Navigate on Autopilot is Useless (2018.42.3)

[whttiger25]

Ah, but you're not looking at your full purchase of FSD . . . it's hardly done yet.

Sure, it's not where it needs to be. We ALL know this.

What you've done is purchased something worth some six-figures--in a few years--for $6,000 now. It's getting better with each OTA update, and one day soon, perhaps as early as next year, it will be Full Self Driving.

Hang in there.

Please please please don’t tell people what they have purchased will be “worth” six figures. Until musk allows you to take your FSD Package to another car (no sign of that happening) it is NOT an investment. By the time it’s “worth six figures” the car it’s attached to will be likely a clunker.

Plus, FSD will never be worth 6 figures to anyone who doesn’t want to lend their car to the robo-fleet.

Plus! Competition is coming. Tesla is ahead but not the only FSD game in town. I’d put real money on prices reversing course and coming down before they Get anywhere near 6 figures.

Lastly, saying Tesla is going to “solve” FSD in a way that allows a driver to feel comfortable giving the car complete control in most situations by next year is hopelessly optimistic. And if you can’t give the car complete control, what is the real value?

 

[drtimhill]

Wow, so that's what confused you? I'll let myself out.

Just because YOU have decided what a term means doesnt mean others either (a) know what you mean by it and (b) agree with your definition. Since there are no industry-agreed terms for what things mean (thank you, marketing departments) its important to clarify what you mean when using some of these terms.

 

[wk057]

Just checking in on this...

Tried this out again on a longish drive the other day on the latest firmware. It still has pretty much all the same major problems as nearly 2 years ago. Phantom braking galore, extremely poor lane change decisions and executions, etc.

I still stand by my original posts contending that NoA is still useless.

They've even made it more useless since with the incessant nagging, poor hand/torque-on-wheel detection, etc. What's the point of letting the car do an autonomous lane change if it'll just sit there like a jackass in the lane with the blinker on and go no where until you jiggle the wheel enough for it to detect your hand? And by the time that actually happens, it either misses its opening, exit, etc.

As far as I can tell there's been pretty much zero improvement to the usefulness of NoA in nearly 2 years, and that's even ignoring that requiring hands-on-wheel for it to take any action completely defeats the purpose of allowing the car to do these things at all. They need to either remove the nagging entirely, or limit the nags substantially. (Ie: nag when confidence is down, or at a reasonable interval on the order of multiple minutes... not every 10-15 seconds). If they did that, at least the other failings would probably be tollerable. But as it sits, I still see zero reason to use it over just using autosteer with manually initiated lane changes.

 

[DrDabbles]

It's because the car makes DIFFERENT choices than what Dr. Dabbles would make. So to him, those choices are useless. To me, they are great. I just have (I guess?) trash driving habits while Dr.Dab is a good driver.

You absolutely couldn't possibly be farther from the truth. But maybe now that you've driven with AP for a couple months, you see how foolish this statement was. The fact that AP makes "different" decisions isn't a concern. The fact that it makes terrible "decisions" is the concern.

It might sound ridiculous when framed like that but it's definitely true. If you don't like the choices the car makes, you won't use it and therefore it's useless. If you really want to go 10 over on city streets but the car refuses, it is the very definition of useless, to you. If you don't mind going the speed limit, maybe the feature has value to you. The car could go 10 over, but it chooses not to, for one reason or another.

More broadly, I mostly hear people complain about how others drive and never about how bad of a driver they are. I attribute this to mostly just disagreeing with the choices other drivers make.

Again, this is far from true. Human drivers make different choices than I would, and I'm completely fine with that. Hell, I drive more conservatively than my fiancée and it drives her crazy, but she doesn't feel unsafe. AP is unsafe in many situations. This isn't a matter of not liking decisions. When it attempts to veer into another lane, or haphazardly phantom brakes, or randomly slows to 20 below the posted speed limit, or crashes into barriers and kills someone, that's not disliking the decision, it's dangerous. When you get enough experience with AP, you'll see these behaviors too.

really good summary and yup, totally agree. Maybe some folks were thinking you meant it is totally worthless to everyone But yea, good point. The car is not driving quite the same as I do, but close enough that I like it. Let's just hope it continues to improve.

The car isn't driving itself, it's aiding you.

