Welcome to Tesla Motors Club
Discuss Tesla's Model S, Model 3, Model X, Model Y, Cybertruck, Roadster and More.
Register

Phantom Braking

This site may earn commission on affiliate links.
TL;DR - times, they are a-changin’.

Longer; may not read - Vehicular complexity is increasing, is here to stay, and our levels of frustration will need to accommodate this paradigm.

Wow, lots of things going on in this thread, which may have outlived its usefulness. My summary, likely oversimplified, is that there are several points of agreement…and a few points of roughness. At a macro level, there seems to be agreement that both building a car at production scale and chosen level of quality is hard, and building software quality is also hard. The joining of the two combines those difficulties into something harder yet. Worse, today’s cars and today’s software are both very complex compared to just a few years ago and that aggregate complexity makes things much harder again. So where does that leave us? It certainly challenges the testing or quality control functions of the manufacturer implementing more thorough hopefully automated testing procedures to meet regulations (another area of changing complexity across the globe) which is also hard and complex. Complexity’s biggest downside is, well, its complexity., i.e., the difficulty in testing each function, and how each function interrelates with each other function. So with all that in mind the manufacturer’s challenge is how to do exhaustive testing, if that is even possible, and if not to determine what absolutely must work flawlessly to meet regulatory and market requirements and test those intensively, and be aware of how much will it cost to do so.

There is another factor at work and that is the imperative to provide “new.” Objectively for me, the iPhone13 mini is very, very similar in almost all regards of importance to me to the previous year’s iPhone12 mini. But there’s enough difference to (a) allow Apple to continue the trend of the latest, greatest offering every year, and (b) recognize many (VERY many) buyers upgrade annually. Even the calculator app on my iPad has about a quarterly ”upgrade” introducing “new improvements” in usability and appearance. Wait what? It’s a calculator. Is the math different? No, just the background colors, numbers rounded or not on the screen, ad tracking, and the like, enough so that it is considered “latest, greatest.”

So where frustrations come in is two-fold: (1) sometimes you just don’t like the way something works or doesn’t as delivered. Experiences with wipers, auto high beams, size of seats, TACC, expectations not met such as FSD, poor trunk lighting, no spare tire, paying extra for a garage door opener, and the list goes on. Some of these you knew about before buying and you decided as you wished; others you found after purchase. You pays your money and you takes your choices/chances. The second and larger frustration is that as more hardware functionality becomes software, that delivered functionality may not be a one-for-one replacement, and the software can and likely will change. So an HVAC control head becomes a series of taps on the screen, a CD player becomes a screen-driven streaming option, and so on. Instead of an experienced team of mechanical engineers designing a control system, that design work has moved to an experienced team of software designers. Perhaps great if they work together, but HVAC controls are new areas for software developers. And the software is built to be changed going back to the extreme complexity and inability to test everything. When you find an “ooops moment“ in software you can reasonably quickly correct it and reissue. If your keyed ignition system doesn’t work in the wild, it’s a massive and expensive recall…ask GM.

But that ability to change results in its own levels of frustration given that we owners can be very resistant to change, and changes can be made that are objectively poor judgment, others can be totally indefensible, and others just plain terrific and much welcomed. Who decides? We do, and we grouse about the changes we don’t like, sometimes praise the ones we do, and make our buying decisions accordingly. Which (finally) brings me to the last point of this tirade: what car you buy and like/dislike will very likely not be the same vehicle throughout your ownership. You will constantly be considering those likes/dislikes until one side takes over and you either trade the vehicle or continue ownership and change your evaluations. That’s a new “thing” in products, and certainly new in the auto industry. I expect we’ll see much more of this, not less.

Perhaps I should switch to decaf.
 
Last edited:
Here is a comment on Reddit. I wonder if this is why I rarely ever see phantom braking and some see it all the time ? I’ve the following distance set to max and FCW to Early.

A friend of mine has 84000 miles in, mostly Autopilot with 5 months of FSD Beta. He says with the following distance set to max (7) and Forward Collision Warning to Early, he's never had phantom braking.​
My settings too (also on Chill) and very rarely experiencing PBing. 2018 M3 on FSD Beta 10.10.2.
 
