Navigate on Autopilot is Useless (2018.42.3)

[StealthP3D]

Frankly I’m most concerned with people who don’t seem to have negative views on how EAP can fail at times. I have to wonder if those people haven’t really come across yet a scenario where it fails badly as an aid — and will they really be prepared to take over or do they just think they are because so far everything has been pretty peachy?

I think you are barking up the wrong tree. It's pretty obvious the vast majority of EAP users are fully aware that it can fail from time to time and are prepared to take over.

The problem comes from those who think it's fully autonomous and perfectly reliable. Or that it should be. Tesla is very clear that it's not. Anyone in this camp probably shouldn't be behind the wheel of a large automobile to begin with. Or any automobile. Except how would Darwin's rule work in modern society if those people were somehow prevented from driving? LOL!

 

[electronblue]

I think you are barking up the wrong tree. It's pretty obvious the vast majority of EAP users are fully aware that it can fail from time to time and are prepared to take over.

The problem comes from those who think it's fully autonomous and perfectly reliable. Or that it should be. Tesla is very clear that it's not. Anyone in this camp probably shouldn't be behind the wheel of a large automobile to begin with. Or any automobile. Except how would Darwin's rule work in modern society if those people were somehow prevented from driving? LOL!

I dislike that group of people as much as you do. I just fear there may be another group that don’t quite appreciate the difficulty of staying sufficiently alert all the time when the computer can basically do ”whatever”.

In fact, I wonder if any of us really do and how much of us not having crashed on Autopilot is sheer luck... Ghost brakings are one phenomenon that I for one have had some close calls with and can totally see cases where I simply would not catch it in time to avoid being rear-ended, because sometimes it can really be a bit violent. It does make me wonder should I simply stop using TACC too. But the thing is: it is not just me I wonder about, I wonder can anyone really reliably avoid ghost brakings causing crashes — just as one example.

I always have two hands on wheel, foot hovering over pedals and use Autopilot only on divided highways.

 

[StealthP3D]

I dislike that group of people as much as you do. I just fear there may be another group that don’t quite appreciate the difficulty of staying sufficiently alert all the time when the computer can basically do ”whatever”.

Inattentive driving has always been a problem. I wouldn't expect it to go away with EAP although I do think the advent of EAP has made the roads more safe by causing people to be more aware of their surroundings as well as the added unknowns of EAP rather than driving down the road on human "auto-pilot".

In fact, I wonder if any of us really do and how much of us not having crashed on Autopilot is sheer luck... Ghost brakings are one phenomenon that I for one have had some close calls with and can totally see cases where I simply would not catch it in time to avoid being rear-ended, because sometimes it can really be a bit violent. It does make me wonder should I simply stop using TACC too. But the thing is: it is not just me I wonder about, I wonder can anyone really reliably avoid ghost brakings causing crashes.

Sheer luck is good. In the end statistics, not uninformed and emotionally charged biases against computers do "a human's job" will prove the safety (or lack of safety) of EAP and it's already apparent that it is more safe than having nothing at all. I'm sure you will disagree with this but I discount the naysayers who don't seem to grasp that slaughter on our highways is a huge problem that needs addressing. People against technological advancement are against one death caused by a computer but illogically comfortable with the thousands of deaths caused by human error every year. Their minds can't seem to grasp that, yes, computers are going to cause an occasional death or injury but it's the overall death rate that is relevant. And EAP is already making the highways more safe. Sure, a small minority misuse it but that is also true of cars driven manually by inattentive drivers.

 

[electronblue]

@StealthP3D My opinion comes from my experience which is running AP2 pretty much since it became available on the market so definitely not from any kind of Ludditism. I guess we will know over time what the Autopilot statistics will be but personally my experiences and several those I’ve read about do give me some pause.

I am personally not confident I am safer by using TACC and Autosteer than I would be without using them and perhaps solely relying on AEB for instance which probably does add to safety in itself. So far this has not been enough to stop using TACC or Autosteer, as I follow the progress (back and forth) with interest, but as said it does give me pause. In many instances I can see — call it a hunch — myself crashing more likely due to TACC/Autosteer than avoiding a crash because of them. Not because of how I control the car but because of how they behave.

It just gives me some pause and introduces some doubt. As it seems to have given the resident Tesla expert @wk057 in this thread as well. In many ways it has not quite been a good product.

This should not be viewed as scepticism towards autonomous driving or even driver’s aids in general, just specifically some pause Tesla’s approach and product has given me.

