Navigate on Autopilot is Useless (2018.42.3)

[clydeiii]

NoA combines all of those things together, plus allows me to focus less on navigation and more on being aware of the cars around me.

 

[rnortman]

NoA combines all of those things together, plus allows me to focus less on navigation and more on being aware of the cars around me.

OK, you can combine all of those things together without NoA, as I do. And in my experience, which matches the experience of quite a few other people clearly, NoA does not allow me to focus less on navigation because on the roads I drive on at least, it is wrong more often than it is right about what lane to be in when to make your exit, and where the exit actually is. On my regular commute is a spot where it will dive into an emergency pull-over lane (which only exists for about 100ft) 100% of the time, forcing me to take over. This does not make me safer.

 

[eli_]

Highway NoA (internally called "Drive on Maps") seems like throw-away code to me, I assume that it will be replaced with a vision system that reads all the lane markings and signage to figure out what lane to be in, like a human would. So the planner will rely on that instead of map data. Either they need that or a massive, expensive effort to accurately tag lane data in their mapping source.

 

[Mardak]

I assume that it will be replaced with a vision system that reads all the lane markings and signage to figure out what lane to be in, like a human would

Yeah, most likely Tesla is using this lane merge map data to automatically label training data on when to yield for merges similar to using traffic light map data to automatically label red and green lights. So the neutral network could additionally output whether the car should yield based on the visual appearance of the lane and other vehicles. Tesla already has something like this for cut-in detection, so this might be well under development already.

Trickier is relying on street signs as so far it seems like Autopilot reacts to what was immediately seen and what is currently visible and not so much on what it has seen a few seconds ago.

 

[rnortman]

Highway NoA (internally called "Drive on Maps") seems like throw-away code to me, I assume that it will be replaced with a vision system that reads all the lane markings and signage to figure out what lane to be in, like a human would. So the planner will rely on that instead of map data. Either they need that or a massive, expensive effort to accurately tag lane data in their mapping source.

HW2's inference hardware is already maxed out, it seems, with the current workload on the neural net. I don't think they have the ability to add this to HW2/2.5. And yet, they promised "On Ramp to Off Ramp" as part of EAP. So either NoA is it, or they somehow manage to cram more onto HW2, or they upgrade EAP purchasers to HW3.

Spoiler: NoA is it.

 

[scottf200]

Highway NoA (internally called "Drive on Maps") seems like throw-away code to me, I assume that it will be replaced with a vision system that reads all the lane markings and signage to figure out what lane to be in, like a human would. So the planner will rely on that instead of map data. Either they need that or a massive, expensive effort to accurately tag lane data in their mapping source.

Have no idea where you got this "Drive on Maps" reference. This video shows the cameras (vision system) picking up many items. Thanks to %green% and company for providing this.

 

[Mardak]

HW2's inference hardware is already maxed out, it seems, with the current workload on the neural net.

Assuming HW2/.5 is indeed close to maxed out, this might be something for the DeepScale acquisition to help optimize the neutral network to slightly increase the capability of cut-in/merge detection but not probably not possible to optimize FSD into the slower hardware.

 

[thewishmaster]

Highway NoA (internally called "Drive on Maps") seems like throw-away code to me, I assume that it will be replaced with a vision system that reads all the lane markings and signage to figure out what lane to be in, like a human would. So the planner will rely on that instead of map data. Either they need that or a massive, expensive effort to accurately tag lane data in their mapping source.

I think for longer-distance planning it still needs to be aware of lanes and other road information so it doesn't make mistakes such as getting stuck in turn-only lanes when it could've made an extra lane change earlier on. This is similar to how we memorize the ideal routes and lane changes after driving them a few times (and can rely on relatively aggressive maneuvers when making mistakes the first time we drive somewhere). I doubt this knowledge is something they'd want to rebuild from scratch, but even if they did, they would arrive at pretty much the same solution as they have now.

 

[eli_]

Have no idea where you got this "Drive on Maps" reference. This video shows the cameras (vision system) picking up many items. Thanks to %green% and company for providing this.

"Drive on Maps" is what it's called in the internal debug tools, also from green: https://i.imgur.com/tZiil6L.jpg

 

[eli_]

I think for longer-distance planning it still needs to be aware of lanes and other road information so it doesn't make mistakes such as getting stuck in turn-only lanes when it could've made an extra lane change earlier on. This is similar to how we memorize the ideal routes and lane changes after driving them a few times (and can rely on relatively aggressive maneuvers when making mistakes the first time we drive somewhere). I doubt this knowledge is something they'd want to rebuild from scratch, but even if they did, they would arrive at pretty much the same solution as they have now.

