Navigate on Autopilot is Useless (2018.42.3)

[TrafficEng]

I have used NOA a number of times (4 or 5 times) in the Atlanta area, driving from up north to the airport down south over the I85, I75, and 400 interstate system, and back. V2018.42.2. It has worked faultlessly for me -- appropriate lane changes; appropriate exits; nothing scary. Perhaps I'm lucky as it was medium traffic that it did not struggle with -- and I'm not trying to push corner cases. I just let the car drive. My experience has been pretty spectacular!

 

[boonedocks]

I have used NOA a number of times (4 or 5 times) in the Atlanta area, driving from up north to the airport down south over the I85, I75, and 400 interstate system, and back. V2018.42.2. It has worked faultlessly for me -- appropriate lane changes; appropriate exits; nothing scary. Perhaps I'm lucky as it was medium traffic that it did not struggle with -- and I'm not trying to push corner cases. I just let the car drive. My experience has been pretty spectacular!

Same experience when there is little to no traffic. Put rush hour into the mix and NoA is a completely different story.

 

[benfrank3]

Please sell us your nag-free, anti-truck lust, lane change upgrades for AP1 cars! Pretty please!!

 

[dmode]

I'm sensing a theme that a bunch of the proponents that aren't haven't as many issues are driving California routes.

The country is more than California, despite what many Californians may believe. This is probably also why rain sensing on AP2+ absolutely sucks.

I tried to look for more explanation of this point in later pages, but couldn’t find anything. Would you mind clarifying this ? Did you test NoAP in inclement weather ? Or are you referring to lane paints and highway construction in general ?

 

[DarkMatter]

My relative lack of issues are not while driving California roads.

 

[davidc18]

+1 to wishing wk would sell his ap1 mods

 

[rnortman]

I'm throwing my hat in with the "NOA is useless" contingent. Just did a nice long road trip for the (US) holidary weekend on beautiful, clearly marked, wide Midwestern interstate highways, in clear skies and no precipitation. NOA did not do one single useful thing for me. When NOA got it right, it was always something I could have more easily gotten right myself -- NOA just added completely unnecessary stress and jerky handling to what is actually quite an easy process.

The real problem comes when NOA gets it wrong, which it did for me (even on these ideal highways in ideal, low-traffic conditions) more than half the time. On at least half of the exits that I wasn't taking it wanted me to get into the left lane for some reason. You can stay in the right lane without taking a right-hand exit, but it apparently thinks otherwise. And worse, when I approached these exits that I was not taking, the car would slow down abruptly, which is a jarring experience and may lead to getting rear-ended in heavier traffic situations. And then, on the other side of the interchange, 100% of the time when the new traffic is merging in from the right, AP wants to center itself in what it perceives to be a very wide lane, leading to me needing to take over to avoid looking like a drunk on a holiday weekend.

So yeah, NOA for me is useless at the best of times, annoying/uncomfortable almost all the time, and dangerous far too often.

 

[scottf200]

I recently did a long roadtrip for the (US) holiday weekend and went from a suburb of IL to around Hickory NC and back using NoA (~1500 miles). It worked very well on picking lanes for upcoming splits and good-to-great for off ramps. One time it turned on the blinker and went to an off ramp and I thought it screwed up. Turns out it was right and I was in a conversation that lasted several minutes and I lost track of the mental distance to my exit. It actually saved me from missing that exit. Not a huge deal but it worked well. I did do mild lane change suggestions or adjust my TACC speed to be just over what the typical lanes were going so it didn't suggest lane changes as often. ie. no reason to have TACC set at 80 MPH when traffic in most lanes was 69-72. Overall I was pretty happy with it on my 1500 mile trip.

 

[ivengo]

Same thing here I'm very happy how NoA works and up to this date I drove a couple of thousand miles on it.

 

[bhzmark]

This works well when traversing through the LA/OC basin on shopping adventures where I may not be driving in the exact lane for an upcoming freeway change when signage is poor. This feature will greatly add confidence for cross country travel through cities never visited before.

Overall I was pretty happy with it on my 1500 mile trip.

I'm very happy how NoA works and up to this date I drove a couple of thousand miles on it

+1. It's a tool with a certain specific limited functionality. If you can deal with that and make it useful for you great. If not turn it off.

 

[rnortman]

+1. It's a tool with a certain specific limited functionality. If you can deal with that and make it useful for you great. If not turn it off.

The disappointment is not that we have this "tool" forced on us (unlike the UI changes, we can just ignore it). The problem is Tesla is far behind on their promises with respect to EAP and FSP, and Elon promised a "new level of autonomy" in V9, and if NOA is it then we all have cause for concern about how much potential AP has in the future, compared to the promises.