  • Like
Reactions: WhiteWi and EVNow
I dunno about anyone else, but if someone starts out with "I had an hour of non congested road to try out AP." I think its safe to assume they had roughly 60 minutes to try it.

Sure there is some wiggle room with the non-congested roads aspect so its not concrete, but safe enough for our needs.

Did the person experience PB? Yes
Did the Phantom braking happen excessively during the duration of the test? Yes
What percentage of the Phantom braking was severe? unknown from the data provided.
Was it FSD Beta? No
Was it Tesla Vision? Yes
Undivided and divided roads? Unknown
Was any of it AEB? Doesn't look like it
What percentage was due to maps issue? Unknown.
What was the FCW set to? Unknown
Was was the following distance set to? Unknown
And again he never said an Hour Either! Yet Sleepdoc made up an entire mathematical story based on Nothing. As are most of his posts.
 
And again he never said an Hour Either! Yet Sleepdoc made up an entire mathematical story based on Nothing. As are most of his posts.

The person who SleepyDoc quoted never challenged the copy and pasted quote that mentioned having one hour.

With that being said there are always assumptions in online communication, and glitches. You assumed the SleepyDoc was male, but I don't think they ever said what their pronouns were. You also glitched in the username and turned the SleepyDoc into a Sleep Doc.

As to honesty I do applaud the Sleepy Doc for being honest about being a Sleepy Doc. It would be kinda scary to be a patient of them though.

"Sorry about the length of that incision. I just zoned out there for a second'

:p
 
I think its important to clarify a couple of things regarding what's being used for TACC, and someone can clarify if I'm mistaken about any of it.

  1. Map Data -> Uses the data to determine where to slow down for curves. Does this regardless of Vision+Radar or Tesla Vision.
  2. Responding to Camera Read Speed Limit changes or map based speed limit changes -> Tesla Vision Only (with Vision+Radar I believe this is AP only)
  3. GPS -> Used only for Map Data
  4. Lanes you should be in -> Not used as its not needed for TACC
  5. Cars in adjacent lanes -> Used to determine if there is a chance they'll cut into your lane, and used regardless of Vision+Radar or Tesla Vision.
All of the above ignores FSD Beta because its an entirely different stack

So I'm purposely limiting it to just TACC on a divided highway. This is something I have thousands of miles of experience with using Vision+Radar and thousands of miles with Tesla Vision.

In the early days of TACC a lot of the PB was caused by #5 with semi's, and I made some posts about really wanting a dumb Adaptive Cruise Control. Over time they improved it. At some point I started to use AP a lot more and most of my PB was related to #1 or #2. Sometimes I'd see the dreaded 45mph limit when passing though Tacoma. During the entire time with Vision+Radar I had a few occurrences of PB due to overhead bridges while there was no traffic in front of me.

So even before the Beta the PB was bad, but not horrible. Then I got the FSD Beta which brought Tesla Vision. I did hear all things people said about it, and I did see the PB poll on it but nothing prepared me for how bad it really was. Being in the Seattle area means a lot of rain, and it did worse in the rain which was really disappointing.

Thankfully the freeway performance did improve from 10.3.1 to 10.9, and PB might be down to the levels of Vision+Radar. I say might because there is a lot of randomness.

What I know is its extremely unlikely that ANYONE with Tesla Vision will be able to do a 400 mile trip from I5 near Everett all the way down to Portland, and back without at least a 3-4 PB events where one will be startling enough to take over to prevent further slowing. This is assuming the person drives 10mph over the limit, and has normal following settings, normal FCW setting, etc.

Would a Tesla Vehicle with Radar + Vision be able to do with? Possibly. I figure it has a better chance than 5%.

Can a 2015 AP1 do it where it has a lot of the same capabilities? Probably. I had one and PB was pretty rare despite also being able to adjust the speed for corners. I did feel like it was a little slow on its cornering speeds.

Can my Jeep Wrangler Unlimited do it where it doesn't have #1, #2, and probably limited #5 do it? Extremely likely.

