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

AP Changing lanes into adjacent vehicle

This site may earn commission on affiliate links.
I have a Tesla Model 3 with FSD and HW3. I have a little over 9000 mi on it. On Jan 11 2020, I was westbound on the San Mateo bridge with NAP engaged. Autopilot just decided to changes lanes (without indicating) causing the vehicle in the adjacent lane to swerve to avoid me. I only got a "Take over immediately" once my vehicle was halfway into the adjacent lane. I realize I have no proof that AP was engaged, and this just looks like I'm driving like an idiot.

I contacted Tesla about this. They told me that the lane markings were unclear and I shouldn't rely on the system. I totally get this, and had my hands on the steering wheel, and took over within a second or so. Had the guy next to me not swerved, this would've been a collision that would've been my fault.

When I told Tesla I don't agree with their assessment of unclear lane markings, and that I would post the video online, they phoned me and told me to stop threatening their employees, and if I have a problem with their response, I should contact their legal department.

I used to be a huge Tesla advocate. After this incident, and they way they've treated me, I'm posting this as a warning to everyone else.

Take a look at the dashcam footage yourself, and see if you think the lane markings are unclear.


NOA always puts the blinkers on and indicates a lane change in advance before doing an auto lane change. So this was NOT NOA changing lanes. More likely, NOA got a little confused about where the lane was because of something on the road. So it tried to steer into where it thought the lane was and then realizing that it was not in the right lane, it gave you the "take over immediately" warning. That is the standard warning when AP or NOA knows it is making a mistake and needs the driver to help by taking over.

Your experience is a reminder that AP and NOA are driver assist systems that are still very much beta because they are definitely not 100% reliable. AP and NOA are far from full self-driving.
 
So my memory is correct. I can exactly repeat this situation. Well, a mirror, the turn was to the right, but same thing.

This morning I was driving on a two lane road with AP engaged. Traffic was light & cooperative so I decided to experiment. I was in the left lane on a road that was in a gentle sweeping turn to the right.

I gently applied some force to disengage AP towards the right. I held the wheel gently to the right as AP disconnected. The car drifted into the right lane. Once it was about 1/2 way into the lane I got the "take over immediately" warning.

I thought I had a recording on the blackvue of the experiment, but I had the audio recording off so it doesn't tell the full story.

Anyway, I suggest people try it (safely) and confirm. I cannot guarantee 100% it's the same as what happened to the OP, but it is one possibility.
 
  • Informative
Reactions: Mrcarcrazy
I gently applied some force to disengage AP towards the right. I held the wheel gently to the right as AP disconnected. The car drifted into the right lane. Once it was about 1/2 way into the lane I got the "take over immediately" warning.

Are you suggesting that you applied just enough force to disengage AP, and then after it disengaged, it warned you to take over immediately, even though you technically had already taken over?
 
The lanes look pretty defined to me but you are still responsible. Technically you should always be checking your blind spot when the car changes lanes even when NOA is on and if you had you would have seen the car and taken over. You did threaten when you made the comment that you would post online but that Tesla managers reaction was a little much if you ask me. I recommend turning off the automatic lane change option on NOA and changing lanes manually with the blinker.
 
Are you suggesting that you applied just enough force to disengage AP, and then after it disengaged, it warned you to take over immediately, even though you technically had already taken over?

Yes, I kept a very slight force after the initial "take over" enough to guide it right but it thought my hands were not on the wheel after the initial bump.

I believe this "take over immediately" after AP disengagement was added a long time ago but was definitely not there from the start. I remember reading about an accident where someone went off the road into a railing. They thought AP was engaged so they were cleaning their front window (ok, that sounds brilliant). While reaching across they hit the wheel enough to disengage but didn't realize it.

Anyway -

Drive on a reasonably straight road with AP engaged. Nudge the wheel to disengage, and then keep hands off the wheel (but very close) and let the car drift. If it crosses a line within some amount of time after disengaging it will warn.

Now if you are just driving without AP and take your hands off the wheel and let the car drift it will not warn (unless the lane warning features are on), so I believe the warning is a safety for the accidental wheel bump (or holding lightly to keep the car happy and accidentally apply too much torque but don't realize it).

Again, not defending Tesla or their behavior. I do like to push the AP as much as possible and learn from everything it does. Just because I'm an engineering geek and I find it fascinating. The best thing Tesla could do in my opinion is to include indicators on the video for AP/TACC status. Either visual or audio.

Edit: Found a reference to the crash I remembered - Bloomberg - Are you a robot? & Another crash on Tesla Autopilot, another driver admits to not paying attention, was cleaning his dash - Electrek. Reason I remember is I drove down that same road around that same week.
 
Last edited:
  • Informative
Reactions: nvx1977
Based on what you're describing (moving to a different lane without a turn signal and also the red hands warning), it seems like AP was thrown off by that huge bump in the highway. The lane markings were fine, but the car might log the event as "I couldn't see the lines properly" and that's what the service tech saw when s/he pulled the logs.

TL: DR If the "hump" on the San Mateo bridge caused issues with NoA, we are in trouble for any short term FSD.

I am pretty sure that there have been thousands if not tens of thousands of trips over the San Mateo bridge in Tesla's so far. While not every Tesla would have been on NoA or have FSD, I believe they are always or almost always evaluating in "Shadow Mode" If a road that has been traveled that much in "shadow mode" still causes this drastic of an issue...

