FIRMWARE UPDATE! AP2 Local road driving...and holy crap

[montefuego]

HI,
This is my first time on this site, I've had my Model S for a month now. In the last couple days, I have tried the autosteer on the freeway, and a variety of roads. i think it never goes more than 10 seconds without doing something that is not just wrong, but potentially hazardous. The speed and acceleration of the car makes it particularly scary: this morning, it suddenly accelerated right toward a wall behind a sharp turn in front of me.
I am not just an owner, but a stockholder like many of you. Believe me, I do not want to find fault if I don't have to. But this software seems very far from becoming usable. When Chris Lattner, the new software engineer from Apple was hired a month ago, he described autonomous driving: "It is also a very, very hard technology problem." Notice the use of two 'very's' by someone who really knows what he is talking about.
I personally now believe that the separation from Mobileye was far more damaging than Elon will admit. Also, the loss of his earlier engineer, that hurt enough to compel Elon to sue him.
Although he is clearly a genius, and perhaps the man I most admire, Elon's confidence that they can quickly do their own autopilot looks like classic hubris. Hubris is the foundation of most tragic literature, and this step in Tesla's journey is starting to look like tragedy (waiting to happen). I will shocked (pleasantly) if anything usable occurs in the next three months, and possibly not this year. I even wonder about his claim that HW2 is capable of level 5. Audi supposedly has two radar sensors, not one. And Elon has already mentioned they will likely need to replace and upgrade the Nvidia processor. So a lot of warning flags have been thrown into the air....
I am certainly not 'angry' and enjoy driving, so I'm not as impatient as some on this thread. And great ambitions deserve some extra dispensation. But I do find this disappointing.

 

[montefuego]

The bike lane had nothing to do with it.

How is it that you do not understand? When AP1 first arrived (7.0) these situations were a piece of cake. Surface road was easy pickings for AP1. Now you need to take over every 10 seconds or you're dead. The point isn't that ap2 hasn't reached parity with ap1. Its that 17.7.2 for surface roads is dangerous. that's literally from every 17.7.2 video I have seen.

Again Ap2 is good on highway/freeway. No one is denying that. but its beyond dangerous on most surface roads.

I completely agree. I think that some here are either blindly allegiant, or have stock holdings that bias them toward overlooking the unpleasant truth: Tesla's autopilot, on its own without Mobileye, is so far a disaster.

 

[bhzmark]

I've seen nothing indicating this early AP2 is worse than the similarly early AP1. And even the more fully baked AP1 now still does not navigate intersections or roads without lane markings or traffic circles very well or at all.

AP is best for well marked divided highways and stop and go traffic surrounded by slow moving cars. And both of those situations are when driving is most tedious and thus when AP is most desirable.

I think Tesla has mismanaged the communication on what to expect from EAP as it progresses to AP1 parity.

Since they acknowledge that AP2 isn't at AP1 parity right now, video showing AP2 not handling situations that AP1 can't even handle isn't a fail. That is just what it means to not yet have surpassed AP1 abilities yet.

 

[Bladerskb]

I've seen nothing indicating this early AP2 is worse than the similarly early AP1. And even the more fully baked AP1 now still does not navigate intersections or roads without lane markings or traffic circles very well or at all.

AP is best for well marked divided highways and stop and go traffic surrounded by slow moving cars. And both of those situations are when driving is most tedious and thus when AP is most desirable.

I think Tesla has mismanaged the communication on what to expect from EAP as it progresses to AP1 parity.

Since they acknowledge that AP2 isn't at AP1 parity right now, video showing AP2 not handling situations that AP1 can't even handle isn't a fail. That is just what it means to not yet have surpassed AP1 abilities yet.

Are you naïve? these are roads that are well marked. Ap1 can handle all of those.

AP2

AP1

there are hundreds more videos on the same road.
are you blind?

 

[bhzmark]

Are you naïve or delusional? these are roads that are well marked. Ap1 can handle all of those.
are you blind?

I don't see a right side lane markings and it is mostly through intersections on local roads. I wouldn't enable my AP1 on that road either.

 

[Bladerskb]

I don't see a right side lane markings and it is mostly through intersections on local roads. I wouldn't enable my AP1 on that road either.

AP1

there are hundreds more videos on the same road.
are you blind?
I wanna know so I can save myself of posting hundreds of videos in the next couple hours.