Sure, but it needs to be vision first and maps second. Vision is your primary source of truth, maps are for "tips and tricks" as Elon put it at the Autonomy Day presentation. Like if there is a sign or pavement mark that says "Right turn only" or "Exit Only", or an overhead sign labeled "I-405 North", those are your sources of truth as to what lane to be in. Right now NoA is the other way around, maps are the only source as to what lane is for what. The vision system doesn't read signs or HOV markings, or turn arrows on pavement, etc, it's all in the map tags, which are frequently wrong.

 

[clydeiii]

OK, you can combine all of those things together without NoA, as I do. And in my experience, which matches the experience of quite a few other people clearly, NoA does not allow me to focus less on navigation because on the roads I drive on at least, it is wrong more often than it is right about what lane to be in when to make your exit, and where the exit actually is. On my regular commute is a spot where it will dive into an emergency pull-over lane (which only exists for about 100ft) 100% of the time, forcing me to take over. This does not make me safer.

Well, yeah, that sucks. Keep reporting that bug via the car. In my 25-mile commute, it is nearly flawless, and of course in the places it isn't, I know exactly where those are and what to do if the car doesn't perform as expected.

 

[wk057]

I honestly don't see how this could be described as "nearly flawless" by anyone anywhere.

I can't even get it to make sane decisions on any local highways in any of 3 different vehicles. This behavior was mirrored when trying on California highways as well. Seriously, it generally can't perform well enough for more than a couple of miles in most cases without intervention (not even counting the useless need for torque on the wheel before it will do anything at all). And I'm not just talking about intervention because it's doing something kind of silly, pointless, or just dumb. Those are plentiful for sure. But, I also mean intervention as in, if I don't prevent what it's wanting to do there may be an actual accident involved. So far, it's not been anywhere near flawless in any situation for me or anyone I personally know.

@clydeiii Really, though, if you can do a 25 mile commute with NoA with minimal interventions, and minimal overrides (like, letting it make and proceed with the decisions it makes to change lanes, interchanges, etc), I'd honestly love to see that in action if you get time and inclination to make a video... like, where the AP/NoA status can be seen in the video the whole time. Will be disappointed if this is on an empty road with no lane changes, or at below traffic flow speed, though.

I think you may be singing some higher than deserved praises when looking at it objectively, but that's just my guess based on now thousands of miles of experience and attempted use of this feature over multiple vehicles in a wide variety of locations.

 

[rnortman]

I honestly don't see how this could be described as "nearly flawless" by anyone anywhere.

I generally agree very much with this sentiment and I'm highly skeptical of anybody who sings the praises of this feature... but playing devil's advocate, I will admit that on long stretches of simple, roomy, well-marked highways without nearby access roads or awkward, space-constrained entry/exit ramps, and generally calm/sane traffic, NoA will only occasionally screw up badly. Perhaps this is what @clydeiii 's commute is like.

To say that it only occasionally screws up badly in such an environment is faint praise of course. Particularly since it is adding no value whatsoever in such an easy, simple, low-stress environment. (Beyond what TACC and Autosteer already provide without NoA -- and TACC at least is a must-have on long stretches of bland highway. NoA not so much.)

 

[APotatoGod]

...On my regular commute is a spot where it will dive into an emergency pull-over lane (which only exists for about 100ft) 100% of the time, forcing me to take over. This does not make me safer.

If you've learned that pattern, why not just hit the turn signal to request the car turn into the lane on the opposite side shortly before that spot? Seems easy to solve, without needing to disengage autopilot? Or just hit the Navigate on autopilot button for a few seconds, let normal AP handle that stretch?

Well, yeah, that sucks. Keep reporting that bug via the car. In my 25-mile commute, it is nearly flawless, and of course in the places it isn't, I know exactly where those are and what to do if the car doesn't perform as expected.

I agree with this, it is nearly flawless for me as well, there's only one exit that I semi-regularly need to take that it misjudges, but it's easy to manually override it when needed. (And that exit is confusing for human drivers that aren't used to it too.)

I've had 60 mile drives that were nearly entirely Nav on AP, and it makes it so much easier and more relaxing to drive! I'm not sure why some people seem to be having such bad results with it? I guess the map quality must vary based on location (I'm in California, for what it's worth). I legitimately would trust Nav on AP's driving in some scenarios more than many of my friends, and in most circumstances I use it, it seems entirely capable of driving unsupervised.

