Auto Pilot Is Dangerous

[vickh]

In my case Auto steering helped avoid a accident. At least a pop up said that and I saw a car in my blind spot.

I don't have AP (for now) but I'm glad the car has cameras for safety in the interim.

 

[AlanSubie4Life]

I absolutely don't understand all the condescension here about "using AP correctly." Obviously, you should keep two hands on the wheel at all times when using Autopilot. But from the videos I cannot conceive that there is any issue here except a hardware/software calibration issue. The car is clearly steering to the right consistently with Autosteer plus TACC enabled, and Autosteer is clearly attempting to constantly correct it back to middle of the lane (the driver's hands are off the wheel in one of the videos, sure - but that is unrelated to the issue, and I'm sure is just meant to demonstrate that the driver is not causing this behavior!). And this is on I-84 W in the middle of East Hartford, CT. This is not a corner case issue, the lanes are well marked, it's a well-traveled corridor...etc.

It sounds like this behavior is intermittent, from what the OP says? That makes it more odd - I'd expect something dramatic like this to be constant when in Autosteer (rendering it unusable). But, it is kind of like an oscillation (instability), and those can be unpredictable, I suppose. It may be dependent on various specific adjustable gains in the hardware or software which may result in marginal stability (no idea how they implement this, but perhaps they are set during calibration), maybe combined with a small offset in the "zero steering angle" input.

@Toppatop55 : When it is NOT happening, does everything appear perfectly normal? The car tracks straight ahead with zero detectable "oscillation" in the steering wheel (when you have your hands on it and are "feeling" its steering corrections)? "Normal" behavior is to not feel any significant correcting steering input from Autopilot, unless the road is not straight. The wheel is "firm," tracking straight ahead, but it should not be moving side to side except in response to significant lane marking movement.

I strongly recommend that if it's repeatable enough that you take a SC employee out for a test drive to demonstrate it (as recommended above). Unfortunately intermittent problems are much harder to get addressed, but at least you have video evidence. But if it's not happening all the time and you can detect steering input oscillations even when it's not fully unstable, that may be something that you can point out to the technician.

When it is finally fixed, do let us know what was wrong!

 

[Toppatop55]

I don't recall your mentioning where you took the car for service initially. In case you weren't aware, one opened in Warwick, RI recently. That's probably a bit closer than the Dedham, MA store. Of course, if you took yours to Warwick for the initial diagnostic, then going to Dedham (or maybe Mt. Kisco, NY) would be better. See Tesla's store list for phone numbers and addresses.

Also, you haven't mentioned what version of the Autopilot hardware your car has. Most of those sold in the last five or six months have the HW3 computer, but older ones have the HW2.5 computer. The two are significantly different, and it's possible that a flaw could affect one but not the other.



You're presenting straw-man arguments. Nobody here has argued that Autopilot's LKA feature is perfect. It does sometimes make mistakes, and as you and others have said, Autopilot is not an excuse to ignore the road or take one's hands off the wheel. Your criticisms of @Toppatop55's driving may be valid, but it's unclear from these short clips precisely how he's actually driving. It is clear from these clips how Autopilot is working, though....

The problem is that @Toppatop55 has documented his car behaving far less reliably than is typical of Autopilot. My own Model 3 has almost 8,000 miles on its odometer, and it's never done what @Toppatop55 has shown his car doing multiple times in a period of just over two weeks. Beta feature or not, that frequency of misbehavior abnormal, even if it's well known that Autopilot sometimes misbehaves in similar ways. Several possible causes occur to me:

• @Toppatop55's hardware (cameras, computer, etc.) may be faulty.
• Autopilot software may be misconfigured for @Toppatop55's car. Perhaps it has HW 2.5 software on HW 3 hardware or vice-versa (if that would work at all; I suspect not, but who knows). Perhaps a software update or other problem left the software in a bogus state. Perhaps the cameras need to be recalibrated.
  
• There may be a flaw in Autopilot's neural network that happens to manifest more strongly on the roads that @Toppatop55 drives than on other roads.
• Something else quirky about the driving conditions (lighting, traffic around the car, etc.) may have caused an Autopilot neural network flaw to manifest more for @Toppatop55 than would be expected.
  
• @Toppatop55 may have been extremely unlucky and happened to run into an issue that affects other people, but with unusual frequency.

In any of these cases (except maybe the last one), Tesla needs to be made aware of the problem. That's what beta-testing is for, after all -- alerting developers of problems with their software. Maybe Tesla already knows about the issue, because Teslas "phone home" with driving data. OTOH, maybe these serious Autopilot errors have gone unnoticed by engineers. If the problem is with @Toppatop55's hardware, or certain software problems, then Tesla should be able to fix it; but so far, he's gotten the runaround from Tesla.

