Navigate on Autopilot is Useless (2018.42.3)

[iTech]

@verygreen has the hardcoded disabling of AEB been removed for bridges/overpasses?

 

[boonedocks]

Big annoying bug still: if the map is wrong so it thinks there is a lane change needed, and you cancel, it just keeps insisting more and more that you need the lane change. This really should result in some feedback. 'Hmm, someone canceled six lane changes in a row to follow the route - maybe my maps are wrong'

I have sent multiple videos attached to emails showing this exact behavior. The responses act as if it is a NOA infancy issue and will be resolved with updates. (which BTW started when NOA was released and still does it as of 2019.8.5) It IS definitely a map issue and not a NOA deciding to change lanes. It is persistent wanting to exit the HOV lane to only insist on getting right back in a 1/2 mile later. Then in 3 more miles it insists on changing lanes to the left in the HOV lane which is a concrete barrier separating the expressway. \i/

I wouldn't mind this behavior until it is fixed IF it would accept the "cancel" when tapped. That lasts for about as long as it takes to move your hand back to the steering wheel. BTW what does cancel mean to 99% of the world.....I would think it would mean stop this particular lane change prompt

 

[Pale_Rider]

Finally got 2019.8.5 on my S last night. I tried it on my commute this morning. My experience was about the same as previous versions of NOA going back to the intial version. It completely failed to merge onto the highway (after engaging on the on-ramp). It dove into an emergency pull-over thinking it was my exit -- the kind that's surrounded by concrete barriers and is only about 5 car lengths long. I quickly took over and re-engaged to see if it could make the actual exit and it couldn't do it. It sort of veered back and forth in a very confused way making me appear drunk, at which point I took over.

So, still useless for me.

Sounds like a map issue. There a ton a construction around Houston on the interstates and NoA is fairly useless in those areas because exits have moved or the interstate is now where a frontage road used to be so it thinks the speed limit is 45 instead of 65. I supposed all of that will eventually be corrected when they update maps but who knows how long it will take all of that to get properly documented. That said, driving from Houston to central Florida worked great as there hasn’t been much change to what the maps expected. They really need to figure out a way to not be so map dependent for speed limits and exits.

 

[eli_]

The map data in general seems pretty outdated, there are ramp revisions made in 2015 here that are still not reflected in the turn-by-turn directions, so NoA always takes the wrong exit.

 

[S4WRXTTCS]

Big annoying bug still: if the map is wrong so it thinks there is a lane change needed, and you cancel, it just keeps insisting more and more that you need the lane change. This really should result in some feedback. 'Hmm, someone canceled six lane changes in a row to follow the route - maybe my maps are wrong'

This is the single biggest issue I have with NoA.

Even in cases where there isn't anything wrong with the navigation maps.

For example on 2019.8.3 it kept wanting me to get over to the right most lane on 405N traveling through Bellevue on my way to Bothell. There is nothing wrong with simply staying in one of the center lanes for most of the journey. The navigation map is completely fine with that, but NoA took issues with that. It didn't matter how many times I canceled the lane change.

I haven't tested the same section with 2019.8.5, but I intend to.

The only road I've tested 2019.8.5 on so far is I5 Northbound past Everett, and it's fine there.

There are still some improvements they could make to situational awareness, and deciding when is a good time to change. But, overall I was happy with how it performed from Everett to Mount Vernon.

So I've finally found one area in WA state where NoA works.

 

[S4WRXTTCS]

The map data in general seems pretty outdated, there are ramp revisions made in 2015 here that are still not reflected in the turn-by-turn directions, so NoA always takes the wrong exit.

What I've noticed is that sometimes the turn-by-turn directions are just plain wrong so there is no way for NoA to work well.

Then there are other times where the turn-by-turn seems fine, but the cars behavior isn't what I'd expect.

This is especially true for closely spaced exits where my exit is one exit past an exit in which is a right lane exit only. The car will want to get in that exit or at least indicate that it does. It's as if it's only knows the general location of an exit, but doesn't know exactly which lane to be in.

Even on 2019.8.5 I've seen it get confused by this. Where it wanted to get in that lane that goes off with that exit, but then a few seconds lane it decided "nah, I guess I won't".

The kind of behavior I expect from a human new to the area, but a maps based driver aid.