 

[DrDabbles]

As far as I can tell there's been pretty much zero improvement to the usefulness of NoA in nearly 2 years, and that's even ignoring that requiring hands-on-wheel for it to take any action completely defeats the purpose of allowing the car to do these things at all. They need to either remove the nagging entirely, or limit the nags substantially. (Ie: nag when confidence is down, or at a reasonable interval on the order of multiple minutes... not every 10-15 seconds). If they did that, at least the other failings would probably be tollerable. But as it sits, I still see zero reason to use it over just using autosteer with manually initiated lane changes.

There have been some real fun regressions, though, like detecting the car is on an offramp instead of on the highway and slowing to offramp or surface road speeds more often. Oh, and driving through construction areas where highway lanes are moved makes it attempt to exit the highway to turn around.

Beyond that, lane selection still seems to be based on varying quality map data, overtaking still has all the same issues of wanting the change lanes when one isn't necessary, or waiting too long before prompting for a change. A partial fix for not handling offramp splits properly has been implemented though- NoAP will attempt to get you out of the right-most lane in North America at least. Which sucks when you actually want to be in that lane.

 

[eli_]

One of my biggest pet peeves is the behavior they added awhile back that slows for adjacent cars merging into your lane. I'm pretty sure that is the cause of many "phantom braking" episodes. It badly needs more tuning. The way it works seems super naive to me, it pretends the adjacent car in a merging lane is actually in your own lane, so it will hard brake to avoid passing them instead of maintaining a constant speed, even if the other car is laterally far away from you in a divided lane, or the merge point is still a quarter mile away, or you have a large speed delta, it doesn't matter. It's also dependent on map data to know what lanes are merging into what, and of course that is unreliable. So it may brake for an adjacent car that will never merge into yours, either because the map data or lane localization is wrong.

 

[Gesteur]

I get phantom breaking with and without NoAP. tACC is really bad , sometimes I think I d preferred to have just a setting to keep the speed constant and let the driver intervene to break.
I almost nevee use NoAP since benefits are not many, it suggest lane changes, but somehow when i action it with the turn signal it wouldnt matter. It s more in it s beta dev i believe.

What i i dont get, reading limit signs should be a piece of cake, why isnt the functionality there yet for AP2 cars?!

 

[emmz0r]

I get phantom breaking with and without NoAP. tACC is really bad , sometimes I think I d preferred to have just a setting to keep the speed constant and let the driver intervene to break.
I almost nevee use NoAP since benefits are not many, it suggest lane changes, but somehow when i action it with the turn signal it wouldnt matter. It s more in it s beta dev i believe.

What i i dont get, reading limit signs should be a piece of cake, why isnt the functionality there yet for AP2 cars?!

LOL the absolute state! It's so bad that people actually want "dumb" cruise control that has existed for ... at least 2 decades, instead of this.
It's very sad, but very understandable. Much of the fault is due to terrible speed limit map data, and the failure to read speed signs.

 

[LowlyOilBurner]

What i i dont get, reading limit signs should be a piece of cake, why isnt the functionality there yet for AP2 cars?!

Patents. Here’s the thing though: I had an AP1 Model S as a loner a month or so ago that could read speed limit signs and adapt to them. It wasn’t as reliable as same tech in my past Volvo’s. It frequently missed signs, and it would pick up school speed limit signs and slow the car when it didn’t need to. I get that, but it was annoying, and people behind me weren’t happy with the car slowing down when it shouldn’t have. Reliable speed limit detection and adaptation would be great. Hopefully we see this in the future.

 

[emmz0r]

"Everyone" is thrilled by how much Tesla is worth. If the patent is the big problem here, just f***ing pay them already.

Sign reading should have been there from day 1 in 2016. I was kinda shocked that it wasn't when I got my car.

 

[DrDabbles]

"Everyone" is thrilled by how much Tesla is worth. If the patent is the big problem here, just f***ing pay them already.

Sign reading should have been there from day 1 in 2016. I was kinda shocked that it wasn't when I got my car.

Software and tech patents in general are beyond predatory, so it's likely not just a "pay them for a license" situation. MobilEye owns most of the patents on this kind of tech, and they aren't in the licensing business. You either use their product, or you get nothing. That was one of the initial rifts between Tesla and MobilEye back in 2016.

To make matters worse, there have been dozens of other companies working in this space before Tesla arrived on the scene, so the minefield that is the patent situation is impossible to navigate. So, Tesla needs to have a "clean room" implementation created by people that have never had any exposure to the patents, briefings, or even talks with other companies doing automated sign reading. This means Karpathy can't be part of the team that works on it, so it needs to be a whole new group. And even then, with all those precautions taken, Tesla is likely to have a suit filed against them and will need to take part in discovery.