Didn’t realize Warwick opened a service center. That’s actually where I went for my test drive when they first opened. I purchase by car august 9 and I was assured it has HW3.0

 

[CuriousG]

That's the problem of being in denial of the very true nature of "beta" even when there have been Autopilot collisions, injuries, and deaths.

By your definition, ALL drivers are beta!

 

[Toppatop55]

I absolutely don't understand all the condescension here about "using AP correctly." Obviously, you should keep two hands on the wheel at all times when using Autopilot. But from the videos I cannot conceive that there is any issue here except a hardware/software calibration issue. The car is clearly steering to the right consistently with Autosteer plus TACC enabled, and Autosteer is clearly attempting to constantly correct it back to middle of the lane (the driver's hands are off the wheel in one of the videos, sure - but that is unrelated to the issue, and I'm sure is just meant to demonstrate that the driver is not causing this behavior!). And this is on I-84 W in the middle of East Hartford, CT. This is not a corner case issue, the lanes are well marked, it's a well-traveled corridor...etc.

It sounds like this behavior is intermittent, from what the OP says? That makes it more odd - I'd expect something dramatic like this to be constant when in Autosteer (rendering it unusable). But, it is kind of like an oscillation (instability), and those can be unpredictable, I suppose. It may be dependent on various specific adjustable gains in the hardware or software which may result in marginal stability (no idea how they implement this, but perhaps they are set during calibration), maybe combined with a small offset in the "zero steering angle" input.

@Toppatop55 : When it is NOT happening, does everything appear perfectly normal? The car tracks straight ahead with zero detectable "oscillation" in the steering wheel (when you have your hands on it and are "feeling" its steering corrections)? "Normal" behavior is to not feel any significant correcting steering input from autopilot, unless the road is not straight. The wheel is "firm," but it should not be moving.

I strongly recommend that if it's repeatable enough that you take a SC employee out for a test drive to demonstrate it. Unfortunately intermittent problems are much harder to get addressed, but at least you have video evidence. But if it's not happening all the time and you can detect steering input oscillations even when it's not fully unstable, that may be something that you can point out to the technician.

When it is finally fixed, do let us know what was wrong!

The issue is intermittent which is why I try and capture video footage as much as I can when it’s safe to do so. As far as the service center, what more can they flern from being in the car when the issue surfaces. The video and bug reports should provide them with all they need. And if they’re unable to pinpoint the problem, Maybe they should replace my car. Clearly I didn’t cause the problem and if they can’t fix it I believe a replacement is warranted. A 3 month old car with 2500 miles should not be a mechanical anomaly that I have to deal with.

 

[CuriousG]

Willing to guess it's some kind of sensor issue. Anyway, I suggest you take it up with the local media if Tesla still doesn't do anything about it. The local media is always hungry for a Tesla story. This would certainly escalate the issue as soon as the media is involved.

 

[AlanSubie4Life]

The issue is intermittent

That's what I thought, ok.

The video and bug reports should provide them with all they need

Sure, it should. But we're just looking for the most expedient, data-driven solution here, trying to find a way to clearly show in person to them what is unusual about your car - this is what is most likely to incentivize them to quickly fix your car - it's just human nature. You have to keep in mind the number of spurious complaints they deal with on a daily basis... This (showing them in person) may not be possible, of course.

That's why I was asking about how your car "tracks", and how the steering wheel "feels" when Autopilot is "working normally."

It's entirely possible with periodic instability like this that the car will feel normal until it doesn't. But I just wanted to get an idea of:

1) Whether steering tracks true when Autosteer is not enabled.

2) Whether steering wheel behaves normally when Autosteering is enabled and the problem is not occurring? Again, it should not be oscillating back and forth in any significant way when Autosteer is engaged on a straight section of highway.

3) When working normally, does the car handle curves in the freeway just fine?

4) Do you feel the incorrect steering inputs when the problem is occurring? Or does it seem like the steering wheel is tracking true in that case, but the car is not? (I would expect with this behavior that the wheel would be oscillating back and forth...I'll take another look at the videos to see if I can see it - EDIT: Checked it, and of course the steering is oscillating as one would expect - it is, after all, directly connected to the steering assembly.)

5) Have you noticed anything in particular that causes the problem to start?

 

[Toppatop55]

Willing to guess it's some kind of sensor issue. Anyway, I suggest you take it up with the local media if Tesla still doesn't do anything about it. The local media is always hungry for a Tesla story. This would certainly escalate the issue as soon as the media is involved.

