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

Tesla Autopilot maps

This site may earn commission on affiliate links.
Tiles aren't being used yet for driving.

Are you really sure? How do we know that without having a clear look at tesla's driving control logic? My guess would have been that the tiles are being used but their quality is inconsistent. At least that would explain the very inconsistent performance of AP on local roads since 10.4. There are videos out there of AP on very windy roads both driving almost flawlessly and failing completely.

Especially this video right here is very curious. At the linked part in the video the road does a series of 3 90 degree turns, but AP only manages to take one of them in one direction (the driver drives through that stretch of road in both directions) while completely failing the other 5 maneuvers.

This is a bit weird because there are tons of videos out there of the recent AP acing way more difficult turns. So I did some digging and checked out the relevant ADAS tiles (c20d6, c20d7) and lo and behold, the quality of the map tiles at that location is very flaky. On that particular stretch of road neither lane has complete splines and only two of the turns have complete and sensible entries in the "local speeds" fields.

But now what's really interesting: the acceleration and deceleration maneuvers AP does correctly in the video match the data from the tiles. On the other hand, the parts of the video where AP is especially awful seem to be consistently on parts of the road that either have no or very incomplete data in the tiles.

Here's another popular 10.4 video. When looking at the tiles of that respective region (u2ede) it's clear that the data in them is a lot more complete with the whole stretch of road being covered, which in turn might explain why AP seems to do a much better job on a pretty difficult road.

Now granted this is only a sample size of two, but from these two examples alone it's pretty obvious that AP performance seems to fluctuate a lot and how well it does seems to be somewhat related to the quality of the ADAS tiles.
 
I mean, they are not using maps for driving, they use maps for localization and speeds but not driving, they are using the lanes and markings for driving through the NN, but the maps aren’t saying hey your on a 3 lane highway in the center lane with a 15 degree curve coming in 200 ft etc.
Ah, yeah, in this sense I am not sure they do have the lanes even, just general road curvature and such (which they do know actually). The new ap2 maps definitely have finer resolution down to individual lanes.
 
Ah, yeah, in this sense I am not sure they do have the lanes even, just general road curvature and such (which they do know actually). The new ap2 maps definitely have finer resolution down to individual lanes.

Yes agree, but they aren’t using it yet, just localization stuff which is how they know to slow for curves etc.

Three months maybe, definitely six
 
  • Love
Reactions: googleiscoul
Waze now has an option for carpool or HOV lane sticker, which in states like CA will give you unique directions that take advantage of (usually faster) carpool lane speeds, often with dramatic benefits to speed and arrival time. Hope Tesla adds this option, especially since we dominate those lanes.
 
I thought this comma ai 'windies' video was pretty good. Their model testing is using HD maps tho AFAIK.
comma ai on Twitter

Completely agree, I watch the scoppie videos every week, really good insights and such, other then the shilling I find it fascinating to see his approach and draw correlations back to Tesla's approach which seem very similar albeit with much different hardware at their disposal.
 
I don’t know about all the discussion on this thread, but I know the new Tesla Maps are far better than the old maps. How? Because they now show the golf course, cart paths, sand traps, AND water hazards as I drive by.

Now I’m not sure how this helps me...but it’s good to have the data if I need it.
 

Attachments

  • 11AD43F0-CD4F-40BD-A441-9AEA9C8A7A71.jpeg
    11AD43F0-CD4F-40BD-A441-9AEA9C8A7A71.jpeg
    737 KB · Views: 126
I don’t know about all the discussion on this thread, but I know the new Tesla Maps are far better than the old maps. How? Because they now show the golf course, cart paths, sand traps, AND water hazards as I drive by.

Now I’m not sure how this helps me...but it’s good to have the data if I need it.
I thought the 17" screen was still Google Maps with a Nav Overlay (or something like that).
 
Yes agree, but they aren’t using it yet, just localization stuff which is how they know to slow for curves etc.