Could I buy another vehicle that has 1, 2, 3, and 5? while having a good chance of no PB? It's a strong possibility. Here is the first google search I did, but there are probably a lot more.

Interesting that the article said the car was the first car that could read speed limit signs - Tesla's been doing it for a while (albeit in beta).

Like I've said before, there are successful systems that use both vision and radar so I don't think we can blame one or the other.

In the past Tesla seemed to respond only to the angle of the steering wheel which clearly would respond too late since you need to anticipate the curve. Later revisions have improved on this and the speed adjustments are much more timely and appropriate. I'm not so sure that TACC uses maps to slow down for curves. That's something that should be very achievable with vision only and given the limited accuracy of maps/GPS that would seem to be a poor choice. Of course that may explain some things, too.

Speed limits definitely use more than just vision, though. There's a road I drive where it consistently reads the 55 MPH speed limit sign, then drops to 35 MPH about 300 feet later. There are no other signs, it just thinks that stretch of road is 35 MPH, so I've concluded it must be inaccurate mapping data.
Does TACC still use some parts of AP, such as all sensors? Could PB just be confusion from some of the sensors when there's overlap? ...

Perhaps TACC would be better if it shutdown, or ignored sensor data from anything other than the forward cameras.

How would other cars with basic adaptive cruise control handle a car/truck on the side or in the blind spot coming/drifting into your lane (potentially causing an accident)? I know they can see the car ahead slowing down, but what about all the cars around you?
I think that's likely the issue with Tesla. They're trying to take too much information, introducing noise into the system and making it worse. Like I said above, you should only add complexity if you improve the system. If this is the case, I think Tesla would be better starting small/old, then expand rather than trying to do everything and failing.
I personally question the accuracy of Tesla Vision at long ranges at night.
...
Odds are the vehicles have more than adaptive cruise control. Regardless of the vehicle I would never count an honest accidental avoidance braking as a PB even if I felt like it was being a little nervous.
Night time is definitely more difficult. It's more difficult for human drivers, too. Judging distance based on headlights is not easy and lines are harder to see so it's easy to understand how the system would think a car might be closer or in the wrong lane. This is a case where radar would likely be better, at least for the distance portion.

And yes, you're totally correct about accident avoidance. There's a ton of conflation of PB and AEB and others have posted reports of AEB in other brands as 'proof' that their adaptive cruise didn't work but they are related but separate systems and issues.
 
  • Disagree
Reactions: Knightshade
My settings too (also on Chill) and very rarely experiencing PBing. 2018 M3 on FSD Beta 10.10.2.
FSD Beta has added a lot of noise to discussions of PB on every thread that talked about the topic since the 10.3 release when so many people got it.

It's noise because when FSD Beta is engaged (like City Streets) it's a different stack so what happens while on FSD Beta can't be mixed in with the results that people get when they have Basic TACC/AP engaged.

In my own posts about my experiences I leave out half of the trip details because of this. I do all these trips to Portland, OR not because I'm going there, but I'm going to the coast to look at land. I can't talk about going to the coast because the FSD Beta clouds all of that due to the road types (there is a lot of switching of stacks).

It's also clouded the NHTSA reports as some owners likely posted PB events with FSD Beta to that site. Which makes sense for the severe AEB events with the broken 10.3, but not so much for stuff outside of that. With the beta people need to realize it's not released stuff and to only report to the NHTSA if its a severe safety issue.

10.3.1 was so ridiculous there were times I had to hold the accelerator down to get through some areas where it wanted to constantly PB.

10.9 wasn't too bad.
 
  • Like
Reactions: sleepydoc
I'm not so sure that TACC uses maps to slow down for curves.
I've been a Tesla owner since 2015 so I actually got to experience the addition of slowing down for curves way back in 2015 with AP1.

The Vision was way less capable back then so map data was being used, and sometimes it did PB due to bad map data. If the system relied only on vision the car would have ended up going way too fast to handle some types of curves.

Humans everywhere use internal map data when handling curves. Where they've experienced the same curve hundreds of times, and they know what to expect. It's not just the angle of the curve, but what types of things to anticipate.

If they've never experienced that curve they're likely going to take it slower than locals.

