Navigate on Autopilot is Useless (2018.42.3)

[Battpower]

What decisions does it make that you disagree with?

I have a couple of examples in addition to the recent post about poor merging into traffic at on ramp.

These relate to UK roads but after 15 years driving in the US I can't see why they could not apply there in some circumstances.

A few weeks back I was on a main dual carriageway (freeway with two lanes in each direction and central divide) with very light traffic and clear conditions. A delivery truck was unloading goods - parked facing oncoming traffic in my lane. As there was so little traffic, I did not intervene until the last moment, but the car made no attempt whatsoever to slow or avoid the truck.

Another time I was on a similar road that I drive often, clear nighttime on a newly refurbished stretch with very clear markings. My off-ramp was about 1/4 mile ahead as I was passing a clearly marked pull-off / lay-by. Suddenly the car applied the turn signal (a little early for the exit I thought) and promptly tried to merge onto the lay-by (which luckily was unoccupied!)

Of course there's the usual range of sudden and unexpected speed changes, most down to poor map data or road layout, but 2 or 3 quite dramatic hard brake applications passing under large signs and exiting tunnels.

Lowest on my list because they are not really a problem as long as you know where you are going are the merge lane instructions. They often pop up at completely irrelevant times and give incorrect guidance. Least significant are the 'move into faster lane' messages when any sensible drive could see that it wouldn't make sense and likely confusing / dangerous for other cars. The 'merge to continue on your route' is more of a problem if you are on unfamiliar routes and several times has had me specifically move out of the lane I needed to be in.

Finally, 'move out of passing lane' has on several occasions had me move into the wrong lane for my route and have to cut back into heavy traffic.

None of that feels like a fantastic driver aid, but it certainly keeps you attentive!

 

[clydeiii]

I have a couple of examples in addition to the recent post about poor merging into traffic at on ramp.

These relate to UK roads but after 15 years driving in the US I can't see why they could not apply there in some circumstances.

A few weeks back I was on a main dual carriageway (freeway with two lanes in each direction and central divide) with very light traffic and clear conditions. A delivery truck was unloading goods - parked facing oncoming traffic in my lane. As there was so little traffic, I did not intervene until the last moment, but the car made no attempt whatsoever to slow or avoid the truck.

Hard to understand this without viuals, but you're saying a truck was unloading goods on the highway and the car would have run straight into it had you not intervened? That does seem shockingly bad, and I wonder why it didn't treat it as just a highway back up and slow down like AP does with near perfect reliablility every day. But I might be misunderstanding what you're saying. Trucks tend to not unload their cargo on US highways (the only place NoA works here).

Another time I was on a similar road that I drive often, clear nighttime on a newly refurbished stretch with very clear markings. My off-ramp was about 1/4 mile ahead as I was passing a clearly marked pull-off / lay-by. Suddenly the car applied the turn signal (a little early for the exit I thought) and promptly tried to merge onto the lay-by (which luckily was unoccupied!)

I can understand that if it does this every single time during every single commute, it makes NoA less useful, but even if that were the case for me, I would simply take over control for this small portion of the road and then use NoA for the rest of the commute. Does that mean it's useless, in your opinion?

Of course there's the usual range of sudden and unexpected speed changes, most down to poor map data or road layout, but 2 or 3 quite dramatic hard brake applications passing under large signs and exiting tunnels.

Yes, agree, those are annoying. Luckily for me, the only time this happened on my commute was a move from 65 to 55, due to bad map data, and eventually that map data was fixed and the commute now features none of this for me. When it was happening, I found that if I modified the speed of the car from 65 to 66 or 64 right before the abrupt change, it would cause the vehicle to not do the break. Does knowing when you must slightly adjust speed of the car make NoA useless, in your opinion?

 

[Battpower]

Hard to understand this without viuals, but you're saying a truck was unloading goods on the highway

Picture a main highway between two cities, passing through a rural area with a large property 50yards off the carriageway down a driveway. For whatever reason, the truck was parked and unloading to the property while effectively parked on one lane of the main carriageway facing oncoming traffic. This would hardly be a regular occurrence, but could equally well have been a broken down car or other obstruction.

The other issues relating to NoA do seriously detract from its utility. Do they imo make it totally useless? For its intended purpose, probably yes. Does it serve a useful purpose? If Tesla were learning from my cars errors, then it would, but I don't believe they do. Yes, after each drive with a dozen disengagements my car does start uploading a lot of data, but from what have read on TMC it is at best unlikely it gets used.

I actually find it keeps me more alert when I drive using NoA because having been caught out several times I have to concentrate on looking out for things that could glitch ap / NoA. There is also the anticipation and hope that it will get better, though that one wears thin pretty quickly.

I feel that NoA as part of FSD at a cost of several thousand is really just an interest free investment in the hope it might pay up one day. For those who financed their cars or sold them without ever seeing anything useful in return, I feel they got fleeced. I hope that I see something better than what I have at present before I change cars.