Lately I often find myself engaging navigation, even when driving to very familiar locations, just to activate NoA, and sometimes even going slightly out of the way to go on a freeway route for the same reason.

 

[rnortman]

If you've learned that pattern, why not just hit the turn signal to request the car turn into the lane on the opposite side shortly before that spot? Seems easy to solve, without needing to disengage autopilot? Or just hit the Navigate on autopilot button for a few seconds, let normal AP handle that stretch?

So you seem to be making the assumption that just because I mentioned only one specific area where it tries to kill me 100% of the time, that that is the only place where it screws up, or that it is always predictable where it will screw up. It's not always predictable, and anyway NOA's greatest value would be using it on unfamiliar roads, where learning its pattern of screwing up is not really an option. On familiar roads I know what lane to be in (much better than it does) and I don't need its help.

But the question being addressed here is not, "Can you work around the bugs and flaws such that you can prevent the car from breaking the law, causing an accident, or taking the wrong path?" Clearly the answer to that question is yes -- you must remain diligently aware and take over when necessary, or else the car will probably kill you or somebody else.

But the question at hand is "Is a trip less stressful for the driver and passengers with NOA enabled or with only TACC and Autosteer?" To that I can definitely answer -- after using this feature extensively in various conditions in both my Model S and my Model 3 -- as a resounding "no". Which is why I have it disabled. I cannot imagine a situation in which NOA adds value.

Well, there is one situation -- if you want to drive traffic to your youtube channel, having NOA videos is probably helpful.

 

[wk057]

The predictability of AP2+ (and by extension, NoA) is actually why I still, 3 years after the release of AP2, greatly prefer using AP1 vehicles.

AP1 is highly predictable. Even in areas I'm not familiar with, I can generally predict what AP1's behavior is going to be like. I know when it will work well, and when it won't. Overall, I still find AP1's general highway lane keeping to be superior to AP2+, and partly for this reason. While AP1 is far from perfect, it's highly repeatable.

AP2+ is all over the place. I can never quite predict how it's going to behave, even on repeated known routes... let alone in some place unfamiliar.

As the post above notes, NavOnAP would be most useful in unfamiliar areas... but these are the same areas where you would need to be triple diligent when using the feature in the first place... which makes the driver workload significantly higher than just not using NoA at all.

 

[scottf200]

AP2+ is all over the place. I can never quite predict how it's going to behave, even on repeated known routes... let alone in some place unfamiliar.

Was AP1 more IF-THEN-ELSE vs AP2+ being more based on the ever updated NN/ML? The more data they train their NN/ML with the more it will 'learn' or have behavior changes I thought.

As the post above notes, NavOnAP would be most useful in unfamiliar areas... but these are the same areas where you would need to be triple diligent when using the feature in the first place... which makes the driver workload significantly higher than just not using NoA at all.

Sorry but that 'unfamiliar areas ... significantly higher' comment did not match my experience in the long trip I mentioned (and showed map for) in some post above (Toronto, Niagara Falls, etc).

I do appreciate your insight on many things. I'm likely more subconsciously forgiving of NoA than you are but my overall experience has been very good. I put 30K plus miles on an AP1 X before I got my AP2.0 X.

 

[emmz0r]

Was AP1 more IF-THEN-ELSE vs AP2+ being more based on the ever updated NN/ML? The more data they train their NN/ML with the more it will 'learn' or have behavior changes I thought.

NN is only used for object recognition here.

 

[DrDabbles]

I've tried NoA a bunch more since my last post, trying to see if it's improved at all. I've even given it a go over about ~500 miles of driving in California recently, since everything Tesla seems to work better in California.

I pretty much only use NoA when alone because anybody else in the car finds it really unpleasant. Now with the TACC changes on highways, I can't even use that feature because it suddenly jerks to a start and stop. While accelerating from a stop behind traffic might have been too slow to begin, at least it wasn't sudden hard acceleration and I could override it.

In summary, it's still pretty useless and just adds more headache to the trips than needed in most cases. It still will wait way too long to overtake vehicles... slowing down significantly before starting a merge. There's so much error in the decision process it's doing that I generally have to abort 2/3 suggestions it makes for lane changes for one reason or another. Things ranging from starting to move directly into other vehicles to general stupidity in lane changes like moving back behind a slower vehicle that it's clearly aware of, slowing down, then immediately wanting back in the passing lane.