Three months maybe, definitely six

I suspect those maps are still being used in Tesla's "fleet learning for radar" approach for stationary object detection. That is, they give greater confidence, when driving along a path in the map shown as blue, when a stationary object is detected ahead that seems to be in the travel path. With that greater confidence, they could give a forward collision warning sooner and/or initiate automatic braking earlier. In fact, if the object is detected at sufficient range, automatic braking could be initiated that is NOT "emergency". If the radar did not detect the truck ahead in the following video (go to the 6:10 timestamp) before it pulled over and stopped (it isn't shown in the instrument panel anyway), then it is evidence they are doing this already. The driver took control and drove around the stopped truck, but it appeared that the Tesla was in process of gradually slowing to a complete stop if the driver did nothing.


As a bonus, notice that the truck was only partly blocking the lane. As NHTSA pointed out in their final report about the Florida fatal accident, at that time every auto OEM they checked had optimized their AEB systems to protect against rear-end collisions, meaning striking the rear end of another vehicle stopped in the lane. A good question is how effective any AEB system is when dealing with a stopped vehicle only partly in the lane, or any obstacle presenting risk of a collision. Tests of automatic braking systems for those situations are not yet being conducted by any safety agencies, as far as I can determine.

Even so, those maps may not be useful for Autosteer, which is what you may have meant by "driving". Even slowing before curves ahead when on a path in the maps is operating of acceleration and braking rather than steering. So, if all that is true, then they are useful for forward collision warning, automatic emergency braking (even though they may be enabling non-emergency braking like in the video), and traffic aware cruise control. For help with Autosteer, it will probably take their high-definition maps and precise real-time location to provide help to the vision system for Autosteer.
 
Last edited:
  • Helpful
Reactions: croman
It was cool to see and relate to this old articles picture.

electrek
ap-debug-7.jpg


Amazing, sir. So the branches are called dev, live, stage.

If I look at stage, there's a veeeeery curious map tile name:
na20180212_localizer_optimal: 2018-02-12 15:12



It's newer than those December-ish live tiles we saw, and "localizer_optimal" sounds interesting too :D


EDIT: Paging @DamianXVI , @verygreen , @BigD0g ….. My local map tiles went from 500KB each to around 1000KB each, in the "stage" subdirectory. I'm gonna try the app tonight to see if we can view them, but just in case some tools need to be updated for this format, or if you can figure out how to take these tiles out for a spin… :D
 
Especially this video right here is very curious. At the linked part in the video the road does a series of 3 90 degree turns, but AP only manages to take one of them in one direction (the driver drives through that stretch of road in both directions) while completely failing the other 5 maneuvers.

That is my video :)

You are spot on though. I have been following this thread and that video accurately shows my experience with incomplete tiles. I drive that road daily and have been getting getting the same results. Hopefully it improves the more I drive it but I do not know how Tesla is improving these.
 
I mean, they are not using maps for driving, they use maps for localization and speeds but not driving, they are using the lanes and markings for driving through the NN, but the maps aren’t saying hey your on a 3 lane highway in the center lane with a 15 degree curve coming in 200 ft etc.

I don't know if there is a definitive answer.

There seems to be some major inconsistencies in observed behavior regarding whether it uses the ADAS tiles for lane-steering.

The recent accidents reinforces the idea that they're not being used for lane-steering. But, there have been videos of AP on while on roads with no visible lane marking (like snow on the ground). Or maybe I just wasn't paying attention enough to the videos, and it was following some other car.

There have also been reports of people experiencing worse lane-steering after an update, but only for a couple of weeks. As if the ap tiles cache was cleared. So I'm wondering if it uses the tiles as an momentary backup for the NN (cresting hills for example).

Where the car doesn't have enough accuracy with the GPS to do anything more with the tiles.

As an aside -

Arstechnica has a pretty good article on a new app that supposedly has 1.5m accuracy which is pretty impressive for cellphone hardware.

This app features super-accurate GPS, and I can’t figure out how it works
 
What's weird is that our AP1 and AP2 car shows different speed on some of the roads.
Different databases?

Probably the same database, but the AP1 cars read speed limits signs as the primary method and only fall back to the database when it can't. AP2 cars currently only use the database. (I would have hoped that they use the AP1 reading signs to improve the database, but it doesn't seem like they do.)