Phantom Braking

[rjpjnk]

I've just had AEB activate on the way to work this morning. Driving at approx 35 mph on a straight section of road, a pedestrian getting into his van at the side of the road activates AEB, the alarm sounds in the cabin and ABS comes on full (the road is dry the tyres warm). The car almost comes to a complete standstill in an instant whilst the van driver is looking at me like I've lost my mind or something. My neck is aching and I expect it will be stiff tomorrow morning. There was absolutely no need for any sudden braking, had there been traffc behind it would have been a heavy shunt with potential for whiplash injury. I'm shocked.

Same exact thing happened to me once. Fortunately only once so far. It was scary and dangerous as you mentioned.

Since then I've read a lot about it and learned it is important to point out that these false AEB events are different than what people typically mean when they say "phantom braking".

Regular phantom braking refers to the common annoying unnecessary decelerations that happen frequently, often daily. AEB slams on the brakes to the point of wheel lock and really wrecks your day. I hope I never experience another one.

 

[drtimhill]

Since then I've read a lot about it and learned it is important to point out that these false AEB events are different than what people typically mean when they say "phantom braking".

Regular phantom braking refers to the common annoying unnecessary decelerations that happen frequently, often daily. AEB slams on the brakes to the point of wheel lock and really wrecks your day. I hope I never experience another one.

Indeed, the "PB" term has come to mean "the car slowed down suddenly when I didn't expect it too". Of course in some cases the PB could have been a real braking event for a danger the car was aware of but the driver missed. In addition (as per some of the posts in this thread) in some cases this is related to using AP on roads for which it wasn't intended.

 

[sleepydoc]

We recently made a 2,000 mile trip from E TN as far north as Ottawa in our 2022 Model 3 LR with whatever version of Autopilot was current a couple months ago. No FSD.

Only a single data point, but we thought the phantom braking was an order of magnitude better than when the car was delivered in February. It used to be unusable on 2 lane roads, now it handled them well. It overreacts to cars crossing some distance ahead, but unexplained phantom braking was virtually nonexistent on the trip - maybe a couple of times in the 2,000 miles.

I can’t explain the difference, but might suggest there’s something wrong with the cars involved in some of the horror stories in this thread.

These differences have been a perpetual frustration. (I can only imagine how frustrating it is for the developers.) I don’t think people are making it up, the possible explanations I can think of are;

• hardware malfunction
• differing roads/topography giving different results
• hardware variation among the cars.
• something else?

Regardless it’s perplexing how inconsistent the performance seems to be. I have to wonder if Tesla’s ‘continual improvement’ approach actually makes things more difficult because you end up with more hardware variations than you would if you restricted hardware changes to once or twice a year.

 

[drtimhill]

• hardware malfunction
• differing roads/topography giving different results
• hardware variation among the cars.
• something else?

There is also the issue of some of these "phantom" events being real .. the car may spot something the human does not. Also, there is no clear definition of what constitutes an FB event .. sure the "holy f*ck" ones are easy to categorize, but at what point is it just the car being (gently) cautious? Where do you draw the line? Clearly this is all very subjective. There also remains the issue of people using AP in unsuitable conditions that really require FSD (which even now is FAR better than AP imho).

 

[DayTrippin]

These differences have been a perpetual frustration. (I can only imagine how frustrating it is for the developers.) I don’t think people are making it up, the possible explanations I can think of are;

• hardware malfunction
• differing roads/topography giving different results
• hardware variation among the cars.
• something else?

Regardless it’s perplexing how inconsistent the performance seems to be. I have to wonder if Tesla’s ‘continual improvement’ approach actually makes things more difficult because you end up with more hardware variations than you would if you restricted hardware changes to once or twice a year.

The interesting thing is I've had the "pleasure" of driving the same roads on over 1k mile trips one way and able to compare them nicely. I drove my Y and my S. Y was one of the first to have radar removed. My S has it still AND I am haven't upgraded to the version were it is purely TeslaVision (and likely will stay on this version forever).

My Y had an insane amount of PB events on the trip. My S not a single one. My S has over 6k miles and only one PB and it was a minor slowdown and could have been legit. I might have missed something. My 3 was never taken on the same trip but it is purely TeslaVision and its behavior has been far better on the local roads where I can compare all 3 cars than the Y but not nearly as good as the S.

I would sell my car S immediately if I somehow have to go to a version of software that would make TACC and AP behave like it did on my Y. My 3 is "just" good enough to lull me into a false sense of security that it is doing great and then it strikes (PB). I think it subscribes to Murphy's law as it does it often at the most inopportune time. I think it waits until someone is pretty close behind me then activates the "collect tailgater protocol" whereupon it brakes or slows and tries to capture the tailgater with the rear bumper of the car.