As I've said before, the decision to pass traffic and the decision to exit a highway are clearly not informing each other. This is frustrating, especially if you're in an area you're not familiar with. To make things worse, if you cancel a lane change for some reason, the car will continue to hound you until you either disengage NoA or get to a place where it can't change lanes. This is annoying when HOV lane is slower than the normal highway, and NoA wants to change in and out of it. It's also annoying when the navigation system recommends a route you know will be worse and you want to pass an exit.

The phantom braking with AP2+ overall is very real still. In the nearly brand-new Model X I was driving around for my California trip recently, the phantom braking was almost a complete show stopper for AP usage in even light traffic.

This is the number one reason I don't use NoA or TACC with passengers. The phantom braking and now the incorrect map interpretation putting the speed limit 30-45 MPH slower than the highway is uncomfortable, and downright dangerous. We've seen TeslaCam videos of hard braking causing collisions behind a Tesla, and it's only a matter of time before someone is seriously hurt. Tesla's blog entry about this from I think 2016 was so spot on that it's bizarre to read now, and it shows how little things have progressed in these three years.

... which really makes me wonder where the rave reviews of this feature are really coming from.

They say love is blind, so fanaticism must be whatever is beyond blind. Some people think being honest about areas for improvement means you're "a short" or somehow "anti-tesla". If you're not fawning over Elon's every word, and ignoring any issues that Tesla has, then you're an enemy. Obviously.

The bottom line for me right now is that even if it did work 100% as it should, it's still more trouble than just using AP alone or not using AP at all.

But if it worked 100% as it should, it wouldn't make useless lane changes, slam on the brakes for people merging onto the highway, and phantom brake. That alone would make it so much more useful for the common case.

If I have to initiate the lane change myself anyway, what the heck is the point.

I presume that a suggested lane change could keep the vehicle's speed more constant and changes timed more smoothly. But this presumes quite a lot about how timing works and how well the vehicle could time things in the future.

Then, with lane change confirmation disabled, the car won't do the lane change unless it detects torque on the wheel (supposedly hands)... which means I could have just steered to make the lane change myself anyway.

This is the difference between autonomy and driver assistance, though. It's not self driving, and it shouldn't behave like self driving because owners are clearly already abusing and misunderstanding the purpose.

Unless you're driving on a road with zero traffic, it's more trouble to babysit NoA than it is to just drive manually or drive with just AP and manually initiated lane changes.

I think your next sentence is the whole point here. They need to improve the system's decision making significantly. And then the babysitting is reduced as reliability is increased. But Tesla's M.O. seems to be to give up user experience now, and only provide a good experience once the features works perfectly. See: rain sensing with a neural network rather than a damn rain sensor.

When I started this thread almost a year ago, I really figured Tesla would have made significant improvements to NoA by now. Instead, it's pretty much just as bad as it was a year ago. That's kind of sad, really.

Hah. Every time I made a comment about how hard the path Tesla has chosen actually is, and how unlikely it would be for them to miraculously have a breakthrough for the entire field of research let alone for themselves, I was told that I was anti-Tesla and had no idea what I was talking about.

Now just imagine, smart summon is based on this same system. And people are using it on purpose.

I super disagree with your post wk057. In Maryland, at least, NoA has become much more aggressive when changing lanes. I drive about 50 miles every day on NoA and the improvements in the feature since its debut have been remarkable. It makes my daily commute so much better. I'll never go back. It's far more than a gimmick, it's literally making my commute safer and less stressful. We all need NoA.

I've driven NoAP through Maryland, and I can say that it only works in extremely well traveled areas, I presume because they're the best mapped. But I've found plenty places driving through MD on my way south where it fails miserably and somewhat dangerously.

This is really the problem with the way Tesla is trying to solve the problem. It fails in subtle ways even in places where it worked previously because conditions changed slightly. This is the weakness of neural nets. And it remains poor until suddenly it works well enough and you stop messing with it.

 

[arghx7]

We're hearing positive reports that 2019.36 has improved lane changes and passing behavior:

Lane changing is much smoother and more assertive. I tried it with Standard and Mad Max. I prefer mad max because it feels more aggressive and less hesitant during lane changes. Moving in and out of traffic was better than before as well. It knows to take the passing lane to pass a car ahead in’s of moving in and out of traffic for no reason like previous updates. I had to take over curved exits since the speed is reduced too much, maybe it’s being overly safe but cars in the back was tailgating me a bit. Lol.