Welcome to Tesla Motors Club
Discuss Tesla's Model S, Model 3, Model X, Model Y, Cybertruck, Roadster and More.
Register

Navigate On Autopilot: automatic lane change results

This site may earn commission on affiliate links.
Hey everyone, just wanted to put the final nail in the coffin for "HW2.5 doesn't exit passing lane, HW3 does exit passing lane," bug as I still have some doubters.

I picked up my Model 3 with HW3 yesterday morning, and SEVERAL times on the first drive using NoA I saw the exiting passing lane message. It was still 2019.40.50.7. The only difference was the HW3 upgrade. Here's even a pic I was so excited
Sadly, that nail may not be the final one. I haven't driven on 2020.x yet, as I just got it this morning, but I have not one time to date seen my car (HW3) exit the leftmost lane for any reason other than it being slower, or my exit coming up ahead, or traffic cones scaring the computer away.
 
Just had a 3 hour there and 3 hour back pretty excellent NOA or AP drive on 2020.4.1. This is HW 2.5. In NOA not once did it voluntarily get out of the passing lane, even after I let it sit at 70 MPH and cars passed on the right.

Now it did do something entertaining. We wuz cruising in the 3 lane on NOA headed south in I680 right after the Martinez bridge, headed to Walnut Creek and then Oakland, when Nicki suddenly decided to urgently prep for exiting onto Highway 4. That's an alternate route but it was a last minute change. The urgency was making life a bit difficult for others, slowing to grab lane changes, squeezing in like his/her life depended on it. I just let it happen to see what next. What next was Doh, sorry 'bout that Chief. Never did get off at HW 4, just worked back into the fast lanes on 680 and ramped up to 70 MPH back in the flow, like it never happened.

No idea of course, but it seemed like it had spontaneously changed the itinerary to run through HW 4 (which is a plausible way if traffic is really bad on 680), and then decided it wasn't worth it, maybe the traffic updated, so it went back to the original track. That's pretty independent thinking. It was pretty funny. You can't call it a "malfunction", but it was the first time I've seen anything like that.
 
Just had a 3 hour there and 3 hour back pretty excellent NOA or AP drive on 2020.4.1. This is HW 2.5. In NOA not once did it voluntarily get out of the passing lane, even after I let it sit at 70 MPH and cars passed on the right.
Dirty Tesla YT'r just got his HW3 upgrade in his TM3 and couldn't stop talking about his pet peeve of getting out of the passing lane is now working.
 
Well as @Fernand noted, firmware 2020.4.1 did not fix the "failure to exit passing lane" issue in my AP2.5 test vehicle. So far, @DopeGhoti says that HW3 (factory installed) did not fix this issue, while @DirtyT3sla says the HW3 upgrade did fix the issue. At this point, I'm just hugely frustrated that NOA is crippled, 7 weeks running. While others may be tolerant of this deficiency (because they either driver faster than most traffic, or believe it's not a safety issue), I'm just reporting the raw performance stats. I also believe that a lay-person would deem it to be a problem, so it's appropriate to flag it.
  • Drive 55: Thankfully the weather was poor and NOA was unavailable for much of the drive, providing relief from the "failure to exit passing lane" problem. Seriously, I still love NOA... but only half the time now.
  • Drive 56: The one driver-abort occurred while entering I-95 S in Santee SC, when Wattson tried to jump 2 lanes straight into the fast lane.
  • Drive 56: "Uncomfortable successes" seem to be on the rise. There were several instances on this drive of cringe-worthy lane change timing (bringing shame, but no immediate collision risk). This may be indirectly related to the "failure to exit" problem. After a failure to exit the passing lane, I manually command a lane change to the slower lane. Then for some period of time (like 1 minute), the car will not change lanes for slower cars at any lower speed. I believe the intent is to sustain my commanded lane change for some period of time without negating it. This creates situations where the car waits to pass for a significant period of time, until this "timer" expires, then it performs the lane change to pass. The timing appears arbitrary to other drivers. I could literally see the drivers in the fast lane behind me toss up their hands in a "seriously, after all this time, you change lanes now?" gesture.
Full results:
upload_2020-2-4_1-6-35.png


Weighted score:
upload_2020-2-4_1-7-39.png


Cumulative results with this firmware:
upload_2020-2-4_1-17-13.png
 
Well as @Fernand noted, firmware 2020.4.1 did not fix the "failure to exit passing lane" issue in my AP2.5 test vehicle. So far, @DopeGhoti says that HW3 (factory installed) did not fix this issue, while @DirtyT3sla says the HW3 upgrade did fix the issue.
Well, I bring a new data point. For my case at least, the combination of HW3 and the 2020.4 firmware did fix the issue-- for the first time yesterday, my car exited the left lane for a stated reason of "Leaving passing lane". However, I did have my first egregious "okay, moving over whoops let's go back and try again" lane change in what seems like a while.
 