On my 3, there were almost no PB events, maybe 1 or 2, in the first 4k miles of driving. For some reason with it, the successive software revisions have made it worse rather than bette with respect to PB. Every day I am grateful that it is nowhere near as bad as my Y was. I sold it and the person who bought it has pretty much given up using TACC/AP with the car, or in only certain circumstances.

 

[sleepydoc]

The interesting thing is I've had the "pleasure" of driving the same roads on over 1k mile trips one way and able to compare them nicely. I drove my Y and my S. Y was one of the first to have radar removed. My S has it still AND I am haven't upgraded to the version were it is purely TeslaVision (and likely will stay on this version forever).

My Y had an insane amount of PB events on the trip. My S not a single one. My S has over 6k miles and only one PB and it was a minor slowdown and could have been legit. I might have missed something. My 3 was never taken on the same trip but it is purely TeslaVision and its behavior has been far better on the local roads where I can compare all 3 cars than the Y but not nearly as good as the S.

I would sell my car S immediately if I somehow have to go to a version of software that would make TACC and AP behave like it did on my Y. My 3 is "just" good enough to lull me into a false sense of security that it is doing great and then it strikes (PB). I think it subscribes to Murphy's law as it does it often at the most inopportune time. I think it waits until someone is pretty close behind me then activates the "collect tailgater protocol" whereupon it brakes or slows and tries to capture the tailgater with the rear bumper of the car.

On my 3, there were almost no PB events, maybe 1 or 2, in the first 4k miles of driving. For some reason with it, the successive software revisions have made it worse rather than bette with respect to PB. Every day I am grateful that it is nowhere near as bad as my Y was. I sold it and the person who bought it has pretty much given up using TACC/AP with the car, or in only certain circumstances.

Thanks for posting your experiences. It would be interesting to see if/how your S behaved on the same roads with the same version of software as your Y but I get why you’re not updating it. I wouldn’t either!

My suspicion is that Tesla made the decision to transition from radar-based to vision-based TACC and like they tend to do with software, they pushed it out when it was ‘ok’ but not fully ready for prime time. (I’ve posted before on my theory that the ability to easily update the software leads to sloppy programming and buggy releases.) Now, 2+ years later they’re finally starting to get the algorithms close to good.

I’ve posted this many times before but we have a 2020 Forester that uses a vision-based adaptive cruise system and it’s been perfect since we bought it. Not a single phantom braking event ever, so it’s definitely possible to use cameras for adaptive cruise - you just have to have good programmers And put the work in to get it right.

 

[IROAR]

For 2+ years my Model S had no phantom braking. In the last week it has happened 5 times with no hazards or other cars anywhere near the vehicle. Service says the hardware is functioning correctly and cancelled my service appointment saying an over the air update will address this problem without really saying it was a problem.

 

[HyperionMark]

Here is another anecdotal data point:
I just got done with a 800 mile round trip on a variety of roads, although most of the miles were done on a major US interstate (90). We have a 2018 Model 3 with the latest version of FSD Beta.
Worst phantom braking per mile that I've ever experienced since owning, as well as being part of Beta for over a year now. Two lane highways were consistently bad as they usually have been for me for most versions, but what really surprised me was the interstate performance. Very very hard phantom braking on 80 mph interstates are not fun and make the drive pretty miserable. In the display during PB the car would show that a person or car would be right in front of me even through no one was obviously there. Happened most frequently while going over a hill with vehicles out a quarter mile or so in my lane.
I ultimately did a hard reset then camera recalibration at a Supercharger stop. Oddly the phantom braking went totally away once I could use just basic autopilot but not FSD yet. It makes me think that it reverts to a much earlier version of a stable autopilot build while the cameras are still fully recalibrating. I did this two different times on this trip and experienced the same thing. Once FSD was ready (even if unchecked) it went back to the same PB issues.
#pleasegiveusdumbcruisecontrol

 

[2101Guy]

For 2+ years my Model S had no phantom braking. In the last week it has happened 5 times with no hazards or other cars anywhere near the vehicle. Service says the hardware is functioning correctly and cancelled my service appointment saying an over the air update will address this problem without really saying it was a problem.

LMAO. That's whats known as "continual improvement/car gets BETTER over time".