I absolutely agree, in fact that was my initial intention. But then I thought about it and didn’t want to add to the bad and negative assumptions that Tesla or electric vehicles for that matter are unsafe. For that reason I came here to see what the community thinks and to make sure it’s not a wide spread issue. If after my second SC appointment it is not addressed or taken seriously, I will be providing all of my recorded videos to the local media as well as the recorded conversation with the service tech in which he completely disregards my concerns and told me I should assume liability if the car injured or kills me or a family member.

 

[Tam]

By your definition, ALL drivers are beta!

You don't have to take my words for it, take it from Consumer Reports:

"Jake Fisher, CR’s senior director of auto testing, says consumers are not getting fully tested, consumer-ready technology. In essence, he says, Tesla owners are being enlisted as beta testers to help fine-tune the technology for the future—even though they’re paying $6,000 up front for the promised automation.

“What consumers are really getting is the chance to participate in a kind of science experiment,” he says. “This is a work in progress.”

 

[AlanSubie4Life]

You don't have to take my words for it, take it from Consumer Reports:

"Jake Fisher, CR’s senior director of auto testing, says consumers are not getting fully tested, consumer-ready technology. In essence, he says, Tesla owners are being enlisted as beta testers to help fine-tune the technology for the future—even though they’re paying $6,000 up front for the promised automation.

“What consumers are really getting is the chance to participate in a kind of science experiment,” he says. “This is a work in progress.”

Have you watched the videos? All of your post is 100% true, but has nothing to do with the issue at hand.

 

[Toppatop55]

That's what I thought, ok.



Sure, it should. But we're just looking for the most expedient, data-driven solution here, trying to find a way to clearly show in person to them what is unusual about your car - this is what is most likely to incentivize them to quickly fix your car - it's just human nature. You have to keep in mind the number of spurious complaints they deal with on a daily basis... This (showing them in person) may not be possible, of course.

That's why I was asking about how your car "tracks", and how the steering wheel "feels" when Autopilot is "working normally."

It's entirely possible with periodic instability like this that the car will feel normal until it doesn't. But I just wanted to get an idea of:

1) Whether steering tracks true when Autosteer is not enabled.

2) Whether steering wheel behaves normally when Autosteering is enabled and the problem is not occurring? Again, it should not be oscillating back and forth in any significant way when Autosteer is engaged on a straight section of highway.

3) When working normally, does the car handle curves in the freeway just fine?

4) Do you feel the incorrect steering inputs when the problem is occurring? Or does it seem like the steering wheel is tracking true in that case, but the car is not? (I would expect with this behavior that the wheel would be oscillating back and forth...I'll take another look at the videos to see if I can see it - EDIT: Checked it, and of course the steering is oscillating as one would expect - it is, after all, directly connected to the steering assembly.)

5) Have you noticed anything in particular that causes the problem to start?

When the autopilot isn’t going haywire, everything works as it should. Steering wheel and display steering are all fine. I even ruled out a dirty or obstructed camera by washing the car. Then I realized if it was obstructed, the car would display a warning. What I find funny is last year my MS85D had connectivity issues and the SC kept it for a week and a half to diagnose and ended up changing the MCU. Now a problem that can be fetal is falling on deaf ears.

 

[CuriousG]

I absolutely agree, in fact that was my initial intention. But then I thought about it and didn’t want to add to the bad and negative assumptions that Tesla or electric vehicles for that matter are unsafe. For that reason I came here to see what the community thinks and to make sure it’s not a wide spread issue. If after my second SC appointment it is not addressed or taken seriously, I will be providing all of my recorded videos to the local media as well as the recorded conversation with the service tech in which he completely disregards my concerns and told me I should assume liability if the car injured or kills me or a family member.

What gets me is Tesla always seem to spin it in a way that it is always the user at fault (I have no say one way or the other) regarding AP accidents. When presented with logs they will say something like "user took hands off why for x amount of seconds". To me that doesn't prove anything as I drive with have my hand lightly on the wheel and it I get the nag message all the time yet my hand(s) are on the wheel. Until they turn on internal camera and catch user with their hands off wheel when accident happens, the onus of proof is on Tesla as I can certainly replicate the nag message repeatedly with my hand on the wheel yet their logs will say I had hands off the wheel.

 

[Tam]

Have you watched the videos? All of your post is 100% true, but has nothing to do with the issue at hand.

Yes! I've ran into AutoSteer's misbehaviors many many times including those as video-taped in this thread and those are included in Bloomberg as:

"13% of owners say Autopilot has put them in a dangerous situation"

Why else would there be Autopilot collisions, injuries and deaths?

But because I know it's beta so I expect those misbehaviors and I've been ready to correct the steering effortlessly.