 

[Ulmo]

This about sums up the latest release for me. I had the identical experience.

In

at 18:50 it shows that the car cannot even tell where the crest of the road is, something we were taught in Driver's Ed. Back when I was taught in Driver's Ed, we were taught how to drive cars on roads with NO PAINT AT ALL. The car ought to be able to drive without any lane markings whatsoever, as well as poor lane markings and "great" lane markings.

 

[_jal_]

I think the lane lines display detracts from having confidence in the system. On roads that the car is navigating fine, the lane lines show confounding info. Why on a straight road with no obstacles does the lane diagram show a curve one second and then a straight line the next, and so on? I know that there are a lot of theories about how the NN is being trained, but like @Ulmo points out, why doesn't the car know how to find the road? I'd think that'd be one of the first things to train because that'd provide a lot of info for other systems.

I'm still optimistic everything will be okay, but it would be helpful to get a little insight into what's going on from Tesla. In the void of info all we have is a diagram that shows our car thinks it may live in a world made of rubber where roads actually move side to side under cars.

 

[bhzmark]

AP1

there are hundreds more videos on the same road.
are you blind or delusional?
I wanna know which one so I can save myself of posting hundreds of videos in the next couple hours.

Do you have AP1 ? Or just a limited vocabulary of insults?

I use AP1 extensively on highways and other good roads with lane markings or in lots of circumstances with stop and go driving.

Through intersections and on roads with poor lane markings and certainly traffic circles and the like, I don't use AP1 as I would have to frequently take over. This based on thousands of miles and 15 months of experience with AP1 from when it was first rolled out in October 2015.

Showing a few select videos of AP1 at its lucky best and currently still being trained AP2 at its worst, and especially while it is clearly and repeatedly stated that AP2 is not yet at parity with AP1 doesn't demonstrate anything interesting.

 

[u00mem9]

I even wonder about his claim that HW2 is capable of level 5.

It is clear that when that claim was made, they could not have known that it was capable.

 

[stopcrazypp]

I've seen nothing indicating this early AP2 is worse than the similarly early AP1. And even the more fully baked AP1 now still does not navigate intersections or roads without lane markings or traffic circles very well or at all.

AP is best for well marked divided highways and stop and go traffic surrounded by slow moving cars. And both of those situations are when driving is most tedious and thus when AP is most desirable.

I think Tesla has mismanaged the communication on what to expect from EAP as it progresses to AP1 parity.

Since they acknowledge that AP2 isn't at AP1 parity right now, video showing AP2 not handling situations that AP1 can't even handle isn't a fail. That is just what it means to not yet have surpassed AP1 abilities yet.

I was also under the same impression that autopilot in general is unpredictable in truly local roads with lots of intersections. This did not seem to be an issue with AP1 as people stuck with primarily using it with divided (and sometimes undivided highways).

It seems their communication was done poorly for AP2 by setting expectations that true local use was an expected use case (as it seems most AP2 users I have seen in this thread seems to expect it to work flawlessly in local roads). I think that was partly driven by the speed limitations.

 

[disagree]

The release notes tell us that we now have:

"Autosteer while on local roads", which is then explained as "On local streets and roads without a center divider..."
​

But we are not supposed to think it will work on local roads. Right, ok.

Is there a number to call to get my Real Meaning Detector Glasses? Seems my DS forgot to give them to me.

 

[u00mem9]

The release notes tell us that we now have:

"Autosteer while on local roads", which is then explained as "On local streets and roads without a center divider..."
​

But we are not supposed to think it will work on local roads. Right, ok.

Is there a number to call to get my Real Meaning Detector Glasses? Seems my DS forgot to give them to me.

Maybe a couple more true believers will weigh in that it's a 'communication issue'. Lol

 

[stefan.bc]

Also be careful of narrow local roads without a center divider(but with a median) and no markings on your right. You will probably end up in the ditch. I was not brave enough to find out.

 

[EinSV]

Here we go again
@EinSV - I don't really care about numbers(I already proved that the results in the report are expected no matter what the car is if they have a decent collision avoidance system), <snip>.

The data that you posted simply tells us about the effectiveness of forward collision avoidance which every manufacturer has and is required by law.

Doesn't say anything about AP1 driving behavior or lack of thereof. <snip>

Sorry those statements are not supported by the data from the NHTSA and IIHS.