 

[DarkMatter]

I wonder how traffic aware cruise has been for other manufacturers. I have not driven anything with it more than casually. It wouldn't surprise me if some of the other systems phantom brake in similar situations. Has anyone else here been getting regular miles in one of the other systems? I would actually expect Cadillac to be the best about this because of their geofencing and HD mapping.

 

[DarkMatter]

You all misread that. 99.5% of the road was clear during the crash. It could have just as well been DIRT during midsummer. Or bad lane paint, or anything else that happens in real life.

So, her usage was fine. AP goofed up, badly, destroying the car. No excuses...

Thanks for the clarification.

 

[StealthP3D]

I wonder how traffic aware cruise has been for other manufacturers. I have not driven anything with it more than casually. It wouldn't surprise me if some of the other systems phantom brake in similar situations. Has anyone else here been getting regular miles in one of the other systems? I would actually expect Cadillac to be the best about this because of their geofencing and HD mapping.

I only have experience with the auto emergency braking on a 2018 Volvo. I was manually driving along a curvy country lane at 40 mph when there was a car parked off the side of the road (outside the white line) and the Volvo slammed on the brakes and scared the crap out of me. Never had that happen with the Model 3 with many more hours driving (about 300 times as much). The most the Model 3 has done is some tentative moderate braking of the kind that I wouldn't worry about getting rear-ended unless the driver behind was texting on their phone or otherwise completely distracted.

 

[SpudLime]

I wonder how traffic aware cruise has been for other manufacturers. I have not driven anything with it more than casually. It wouldn't surprise me if some of the other systems phantom brake in similar situations. Has anyone else here been getting regular miles in one of the other systems? I would actually expect Cadillac to be the best about this because of their geofencing and HD mapping.

My wifes odyssey has TACC and it does have the phantom breaking on occasion. FWIW

 

[DarkMatter]

Hopefully they will fix that in the next OTA update for the Volvo and Honda.

 

[electronblue]

I have owned and driven several cars with TACC for years and Tesla is the first one with serious and repeated ghost braking. Earlier it was mostly perhaps when a car in front of you slowed down and turned away kind of stuff but Tesla mistakes overhead bridges for emergency stops so that’s a new one...

It has gotten somewhat better lately of course. Early days of AP2 were horrible.

That said I still consider Tesla’s TACC much more clearly a safety benefit than Autosteer where the jury feels more out in my view.

 

[S4WRXTTCS]

Well...I have never had to disengage Autosteer to avoid a disaster so maybe that is the difference. Commuting 100+ miles a day almost all on the expressway, autosteer routinely either won't change lanes because of traffic conditions, does it too slowly to be relied upon in traffic, or won't auto change lanes because of obscure lane markings. Between all of those situations I have to make lane changes just as you would with TACC. In my opinion it should remain at the set speed as TACC would. How many of you have actually had to disengage autosteer " presumably due to avoiding some kind of disaster" I can honestly say after some odd 55k+ miles of using autosteer that has not happened once to me.

90% of the time when I have to disengage auto steer is during an aborted lane change. It lets me know its aborting because it kicks the steering back to the lane I was in. I'm always paying attention to traffic so I clearly know everything is fine. Sometimes it gets confused by traffic in the lane next to the lane I'm going into. Like it's more susceptible to aborted lane changes if there are big semi's in that lane.

This is during long stretches down I5 between Seattle, and Portland where I likely do 40-50 lane changes. Where it will abort at least 10% of them for some reason. I'm a little concerned that it's aborting them due to rain causing reflections, and the neural net is falsely identifying cars next to me when there are no cars. Like sometimes I see a ghost car passing me when there is no car passing me.

I don't see why taking over should change the set speed. I don't believe it should and I'm still convinced that it's a bug.

Why am I being punished with how I take over? If I cancel it any other way it doesn't change the set speed.

The other times of canceling autosteer is when I'm in the right lane, and there is a merging lane. In WA state this isn't painted so its effectively a really large lane so autosteer will steer itself into the middle. I used to fight autosteer, and it would cancel. Now days I just leave it, and accept that it makes me look like an idiot.

There have only been a handful of times where I had to take over to prevent a disaster.

With AP1 it was susceptible to truck lust, and occasionally being in the wrong position after cresting a hill.

With AP2 I'm less likely to take over autosteer, but more likely to take over TACC. Where it's not prone to truck lust at all (at least I haven't experienced anything), and it's not prone to being in the wrong position after cresting a hill. I've only had to take over a couple times where it go too close to a barrier for comfort.