 

[DarkMatter]

So I've finally found one area in WA state where NoA works.

It's great once you get South of the construction in Tacoma all the way to Portland. Of course there are many feetf Teslas following that route than there are on 405.

I stand by what I said: they should be gathering cancel data and using it to update the map. Take advantage of the human perception.

 

[S4WRXTTCS]

It's great once you get South of the construction in Tacoma all the way to Portland. Of course there are many feetf Teslas following that route than there are on 405.

I stand by what I said: they should be gathering cancel data and using it to update the map. Take advantage of the human perception.

These days there are Tesla's everywhere in WA state so we don't need to worry about coverage.

Your suggestion if implemented would go a long ways in fixing the problem. It's pretty easy for the system to determine what is canceled navigation lane change. and a canceled speed based change.

 

[timewasted]

I don't have 2019.8.5 so I can't try out the new amazing "@wk057 is a threat and must be dealt with" neural net, but 2019.8.4 is still as worthless as the last time I posted here about NoA. If I'm sitting in the middle lane on the highway with my cruise set to 75mph with the person in front of me going 65mph and the lane to my left is completely empty, my car will happily continue following this person regardless of what my aggressiveness setting is at. It constantly suggests I merge to the right to prepare for my exit despite said exit still being 4+ miles away. Merging on and off freeways is still scary as hell because the car doesn't seem to care much about cars beside it as the lanes start to condense.

I'm glad other people are finding it useful, but it's just simply been a liability for me. I'll give it a try again when I finally get 2019.8.5, but probably only when I'm driving alone in the car.

Going from 2019.8.4 to 2019.8.5 resulted in exactly zero improvement in navigate on autopilot. My car shows no real desire to make an intelligent decision, and it’s certainly in no rush to make it to its destination. I turned off the need for confirmations, and each time I’ve left work the need for confirmations was turned back on.

The stop and go traffic handling seems to be smoother in this version (did we finally get our “silky smooth” update?!), but the marquee feature is still proving useless to me.

 

[S4WRXTTCS]

I turned off the need for confirmations, and each time I’ve left work the need for confirmations was turned back on.

That kinda defeats the entire purpose of the update.

Have you opened a support ticket to figure it out as it's really odd behavior?

Personally I'd reboot, and try again.

I've also noticed the stop and go handling is smoother, but I think it's because 2019.8.3 (and maybe 2019.8.4) regressed. I almost drove without even TACC on a freeway because of how bad it was in 2019.8.3.

 

[scottf200]

Just ran across this well done video on NoA improvements in v10 over v9. First 2.4 minutes and starting at 8:05 for the exit show a few examples of how it has improved. After watching remembered a couple conversations about it and different threads.

 

[eli_]

The improvements are great but the weakest link is still bad/inaccurate/outdated map data IMO. NoA is pretty helpless if the map data isn't correct, and unfortunately there's a lot of local highways here that fall into that category. Some junctions it will absolutely flip out, like I-405 N -> WA-520 East, it will choose the wrong lane, then try to change lanes at the last second, then it will try and fail to change lanes while on a single-lane ramp! In some cases ramp revisions that happened 5 years ago are still not present and result in NoA taking the wrong exit.

 

[Pale_Rider]

The improvements are great but the weakest link is still bad/inaccurate/outdated map data IMO. NoA is pretty helpless if the map data isn't correct, and unfortunately there's a lot of local highways here that fall into that category. Some junctions it will absolutely flip out, like I-405 N -> WA-520 East, it will choose the wrong lane, then try to change lanes at the last second, then it will try and fail to change lanes while on a single-lane ramp! In some cases ramp revisions that happened 5 years ago are still not present and result in NoA taking the wrong exit.

Agreed, there are large swaths of major roads in Houston, that are significant interstates, that while work has been done in the past few years, is out of date and I just have to turn it off when I drive across town. It works great when the map data is correct but when it isn’t it is pointless.=\

 

[wk057]

And this right here is a huge part of why I just don't bother to use NavOnAP in most cases (time around 2:30, since the forum removes this jump from the embedded version below):

The car in front is going slower, so a lane change should be made. There are no cars to the right, yet the car slows down by six MPH before it starts to change and accelerate again. This is both stupid and inefficient. What's worse, is this is the behavior in the vast majority of cases. I rarely see NavOnAP make a sensible lane change to overtake, despite it having more than sufficient information to do this without an inefficient deceleration and re-acceleration.

