What are the guidelines for using Autopilot on roads with cross traffic?

[diezel_dave]

I think this thread can be put to bed. Its a clear answer at this point. The guideline is just as the user manual states. TACC is not to be used on roads where cross traffic can occur and if it is used, the driver must be prepared for unexpected behavior. Additionally, at least a few of us hope that Tesla improves the system's capability so that it can be used around town without unexpected behavior because it is kind of silly that cruise control can only be used on the freeway.

 

[HighZ]

having deliberately used TACC/AP in scenarios that push the envelope, you need to be careful in how you report its behavior

Avoid saying things like "should have done X"

100% agree - OK to say, I would like it to have done or I hope they will improve it to do. Even when it does something correctly could be intermittent based on differing conditions (traffic, sun position, line visibility)

People dont read manuals.

While I know you are correct, this is an area that scares me. I read the manual multiple times before my car was delivered like a kid waiting for Christmas. You and many others on this forum probably did that as well. But for those that get only a 30 min overview at pickup on the operation of the systems AND don't read the manual, that will be scary. It is good that the manual is available in the car, but as this gets out more and more to the masses, do they even know how to find it? Also, some people learn better by hearing and seeing than reading.

I read about the 12/31 Fremont delivery and that a local Tesla owners club helped. Are those everywhere? It would be good if local clubs set up free training sessions for people to attend. And its always fun to meet other drivers. YouTube is a big help, but that can be overwhelming with the millions of videos on Teslas. Maybe there are already good New Owner Overview videos.

 

[Knightshade]

I don't think anyone is saying that the currently "should" work correctly

Dave has said exactly that a bunch of times...some example quotes from him earlier in this thread-


"40.50.7 still unnecessarily brakes for cross traffic"

" just pointing to the user manual that says "oh, that might happen" is not a good excuse for the system's behavior."

"I have to wonder when these guys will give up defending bad software?"

" Just because it operates a particular way now, doesn't mean that I accept it "



He clearly keeps thinking the system is doing something wrong, when it's obvious to everyone else it's the driver doing something not only wrong, but explicitly contradicted by the owners manual.

He even goes so far as to insist just because the owners manual told him it's not meant to work that way is "no excuse" for it... not working that way...just like the manual told him....

We get that Tesla says that the systems as designed have not been programmed to handle most of the situations involved on city streets.

Well, everyone but Dave anyway...

If you say that it is just a coincidence, then I will maintain that it is not. And again, unless you are a Tesla programmer, you don't know what they have baked in for reactions, despite what a hacker might think is happening.

I'm not talking about what a hacker "thinks" is happening.

I've talked about what one has shown us is happening.

Including raw feeds of what the AP computer is seeing and doing.

Hence his explicit advice to not use it outside of freeways because it does not have code to handle those situations

One example: I have taken a certain two-lane road that widens to include a left turn lane (no right allowed). A few updates ago, the car would struggle to determine what to do as the middle yellow line curved left while the white shoulder line stayed straight. It handled it, but with a little zig. As of at least the last update, it handles it as best as I would. I've tried it at different speeds and traffic conditions.

That's the system simply being better at handling where to go when a lane widens.... same as happens on many freeways when a lane widens to add in a new lane... and as you'd expect many used to report that was handled poorly on freeways, and a recent update greatly improved it.

The fact it does the same thing when a lane widens (to add a new lane) outside freeways isn't surprising- but it's clearly NOT indicating they've added "understanding local turn lanes"

I maintain that while the disclaimer will not change until Tesla is willing to say that they are now officially covering those situations, we don't know what is being baked into AP or TACC.

Not every line of code certainly- but we do have direct observation of what it does in such situations thanks to folks with access to the AP computer who publish the data.

It shows what things it recognizes, what it labels them as, what places it considers "driveable", projected paths, etc...

It does not support your thinking on this stuff that there's anything doing anything active in the code today for city driving... though there's certainly code to RECOGNIZE some things that'll be relevant when they roll out some "actually does something" code... like lights, stop lines, etc...

This could be for a different thread, but I have been thinking about what they could implement to start. Certainly start/stop at stop lights while going straight.

That is probably the "easiest" but based on current level of correct reading/detection of lights they're not there yet.

Especially as this is one of those "get it wrong once and people can die pretty easily" kinda things.

