Is Autopilot more reliable in cities with lots of Teslas traveling the highways?

[calisnow]

Haven't seen discussion on this yet - but I'm just wondering if this is true.

Here in Southern California I expect we have higher Model S density than any other city on the planet with perhaps the exception of the Bay area.

Yet there are still certain conditions on I-10 and the 91 freeway which consistently flummox autopilot - low sun, high glare, faded lane markings and no cars around for the autopilot to help lock its position.

I'm not saying the car wants to drive off the freeway but the wandering can get so bad I turn it off - not because I am scared (I'm more fascinated and want to see what it will do) but because I'm afraid I'm frightening the cars behind me, perhaps so much that someone will call 911 and my plates in and tell the cops I'm drunk.

If these high definition maps Tesla claims the fleet is building, and the vaunted fleet learning are both in fact actually working - I would have expected that the Los Angeles freeways would be almost 100% bug free by now, even in the glare conditions - because in theory the fleet of local Teslas should have mapped out all the lanes by now.

For those of you who do not visit So Cal frequently, Teslas seem to be everywhere - even 80 miles east of Santa Monica all the way out in Riverside and San Bernardino.

Also remember how Elon told the world that Autopilot was delayed in the final months leading up to 7.0 because Tesla was trying to improve reliability in the "corner cases" of poorly marked sections of the 405.

Clearly the problem is not completely solved yet 6 months later despite thousands of local Model S's tracing the same routes day after day on our freeways.

Thoughts?

 

[calisnow]

More theories:

1 - High definition gps maps are not as precise as I am imagining they are - and so maybe they prevent outright travel into the next lane but still give only rough guidance to the car's brain - and thus when other markers are lacking the car still wanders. If this is the case perhaps AP 1.0 will never completely reliable on high glare, no-traffic, poorly marked freeways - even in high-tesla-density areas.

or

2 - Maybe the fleet simply needs more time.

Having said all this - I am still delighted with Autopilot and use it 100% of the time to do the driving on freeways for me unless it starts wandering in bad conditions.

 

[Todd Burch]

I think the cars are collecting a lot of precision map data, but none of it is being used yet.

 

[CHG-ON]

I seem to recall that Tesla is only pushing certain improvements to AP. I think the most (I can't think of the word) obvious (?) fixes. They are still crunching all that data. I can say that I have seen mine incrementally improve in specific areas that it could not handle before. I don't know if that is because I have tried to be very consistent in my corrective actions and it is learning from those actions or if the Mother Ship is quietly correcting her.

 

[S4WRXTTCS]

I doubt it, and even in perfect conditions the autopilot can do some weird things.

I've tried ignoring the display console lines because they constantly shift a little left, and right as it updates.

I haven't had too much issue with it wondering, but I do think it could use some more dampening so it didn't correct so often.

I can go for hundreds of miles without any serious screw up. The last screw up happened at night (no rain, no fog), and I was in the center lane. There was a semi in the right lane perfectly centered, but for some reason the autopilot shifted close to it when I got near it. Which was really weird because usually it shifts a bit away from it.

So I took control to push it to the left, but realized I over corrected (I should have just held it like I normally do, but I panic'd a bit). So of course I go slightly to the left lane, but there was a car coming up (some idiot that kept the same pace as me, but in the left lane). So then I go back. Of course this back and forth creates oscillation. Nothing bad happened, but left me feeling uneasy.

Then I realized I left the steering in sport mode (which I prefer for driving around), but medium is best for freeway.

 

[kort677]

Haven't seen discussion on this yet - but I'm just wondering if this is true.

Here in Southern California I expect we have higher Model S density than any other city on the planet with perhaps the exception of the Bay area.

Yet there are still certain conditions on I-10 and the 91 freeway which consistently flummox autopilot - low sun, high glare, faded lane markings and no cars around for the autopilot to help lock its position.

what you seem to be complaining about is hardware failings, not software. the number of tesla "learning" the roads in your area isn't relevant to any shortcomings of the hardware. "blinded" cameras, poor road markings are not something that "learning" would correct. remember, it's only a beta

 

[brkaus]

Per the U.S. Geological Survey there would be lots of wondering with gps.

