And that's USA!
Imagine the rest of the world, lol.As I am in the 26k VIN range, I wonder if they will go through your VIN before mine anyway. That does not bode well, lol.
Autopilot getting worse
- Thread starter dano9258
- Start date
And that's USA!
Imagine the rest of the world, lol.As I am in the 26k VIN range, I wonder if they will go through your VIN before mine anyway. That does not bode well, lol.
W
Was this alignment done at your expense, or under warranty?
How old is your car and what is your AP version?
Was this alignment done at your expense, or under warranty?
How old is your car and what is your AP version?
Hi everyone, first post here on TMC, but I'm on TOO and Reddit under the same username. I agree with everyone in this thread about AP having some notable regressions in the past 2 months. (I'm on HW 2.5) It's almost as if all the computing power that went to deep rain got taken away from basic lane centering and highway speed curve prediction. I noticed a drastic reduction in AP's "attentiveness" around 2019.40.1 that has yet to return; I override AP about once per 20mi now on the interstate for things such as forgetting to turn until it's scary, hunting the double width right lane after a merging on-ramp lane, and just refusing to stay in the center of the lane.
AP used to be the feature I would show off the most to passengers in my car and now I rarely use it with passengers because I don't know what it's going to do wrong anymore, I just know it has a much higher chance of being embarrassing than it was on 2019.36 and previous builds. Understeering the same curve multiple times which results in soft bouncing off the outside lane line or sling-shotting off the outside lane line are not features you want to demonstrate to others.
My car will occasionally get about a 6" left to right and 4-5 second sinusoidal sway, it is very predictable and repeats 10+ cycles of slowing bouncing from left of center to right of center and back again in the lane. It's like the lane is just a touch too wide for either camera to care until the car drifts into it's territory and then it'll redirect the wheel to steer the car back across the no-man's land in the center of the lane until it encroaches on the other side's territory.
On my commute home today I had a new experience. I don't know if this is a PSA for general autopilot use or if it is part of the recent lane centering regressions. It was raining and I was passing a 1-ton dually with a gooseneck trailer that was kicking up quite a bit of road spray. When I was about even with the front cab of the truck AutoPilot decided to dive bomb right towards the truck next to me. The right pull on the wheel was as aggressive as a modern AP lane change or re-centering into the double width merging lane pull that we've all experienced.
This was the first time that I'm fairly certain AP would have hit the car next to me if I didn't prevent it from doing so. My best guess is that the spray from the truck confused the ultrasonic sensors and the camera's lost track of the truck and the center lane line (it was clearly painted BTW). After the car lost track of the center line and the truck due to the spray I think it continued to see the outside lines of each lane and started to interpret the two lane highway as a single double width lane. I think the violent pull towards the truck was the same centering behavior as a merge lane, just with a 70mph pickup 3ft into its chosen path. I saved the TeslaCam clip but of course forgot to properly pause the CAM before ejecting it when I got home so my saved clips are the dreaded 594 byte nothing files. Has anyone else had this happen?
AP used to be the feature I would show off the most to passengers in my car and now I rarely use it with passengers because I don't know what it's going to do wrong anymore, I just know it has a much higher chance of being embarrassing than it was on 2019.36 and previous builds. Understeering the same curve multiple times which results in soft bouncing off the outside lane line or sling-shotting off the outside lane line are not features you want to demonstrate to others.
My car will occasionally get about a 6" left to right and 4-5 second sinusoidal sway, it is very predictable and repeats 10+ cycles of slowing bouncing from left of center to right of center and back again in the lane. It's like the lane is just a touch too wide for either camera to care until the car drifts into it's territory and then it'll redirect the wheel to steer the car back across the no-man's land in the center of the lane until it encroaches on the other side's territory.
On my commute home today I had a new experience. I don't know if this is a PSA for general autopilot use or if it is part of the recent lane centering regressions. It was raining and I was passing a 1-ton dually with a gooseneck trailer that was kicking up quite a bit of road spray. When I was about even with the front cab of the truck AutoPilot decided to dive bomb right towards the truck next to me. The right pull on the wheel was as aggressive as a modern AP lane change or re-centering into the double width merging lane pull that we've all experienced.
This was the first time that I'm fairly certain AP would have hit the car next to me if I didn't prevent it from doing so. My best guess is that the spray from the truck confused the ultrasonic sensors and the camera's lost track of the truck and the center lane line (it was clearly painted BTW). After the car lost track of the center line and the truck due to the spray I think it continued to see the outside lines of each lane and started to interpret the two lane highway as a single double width lane. I think the violent pull towards the truck was the same centering behavior as a merge lane, just with a 70mph pickup 3ft into its chosen path. I saved the TeslaCam clip but of course forgot to properly pause the CAM before ejecting it when I got home so my saved clips are the dreaded 594 byte nothing files. Has anyone else had this happen?
On a recent long drive through France, I experienced very pronounced weaving on the Autoroutes. It seemed most obvious when lanes were slightly wider than on other roads, and withe very clear lane markings. I suspect that the effects we see are typical of any servo controlled oscillation, where you are trying to balance between sharp responses and (in this case) allowing the car to steer around sharp bends, with precise lane positioning and smooth movement. I imagine the ideal would be a critically damped oscillation, but in the case of our cars, there are so many variables that this is usually impossible to achieve. I felt quite often that the ping pong motion got so large that it was going outside of the lane markings.
I haven't had it quite this bad in freeway driving, but in the UK there is a huge variety of road layouts especially fairly busy main roads with parked cars along the sides. AP must make assumptions about what type of road it is on and use those to determine how it drives. I find it very odd that AP will move over for a large truck in an adjacent lane, but not flinch to avoid a parked car protruding into your lane on the roadside. If it sees enough of the parked car, it just treats it as a traffic queue and stops behind it.