This is why Europe is so strict about the kinds of "e-patents" that can be file for.

 

[emmz0r]

Software and tech patents in general are beyond predatory, so it's likely not just a "pay them for a license" situation. MobilEye owns most of the patents on this kind of tech, and they aren't in the licensing business. You either use their product, or you get nothing. That was one of the initial rifts between Tesla and MobilEye back in 2016.

To make matters worse, there have been dozens of other companies working in this space before Tesla arrived on the scene, so the minefield that is the patent situation is impossible to navigate. So, Tesla needs to have a "clean room" implementation created by people that have never had any exposure to the patents, briefings, or even talks with other companies doing automated sign reading. This means Karpathy can't be part of the team that works on it, so it needs to be a whole new group. And even then, with all those precautions taken, Tesla is likely to have a suit filed against them and will need to take part in discovery.

This is why Europe is so strict about the kinds of "e-patents" that can be file for.

Just do it anyway, and let the court cases drag on. Tesla has the money to go on for years in the courts.
Fact is - that this is 100% necessary to get anywhere at all with FSD. So it is forced to happen sooner than later.

 

[S4WRXTTCS]

Just checking in on this...

Tried this out again on a longish drive the other day on the latest firmware. It still has pretty much all the same major problems as nearly 2 years ago. Phantom braking galore, extremely poor lane change decisions and executions, etc.

I still stand by my original posts contending that NoA is still useless.

They've even made it more useless since with the incessant nagging, poor hand/torque-on-wheel detection, etc. What's the point of letting the car do an autonomous lane change if it'll just sit there like a jackass in the lane with the blinker on and go no where until you jiggle the wheel enough for it to detect your hand? And by the time that actually happens, it either misses its opening, exit, etc.

As far as I can tell there's been pretty much zero improvement to the usefulness of NoA in nearly 2 years, and that's even ignoring that requiring hands-on-wheel for it to take any action completely defeats the purpose of allowing the car to do these things at all. They need to either remove the nagging entirely, or limit the nags substantially. (Ie: nag when confidence is down, or at a reasonable interval on the order of multiple minutes... not every 10-15 seconds). If they did that, at least the other failings would probably be tollerable. But as it sits, I still see zero reason to use it over just using autosteer with manually initiated lane changes.

On longish drives in fairly rural areas it used to be useful because it would automatically change lanes, pass a vehicle on the left, and then get back over. Sure it wasn't incredibly useful, but it did do something. It didn't have any more false braking than TACC has itself (unless the maps were wrong). The autolane changes seemed to work for me at least. Sure sometimes you had to apply a little torque if you sensed a little longer delay.

But, then they ruined it.

It got so bad that some people reviewing it didn't even give it a score. They just said "Due to major bugs I can't review this". Bugs like not getting out of the passing lane or long delays in changing lanes no-confirmation mode.

Then there was the HW2/HW2.5 to HW3 jump where we had people with HW3 reporting different behaviors. I went from HW2.5 to HW3, and it seemed to clear up the issue with not getting out of the passing lane.

But, it still has this weirdly long delay when changing lanes. When mixed with the torque sensor issue it just becomes a mess. You have to look at the screen to see it its wanting wheel torque.

So I would list the desperately needed fixes as:

The no-confirmation lane change delay has to be either eliminated or allowed to be user settable. Even Miss Daisy would complain about how long it was taking to start the lane change.
Completely decouple steering wheel torque from auto-lane change. They advertised it as no-confirmation so why couple it to wheel torque?

As a medium term goal they need to get rid of the torque sensor entirely. Having a true driver monitoring system will be a requirement in certain regions very soon so why keep delaying the enviable? Tesla needs to accept that the steering torque sensor is dead. it was a dumb idea that has finally died. Just give up, and give us an option to buy an upgrade to rid our vehicles of such awfulness.

 

[drtimhill]

But, it still has this weirdly long delay when changing lanes. When mixed with the torque sensor issue it just becomes a mess. You have to look at the screen to see it its wanting wheel torque.

While I agree with most of your points, you can (and I have) switch to have it vibrate the wheel, which at least gets around the "looking down at the screen" bit. But yeah I agree it takes FAR too long to actually make the lane change.