Well, I bring a new data point. For my case at least, the combination of HW3 and the 2020.4 firmware did fix the issue-- for the first time yesterday, my car exited the left lane for a stated reason of "Leaving passing lane". However, I did have my first egregious "okay, moving over whoops let's go back and try again" lane change in what seems like a while.

Just to add to this my HW3 install was originally on 2019.40.50.7 and I had over 10 passing lane exits just on the drive to work from the service center. And then several more in my video that was already posted on this thread.
 
@Enginerd I think you really need to change that reporting format.

We all KNOW there's a bully behavior on HW2.5; it hogs the passing lane. Frankly, it doesn't affect my driving comfort, I just tap the blinker lever to get out of the passing lane, if I feel like it, and it does it competently and safely. Note that in many jurisdictions there is no statutory obligation to vacate the leftmost lane.

I'd suggest excluding bully hogging, which isn't an event but rather a judgement as to when an event SOULD have occurred, and logging that in a separate table, so we can see both it and the actual failure events.

The problem is that by including these non-events in your tabulation you're obscuring the core results we'd like to know about: how's the NOA, how is it handling lane changes, is it getting better. Your chart implies the NOA is going to hell, from the high 90s to under 30%, and that's not the case at all.

You also don't have a way to convey how slick a successful lane change was. With V10 the confidence and tightness of fit keep improving. In fact I'm constantly amazed at how intelligently it works, how appropriately aggressive at times, and how deferential at others. Some might judge a tight squeeze or an occasional game of chicken poker as "uncomfortable success". At a freeway onramp where 2 feeder lanes marge, I LOVE sitting back and watching Nicki work it out. It's what it takes. Bona fide AI in action.

I love to hear how others are experiencing it, so I regret that your chart isn't more relevant to my NOA experience. I find the system way way better than a few months ago.

Let's agree we're eagerly waiting for HW3 or whatever will fix the bullying, and in the interim please continue to tabulate and display actual lane change problem events :)
.
 
Last edited:
Let's see if I can say this better. I think I can guess at why the passing lane hogging happens on HW2.5.

The core method of lane change in the early stages was waiting for a clear spot and time and moving into it. Fine. Been there, done that.

The next problem is that if the AI car always lets the humans through, always grants them priority, in many dense traffic cases it will never "get there". It requires a different tactic.

We humans always play chicken as we squeeze in. There's a fair element of push and aggression. The implication is "if you don't make a little room for me, we will both die". The Tesla AI is learning how to do that very well. As of around V10 we're onto this next phase. It requires a lot more intelligence.

For instance, look at getting on the freeway with two bumper to bumper feeder lanes creeping and merging into a single lane on-ramp. If the AI is too polite, and lets humans always have right of way, it will never get through. But not only that. It will also cause more logjam because most people expect an alternating give and take, sometimes more give, sometimes more take. And some people are just bullies. Watching how my car does it as of 2020.4.1, it's amazing. It's clear how much AI goes into it, and how, in less dramatic form, this now also applies to ALL lane changes. And as @Enginerd points out, there are other tricks and considerations at work.

I suspect that with the required degree of processing for these smarter lane changes, including detecting rear oncoming and passing traffic, HW3 has no trouble, but HW2.5 is stretched on processing power. So it's probably programmed to sacrifice the tricky detections and maneuvers needed to implement a smart "leaving passing lane". Since all NOA cars will soon have HW3, it's better for Tesla not to waste development on trying to make the old HW2.5 chicken fly. The issue will soon be long forgotten. I doubt we'll suddenly see HW2.5 acquire a "leaving passing lane" capability, unless it's very rudimentary. That's my guess anyway.
.
 
Last edited:
I love to hear how others are experiencing it, so I regret that your chart isn't more relevant to my NOA experience. I find the system way way better than a few months ago.
Still on 2.5 here:

I find that the firm lane changes (sufficient lateral g-force to induce passenger discomfort) no longer have any context.

Merging in rush hour traffic on the 401 in Toronto, a firm lane change is expected and welcome.

Two hours east of Toronto, in the middle of nowhere with no traffic around except for the one 18-wheeler you want to pass, the lateral g-force is reminiscent of being in the back seat during day one of drivers ed.