Commercial grade handheld GPS units are able to obtain coordinates with a horizontal accuracy of approximately 3 meters if the unit can receive a wide area augmentation system (WAAS) signal; otherwise, the accuracy is approximately 10 meters. This type of GPS handheld unit provides elevation data with poor accuracy.

 

[ShotgunF15E]

Haven't seen discussion on this yet - but I'm just wondering if this is true.

Here in Southern California I expect we have higher Model S density than any other city on the planet with perhaps the exception of the Bay area.

...

If these high definition maps Tesla claims the fleet is building, and the vaunted fleet learning are both in fact actually working - I would have expected that the Los Angeles freeways would be almost 100% bug free by now, even in the glare conditions - because in theory the fleet of local Teslas should have mapped out all the lanes by now.

...

Clearly the problem is not completely solved yet 6 months later despite thousands of local Model S's tracing the same routes day after day on our freeways.

Thoughts?

I think your thoughts are idealistic and unfortunately don't fully understand the limitations of the autopilot HARDWARE in the MS. Driving over the same piece of highway a million times and you're not going to overcome a poorly marked road. The best you could hope to accomplish is that the autopilot, based on it's known GPS location, would know in advance it's approaching a section of road that is poorly marked and coast through it and warn you in advance to take over the controls. It CAN'T learn, with it's current hardware, to magically lane keep, without proper markings.

I've driven the same piece of road every day with the autopilot on and it behaves the exact same way each day. I know why and the system works as advertised and within it's limitations. Realize GPS is not all that great for real time navigation. At highway speeds, it may not even be able to know which road you're on much less what lane.

Also, let's keep it in perspective....it's light years ahead of anything else out there on the market that I'm aware of and I never get tired of people looking over at me in the cars and I'm on autopilot....#Teslagrin

 

[sandpiper]

Per the U.S. Geological Survey there would be lots of wondering with gps.

Commercial grade handheld GPS units are able to obtain coordinates with a horizontal accuracy of approximately 3 meters if the unit can receive a wide area augmentation system (WAAS) signal; otherwise, the accuracy is approximately 10 meters. This type of GPS handheld unit provides elevation data with poor accuracy.

This. The current GPS implementation cannot reliably be used for guidance. It's too imprecise. I believe another upgrade is in process that will improve it to within cm, but I don't know when.

 

[calisnow]

what you seem to be complaining about is hardware failings, not software. the number of tesla "learning" the roads in your area isn't relevant to any shortcomings of the hardware.

Commercial grade handheld GPS units are able to obtain coordinates with a horizontal accuracy of approximately 3 meters...

I think your thoughts are idealistic and unfortunately don't fully understand the limitations of the autopilot HARDWARE in the MS. Driving over the same piece of highway a million times and you're not going to overcome a poorly marked road...

Actually Shotgun, kort677 and brkaus - the high precision GPS maps that Tesla and Mobileye (Mobileye's crowdsource project being entirely separate from Tesla's own efforts, to be clear - this is according to Mobileye's CEO) claim to have in development (multiple articles in multiple sources over the last six months) are in fact supposedly going to help overcome poorly marked roads. I'm not the one making this up - it's been in the press and Elon Musk has said it himself. Is it only hype? Maybe - but it's not hype that started with me. The idea is that the car will know where it is within inches - not ten feet - and thus the car will have one more method of precisely determining its position - even in the absence of high quality lane markings. Either Musk is full of crap or current GPS hardware is in fact capable of more accuracy than 3 meters (maybe it uses a combination of GPS and dead reckoning from its last highly confident position with good lane markings, wheel input etc).

It *is* supposed to be something coming in the current generation of hardware - I'm not the one who made this claim.