In fact Green posted just yesterday after having upgraded his own system to HW3 that the current code doesn't appear to be any better TODAY than it was back in March 2019 running on HW2 at correctly IDing what color a light is.... (that is, it often gets it right, but not NEARLY close enough to 100% you'd want to trust it to actually take actions)

 

[Mardak]

at least a few of us hope that Tesla improves the system's capability so that it can be used around town without unexpected behavior because it is kind of silly that cruise control can only be used on the freeway

It'll be interesting to see how Tesla improves (or not) Autopilot behavior on city streets for those who don't purchase FSD software or those without FSD hardware. On one hand, highway driving is pretty good right now on HW2+, so generally changing the Autopilot behavior to handle cross traffic could cause regressions, so a separate CityStreetsBehavior could be less risky (at least initially). On the other hand, Tesla probably wants to improve Autopilot safety features for city streets, e.g., predicting when cross traffic would result in a collision. I suppose even if HW2/.5 continues to only sample frames from pillar cameras (roughly every 4th frame because it's already reaching hardware limitations) when HW3 processes full frame rate to correctly enter an intersection, Autopilot safety feature could still predict collisions and avoid an accident in some cases.

 

[diezel_dave]

It'll be interesting to see how Tesla improves (or not) Autopilot behavior on city streets for those who don't purchase FSD software or those without FSD hardware. On one hand, highway driving is pretty good right now on HW2+, so generally changing the Autopilot behavior to handle cross traffic could cause regressions, so a separate CityStreetsBehavior could be less risky (at least initially). On the other hand, Tesla probably wants to improve Autopilot safety features for city streets, e.g., predicting when cross traffic would result in a collision. I suppose even if HW2/.5 continues to only sample frames from pillar cameras (roughly every 4th frame because it's already reaching hardware limitations) when HW3 processes full frame rate to correctly enter an intersection, Autopilot safety feature could still predict collisions and avoid an accident in some cases.

I am super curious how they are going to handle this feature split as well. I have HW3 but not FSD so I will be watching to see how the behavior of my car changes parallel with developments towards their goal of FSD.

 

[Knightshade]

It'll be interesting to see how Tesla improves (or not) Autopilot behavior on city streets for those who don't purchase FSD software or those without FSD hardware.

Not at all would be my expectation.

That system is explicitly not meant to be used on city streets after all. If you want that functionality you'd need to buy FSD.

 

[diezel_dave]

Not at all would be my expectation.

That system is explicitly not meant to be used on city streets after all. If you want that functionality you'd need to buy FSD.

That would be the expectation but doubtful that will be the reality. People still want to be able to use cruise control around town and if Tesla can't figure that out, that will push some customers, like myself, to look at other vehicles that are able to handle a seemingly simple task.

 

[Mardak]

Not at all would be my expectation.

I don't expect much either, but over the last year, Autopilot has improved quite a bit on city streets already. Before it would consistently fail on various 25mph curves especially without a lane line on each side as well as frequent hard braking for parallel parked vehicles or crossing an intersection into the wrong shifted lane. Those types of failures still happen today, but definitely not as often now.

And even if traffic light and stop sign warnings are primarily for Autosteer that was active before a highway exit or highway that ends, those warnings can be useful for those who do use Autosteer without FSD on city streets. So Tesla has already improved Autopilot on city streets through minor general improvements as well as explicit features. But especially with the hardware difference with FSD computer, it seems practical for engineering effort to focus on the new hardware. There's roughly 400k HW2/.5 vehicles, and there'll likely be at least 400k HW3 vehicles by the end of this quarter (ignoring retrofits) and maybe double that this time next year.

 

[diezel_dave]

I don't expect much either, but over the last year, Autopilot has improved quite a bit on city streets already. Before it would consistently fail on various 25mph curves especially without a lane line on each side as well as frequent hard braking for parallel parked vehicles or crossing an intersection into the wrong shifted lane. Those types of failures still happen today, but definitely not as often now.

And even if traffic light and stop sign warnings are primarily for Autosteer that was active before a highway exit or highway that ends, those warnings can be useful for those who do use Autosteer without FSD on city streets. So Tesla has already improved Autopilot on city streets through minor general improvements as well as explicit features. But especially with the hardware difference with FSD computer, it seems practical for engineering effort to focus on the new hardware. There's roughly 400k HW2/.5 vehicles, and there'll likely be at least 400k HW3 vehicles by the end of this quarter (ignoring retrofits) and maybe double that this time next year.

There have been way too many improvements to the car's ability to drive around the city on AP to claim that Tesla isn't working on its abilities in those environments or to try and claim that those new behaviors are just a result of improvements to allow AP to handle the freeway better because so many of the things that it does "well" around the city are simply not something that would be expected on a freeway.

 

[turnem]

There have been way too many improvements to the car's ability to drive around the city on AP to claim that Tesla isn't working on its abilities in those environments or to try and claim that those new behaviors are just a result of improvements to allow AP to handle the freeway better because so many of the things that it does "well" around the city are simply not something that would be expected on a freeway.

Agree. I have seen some great improvements over the past 6 months or so. The first software version 10 release was probably the best and then we seem to have slowly regressed from there.

I have HW3 and I purchased FSD so I'm eagerly waiting for improvement.