 

[2101Guy]

Here is another anecdotal data point:
I just got done with a 800 mile round trip on a variety of roads, although most of the miles were done on a major US interstate (90). We have a 2018 Model 3 with the latest version of FSD Beta.
Worst phantom braking per mile that I've ever experienced since owning, as well as being part of Beta for over a year now. Two lane highways were consistently bad as they usually have been for me for most versions, but what really surprised me was the interstate performance. Very very hard phantom braking on 80 mph interstates are not fun and make the drive pretty miserable. In the display during PB the car would show that a person or car would be right in front of me even through no one was obviously there. Happened most frequently while going over a hill with vehicles out a quarter mile or so in my lane.
I ultimately did a hard reset then camera recalibration at a Supercharger stop. Oddly the phantom braking went totally away once I could use just basic autopilot but not FSD yet. It makes me think that it reverts to a much earlier version of a stable autopilot build while the cameras are still fully recalibrating. I did this two different times on this trip and experienced the same thing. Once FSD was ready (even if unchecked) it went back to the same PB issues.
#pleasegiveusdumbcruisecontrol

Copy my prior reply, and insert here as well.

 

[Knightshade]

Copy my prior reply, and insert here as well.

I'll take "What is literally every post you've ever made here" for $1000 Alex!

 

[Jay Smith]

I guess it boils down to the fact that we chose to be beta testers.
Let's persevere and continue contributing to the hive.

 

[DrChaos]

I’ve posted this many times before but we have a 2020 Forester that uses a vision-based adaptive cruise system and it’s been perfect since we bought it. Not a single phantom braking event ever, so it’s definitely possible to use cameras for adaptive cruise - you just have to have good programmers And put the work in to get it right.

And use stereo cameras with dedicated hardware support to get 3d from parallax, which is what Subaru does. There is a Japanese professor who was the core originator of the technology decades ago.

https://www.eetimes.com/subaru-eyesight-father-returns-in-stereo-vision/

 

[drtimhill]

My Y had an insane amount of PB events on the trip. My S not a single one. My S has over 6k miles and only one PB and it was a minor slowdown and could have been legit. I might have missed something. My 3 was never taken on the same trip but it is purely TeslaVision and its behavior has been far better on the local roads where I can compare all 3 cars than the Y but not nearly as good as the S.

I would sell my car S immediately if I somehow have to go to a version of software that would make TACC and AP behave like it did on my Y. My 3 is "just" good enough to lull me into a false sense of security that it is doing great and then it strikes (PB). I think it subscribes to Murphy's law as it does it often at the most inopportune time. I think it waits until someone is pretty close behind me then activates the "collect tailgater protocol" whereupon it brakes or slows and tries to capture the tailgater with the rear bumper of the car.

Interesting observations .. thanks for the info. I'm curious, and dont take this the wrong way ... would you be happy if the car didn't do any PB but also had all the safety features disabled? That is, no AEB, no lane warnings etc etc?

 

[CyberGus]

Interesting observations .. thanks for the info. I'm curious, and dont take this the wrong way ... would you be happy if the car didn't do any PB but also had all the safety features disabled? That is, no AEB, no lane warnings etc etc?

It's disconcerting to drive a car without autonomous features after having done so. It's like the car doesn't even care

 

[drtimhill]

It's disconcerting to drive a car without autonomous features after having done so. It's like the car doesn't even care

Indeed, and when people say "get rid of PB" that's easy .. turn off all the safety features. It's like minimizing your taxes .. all you need to do to pay ZERO tax is to earn ZERO income. When people say they want to "minimize their taxes" what they really mean is "maximize their post-tax income" which is quite a different thing.

So what they are really saying is "get rid of PB while retaining genuine emergency braking". And this, of course, is much harder. I'm sure Tesla could reduce PB dramatically tomorrow just by (e.g.) de-tuning the AEB so it is less sensitive. But try explaining that decision to someone who has just had their car fail to save them from a horrible accident "Sorry, we realize it didnt brake, but users didn't like that it occasionally scared them by braking when they didn't expect it to". Yeah, right.

In fact, I'm not sure anyone has solved this in the industry. There are lots of posts here of people saying "my XXX model car had AEB [etc] and it never did PB". But would the car have braked in a real emergency? If not, then basically my suggestion to turn off the safety features is basically what they are asking for without realizing it, since that car they point to as an example of what they want didn't break when it was supposed to.

There have been several examples of unbiased independent testing of system like AEB (IIHS for example), and many cars did VERY badly, while Tesla did pretty well. The reality is that to brake well in an emergency is always going to risk some level of PB .. even a "perfect" AEB system will generate some "PB" events because the car is bound to sometimes see events that humans miss (and then assign falsely as a PB event).

I'm not saying the current level of PB is acceptable, it's not, and needs to be improved. That's part of the FSD project, since by definition FSD needs far better acuity when it comes to driving on regular (non-divided) roads. But its never going to go away, partly because people will continue to use the car in unintended ways (witness the people here using AP on two-way single lane roads).