At the very least you should be able to transfer FSD to a subsequent vehicle while the system is so much still in development.

 

[clydeiii]

Picture a main highway between two cities, passing through a rural area with a large property 50yards off the carriageway down a driveway. For whatever reason, the truck was parked and unloading to the property while effectively parked on one lane of the main carriageway facing oncoming traffic. This would hardly be a regular occurrence, but could equally well have been a broken down car or other obstruction.

Ok, so from what I can tell, there was a truck on the highway, completely stopped, and your car was just going to barrel into it? Yeah, obviously a huge problem if so. I can't help but wonder what made that scenario different from normal traffic that was just at a standstill though, from NoA's perspective. And hoping that in the near future, the rewrite of the car's eyesight from 2.5D to 4D will fix this sort of issue, but that remains to be seen. Even so, this scenario is a relatively rare occurrence, so the question becomes, given how rare it is, does that render the entire NoA system useless? For me, no.

For its intended purpose, probably yes.

I do wonder what you think its intended purpose is. I have never assumed NoA would be 100% safe or make 100% correct decisions, because I don't think anyone or any computer will ever be 100% safe or ever be 100% correct. It's just impossible. It all comes back to our own personal thresholds.

 

[Daniel in SD]

Like for me the primary cause of phantom braking seems to be semi truck trailers. This seems to happen just as easily in TACC as in AP. I don't see the need for it under TACC. Adaptive cruise control these days is pretty widely available on new cars so there is no reason for Tesla to make it more complicated than it needs to be.

I think TACC is exactly the same whether or not Autosteer is enabled or not. Tuning TACC to be less conservative when used without Autosteer would be an admission that Autosteer makes people less vigilant. If someone crashed into the back of a stopped truck while using the TACC only tuning they would complain that it wouldn't have happened had they been using the "safer" TACC+Autosteer tuning.
The speed limit issue is annoying but I've never had it slam on the brakes for that. I've actually only had it slam on the brakes (50% braking?) once.

 

[Battpower]

Yeah, obviously a huge problem if so. I can't help but wonder what made that scenario different from normal traffic that was just at a standstill though,

I believe this is a fairly well recognized shortcoming that stationary objects are difficult to deal with in certain scenarios.

I do wonder what you think its intended purpose is.

I don't think it is expecting too much for any system to be consistent and provide the functions it sets out to.

I feel that in my experience so far it falls short on both counts. Merging from on-ramp is its first move, so to fail at that is a bad start. When the system consistently issues incorrect or misleading directions that is another fail. Even so, personally I'm happy to use it as a toy, gimmick and alertness aid until such time as it matures into something nearer a finished (non beta) product.

 

[DrDabbles]

The definition of useless here is whether NoAP does anything that AP doesn't already do. And the answer is effectively no. If you had EAP in 2017/2018, then you already had a car that would change lanes using the blinker. Since NoAP uses map data and not the vision system, you're 100% relying on that data being accurate and there are massive swaths of the planet where it is not.

The fact that experiences vary so dramatically is exactly the problem. You can leave an area where map data is accurate, end up where it isn't, and suddenly NoAP is making incorrect lane changes, not properly handling interchanges, missing exits, trying to exit when it shouldn't, etc. And that's on top of all of the AP failures like phantom braking, improper merges, improper speed changes, etc.

 

[Battpower]

If you had EAP in 2017/2018, then you already had a car that would change lanes using the blinker.

I'm really not clear what (if any) significant differences there are even now between EAP and FSD. It appeared to me that not so long ago you could have paid several thousand to upgrade EAP to latest FSD (with possibility of HW3 when available) and be practically no better off as far as actual use experience.

Obviously with Stop and traffic light and some of the latest AP features, the overall package can now do more, but they are not specifically NoA features. It becomes rather difficult to separate AP, NoA and FSD features as there are different overlaps and interdependencies involved, and it's even harder if EAP is considered.

 

[emmz0r]

What is "Automatic driving on city streets" ? I suspect it's just automatically stopping and starting at stop lights - presto FSD delivered.

 

[diplomat33]

What is "Automatic driving on city streets" ? I suspect it's just automatically stopping and starting at stop lights - presto FSD delivered.

No. "Automatic driving on city streets" will be more than that. We know "turning at intersections" and "roundabouts" are in development since Elon has said that he is testing them on his car and that the features should be delivered to the fleet later. So at a minimum, we can expect turning at intersections and roundabouts to also be part of "automatic driving on city streets".

 

[emmz0r]

No. "Automatic driving on city streets" will be more than that. We know "turning at intersections" and "roundabouts" are in development since Elon has said that he is testing them on his car and that the features should be delivered to the fleet later. So at a minimum, we can expect turning at intersections and roundabouts to also be part of "automatic driving on city streets".