The question I was asking in my post was *assuming* that we are all on the same page with what Musk has said the fleet is doing - building high precision GPS maps that can tell the car which lane it is in - and that the way those maps are being built is by thousands of customer-driven Teslas tracing the routes over and over as they go about their daily driving. Therefore, if those high precision maps are being built it would seem to make sense that in situations with poor lane markings the GPS precision maps would help compensate for the lower precision of the sensor input to keep the car confident about its path.

Listen to Elon in this call - start the video at the 7 minute mark where he talks about the 405 - it's fascinating.

 

[calisnow]

...Driving over the same piece of highway a million times and you're not going to overcome a poorly marked road.... It CAN'T learn, with it's current hardware, to magically lane keep, without proper markings. ...
I've driven the same piece of road every day with the autopilot on and it behaves the exact same way each day.

Again - Musk seems to claim otherwise - listen to the Youtube link the post above, starting 6:55 - Musk talks specifically about building maps for the 405 with poor lane markings by driving Teslas over and over - it's quite fascinating. When talking about how they solved the 405 - at the 9:00 minute mark Musk claimed that after their test drivers were done driving each lane of the poorly marked 405 "...now the car will hold its position in the lanes correctly...it will actually do better than a person." And Musk says this is being accomplished via GPS lateral guidance.

Which is why I asked the question - because like you, Shotgun, I don't see my Tesla improving at this point in these poorly marked sections - although it *has* improved a lot from 7.0 on poorly marked lanes (Tesla told us themselves in the release notes and it seems to be true) - what I am wondering is the improvement traveling poorly marked lanes in 7.1 due to some build-out of these high precision maps, or some additional neural network training for visual cues?

And have we seen all the improvement we will see on lane precision with the current hardware suite - or is there more improvement yet to come with additional neural network training and/or the deployment of the high precision maps to the fleet as an additional factor in the car's decision making on whether or not it is "in the lane"?

Has Tesla perhaps deployed the high precision map data to limited areas around major metros - but not others? Or has Tesla not deployed it at all? AFAIK they've been silent on this so far.

As for the maps, see this quote from Teslarati and then look at the picture Musk posted as an example of areas that Tesla customers have already mapped with high precision in the SF Bay area:

"Musk focused on a section of Interstate 405 in California, a heavily traveled highway with lanes that are often poorly marked. Using information provided by other Tesla drivers who have traveled this specific section of road, Tesla Autopilot can still function with confidence, even in the absence of clear lane markings."

Tesla Building Next Gen Maps through its Autopilot Drivers

 

[kort677]

what musk says and what is in the cars today are two different things.

 

[donv]

GPS is not the whole story. Clearly it's not, or autopilot (and navigation) wouldn't work in downtown areas with lots of tall buildings. There are plenty of ways to augment GPS to determine position, the most obvious being to use the very accurate track and speed information available in a car (I hate to call this dead reckoning, since it's much more accurate than that term implies-- more like a simplified version of inertial navigation). Furthermore, the machine vision can also add data into that mix.

 

[ShotgunF15E]

View attachment 173482





Actually Shotgun, kort677 and brkaus - the high precision GPS maps that Tesla and Mobileye (Mobileye's crowdsource project being entirely separate from Tesla's own efforts, to be clear - this is according to Mobileye's CEO) claim to have in development (multiple articles in multiple sources over the last six months) are in fact supposedly going to help overcome poorly marked roads. I'm not the one making this up - it's been in the press and Elon Musk has said it himself. Is it only hype? Maybe - but it's not hype that started with me. The idea is that the car will know where it is within inches - not ten feet - and thus the car will have one more method of precisely determining its position - even in the absence of high quality lane markings. Either Musk is full of crap or current GPS hardware is in fact capable of more accuracy than 3 meters (maybe it uses a combination of GPS and dead reckoning from its last highly confident position with good lane markings, wheel input etc).

It *is* supposed to be something coming in the current generation of hardware - I'm not the one who made this claim.

The question I was asking in my post was *assuming* that we are all on the same page with what Musk has said the fleet is doing - building high precision GPS maps that can tell the car which lane it is in - and that the way those maps are being built is by thousands of customer-driven Teslas tracing the routes over and over as they go about their daily driving. Therefore, if those high precision maps are being built it would seem to make sense that in situations with poor lane markings the GPS precision maps would help compensate for the lower precision of the sensor input to keep the car confident about its path.

