Tesla Autopilot maps

[lunitiks]

 

[scottf200]

I know someone with an AP1 car and this east to west road is 70 MPH earlier on this road per signage. He indicated that the Tesla would just take this corner at 70 MPH if he let it. I looked up the tile data and *clearly* all the data about this section of road shows the speed that should be driving and are is *way* less than 70 MPH. How could Tesla not use this data to slow down instead of assuming 70MPH? Clear the majority of the traffic is making this 90 degree corner.

Here is some of the text data about the section of road that is 869 feet long right before that corner. You can see they believe the speed limit is 70 but samples show/tell them the speed is way less in the real world.

Position: 0.7694 (265.10m/344.56m) (869.76ft/1130.44ft)
Location: 47.64029200, -111.92254389
Local Speed: 30.88mph (49.70km/h) (13.8066m/s)
Local Accel.: -4.0153
Next Id: 60301102
Trust: 0
Width: 6.668m (21.88ft)
Ramp Type: -
Speed Limit: 70.00mph (112.65km/h) (31.2928m/s)
Autosteer Restricted: Yes
Autopark Enabled: No
Radar Braking Enabled: No
Pedal Misapp. Mitigation Enabled: No
Unknown Speed1: 42.55mph (68.48km/h) (19.0216m/s)
Unknown Speed2: 50.50mph (81.27km/h) (22.5755m/s)
Unknown Speed3: 32.35mph (52.06km/h) (14.4617m/s)
Width1 (Left?): 2.766m (9.08ft)
Width2 (Right?): 3.901m (12.80ft)
F13 (Confidence?): 6%
F18 (Unknown): No
F24 (Highway?): No
F26 (Highway?): No
F27 (Unknown): 2
Local Speeds (7):
------------------------------------
1:
Speed: 55.83mph (89.85km/h) (24.9574m/s)
Accel.: -0.464
2:
Speed: 52.72mph (84.84km/h) (23.5679m/s)
Accel.: -1.508
3:
Speed: 48.11mph (77.42km/h) (21.5051m/s)
Accel.: -2.106
4:
Speed: 43.38mph (69.82km/h) (19.3939m/s)
Accel.: -2.365
5:
Speed: 36.92mph (59.41km/h) (16.5039m/s)
Accel.: -2.709
6:
Speed: 27.13mph (43.66km/h) (12.1279m/s)
Accel.: -4.829
7:
Speed: 18.39mph (29.60km/h) (8.2222m/s)
Accel.: 2.793

 

[robertvg]

How could Tesla not use this data to slow down instead of assuming 70MPH? Clear the majority of the traffic is making this 90 degree corner.

Good question. I think AP1 cars will never use these 'local speed' as they have no reliable way to determine their exact distance to the corner: you need AP2 with their multiple camera's to create a stereoscopic image for that. Same reason why AP1 cars can already for a long time recognize traffic lights afaik but without distance they can't do anything with that information.
Still wondering what 'Unknown Speed 1/2/3' are.

 

[calisnow]

Good question. I think AP1 cars will never use these 'local speed' as they have no reliable way to determine their exact distance to the corner: you need AP2 with their multiple camera's to create a stereoscopic image for that. Same reason why AP1 cars can already for a long time recognize traffic lights afaik but without distance they can't do anything with that information.
Still wondering what 'Unknown Speed 1/2/3' are.

You can measure distance with only one camera using frame by range comparison of the relative position and size of objects.

 

[AnxietyRanger]

I think AP1 and traffic lights is more about the limited FoV. EyeQ3 can process the lights, but with a single, fairly narrow camera seeing enough when stopped at lights can simply be impossible.

I speculate this is what the "AP 1.5" dual camera setup in early Model Xs was supposed to fix, amongst other things, with a second wider front camera.

 

[robertvg]

You can measure distance with only one camera using frame by range comparison of the relative position and size of objects.

Interesting.
Still I doubt how much AP1 cars will benefit from this data, how much 'self-driving' will improve from the current official 'highway only' scope.

I'm concerned about some obvious mistakes I have found in these maps: incorrect maximum speed limits, which makes the current cars suddenly brake at specific splines. Which makes me wonder: -how is the data is gathered, -how did these mistakes got into the maps, -most important: how can we get them corrected ?

 

[scottf200]

Good question. I think AP1 cars will never use these 'local speed' as they have no reliable way to determine their exact distance to the corner: you need AP2 with their multiple camera's to create a stereoscopic image for that. Same reason why AP1 cars can already for a long time recognize traffic lights afaik but without distance they can't do anything with that information.
Still wondering what 'Unknown Speed 1/2/3' are.