What I find really dangerous is that when you intervene to take control and AP cuts out, TACC remains engaged until you brake.
2020.4.1
2018.06 build.
I had to pay for the alignment.
The weaving seems a lot better but still isn't completely gone.
It's gone enough that my wife doesn't complain anymore.
I think it was the camera alignment that did the most good.
I was watching on my app as the service person was driving the car up and down a road near the watertown service center and I assume he was adjusting something.
Last edited:
Enginerd
Member
Regarding the weaving on the highway... while it's possible that ultrasonic sensor calibration is responsible for some of the weaving on the highway when other vehicles are present, I'm sure that's not the only contributing factor. On I-95 today with no traffic for a couple hundred yards, the car was weaving. There were no guardrails or other objects for the sensors to pick up. I stick by the theory that it's at least partially just an underdamped feedback loop for the visual-to-steering control system.
I agree with you, I don't think it's a hardware issue. This is a recent regression that originated with a software update so I'm confident it can be fixed with a software update. I'm not positive if the late and understeered corners are the same root problem as straight lane weaving but they are equally annoying on our current firmware. The underdamped feedback loop theory seems solid since you can really notice AP going way past center when exiting curves, resulting in a very noticeable soft bounce off the previously inside lane line. I tweeted Elon to rededicate some of the resources that went into deep rain back into deep lane centering...I'm not holding my breath for a response!
Sorry, I still rate being able to see where I'm going as pretty high on the list. Although I suspect Elon would much prefer it if I just had 'deep trust' that I am in the center of the lane! That uses least processing power.
I'm 052xxx and have my HW3 install scheduled for March 4 in Maryland.And that's USA!
As I am in the 26k VIN range, I wonder if they will go through your VIN before mine anyway. That does not bode well, lol.
I recently got HW3 retrofit installed on my Model 3 and have had some concerning AP behavior. In the morning I'm traveling south so if the sun is out, it's shining directly into the drivers side B pillar camera. There's only been 2 mornings so far since getting HW3 installed where the sun it out. On both those days driving on a 2 lane divided highway, AP will harshly dive left (towards the direction of the blinded camera) ride the line and recenter. The IC still shows the lines perfectly (not bouncing like when it can't determine where they are) and knows it's right on the line or going over. It dives so abruptly and continually it makes AP completely unsafe and unpredictable.
It seems like there is a correlation to the B pillar cameras affecting lane centering. In my case to a pretty significant degree. I'm wondering if it's also affecting these cases where it's ping-ponging to lesser degree.
Next sunny morning, I want to put some tape over the drivers side B pillar camera and see if the behavior changes.
It seems like there is a correlation to the B pillar cameras affecting lane centering. In my case to a pretty significant degree. I'm wondering if it's also affecting these cases where it's ping-ponging to lesser degree.
Next sunny morning, I want to put some tape over the drivers side B pillar camera and see if the behavior changes.
diplomat33
Well-Known Member
I've noticed some regressions on AP2.5. On a 2 lane "city street", the car seems to lose the lane lines because of faded asphalt and will freak out mid auto lane change and throw me back into the right lane. On NOA, I've had several instances where it freaked out on an exit that it used to be able to do smoothly before. Is the software being optimized now for AP3 and not working as well on AP2.5?
I strongly believe this to be the case. I think they switched over their programming from HW2.5 base to HW3 base with 2019.40 and later firmwares. I've noticed the regressions you mention and others that showed up for the first time in 2019.40 and have had no signs of going away. Sections of my commute that the car did perfectly every time in November now require multiple disengagements. Lane re-mergering when on-ramps join the highway has gotten much worse as well as general lane centering. It's like HW2.5 is too slow to make decisions on the new firmware even though it was doing great in 2019.36.
NOAP tried to exit at 60mph onto lay-by
My MS Raven HW3 FSD is not obviously different / better than the 2017 MX 100D loaner that I've been driving for a couple of weeks.
All the behaviors mentioned by @diplomat33 feel like business as usual to me.
My MS Raven HW3 FSD is not obviously different / better than the 2017 MX 100D loaner that I've been driving for a couple of weeks.
All the behaviors mentioned by @diplomat33 feel like business as usual to me.
I've been a vocal critic of lane centering on HW2.5 ever since 2019.40 firmware. Today was my first drive on 2020.8, still with HW2.5, and it appears that lane centering has gotten much better on this release. I still felt an occasional soft bounce off the lane lines but the system steering damping feels better now, no more 4 or 5 over correction/laziness bounces.
In addition to getting lane centering almost back to where it used to be I noticed less re-merging tendencies with on-ramps, again closer to how the car drove with 2019.36 and before. Lane changes are slightly slower but smoother and more natural feeling whereas recently they've been fairly dive bomber in style.
Overall the AP system appears to be behaving slightly more cautious than the previous few updates which is good. The acceleration and deceleration isn't as butter smooth as it was the past few updates but that's a perfectly fine trade off for better lane centering to me.
In addition to getting lane centering almost back to where it used to be I noticed less re-merging tendencies with on-ramps, again closer to how the car drove with 2019.36 and before. Lane changes are slightly slower but smoother and more natural feeling whereas recently they've been fairly dive bomber in style.
Overall the AP system appears to be behaving slightly more cautious than the previous few updates which is good. The acceleration and deceleration isn't as butter smooth as it was the past few updates but that's a perfectly fine trade off for better lane centering to me.
I used to have really similar sway, just not as bad as you. It was particularly bad at night, especially when driving by semi trucks. AP is back to being solid as a rock on 2020.8.1 for me. I have a 2017 MX 75D. Hope it fixes things for you after you get it if you haven’t yet.
Similar threads