 

[S4WRXTTCS]

I have owned and driven several cars with TACC for years and Tesla is the first one with serious and repeated ghost braking. Earlier it was mostly perhaps when a car in front of you slowed down and turned away kind of stuff but Tesla mistakes overhead bridges for emergency stops so that’s a new one...

It has gotten somewhat better lately of course. Early days of AP2 were horrible.

That said I still consider Tesla’s TACC much more clearly a safety benefit than Autosteer where the jury feels more out in my view.

Tesla attempt to make TACC smarter than it really is has ruined TACC.

In the old days with TACC with AP1 there was no false braking or even braking for corners. It simply braked for traffic, and only occasionally would it false brake in a corner when it got confused by a car next to it.

With AP2 TACC it will false brake for overhead signs, and during lane changes. Like yesterday it false braked while I was passing someone so it made me look like I brake checked someone.

Yesterday I noticed it braking not because of a false braking event, but because it got confused about where I was location wise on the maps. It was in Tacoma and where I5 is in Tacoma has no correlation with what's on the maps. So suddenly it has me slowing down.

I really wish TACC had a couple settings where I could simply say I want dumb TACC. Where I want it to be radar only, and only track/see moving vehicles. I don't want anything fancy.

 

[StealthP3D]

I used EAP for about 80 miles in heavy rain today, the kind where everything is just grey mist, on three different kinds of roads (Interstate, 2 state highways and some county roads) and it did a fantastic job even though it was difficult to see the car ahead and the lane edges. Simply phenominal how well it performs in low visibility situations.

 

[S4WRXTTCS]

I wonder how traffic aware cruise has been for other manufacturers. I have not driven anything with it more than casually. It wouldn't surprise me if some of the other systems phantom brake in similar situations. Has anyone else here been getting regular miles in one of the other systems? I would actually expect Cadillac to be the best about this because of their geofencing and HD mapping.

All adaptive cruise control systems are susceptible to false braking.

Even human beings are susceptible to false braking in a rare occurrence.

It's also highly dependent on the type of system, and the design constraints.

My first experience with adaptive cruise control was Tesla with AP1. That experience was a lot better than AP2 TACC in that AP1 had very few false braking events. It was based on mobile-eye technology, and was quite good at having minimal false braking events.

There was a recent study done that showed that AP2 had far more false braking events than AP1. Although that study was frustrating in that they didn't upgrade the firmware of the AP1 system to a newer versions so it was hard to tell if it was the ADAS hardware (radar, camera, detection, etc) or what it was tasked with (more of the firmware element).

With the Subaru eyesight system it used to be more susceptible to false braking events, but recent versions haven't had as nearly as many.

I think it's acceptable to have them in a rare occurrence, but they happen far too often with AP2 than I'm comfortable with. Sure I know generally what situations to be on the look out for, and I have pretty good reaction to the feeling. So I pretty much know instantly when it's going to happen. But, do I allow someone to borrow my car for a long drive with TACC? Hell no.

 

[S4WRXTTCS]

I used EAP for about 80 miles in heavy rain today, the kind where everything is just grey mist, on three different kinds of roads (Interstate, 2 state highways and some county roads) and it did a fantastic job even though it was difficult to see the car ahead and the lane edges. Simply phenominal how well it performs in low visibility situations.

As frustrating as I find NoA I have to credit Autosteer for how it does in poor visibility areas in staying in the lanes. Or at least making it easier on me as the driver.

That's extremely important in the PNW where grey mist is 9 months out of the year.

 

[DarkMatter]

In the old days with TACC with AP1 there was no false braking or even braking for corners. It simply braked for traffic, and only occasionally would it false brake in a corner when it got confused by a car next to it.

I think that might be a bit of rose colored glasses. The first AP1 loaner I had in 2016 was impressive, but also freaked me out with a phantom braking event on a metal bridge. The most recent AP1 loaner I had could not make it through a corner I have been using as an AP2 test corner without crossing the center line and hitting the brakes. I would trust AP2 in more situations than AP1, though in the situations that AP1 has down, it does tend toward more confidence. Caveat: I only have a few hundred miles on AP1, compared to thousands on AP2.

 

[electronblue]

I think that might be a bit of rose colored glasses. The first AP1 loaner I had in 2016 was impressive, but also freaked me out with a phantom braking event on a metal bridge. The most recent AP1 loaner I had could not make it through a corner I have been using as an AP2 test corner without crossing the center line and hitting the brakes. I would trust AP2 in more situations than AP1, though in the situations that AP1 has down, it does tend toward more confidence. Caveat: I only have a few hundred miles on AP1, compared to thousands on AP2.