Also, there are legitimate regressions in AP unrelated to NOA. It wobbles in the lane, oversteers in curves, and slams on the brakes for no reason more often than V8 did. The constant subtle wobbling made my daughter carsick during our trip. But that's not what this thread is about...

 

[DarkMatter]

I have not noticed the wobbling at all. I wonder if, like past versions, it works better with the styles of lane markings used in some locations.

 

[rnortman]

I have not noticed the wobbling at all. I wonder if, like past versions, it works better with the styles of lane markings used in some locations.

This is entirely possible. This is a common problem with deep learning approaches; things that should not make a difference end up making a difference, depending on what is over- or under-represented in your training data set, or the quality of the labeling in different images.

 

[josher32]

I have not noticed the wobbling at all. I wonder if, like past versions, it works better with the styles of lane markings used in some locations.

I have notice more wobbling in both my X and 3 since v9

 

[DarkMatter]

This is entirely possible. This is a common problem with deep learning approaches; things that should not make a difference end up making a difference, depending on what is over- or under-represented in your training data set, or the quality of the labeling in different images.

There was a period where I-5 through Tukwila would absolutely throw AP2 for a loop, even though the stripes looked fine to a human. I suspect we are going up keep seeing these regressions each time there is a big feature add. It's a very real downside to machine learning.

 

[rnortman]

There was a period where I-5 through Tukwila would absolutely throw AP2 for a loop, even though the stripes looked fine to a human. I suspect we are going up keep seeing these regressions each time there is a big feature add. It's a very real downside to machine learning.

This is why the big boys don't rely on deep nets to find the lanes. They have maps.

 

[mtndrew1]

Has anyone else noticed that the target car choice on AP is super wonky now?

In the past if I was in a single lane and following a car, that particular car was highlighted on the screen. As of the NoA update the “target” car is all over the map. Next to me, *behind* me, ahead but adjacent lane, etc.

This morning on my commute I watched the highlighted car flipping all around and noticed that it corresponded with weird braking and following distance behavior. The car ahead of me in the lane to my left would slow and so would I. Etc. Is anyone else seeing this?

 

[rnortman]

The car ahead of me in the lane to my left would slow and so would I. Etc. Is anyone else seeing this?

I have definitely seen this. The new DLN is much more likely to think that a car in an adjacent lane is coming into your lane and start ACC'ing off of it. I think this is probably caused by the fact that the bounding boxes are imprecise and unstable, which you can also see in your IC display with the cars jumping around.

I suspect V8 using only one (or 2?) forward camera(s) for this task was much less likely to pick up on nearby cars in adjacent lanes. The downside with V8 of course is that sometimes those cars really are coming into your lane, and V8 was bad at noticing that (despite the bogus claims about "side collision avoidance"). V9 tries to notice but because the bounding boxes are jumping all over the place it gets false positives.

This is among the regressions in V9. Hopefully they will improve the situation over time, but they had a long time to work on V9; why wasn't it already good enough to be free of significant regressions prior to release? Do they need HW3 to get those bounding boxes to quit jumping around?

 

[BigD0g]

I have definitely seen this. The new DLN is much more likely to think that a car in an adjacent lane is coming into your lane and start ACC'ing off of it. I think this is probably caused by the fact that the bounding boxes are imprecise and unstable, which you can also see in your IC display with the cars jumping around.

I suspect V8 using only one (or 2?) forward camera(s) for this task was much less likely to pick up on nearby cars in adjacent lanes. The downside with V8 of course is that sometimes those cars really are coming into your lane, and V8 was bad at noticing that (despite the bogus claims about "side collision avoidance"). V9 tries to notice but because the bounding boxes are jumping all over the place it gets false positives.

This is among the regressions in V9. Hopefully they will improve the situation over time, but they had a long time to work on V9; why wasn't it already good enough to be free of significant regressions prior to release? Do they need HW3 to get those bounding boxes to quit jumping around?

The bounding boxes are jumping around on the IC, because they are poorly rendering them, not because they are jumping around on the ape. You can see clearly on @verygreen dev ape videos that the bounding boxes don't jump around compared to the crap they show us on the IC. This is just Teslas poor rendering of the data provided.

 

[AWDtsla]

The bounding boxes are jumping around on the IC, because they are poorly rendering them, not because they are jumping around on the ape. You can see clearly on @verygreen dev ape videos that the bounding boxes don't jump around compared to the crap they show us on the IC. This is just Teslas poor rendering of the data provided.

They're jumping around quite a bit there, but the view angle is entirely different. Are you sure what you're seeing is not just a result of the 3d transform the data is going through? In effect an axis with less accuracy is projected more prominently.