No, that is a completed job. The state tree here in Ohio is the orange construction barrel. You see them in full bloom in a construction zone. The mere absence of the state tree in the photos shows it is not a construction zone.
-
Want to remove ads? Register an account and login to see fewer ads, and become a Supporting Member to remove almost all ads.
Autopilot does not like road repairs
- Thread starter Atari2600
- Start date
StealthP3D
Well-Known Member
Let me counter with a very specific statement:
Your statement is very trollish. It's almost like you wanted to create a negative sounding thread title and then push peoples buttons to keep the thread visible as people responded to you.
I don't care how high the resolution is, that looks like a two-way highway with no passing allowed. All the resolution in the world is not going to fix the paint on the highway. Humans get confused by things like this. People take crack, get drunk and fail. In this case, no crack is necessary. But let's see how long you can keep this negative sounding thread title active. You have worked hard to get this thread to this point.
SammichLover
Banned
If it does can I order it as an implant for my head because I was confused as hell about what the issue was when I first looked at this picture.
"Autopilot does not like road repairs" is a negative sounding thread title? If I wanted negative I could do a lot better.
His statement was totally open ended with no way to even debate it. So I replied with something very specific one could debate. And instead of debating it yourself you attacked me instead. I stand by my posts.
I stand by my comments that the yellow line on the left looks nothing like the black sealer with the sun reflecting on it.
I stand by my comment what people are referring to as painted lines is in fact road sealant to patch the repair.
I stand by my comment that it is a road repair, it is not a construction zone.
I started this thread debating if it was considered a total failure or not and some people including you give it a total pass even though it is not working as it should?
Listen dude, I drive over 15,000 miles a year on this road. 15 ****ing thousand miles. As someone who paid for EAP/FSD, specifically to help me drive this road, do you not think I have a right to be unhappy about it?
In real life it does not look that confusing. Hopefully Kessler can confirm or deny. To my eyes it looked like white dashed lines on a patch job.If it does can I order it as an implant for my head because I was confused as hell about what the issue was when I first looked at this picture.
I don't even think it looks confusing in the pictures. Then again, I'm also in Ohio (the other end of I71) and used to this kind of stuff. (but I do see why AP acts the way it does.)
Last edited:
SammichLover
Banned
Thanks, I agree on all points. I hope they fix it via software or when I get the HW3 upgrade with the better image processing stuff.
I did submit a written support incident with Tesla the other day.
Derek Kessler
Active Member
That is correct. The seam between the two lanes of asphalt was crumbling so they put down a new strip of asphalt bounded by tar seals down the middle to at least temporarily arrest further decay. This is meant to be a permanent fix (at least until it falls apart again, because Ohio). It has reflectors and lane markings (I believe they're presently temporary lane markings until they can close down the highway over a non-holiday weekend evening to install permanent ones).
All of the confusion here, and I mean ALL OF IT — humans and computers alike — is from the way the low sun is reflecting off the new tar strips. This has long been a known weakness in all iterations of the Autopilot suite — it sees a bright solid line on the road roughly where it expects there to be lane markings. The decision engine currently lacks the logic and bandwidth to handle this specific edge case so it errs on the side of caution and assumes that it's an uncrossable line.
This will be solved by a few things: Time will dull the reflectiveness of the tar strip. ODOT will install brighter, more reflective, and more distinct lane markings. HW3+ bandwidth and an advanced neural network will allow Autopilot to discern the difference between tar lines and lane markings.
I'm sure you've all at some point been confused by tar lines on a road. I know I hit a highway in California once that had so much tar sealing all the cracks that I couldn't even tell where the lanes were supposed to be. It's worth remembering that while you are familiar with the road, Autopilot has no such memory. It's driving every road as if it's the first time with no idea what is around the bend. I look at that road and I see tar lines on a center patch because I know that's what it is. Autopilot today doesn't have the processing capacity to even understand that concept.
This is good. They will probably pull relevant footage from your car (if they haven't already) to use in training the neural net. I won't venture a guess as to how long it will take, but I am confident that the system will improve to properly recognize what that road is doing.
The 3 is my first Tesla so we are getting to know each other and are in the honeymoon phase But I have noticed that for the most part autopilot is good.. however, I happen to use a reversible express lane (for those that don't know what this is, it's a reversible road on a highway that goes one way in the morning, the opposite way in the afternoon). So, you can imagine there are crisscrossing lines that totally freak out autopilot. But what I've learned is that ME, the driver, needs to understand when I can use it safely and when I cannot. NOA is a bit rough at the moment, but it's better than it was in the last release.
But I have noticed that for the most part autopilot is good.. however, I happen to use a reversible express lane (for those that don't know what this is, it's a reversible road on a highway that goes one way in the morning, the opposite way in the afternoon). So, you can imagine there are crisscrossing lines that totally freak out autopilot. But what I've learned is that ME, the driver, needs to understand when I can use it safely and when I cannot. NOA is a bit rough at the moment, but it's better than it was in the last release.
Derek Kessler
Active Member
Drove this same stretch of I-71 today after a rain and got the same results: AP algorithms interpreted the dual tar stripes from the seam patch strip as an uncrossable double line.
Living in Ohio and just coming back from a road trip to North Carolina I've just determined ODOT sucks.