 

[S4WRXTTCS]

While I agree with most of your points, you can (and I have) switch to have it vibrate the wheel, which at least gets around the "looking down at the screen" bit. But yeah I agree it takes FAR too long to actually make the lane change.

Thanks. I've used the "vibrate the wheel" setting before for confirmed lane changes, but I never used it for non-confirmed lane changes.

 

[PhaseWhite]

There have been some real fun regressions, though, like detecting the car is on an offramp instead of on the highway and slowing to offramp or surface road speeds more often. Oh, and driving through construction areas where highway lanes are moved makes it attempt to exit the highway to turn around.

Beyond that, lane selection still seems to be based on varying quality map data, overtaking still has all the same issues of wanting the change lanes when one isn't necessary, or waiting too long before prompting for a change. A partial fix for not handling offramp splits properly has been implemented though- NoAP will attempt to get you out of the right-most lane in North America at least. Which sucks when you actually want to be in that lane.

The lane selection issue still seems like low hanging fruit to me. The car can tell if it's in the right most lane based on vision, but if the map data is wrong, it will sit in the wrong lane approaching an exit it knows is on the right side and miss the exit. There's a stretch of road my my house I used to travel weekly that the map data still doesn't know exists and it NOA fails.

Given the almost complete lack of improvement on NOA I can only conclude that Tesla has not been spending much time there recently. Perhaps because they are working on City NOA or because they are planning to implement the major re-write (4d vs 2.5d Elon talked about on Q2 earnings call) which fixes a lot of these issues.

It's frustrating because even a rules based algorithm for lane changes could have made this so much better than it is now. The dumb logic we have now appears to be unchanged since it went live. It works something like for every mile you are away from the exit, make sure you are in the X minus miles away from exit lane. The car always immediately turns on the blinker whether the adjacent lane is clear or not a that point.

A simpler but easy approach to improve it would be something like:

If within 2 miles of the exit start trying to move to the exit lane but only turn on the blinker and move over when a gap in traffic appears, then if you could not move by the time you hit within 1 mile, turn on the blinker and decide to speed up (within some defined range) or slow down for the nearest gap.

 

[DrDabbles]

Just do it anyway, and let the court cases drag on. Tesla has the money to go on for years in the courts.
Fact is - that this is 100% necessary to get anywhere at all with FSD. So it is forced to happen sooner than later.

This is just about the worst, slowest way to get anything done. I'm presuming you don't have much experience with the US court system, and aren't considering every single country has its own to navigate.

The car can tell if it's in the right most lane based on vision, but if the map data is wrong, it will sit in the wrong lane approaching an exit it knows is on the right side and miss the exit.

The major problem is that these cars do not use the vision system to localize within a map location. What needs to happen is the vision system needs to be used to glean the exact position in space where the car is. This would get rid of the stupid behavior where the GPS says the car is on an offramp, it would get rid of the annoying behavior where maps don't show a lane or construction has physically moved a highway, etc. But this is not an easy problem, and I don't see a meaningful fix coming within the next couple years.

I can only conclude that Tesla has not been spending much time there recently.

Almost certainly true. Same with smart summon. There have been no appreciable changes to behavior of AP, NoAP, or Smart Summon since before December.

Perhaps because they are working on City NOA

Given how poor highway NoA works, let's hope not. If I can't trust my car to remain properly positioned in a splitting or merging lane, there's zero chance it will handle unprotected left turns without someone being killed.

they are planning to implement the major re-write (4d vs 2.5d Elon talked about on Q2 earnings call) which fixes a lot of these issues.

They've been working on that since last year, Elon has talked about it for almost a year now, and I expect this will be the same as every other rewrite. We absolutely can not say it solves any problems yet, because it's not in production use yet.

 

[Gesteur]

With the lastest update 2020.24.6.11 even tACC is useless. On highways I get scary phantom breaking, tesla should get their s#** together. It s dangerous for driver and cars following you.

 

[Silicon Desert]

With the lastest update 2020.24.6.11 even tACC is useless. On highways I get scary phantom breaking, tesla should get their s#** together. It s dangerous for driver and cars following you.

Then I suggest you don't use it until they fix it. Then it won't be dangerous. Sounds like you continue to use it anyway, or am I wrong about that?

 

[LowlyOilBurner]

I’m on 2020.28.6, and phantom braking has improved greatly. I get it in a few instances, much better than it has been.