For those not from the SF Bay Area, this bridge is 23 miles from the Fremont Factory, and is one of the few commute routes from the SF Peninsula to the East Bay. It looks like the last major work to the bridge was repaving in 2015, and it's daily traffic is listed as 93,000 cars on the wikipedia page.[/QUOTE]
 
Last edited:
NOA always puts the blinkers on and indicates a lane change in advance before doing an auto lane change. So this was NOT NOA changing lanes. More likely, NOA got a little confused about where the lane was because of something on the road. So it tried to steer into where it thought the lane was and then realizing that it was not in the right lane, it gave you the "take over immediately" warning. That is the standard warning when AP or NOA knows it is making a mistake and needs the driver to help by taking over.

Your experience is a reminder that AP and NOA are driver assist systems that are still very much beta because they are definitely not 100% reliable. AP and NOA are far from full self-driving.

Also pays to be alert in construction areas where lanes have been redirected. I had to take command quickly on I-10 last week when my car tried to follow old lane marking that was still very visible.
 
TL: DR If the "hump" on the San Mateo bridge caused issues with NoA, we are in trouble for any short term FSD.

I am pretty sure that there have been thousands if not tens of thousands of trips over the San Mateo bridge in Tesla's so far. While not every Tesla would have been on NoA or have FSD, I believe they are always or almost always evaluating in "Shadow Mode" If a road that has been traveled that much in "shadow mode" still causes this drastic of an issue...

For those not from the SF Bay Area, this bridge is 23 miles from the Fremont Factory, and is one of the few commute routes from the SF Peninsula to the East Bay. It looks like the last major work to the bridge was repaving in 2015, and it's daily traffic is listed as 93,000 cars on the wikipedia page.
[/QUOTE]

Fun fact from Former Tesla Programmer's anecdotes about problems : EnoughMuskSpam
  • the first time we turned on real time telemetry for the dev fleet we caught somebody going 130mph over the san mateo bridge
 
  • Like
Reactions: ZappCatt
At 0:05, the car goes over a bump or drop that makes it bounce. (The passing car bounces immediately afterwards.)

This car has a fairly stiff ride. A bounce like that is enough to cause the weight of your hand on the wheel to take it out of AP, after which the weight of your hand could be causing the car to drift left.
 
I believe that AP in the latest software versions is tracking the lane lines not only with the front camera(s), but also with the side cameras. Here's what I think happened here:

1. You're on a curve that bends slowly to the left.
2. The sun is directly to the left of your car, at around a 35-40 degree angle above the horizon.
3. During the turn, at some point the sun aligned with the car to do one of two things:

a. Washed out the left pillar camera, causing it to lose sight of the lane lines.
b. Reflected off your window (are your windows tinted?) and onto the pavement, illuminating the road surface enough to wash out the lane lines.

4. Once the left lane lines were lost in the camera feed, AutoPilot attempted to center the car between the left and right concrete barriers, which are the next best thing to define the "lane" that the car is supposed to be in.

This wasn't a lane change, this was a drift similar to what we see the car do if you're on the freeway in the right-hand lane and pass an entrance ramp. The car attempts to center itself within the very wide merge area, thus moving to the right.

I had some oddities happen the other day as a result of sun directly abeam of the car. I was unable to initiate a lane change twice on two different occasions because AP thought that there was not a lane available to that side of the car. This occurred on each side of the car on the same road, once eastbound, and once westbound. The sun was directly abeam on both occasions. (At my latitude, noon sun in the winter sits about 45-50 degrees of elevation to the south).
 
The surprise to me is that it would have even made the lane change. For me NoA always blinks three times before it ever attempts to change the lane. The time, to me, is always interminable. If a vehicle comes up in that time on the left the lane markers usually red. I could see if maybe you pulled too hard to confirm, maybe you came out of autosteer? That's happened to me before.

But I agree with others, there's tons of times where AP doesn't behave as expected. I simply just submit a bug report an hope for the best.
 
This system does weird things sometimes, I’ve seen it first hand..

While I like the idea of being able to toy with it while still in beta, for the price we pay for the feature there should definitely be a higher reward than knowing you’re contributing to the big ball of data that will eventually make “JohnnyCab” a reality. I’ll take free supercharging. :D
 
They are just trying to intimidate you to keep from you exposing them, the same way Enron would do that to analysts right before their crash. If your supercharging stops working, you'll know why.
Dude do you even own a Tesla? Seems like every post i see from you is trying to provoke a negative response. “......right before their crash” Hhhmmm. Watcha trying to insinuate there i wonder?
 
I've now been able to recreate the situation with a quick bump to the steering wheel in both a 2017 S and a more recent 3 with my hand resting on one side of the the wheel. I'm betting on the simple explination - accidental disengagement of AP.

Weird that Tesla didn’t tell the OP that then instead of telling them the lines were the cause. Unless they didn’t actually review the logs, which is possible.
 
I would be very surprised if they go into some sort of extensive review process for every video presented. That would require matching time stamps and a good bit of time. Reviewing the logs probably just looks for major hardware errors. Especially if presented in a not very polite manner.
 
Lane markings look fine to me, however, I don't think this was NOA changing lanes. There wasn't any reflections showing the turn signal. Also, NOA won't change to the left lane when therw isn't any slow cars in front of you. NOA don't like the left lane even when AP is set to 90.

Looks like AP went from following the white lines to the dark black tire marks as it tried to center the car in between. Maybe that is why the red hand of death popped up on your screen.

Also, with my experience I don't think your hands were on the wheel or just have slow reflex. At least for me when AP does stupid stuff like this it breaks right away as my hands will provide enough grip before the car veer too far. My car does stupid stuff like this too where it will try to curb the car on left side island but my hand always breaks AP as soon as it veers off an inch.

Not saying it's your fault. I'm mad as well that AP is not perfect, so be careful. It works 98% of the time but that 2% could hurt or kill you. Keep an eye on the road and both hands on the wheel when using AP.
I agree with this assessment. It looks more like a lane tracking problem than NoA changing lanes. The lack of a car in front and no blinker are give always.