I had more autopilot issues in the 1.5 hours of Ohio than the other 10 hours in West Virginia, Virginia, and North Carolina combined.
Their construction zones are a mess, their line markings are terrible, they don't continue dotted lines for exits and entrance ramps, and they'll call a project like this done even though it's at least visually confusing to humans.
I had more autopilot issues in the 1.5 hours of Ohio than the other 10 hours in West Virginia, Virginia, and North Carolina combined.
Their construction zones are a mess, their line markings are terrible, they don't continue dotted lines for exits and entrance ramps, and they'll call a project like this done even though it's at least visually confusing to humans.
Derek Kessler
Active Member
This picture made things a lot clearer about what’s happening here. But at the same time I’ve had AP fail on far less confusing examples than this where it required intervention, so it is no surprise at all that it interprets this as a solid line.
The perils of being an early adopter. These situations *will* improve, but right now NoAP is still fairy rudimentary and requires driver supervision.
The more you drive, the more you will understand what the system struggles on.
I have driven over 14,000 miles and this is THE ONLY thing I’ve seen this struggle over.
I’ve had dirty camera issues but that’s not AP system issues... that’s hardware issues.
This is not a NoA issue either as it is simply a pattern recognition AP issue. Try to initiate a lane change in regular AP and it doesn’t work.
As a previous poster stated - ODOT sucks. But the system needs to be able to interpret the roads correctly.
I’ve had dirty camera issues but that’s not AP system issues... that’s hardware issues.
This is not a NoA issue either as it is simply a pattern recognition AP issue. Try to initiate a lane change in regular AP and it doesn’t work.
As a previous poster stated - ODOT sucks. But the system needs to be able to interpret the roads correctly.
For some reason I was not able to process this video properly with the teslacam software but I got it right today. This video is from the same trip in my original post. What you see in the right repeater view is what I visually see. The front camera view is garbage.
Very clearly looks like double yellow (Atari2600's footage/photo with sun reflecting) or faded white (Derek Kessler's recent photo), if you lack any previous knowledge. Even knowing what to look for, it is difficult to pick out the dashed lines in the video when looking through the forward camera. To be fair, I don't know if the NN's process the images in HDR or not, although the FSD chip does do tone mapping of the HDR images before they're passed to the NN I don't know if this means they're only getting a tone mapped SDR image or some kind of color / brightness corrected HDR image - so the NN may or may not be working with more than we see through the dashcam footage.
Even then, seeing some white dashed lines mixed in means nothing, since they're right on top of the reflective strips and when the light is reflecting off them they drown out the visible dashed lines to at least the point of confusion, if not loss. Keep in mind many places may have solid lines painted over dashed lines, so when the two are mixed, it is safer to assume a solid is intended. If dashed was intended, the solid should be removed. Derek's photo a human can probably make the jump to this being a single dashed line and not double white, but even then, if you weren't paying attention, you might be fooled momentarily.
Clearly in the repeater cams the dashes are only visible as the sunlight is not reflecting off them in the same way, but for the purposes of determining whether the lane is blocked off by solid lines, only the forward looking cameras are consulted (as unless you're driving in reverse, whether the lines behind you are dashed or solid is pretty irrelevant).
Perhaps some day it will use the rear facing cameras to determine that it must be reading the forward cameras wrong, but that requires a significant jump in logic and you must also be concerned about what if it was wrong, but then becomes right, but you're still treating it as wrong? You might attempt an illegal/dangerous maneuver in the time it takes to pass the point where you began crossing the lane boundary before the future markings become past markings that you can double check with the rear facing cameras.
It is unreasonable to expect perfection in this situation as most of the humans (well, I'm assuming) on who have commented they also see primarily a double line. Sure, it's annoying to get stuck in a lane and have the car fight you to change to the other lane, but as far as "failures" go it's a pretty human one.
Even then, seeing some white dashed lines mixed in means nothing, since they're right on top of the reflective strips and when the light is reflecting off them they drown out the visible dashed lines to at least the point of confusion, if not loss. Keep in mind many places may have solid lines painted over dashed lines, so when the two are mixed, it is safer to assume a solid is intended. If dashed was intended, the solid should be removed. Derek's photo a human can probably make the jump to this being a single dashed line and not double white, but even then, if you weren't paying attention, you might be fooled momentarily.
Clearly in the repeater cams the dashes are only visible as the sunlight is not reflecting off them in the same way, but for the purposes of determining whether the lane is blocked off by solid lines, only the forward looking cameras are consulted (as unless you're driving in reverse, whether the lines behind you are dashed or solid is pretty irrelevant).
Perhaps some day it will use the rear facing cameras to determine that it must be reading the forward cameras wrong, but that requires a significant jump in logic and you must also be concerned about what if it was wrong, but then becomes right, but you're still treating it as wrong? You might attempt an illegal/dangerous maneuver in the time it takes to pass the point where you began crossing the lane boundary before the future markings become past markings that you can double check with the rear facing cameras.
It is unreasonable to expect perfection in this situation as most of the humans (well, I'm assuming) on who have commented they also see primarily a double line. Sure, it's annoying to get stuck in a lane and have the car fight you to change to the other lane, but as far as "failures" go it's a pretty human one.
Similar threads
