydoc]

Yes, actually hitting a car is significantly better, given the acceleration you experience from a head on collision varies by the relative mass of the two vehicles, so hitting a truck (especially a semi truck) is far worse than hitting a car. It is very basic physics:
Truck Collision Example
At the extreme end, you can think of trains vs cars. As long as the train is not derailed, most of the time the train conductor is fine even though they don't typically wear seat belts or have airbags. However, the car tends to be completely destroyed.

This is even ignoring that trucks tend to be higher and thus make the crash structures in Teslas far less effective (or even useless in the situation of taking off the roof as described later below).

There is also the fact that there have been instances where people in Teslas were decapitated from crossing under a truck (first article describes two examples):
Tesla driver killed after smashing into truck had just enabled Autopilot – US crash watchdog
This one the driver/passenger was "lucky":
UPDATE: Tesla Crashes Into Another Semi: Federal Investigation Underway
Or killed after smashing into or being smashed by a truck:
A man dies after his Tesla Model S gets hit by a dump truck, likely the first death in the Model S under "regular traffic conditions"
Tesla driver killed in crash posted videos driving without his hands on the wheel
1 dead in Paradise Valley crash involving Tesla and truck

When they slam into a car, usually Teslas end up killing the people in the other car (while those in the Tesla do better), due to Teslas being heavier on average and having better crash structures:
Tesla head on collision with a Honda
A Tesla driver is charged in a crash involving Autopilot that killed 2 people

The car being tuned to be more "scared" of trucks makes sense in this context.

I completely understand the physics and figured someone would try make this point. Maybe I should phrase it like this "is it any more acceptable to hit a truck than it is to hit a car (or a 2 ton pickup, since it doesn't slow for those, either)?" You can argue that you're less likely to die in an accident with a car, but the point is to avoid accidents completely not just avoid the 'worst accidents.' Colliding with a car or pickup or tow truck will still cause massive damages and physical injury so as a practical matter, there's no difference. Unless you say you're ok with colliding with a car going 45 MPH.

 

[stopcrazypp]

I completely understand the physics and figured someone would try make this point. Maybe I should phrase it like this "is it any more acceptable to hit a truck than it is to hit a car (or a 2 ton pickup, since it doesn't slow for those, either)?" You can argue that you're less likely to die in an accident with a car, but the point is to avoid accidents completely not just avoid the 'worst accidents.' Colliding with a car or pickup or tow truck will still cause massive damages and physical injury so as a practical matter, there's no difference. Unless you say you're ok with colliding with a car going 45 MPH.

I'm not "ok" with either, but Tesla likely cares a lot more about eliminating fatalities that are headline making and generates federal investigation, than focusing on eliminating all crashes (which is largely impossible and may make false positives worse). The tuning of tolerance for false positives is a balancing act (with the two extremes of completely ignoring everything that is not in same lane vs slowing down for anything that passes in other lanes).

I also made this point elsewhere, but the simple move of slowing the car down a few mph allows the system more reaction time to consider what to do (especially with the production system that is based on single frames), plus the fact it very effectively puts the driver on alert. The risk overall of a crash from the rear however is lower (and can be made even lower if Tesla combines rear camera monitoring into the decision to slow down).

Also, I've been around the forums long enough to remember the "truck lust" (car veering toward trucks) that AP1 was said to be subject to and the slowing for trucks may be related to this also.
Truck Lust and the myths that accompany it!