Right. He said roundabouts is a year out, but robotaxi is end of this year

 

[diplomat33]

Right. He said roundabouts is a year out, but robotaxi is end of this year

By the way, I don't expect "automatic city driving on city streets" to enable robotaxis. It won't be FSD since it won't be able to react to every situation. I expect it will just be L2 driver assist because I expect it will just be steering and braking to navigate on city streets. Specifically, I expect it do lane keeping, lane changing, making turns, handling a roundabout, and traffic lights and stop signs on city streets but that's pretty much it. The driver will still need to supervise and take over for cases it can't handle.

 

[Daniel in SD]

Right. He said roundabouts is a year out, but robotaxi is end of this year

Sounds like the robotaxis will be geofenced to streets without roundabouts. This should work in most American cities. I'm thinking that regulatory issues will delay release in Europe anyway so it makes sense not to prioritize roundabouts until next year.

 

[emmz0r]

By the way, I don't expect "automatic city driving on city streets" to enable robotaxis. It won't be FSD since it won't be able to react to every situation. I expect it will just be L2 driver assist because I expect it will just be steering and braking to navigate on city streets. Specifically, I expect it do lane keeping, lane changing, making turns, handling a roundabout, and traffic lights and stop signs on city streets but that's pretty much it. The driver will still need to supervise and take over for cases it can't handle.

Sounds more stressful than it's worth.

 

[totteraptor]

Tesla could easily get rid of phantom braking but then the car would run into more stuff. I bet your Mercedes would run right in to stopped cars way more often if left unattended.
Autopilot needs to be safer in real world usage so Tesla has to compensate for all the people who abuse it by making it err on the side of caution. Simple adaptive cruise control doesn't have that issue so it doesn't have to try to "see" stationary objects.

Not really buying that being true but even if it was then it is a really sad trade off tesla made. Taking basic stuff below standard to make AP work - not ok.

 

[totteraptor]

I guess if we're bumping this thread, it's still useless for me. I got my HW3 upgrade finally, and as expected AP still has all of the same bad behaviors. NoAP still makes terrible lane choices, and TACC is still not smooth enough.

The car itself is great though.

Precisely how I feel!

 

[Daniel in SD]

Not really buying that being true but even if it was then it is a really sad trade off tesla made. Taking basic stuff below standard to make AP work - not ok.

There are very few roundabouts in most of the United States.

I was joking. American drivers may find roundabouts difficult but I don't think they are one of the difficult self driving car problems. I'm confident that any robotaxi will be able to handle roundabouts. I'm also very skeptical that Tesla robotaxis will be operational any time soon.

 

[DrDabbles]

I'm really not clear what (if any) significant differences there are even now between EAP and FSD. It appeared to me that not so long ago you could have paid several thousand to upgrade EAP to latest FSD (with possibility of HW3 when available) and be practically no better off as far as actual use experience.

Obviously with Stop and traffic light and some of the latest AP features, the overall package can now do more, but they are not specifically NoA features. It becomes rather difficult to separate AP, NoA and FSD features as there are different overlaps and interdependencies involved, and it's even harder if EAP is considered.

Basically if you had EAP and you bought FSD, then you get HW3 upgrade. I don't know if EAP is doing stop and go at traffic control devices on streets, but that's pretty much the only difference.

No. "Automatic driving on city streets" will be more than that. We know "turning at intersections" and "roundabouts" are in development since Elon has said that he is testing them on his car and that the features should be delivered to the fleet later. So at a minimum, we can expect turning at intersections and roundabouts to also be part of "automatic driving on city streets".

Yeah, some day in the future those are the plans. It's exceptionally unlikely that these get delivered in a meaningful way within a year's time. Unprotected left turns is a nightmare scenario, and I flat out don't see Tesla solving that any time soon. Roundabouts are a little simpler as long as it's just one lane. Two lane roundabouts, I have zero faith that they'll get right within even a couple years because their reliance on map data is just too high for those scenarios.

Also keep in mind that Elon has said for several years now that he's been testing self driving on his car and that he's been saying it's working perfectly or almost perfectly since the very beginning. The CEO of a company isn't going to say "Yeah, this thing we've sunk billions into sucks".

Sounds like the robotaxis will be geofenced to streets without roundabouts. This should work in most American cities. I'm thinking that regulatory issues will delay release in Europe anyway so it makes sense not to prioritize roundabouts until next year.

Roundabouts are prevalent in some areas of the US, and they're growing in prevalence since they're less expensive to build and increase traffic flow while reducing collisions. Tesla would also need a much better routing system if they're going to be routing around things like road fixtures.

But hey, who knows. Maybe they'll solve their left turn problem by adopting UPS's right turns only policy.

 

[emmz0r]

Sounds like the robotaxis will be geofenced to streets without roundabouts. This should work in most American cities. I'm thinking that regulatory issues will delay release in Europe anyway so it makes sense not to prioritize roundabouts until next year.

I think you are right! I see no reason to buy FSD in Europe. Also - they raised the price $1000 .

 

[Gesteur]

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?

I usually give it a try after I get an update, how else would we know if it has been fixed?
Not sure Tesla acknowledges it as an issue, so not expecting a release notes saying "hey driver, the dangerous phantom braking has been fixed"