Also, (I suspect it's a 2.5 thing) I find the software applies a "hack" that puts me in the left lane when approaching any interchange and thus avoiding any issue about dealing with merging traffic...again, with NO traffic around and approaching an interchange in the middle of nowhere, it insists on going into the left lane...unless I force it to stay in the right lane.

Note that in many jurisdictions there is no statutory obligation to vacate the leftmost lane.

Although technically correct, one could argue that "best practices" is to treat the left lane as a passing lane and use it as such.

What I would like to see in future software developments: the car will accelerate and shift to the left lane and "match" the left lane traffic flow versus the current practice of becoming a rolling roadblock unless one puts their foot into the accelerator to match the left lane traffic flow speed.
 
Last edited:
What I would like to see in future software developments: the car will accelerate and shift to the left lane and "match" the left lane traffic flow versus the current practice of becoming a rolling roadblock unless one puts their foot into the accelerator to match the left lane traffic flow speed.
I for one to not want my car to autonomously go faster than I have directed it to go.
 
We all KNOW there's a bully behavior on HW2.5; it hogs the passing lane. Frankly, it doesn't affect my driving comfort, I just tap the blinker lever to get out of the passing lane, if I feel like it, and it does it competently and safely. Note that in many jurisdictions there is no statutory obligation to vacate the leftmost lane.
I understand this viewpoint, and I respectfully disagree with it. In order to reduce my driving stress and provide on-ramp to off-ramp functionality without driver intervention, I believe exiting the passing lane is a required feature. Whether or not it's legally "required" is another matter, but everywhere I've driven, this etiquette is at least "expected". Given that this feature used to work and now doesn't (and does again w/ HW3), indicates that it is a bug and not a design choice. Also, on a personal level, I feel like an a-hole when the car does this, and this behavior does not reduce my stress. Using my wife as a litmus test, if it would make her lose trust in the system, it's a problem. To each their own, but I believe this is a required waypoint on the way to "feature complete".

Let's agree we're eagerly waiting for HW3 or whatever will fix the bullying, and in the interim please continue to tabulate and display actual lane change problem events
Fair. I think the charted data still provides this. The "driver abort" category includes driver intervention for safety reasons. Recent data shows there is a finite small quantity of these events, and there has been improvement over time. I think we both agree that for system maturity, avoiding driver intervention for safety reasons is the most important thing.

Since all NOA cars will soon have HW3, it's better for Tesla not to waste development on trying to make the old HW2.5 chicken fly. The issue will soon be long forgotten. I doubt we'll suddenly see HW2.5 acquire a "leaving passing lane" capability, unless it's very rudimentary.
It is not true that all NOA cars will get HW3. Everyone with HW2.5 who bought Enhanced Autopilot but not FSD is going to keep their current hardware, which should not be saddled indefinitely with a bug that is serious at least for some.
 
I for one to not want my car to autonomously go faster than I have directed it to go.
Agreed. That is why there would be an option within the UI that allows one to blend into the current speed of traffic within the passing lane.

I must ask the question though, if you allow autopilot to pull into a passing lane where the traffic is travelling faster than your current set speed, why would you wish to hold up that traffic?
 
Agreed. That is why there would be an option within the UI that allows one to blend into the current speed of traffic within the passing lane.
We have that already when the outside lane is moving slower; it even will show you that your car is going slower than its set speed to more closely match the speed of the adjacent lane. If the lane is moving faster and you want to merge into it nonetheless, I would suggest first contemplating why. If you still want to, you can use the dial on your steering wheel to adjust the set speed to match the adjacent lane before starting a merging maneuver with the turning signal lever.
 
We have that already when the outside lane is moving slower; it even will show you that your car is going slower than its set speed to more closely match the speed of the adjacent lane. If the lane is moving faster and you want to merge into it nonetheless, I would suggest first contemplating why. If you still want to, you can use the dial on your steering wheel to adjust the set speed to match the adjacent lane before starting a merging maneuver with the turning signal lever.
I don't want us to talk over each other.

What you describe are hacks that we must do because this car is not nearly "self driving" as some folks wish it to be.

As long as I have to force the car to speed up above my set point to merge into and with faster traffic then the software is not yet mature.

I believe this is the point @Enginerd is making with his approach to tracking the results he sees.

I agree with it, you don't. Fair enough.
 
Oh, I think this is all wonderful. The discussion, I mean. And the way I see it, we're discussing subtleties that we wouldn't be talking about unless the basics were damned good. I can handle etiquette differences OK in my "supervisory" role, since the car is handling ever more of the lower level driving just fine.