 

[sleepydoc]

Indeed, and when people say "get rid of PB" that's easy .. turn off all the safety features. It's like minimizing your taxes .. all you need to do to pay ZERO tax is to earn ZERO income. When people say they want to "minimize their taxes" what they really mean is "maximize their post-tax income" which is quite a different thing.

So what they are really saying is "get rid of PB while retaining genuine emergency braking". And this, of course, is much harder. I'm sure Tesla could reduce PB dramatically tomorrow just by (e.g.) de-tuning the AEB so it is less sensitive. But try explaining that decision to someone who has just had their car fail to save them from a horrible accident "Sorry, we realize it didnt brake, but users didn't like that it occasionally scared them by braking when they didn't expect it to". Yeah, right.

In fact, I'm not sure anyone has solved this in the industry. There are lots of posts here of people saying "my XXX model car had AEB [etc] and it never did PB". But would the car have braked in a real emergency? If not, then basically my suggestion to turn off the safety features is basically what they are asking for without realizing it, since that car they point to as an example of what they want didn't break when it was supposed to.

There have been several examples of unbiased independent testing of system like AEB (IIHS for example), and many cars did VERY badly, while Tesla did pretty well. The reality is that to brake well in an emergency is always going to risk some level of PB .. even a "perfect" AEB system will generate some "PB" events because the car is bound to sometimes see events that humans miss (and then assign falsely as a PB event).

I'm not saying the current level of PB is acceptable, it's not, and needs to be improved. That's part of the FSD project, since by definition FSD needs far better acuity when it comes to driving on regular (non-divided) roads. But its never going to go away, partly because people will continue to use the car in unintended ways (witness the people here using AP on two-way single lane roads).

A couple of points -
First, to everyone who says ‘I wish I could just have dumb cruise control,’ I used to say that. Then I drove our 2011 odyssey with dumb cruise and missed smart cruise from my Tesla, even with its flaws.

This also presumes that PB events are caused by the safety features. Sometime back I disabled all safety features on my car and still had PB events meaning that to the extent I could check, the safety features were not the cause (or sole cause) of phantom braking.

Finally, all of this is not black and white. I presume you’re familiar with ROC curves. Assuming PB is caused by over sensitive emergency braking, it should be possible to back off on the sensitivity without completely disabling it. Since these features are generally intended to be a backup to the human the actual impact of such a change should be minimal. Of course Tesla has the problem of trying to develop a totally autonomous system in which there is no backup. They also have the problem of people testing the backup emergency system and screaming bloody murder if it fails 10% of the time (as opposed to being happy it succeeded the other 90%)

 

[ApexProxy]

These PBs make me want to get a dash cam specifically to record my screen and the road together so that I can review and see if anything pops up on the display and if there was some event nearby that I missed. I will also do my best to hit save on the teslacam to see if it caught anything as well. This shouldn't be hard to catch because I get a PB on average once every 500 or so miles.

I want to give the car the benefit of the doubt but I think there is a gross miscalculation happening somewhere that they are having issues rooting out.

 

[Bouba]

In any earlier post where someone was complaining about FB they used the word ‘mirage’. But mirages are real...long black highways with gasoline vapor and the sun beating down

 

[Tronguy]

In any earlier post where someone was complaining about FB they used the word ‘mirage’. But mirages are real...long black highways with gasoline vapor and the sun beating down

Just so we’re clear: It’s not gas vapors. It’s two main things:
1. A layer of hotter or cooler gas near a surface with other gas or a solid. Light hitting at thin angles gets refracted. Let there be a couple of these in the line of sight and one can get palm trees, sometimes upside down, from an oasis miles off. Or a ship 10 miles off or even beyond the horizon floating in mid air. No sun required.
2. Mirrors obviously reflect; there’s a free-flowing electron gas that reacts to electromagnetic E fields by accelerating. Accelerating electrons give off electromagnetic fields and, well, that’s what a reflection is. If the electrons can’t freely move, like with a non-conductor, then other things happen, but not typically a reflection. Except.. electrons in a non-conductor may not move much, but they do wiggle in place. If one shoots an EMF wave at a shallow enough angle, it can interact with a bunch of electrons, each wiggling a bit, so a reflection can occur anyway. Hold up a flat surface at a narrow angle to a light, and it usually reflects quite well, thank you very much.

All of which simply means that even looking at standard stuff with one’s eyeballs is rather like wandering around inside a tent full of funhouse mirrors. Ask people who wear glasses how long it took them to get used to all the reflections. But we got a zillion years of evolution discerning what’s real and what’s not. (Didn’t spot that tiger? Now you get to contemplate your failure from the inside..) Computer vision is still an exciting field of study and not that old, either.