However, "effortlessly" riding with me is not for most of my passengers who feel Autopilot should be banned until it passes a neutral third party validation because they are ready to flunk Autopilot right now!

I understand them because they are not beta testers.

 

[CuriousG]

You don't have to take my words for it, take it from Consumer Reports:

"Jake Fisher, CR’s senior director of auto testing, says consumers are not getting fully tested, consumer-ready technology. In essence, he says, Tesla owners are being enlisted as beta testers to help fine-tune the technology for the future—even though they’re paying $6,000 up front for the promised automation.

“What consumers are really getting is the chance to participate in a kind of science experiment,” he says. “This is a work in progress.”

So where's the uproar when cruise control, auto emergency braking came out? This is just something people got accustomed to over time and didn't think twice about it later on. I could count on one hand how often I used cruise control on my last car. If it was traffic aware cruise control I'm sure I would have used it more frequently but then again defeats the purpose when driving manual transmission.

Because it is Tesla, everything is hyper sensitive. There are multiple manufacturers now that employ similar to AP yet we never hear negative news about these systems. Are they more mature? Doubtful.

 

[AlanSubie4Life]

misbehaviors many many times including those as video-taped in this thread

Weird. Guess your quality threshold is different than mine. If my car even once tried to repeatedly drive me off a perfectly straight and well-marked road, as repeatedly demonstrated here, I would be asking questions. But you have confirmed that you have frequently experienced exactly that, and you view that behavior as normal and expected for most users.

Not once have I seen this behavior while I have had my car. All sorts of veering and incorrect behavior in much more understandable situations, where incorrect behavior would be expected, and perhaps I may have even crashed had I not been monitoring as required, but nothing like this.

 

[Tam]

...Weird...

"13% of owners say Autopilot has put them in a dangerous situation"

I don't think those in 13% group lied.

And I don't think those who died while had their Autopilot verified by Tesla as activated during the deadly collisions lied either.

 

[CuriousG]

I didn't take the survey but could the question be worded in a way to skew the results? Are those the only options where AP made them safer and AP put me in danger? What about option like AP didn't perform as expected but I didn't feel it was a big deal.

People saying that it put them in danger. Examples could be phantom braking (sometimes under the overpass), getting too close to semi for comfort, going too slow to flow of traffic, followed a long crack on the freeway rather than actual line, etc. I would discount those in voting.

Now if it was something like swerved me into oncoming traffic, steered into divider, almost rear ending driver in front and not due to sudden stop of traffic, then yes I would consider those legit complaints.

 

[AlanSubie4Life]

"13% of owners say Autopilot has put them in a dangerous situation"

I don't think those in 13% group lied.

I don't think people lied either. But some people wouldn't set foot in a car on an autocross course, and freak out when the car phantom brakes, or fails to complete a lane change successfully at 70mph and veers back into the original lane. My point is that people have different tolerance levels. Phantom braking is a completely a legitimate complaint, and a poll would indicate that those people think that Autopilot endangered them (and they would be right, but we could quibble about the actual danger in this specific case). But that's a poll and a question of perception.

But that's totally different than the subject of the OP.

I want to confirm: you have seen this exact same behavior as the OP, and you confirm that you believe it is normal and expected? Clearly marked road, straight, repeatedly veering out of the lane of travel. And this has happened to you multiple time?

 

[AlanSubie4Life]

I didn't take the survey but could the question be worded in a way to skew the results? Are those the only options where AP made them safer and AP put me in danger? What about option like AP didn't perform as expected but I didn't feel it was a big deal.

People saying that it put them in danger. Examples could be phantom braking (sometimes under the overpass), getting too close to semi for comfort, going too slow to flow of traffic, followed a long crack on the freeway rather than actual line, etc. I would discount those in voting.

Now if it was something like swerved me into oncoming traffic, steered into divider, almost rear ending driver in front and not due to sudden stop of traffic, then yes I would consider those legit complaints.

Yeah, we need to unskew the poll.

 

[hmgolds]

Wrong is the nature of beta.

It's just like wet is the nature of rain.

If I go into a house that is not finished and I close the door and I am soaking wet inside the unfinshed house.

Something must be really wrong and I look up, there's no roof! The rain just falls down hard on me. A house without roof can soak me up during a storm.

And that's what beta pilot is like. The feature is not perfected nor proven by any third party tester.

If "wrong is the nature of Beta", what's the purpose of beta? I always thought it was to find the inevitable problems and correct. Not to find the inevitable problems and (assuming the OPs description is correct) ignore.

The issue here is that you have software that is operating outside the expected norm. That's an ideal opportunity to find out why and apply that learning.

A more accurate analogy is going inside a house with a finished roof of a new design and getting wet. You look for the leaks and fix them.