FSD Beta is likely really clouding things when it comes to what's being used for braking while under TACC. When that's in play its not longer using the "legacy" code for handling curves.
 
  • Like
Reactions: sleepydoc
Limited resources at Tesla, right. So have to work on what is important. Would think PB is. SO glad to see they have 2 teams (Seattle and Austin) working on intergating Steam Games into the system directly. Such a good use of engineering talent and money /s (end sacrasim)
i know that can seem like a waste of resources, however keep in mind that the guys creating Steam integrations are probably useless when it comes to programming their AI, so there really isn't a resource cost since those guys couldn't help if they wanted. programming tends to be a highly specialized field.
 
  • Like
Reactions: WhiteWi
I drive a lot of 2 lane highways in rural alabama and it seems to me that the phantom braking events happen more frequently when there is a dotted center line as opposed to when there is double solid center lines. As if the car is unsure which direction the opposite lane is traveling when center line isn’t solid. Anyone else notice this?
 
I drive a lot of 2 lane highways in rural alabama and it seems to me that the phantom braking events happen more frequently when there is a dotted center line as opposed to when there is double solid center lines. As if the car is unsure which direction the opposite lane is traveling when center line isn’t solid. Anyone else notice this?


Are you talking AP or TACC? Because AP is explicitly not intended to be used on such roads, and assumes all traffic is going the same direction you are- so being confused by oncoming traffic would be expected.

The beta is the solution being worked on for that- and while it can slow for oncoming traffic (esp. when hills or curves are involved) it's gotten extremely mild at this point when it does (2-3 mph typically)
 
My 2021 Model Y was delivered in March 2021. While I don't recall any issues with AEB early in our ownership, phantom braking has become common and scarier since about mid-summer of last year. The car does not have FSD. I would really like to know how I can tell if the car does have radar.

I can reliably get the car to demonstrate phantom braking in autopilot by driving under high tension power lines along one of my usual routes. By "reliably", I mean 100% of the time. This was reported to Tesla when I asked for a service appointment, their response was that this was a software issue and that I ought to press the accelerator (or words to that effect) to make the car record the problem. They didn't want to see it or my evidence that outside interference might be a cause.

More recently I had an incident on an interstate highway at relatively high speed (65+) in autopilot, passing a semi, where the car braked suddenly in heavy traffic. It went from 65 to about 50 before I could catch it and override the system. I don't recall seeing other oncoming semis on the other side of the highway, but I was occupied trying to avoid a collision and might not have seen them, had they been present.

Then, within the past few days on 2-lane roads in the area, the car began to have what I call "spasms" of phantom braking, where it must have seemed that I was brake-checking cars that were following me. Embarrassing.

If I MUST use adaptive cruise control, my $26k 2021 Subaru Forester is the car I drive. With EyeSight®, it works perfectly.

I have finally complained to NHTSA about this problem with the Model Y. In all other respects, this is the best of the 50 cars I have owned, but this problem is serious.

Chris
 
  • Like
Reactions: xls_
Are you talking AP or TACC? Because AP is explicitly not intended to be used on such roads, and assumes all traffic is going the same direction you are- so being confused by oncoming traffic would be expected.

The beta is the solution being worked on for that- and while it can slow for oncoming traffic (esp. when hills or curves are involved) it's gotten extremely mild at this point when it does (2-3 mph typically)
Good to know, so how when is autopilot appropriate to use? I’ll try and use TACC more and see if there is a significant difference
 
Good to know, so how when is autopilot appropriate to use? I’ll try and use TACC more and see if there is a significant difference


Manual covers this-- it's intended for controlled-access highways (so all traffic going one direction, access controlled by things like on/off ramps, no at-grade intersections at all)

It also mentions it's not intended to be used in construction zones, or around pedestrians/cyclists (though the later 2 are rare on freeways)
 
Manual covers this-- it's intended for controlled-access highways (so all traffic going one direction, access controlled by things like on/off ramps, no at-grade intersections at all)

It also mentions it's not intended to be used in construction zones, or around pedestrians/cyclists (though the later 2 are rare on freeways)
To clarify, I was talking about while using in FSD beta. Does that make a difference?