That’s because the ghost brakings due to overhead structures started in 2016. They have became less for both AP1 and AP2 since that time, fortunately. The bad thing is that they seem to come back in some updates more so than in others. (The ability to take corners is completely unrelated, AP1 has one narrow camera unable to see tight corners. So let’s stick to radar ghost brakings.)

Let me explain the historical and technical context behind Tesla’s radar ghost brakings that started in 2016:

As @S4WRXTTCS explains above, Tesla uses the adaptive cruise control radar in an overly ambitious manner. The reason for this was the Joshua Brown incident and the ”radar point cloud / whitelist” Tesla implemented. This is what makes Tesla’s TACC so much more susceptible to ghost braking compared to other radar systems which mostly use radar to track moving objects. The competition’s radar ghost brakings thus are usually limited scenarios like when the moving object in front of you turns away and there is a delay in radar acknowledging this.

But Tesla nowadays and already in 2016 uses their radar to track non-moving blocking objects as well, again due to the Joshua Brown incident where such a feature would have helped. Tesla has implemented a very ambitious Autosteer (given the otherwise limited autonomy of their cars) and are more susceptible to incidents like Joshua Brown where the driver loses attention and this is why they did this. Another thing Tesla of course implemented were the nags which were mostly missing before that time as well. This applies to both AP1 and AP2 as both use radar in a similar manner and these things are currently mostly unrelated to their vision systems.

However this does not apply to the early days of AP1 because originally Tesla did not implement the system to brake for non-moving radar hits. Once they did, the problem then became that radar also takes hits from overhead structures like bridges and those are the one’s that have caused the nastiest ghost brakings because they are unrelated to the flow of traffic and thus the most unexpected (unless you become paranoid at every overhead structure which indeed some AP users probably do).

Audi has a similar system in their new Audi A6/A7/A8 family (already shipping) but they use Lidar instead which is not susceptible to such ghost readings. Most other manufacturers ignore non-moving objects in their radar and instead rely on vision to notice non-moving objects or ignore them entirely. (Of course the fact that Tesla has had a more validated, stable vision in AP1 in MobilEye has meant that AP1 has been somewhat less susceptible to issues over the past couple of years due to additional vision-related incidents being less. Within the limits of AP1s one narrow camera of course — not seeing tight bends etc).

 

[StealthP3D]

However this does not apply to the early days of AP1 because originally Tesla did not implement the system to brake for non-moving radar hits.

It's pretty impressive how quickly EAP is improving. We didn't take delivery of our first Model 3 until May and it's so much more consistent and refined now then it was just 7 months ago. The speed of development is stunning and actually seems to be accelerating in the last couple of months. I've never driven AP1 but it sounds like it was the right move to ditch it. Tesla moves at lightning pace compared to most car makers.

 

[electronblue]

It's pretty impressive how quickly EAP is improving. We didn't take delivery of our first Model 3 until May and it's so much more consistent and refined now then it was just 7 months ago. The speed of development is stunning and actually seems to be accelerating in the last couple of months. I've never driven AP1 but it sounds like it was the right move to ditch it. Tesla moves at lightning pace compared to most car makers.

Your view might be different had you used AP2 longer. But it is true Tesla has a very aggressive approach to software updates and feature changes compared to other manufacturers. The momentum is not always forwards though.

 

[StealthP3D]

Your view might be different had you used AP2 longer. But it is true Tesla has a very aggressive approach to software updates and feature changes compared to other manufacturers. The momentum is not always forwards though.

I think the nature of neural net learning is "two steps forward, one step backward". It's pretty clear that EAP is more capable and more safe than ever. I can't believe how fast it's improving, it was pretty rough back in May. I use it in a lot of unsupported situations, curvy 2 lane highways, country roads, inclement weather, etc. and it just keeps impressing me more every month. We have two Model 3's with EAP and I keep hearing about these phantom braking events but have yet to experience any HARD phantom braking events.

The only time either of our Model 3's have automatically braked with any force was when a car pulled RIGHT out in front of us (from behind a big box truck). I could see the wheels of the car under the truck (but they couldn't see me). Instead of manually braking (I was on EAP at the time) I decided to wait until the last minute to see how the car would handle it (I was only going about 30 mph). Sure enough, right before I slammed on the brakes, AEB beat me to it and hit the brakes HARD. It let off when we were going about 10 mph and the threat had cleared the front of our car. It was a certain collision with no braking.

Yes, I'm pretty impressed with AEB and it is certainly already making our roads a safer place to be. That's why insurance companies endorse these systems.