It knows the speeds of the vehicles ahead in all lanes, and knows the lane to the left is clear... why can't it just smoothly change over to the left without slowing down? If I had to venture a guess, I'd say some of the lane change lag is due to the car verifying torque ("hands") on the wheel for multiple seconds (even while hands are constantly on the wheel applying torque) before it will actually make a lane change instead of, you know, doing what it's supposed to do and not drive like a jackass and just making the lane change. In the situation in this video, too bad you can't see traffic to the rear because many times when NavOnAP does this to me in traffic the car behind me will start to move to overtake me as my car is stupidly slowing down, and my car ends up moving to cut them off before accelerating again... making me look like a complete a**hole.

I've tried v10 NavOnAP in 3 different cars now, to see if it's improved. While some situations are less dumb than they were previously, it still has pretty much all of the same issues as before, including tons of phantom braking, dumb lane change decisions (in some cases even worse than before), etc.

What's sad is that my original assessment from nearly a year ago still pretty much stands. NavOnAP is pretty useless.

 

[Electroman]

Agree with that. In many situations humans tend to accelerate and change lanes. You tend to increase speed in your current lane and then do a smooth lane change all the while increasing the speed during the change.

NoAP almost always reduces the speed, starts to change lane and the 3/4th of the way through increases the speed. Very annoying and useless

 

[DarkMatter]

In traffic with even flows it works pretty well.

However I have found a new regression that is frustrating. It seems to completely lose track of what is going on if it changes lane in to a wide merge lane. Of course there is one of those on my route home, so I get to take over a mile before the exit every time.

And yeah, it's totally flustered by map errors.

 

[Jeff Hudson]

In traffic with even flows it works pretty well.

This sentiment pretty much sums up my thinking about NoA even though I seldom use it because my local driving circumstances do not provide the opportunity. Approaching the NoA shortfalls topic from the direction of suggested usage advice concerning Autopilot shortfalls goes a long way to preventing the dissatisfaction with NoA most commenters are experiencing because the best NoA advice I would give is to not use NoA when road and/or traffic conditions call for a human directing the vehicle. NoA works pretty well everywhere the highway is properly maintained (very important) excepting highway interchanges. As far as I can determine NoA through highway interchanges will be one of the very last things Tesla achieves before applying to the authorities for Level 4 Autonomous driving designation. I applaud all of you contributing commenters for being unpaid Tesla testers vigilantly exercising the NoA feature set and I would be testing right there alongside you if I had the proper roadways near my location. In parting, be safe out there and keep your hands on the steering wheel.

 

[mhan00]

I’ve used NoAP for about 2300 miles of road trip over the past month, and loved it. In El Paso it failed a couple of times since they’re doing a bunch of construction there (also encountered an instance where the max speed changed to 35 from 75 while passing under a bridge there which I’ve heard about happening but had never experienced myself - I can see how that would be really dangerous if you weren’t paying attention at all or were insistent on letting NoAP handle things on its own for whatever reason, but for me as soon as I felt it slowing for no reason I could see I just hit the accelerator and never dipped below 70), but everywhere else it was pretty damn good, especially in California. Definitely nowhere close to being able to sleep while driving on a freeway, but as a supplemental driver aid it is remarkable.

 

[AwlBidnz]

Approval of lane changes and, old fashioned road racing needs to be a human process... FSD is not as close as some are wishing.. watch advanced summons for clues on when it might be getting closer.

 

[scottf200]

So at about 8:16 when NoA wants to take the exit and the Tesla is slowly overtaking the pickup truck in the adjacent lane. When it is *right* next to the Tesla, the car appears to 'do the math' and realize it is not going to overtake the pickup truck without going over the cruise control set speed (70 mph) ... it then slows down and makes the lane change. Now, I've had the Tesla/NoA do with while was just at the rear bumper of the adjacent car but this action in the video seems better on v10 than my past experiences of letting NoA do it's own thing. Aside: yes, I have pressed the accelerator pedal and gone 75ish in this equivalent example and the Tesla then made the automatic lane change. Humans do that. I don't think Tesla/NoA goes past the cruise control speed 'like a human' would.