Some useful tweaks I think include how far above or below the target speed you initiate, that's a setting I prefer at negative 5 or 10 MPH. Another is to sometimes roll the top speed up to far higher than is achievable, which seems to bring about more aggressive lane seeking, kind of like using your spurs. I think leaving the top speed set like in a 1970 cruise control is too passive. How's the poor car to know what you're thinking? working that scroll wheel, and sometimes a touch of the accelerator, is a great way to conduct the computer ensemble, to communicate your wishes. Whoa! Giddyup! Easy, boy! A good horse thrives on interaction. I've come to see Nicki's behaviors as adding up to a creature of sorts.

But I persist in my belief that it's much too easy to inadvertently jerk out of AS, with potentially tragic consequences. I understand why Tesla does it this way for legal reasons given average drivers. A small 100g counterweight on the right spoke of the steering wheel makes my response to the Nags much smoother, with much less risk of unconsciously dropping out of autopilot.
 
Last edited:
  • Like
Reactions: VValleyEV
It is not true that all NOA cars will get HW3. Everyone with HW2.5 who bought Enhanced Autopilot but not FSD is going to keep their current hardware, which should not be saddled indefinitely with a bug that is serious at least for some.

That's a very good point, I'd forgotten about the "nein, danke" contingent. If that proves to be a real number long term, then I'd expect a rudimentary "leaving passing lane" to be added, kind of like what we had before V10, without much rear-view analysis.
.
 
Oh, I think this is all wonderful. The discussion, I mean. And the way I see it, we're discussing subtleties that we wouldn't be talking about unless the basics were damned good. I can handle etiquette differences OK in my "supervisory" role, since the car is handling ever more of the lower level driving just fine.

Some useful tweaks I think include how far above or below the target speed you initiate, that's a setting I prefer at negative 5 or 10 MPH. Another is to sometimes roll the top speed up to far higher than is achievable, which seems to bring about more aggressive lane seeking, kind of like using your spurs. I think leaving the top speed set like in a 1970 cruise control is too passive. How's the poor car to know what you're thinking? working that scroll wheel, and sometimes a touch of the accelerator, is a great way to conduct the computer ensemble, to communicate your wishes. Whoa! Giddyup! Easy, boy! A good horse thrives on interaction. I've come to see Nicki's behaviors as adding up to a creature of sorts.

But I persist in my belief that it's much too easy to inadvertently jerk out of AS, with potentially tragic consequences. I understand why Tesla does it this way for legal reasons given average drivers. A small 100g counterweight on the right spoke of the steering wheel makes my response to the Nags much smoother, with much less risk of unconsciously dropping out of autopilot.

The horse metaphors are interesting.

I was not aware that accelerator touches would change NOA lane change behavior and aggressiveness level. Is this a documented fact, a conjecture, or a suggestion?

Cruise speed setting changes seem more likely to affect NOA lane change behavior and aggressiveness. Has this in fact been documented?

Thanks.
 
No, I didn't mean using the accelerator except to push at certain times. I raise the top speed way above ambient speed at times with the right scroll wheel, to make the autopilot more aggressively seek lanes where it hopes to reach that higher target speed. It's nothing to document, it's pretty self-evident that at equilibrium, with everyone steady at their 70 MPH cruise control setting, if the Tesla's target top speed is also set to 70 MPH, it's not going anywhere. Raise it to 85 and watch it look for opportunities !

In the animal metaphor, it's like hanging the carrot in front of the mule's nose :)
.
 
Last edited:
So, uh, tell car to go faster and it goes faster. Math checks out.

If I have my car set to 73, and I am not in the passing lane, and people are passing me, am I that strange in not feeling a need to join all those folks in breaking the law even more than I already am (presuming a limit of 65)?
 
No, I didn't mean using the accelerator except to push at certain times. I raise the top speed way above ambient speed at times with the right scroll wheel, to make the autopilot more aggressively seek lanes where it hopes to reach that higher target speed. It's nothing to document, it's pretty self-evident that at equilibrium, with everyone steady at their 70 MPH cruise control setting, if the Tesla's target top speed is also set to 70 MPH, it's not going anywhere. Raise it to 85 and watch it look for opportunities !

In the animal metaphor, it's like hanging the carrot in front of the mule's nose :)
.

Yeah I quite like the “giddy up” metaphor.

Of course if your cruise speed is set to the same speed as the car you are rolling, it won’t try to pass, and it will if you bump the cruise speed up. I guess my question is:

If your cruise speed is set at 5mph above the car you are following, will it get more aggressive at looking for an opening to pass if you move it to 10mph above? More yet at 20mph above? Have you seen this kind of behavior?