Listen to Elon in this call - start the video at the 7 minute mark where he talks about the 405 - it's fascinating.

I've got a bridge to sell you....

 

[ShotgunF15E]

Again - Musk seems to claim otherwise - listen to the Youtube link the post above, starting 6:55 - Musk talks specifically about building maps for the 405 with poor lane markings by driving Teslas over and over - it's quite fascinating. When talking about how they solved the 405 - at the 9:00 minute mark Musk claimed that after their test drivers were done driving each lane of the poorly marked 405 "...now the car will hold its position in the lanes correctly...it will actually do better than a person." And Musk says this is being accomplished via GPS lateral guidance.

Which is why I asked the question - because like you, Shotgun, I don't see my Tesla improving at this point in these poorly marked sections - although it *has* improved a lot from 7.0 on poorly marked lanes (Tesla told us themselves in the release notes and it seems to be true) - what I am wondering is the improvement traveling poorly marked lanes in 7.1 due to some build-out of these high precision maps, or some additional neural network training for visual cues?

And have we seen all the improvement we will see on lane precision with the current hardware suite - or is there more improvement yet to come with additional neural network training and/or the deployment of the high precision maps to the fleet as an additional factor in the car's decision making on whether or not it is "in the lane"?

Has Tesla perhaps deployed the high precision map data to limited areas around major metros - but not others? Or has Tesla not deployed it at all? AFAIK they've been silent on this so far.

As for the maps, see this quote from Teslarati and then look at the picture Musk posted as an example of areas that Tesla customers have already mapped with high precision in the SF Bay area:

"Musk focused on a section of Interstate 405 in California, a heavily traveled highway with lanes that are often poorly marked. Using information provided by other Tesla drivers who have traveled this specific section of road, Tesla Autopilot can still function with confidence, even in the absence of clear lane markings."

Tesla Building Next Gen Maps through its Autopilot Drivers

View attachment 173494

Musk also said the car will do a 1/4 mile in 10.9

 

[CHG-ON]

This. The current GPS implementation cannot reliably be used for guidance. It's too imprecise. I believe another upgrade is in process that will improve it to within cm, but I don't know when.

Agreed.

 

[DFiveK2]

in addition to inertial data which is probably the most important to nav as its realtime and always available... I'm sure triangulation data and services are being considered via fixed cell towers and known wireless access points. The beauty of an always connected car.

GPS is too unreliable to make decisions like this.

 

[1101011]

Has Tesla perhaps deployed the high precision map data to limited areas around major metros - but not others?

Certainly this makes the most sense.

 

[GSP]

It is one thing to have a high precision map. It is a completely different thing to know where you are on that map.

GPS cannot tell the car if it is in the center of the lane, or to one side, or even in the next lane. If, for example, the camera is blinded by sun, it cannot see lane markings or landmarks on the map. Better lane markings and landmarks are needed. Start lobbing your local government now.

GSP

 

[calisnow]

It is one thing to have a high precision map. It is a completely different thing to know where you are on that map.

GPS cannot tell the car if it is in the center of the lane, or to one side, or even in the next lane. If, for example, the camera is blinded by sun, it cannot see lane markings or landmarks on the map. Better lane markings and landmarks are needed. Start lobbing your local government now.

GSP

Totally agree that better lane markings are needed - but Musk claims GPS is capable of solving the "which lane am i in" problem in cases of poor lane markings and has in fact done so on Los Angeles' 405 freeway - which is part of his daily commute to SpaceX from his home. Do you think he's lying or do you think there's a further technical explanation behind his claims that we haven't thought of yet?

.When talking about how they solved the 405 - at the 9:00 minute mark Musk claimed that after their test drivers were done driving each lane of the poorly marked 405 "...now the car will hold its position in the .anes correctly...it will actually do better than a person." And Musk says this is being accomplished via GPS lateral guidance.