The NHTSA reported a 40% decrease in serious accidents after AP1 was enabled. The IIHS's extensive analysis of AEB and FCW covering five manufacturers' systems (Mercedes, Volvo, Acura, Honda and Subaru) over a total of four years concluded that the overall decrease in accidents from both AEB and Forward Collision Warning Systems was only 6%, which did not even reach the level of statistical significance. http://orfe.princeton.edu/~alaink/S...s/IIHS-CicchinoEffectivenessOfCWS-Jan2016.pdf at page 15.

I think the confusion arises because the IIHS report also concluded that AEB plus FCW reduced the number of rear-end collisions by 39%. But rear-end collisions are only about 25% of total collisions, and the IIHS only looked at 1/2 of rear-end collisions -- those in which the car with AEB/FCW rear-ended another car.

Bottom line: the IIHS did an extensive analysis and found that other manufacturers' FCW/AEB systems reduced overall accidents by only 6% -- so small that it was statistically insignificant. So the data just don't support your conclusion that the 40% reduction in serious accidents achieved by AP1 is something that has been obtained by other manufacturers' collision avoidance systems. The IIHS looked deeply into this very issue and came to a different conclusion. In fact, the 40% reduction in serious accidents the NHTSA reported for AP1 is many times higher than the accident avoidance the IIHS found was achieved by other manufacturers' FCW/AEB systems.

 

[mhan00]

Meh, I remember these exact types of threads and complaints when people were waiting for AP1 to activate, which took like a year, iirc. Even had a couple guys threatening to take legal action too. The more things change . . .

 

[bhzmark]

With AP1, and likely with AP2, it is primarily about the lane markings and secondarily whether you are following another car, whatever the type of road.
Missing lane markings on one or both sides, which almost always happens at intersections, will often cause problems.. even still with AP1. This can be mitigated somewhat if you are following another car.

Missing lane markings occur more frequently on local roads. Hence even though autosteer is now able to be engaged on local roads, if those roads have deficient lane markings, and especially if there isn't a target car ahead to show the way, AP will not work well. This is true on divided highways with deficient lane markings too.

It looks like AP2 is still struggling to reach parity with AP1 on even following some lane markings on curves. That's probably why EM said AP2 still doesn't have parity with AP1. Hopefully it will come sooner for the first AP2 buyers than it did for the first AP1 buyers.

 

[sdorn]

I don't consider a system safe when at some overpasses or some other metal structures that "bend" over the highway, I have to pay attention if someone is behind me.

You should always know whether someone is behind you. You are the driver. You should always know whether someone is in front of you, beside you, in your blind spot, behind you, at the intersection that is coming up, at the intersection you just passed, etc. You should know where every vehicle and every pedestrian and every stationary object is that is in your range of sight. That is how driving works.

 

[drshan]

You should always know whether someone is behind you. You are the driver. You should always know whether someone is in front of you, beside you, in your blind spot, behind you, at the intersection that is coming up, at the intersection you just passed, etc. You should know where every vehicle and every pedestrian and every stationary object is that is in your range of sight. That is how driving works.

AP2 places an enormous effort on the driver. I flick it on, on well marked local roads out of curiosity w the assumption that it is trying to kill me. And then give up after exhaustion. Hopefully there rapid improvements that gain the trust of the driver. Personally shocked this was released to the public (based upon my lofty view of the corp).

Again...great car...happy w every other aspect. But BAD software so far...

 

[Zhelko Dimic]

Are you naïve? these are roads that are well marked. Ap1 can handle all of those.

AP2

AP1

there are hundreds more videos on the same road.
are you blind?

I am brand new owner of AP1 car (3-4 days of driving), and I don't understand AP1 capabilities yet.
From that viewpoint, I would not be surprised if AP1 was as bad as your videos. You're saying it's not, that's great to hear. But it speaks of my expectations for AP in general...

While disappointing experience, your use case: impatient, previous AP1 owner that knows what he's lost, is somewhat niche and unlikely to damage Tesla reputation much, especially if it get solved in the next couple of months...

Please take a solace in how cheap your education on 'Tesla time' was... It cost me 50% of all my retirement savings until I understood how different Tesla is in communicating timelines, compared to most any other company. That was few P100Ds gone...