Wait a second ... what am I missing. Those tiles have GPS coordinates in them so they know where (coord) the corner is (and where people are slowing to 10 MPH) ... they don't need the camera at all. They could do it in a MT blizzard

 

[DamianXVI]

There is definitely something going on around the maps at the moment, because the maps from "dev" source in the US got updated again recently (as in contrast to the more than 5 months of silence in version string previously). This time it looks like the maps in the whole US got updated, not just in California. New version string for them is "na20171216: 2017-12-16 06:03".

It looks similar to the previous update in California, although this time there are a bit less splines. But it does not look like simple adding or removing of splines, rather like if the whole map was fully regenerated (with new IDs and some details).

A more notable change is in the way the splines creating ramps are marked. Previously it was very simple, where ramps usually did not fork into multiple branches, and where a single ramp could not change from off-ramp to on-ramp on an intersection of highways. This time the ramps labeling is more complex. Not always perfect, but definitely more advanced.

Also, the "stage" source from Tesla HQ area contains another intermediate version: "ca20171214_old: 2017-12-14 15:29". This adds up to three tile versions I saw in December.

To be clear - this is an update in the development state of the maps, not affecting the behavior of AP in the public.

I looked up the tile data and *clearly* all the data about this section of road shows the speed that should be driving and are is *way* less than 70 MPH. How could Tesla not use this data to slow down instead of assuming 70MPH? Clear the majority of the traffic is making this 90 degree corner.

We don't really see a whole picture about it, just a couple examples from the fragments of the map, and we are not understanding how these data are collected or generated. Maybe when testing it with bigger sets, there are many places when it would not work, maybe it is not ready yet in many places, maybe there are other, more general method in works.

I think AP1 cars will never use these 'local speed' as they have no reliable way to determine their exact distance to the corner: you need AP2 with their multiple camera's to create a stereoscopic image for that.

I don't think that the camera setup is an issue in determining the position or distance. For example, both generations of Autopilot are now able to determine if their position is in the radar "false echo" zones. More over, as far as we know, stereo-vision is most likely not used in AP, and AP2 looks not to be build with stereo-vision in mind. The distance to the objects and constructs ahead is most likely determined using perspective and reference to the ground plane (that is not fully compatible with kangaroos). Triangulating a very precise position around known point of interest is something that MobilEye CEO was explaining when talking about "high precision maps". Those tiles seem not to be "high precision maps", at least not yet.

 

[verygreen]

More over, as far as we know, stereo-vision is most likely not used in AP, and AP2 looks not to be build with stereo-vision in mind

Actually somewha topposite. What we know about AP2 seems to imply it does use stereo vision between the main and narrow cameras.

The high precision maps is a totally different thing. They could be used from ape computer itself and the format is different that these adas tiles, but the availability is minuscle. They only cover bay area, bits of Los Angeles and Edgewood (also in Bay area, where Elon has a house it seems?).... hm, Actually, it seems they uploaded whole Calfornia yesterday, dated Dec 15th and some other stuff that's not totally obvious right now but also big.

 

[BigD0g]

Actually somewha topposite. What we know about AP2 seems to imply it does use stereo vision between the main and narrow cameras.

The high precision maps is a totally different thing. They could be used from ape computer itself and the format is different that these adas tiles, but the availability is minuscle. They only cover bay area, bits of Los Angeles and Edgewood (also in Bay area, where Elon has a house it seems?).... hm, Actually, it seems they uploaded whole Calfornia yesterday, dated Jan 15th and some other stuff that's not totally obvious right now but also big.

He’s got 5 or 6 homes in bel air.

 

[DamianXVI]

Actually somewha topposite. What we know about AP2 seems to imply it does use stereo vision between the main and narrow cameras.

Thanks for correction. I saw the discussion about stereo vision between main and narrow cameras, but I thought it was only about theoretical possibilities, not actual implementation.

 

[scottf200]

Actually somewhat opposite. What we know about AP2 seems to imply it does use stereo vision between the main and narrow cameras.

Did you guys see KmanAuto's recent video where he was covering the cameras with tape and testing autopilot? It seems he could use one front camera at a time and still enable autopilot. Meaning they work as some level of redundancy.

Thread post that shows youtube. Tip click on video and then the youtube icon in the lower right to actually watch on youtube.com. Use the gear icon to watch at x1.5 speed:
Kman's Tesla Videos

 

[lunitiks]

What we know about AP2 seems to imply it does use stereo vision between the main and narrow cameras.

Please refresh our memories @verygreen. I can't seem to remember any references to stereo vision, like @DamianXVI I've always thought this was a mere (theoretical) possibility given the separation between the two cameras. Have you seen anything concrete in the code implying stereo vision?

 

[daktari]

Can we now expect them to update maps on a trans America route first?

 

[lunitiks]

Major AP update with V8.0

TESLA UPDATES: Elon Musk: Autopilot's v8.0 will be released in two weeks to Tesla vehicles

 

[J1mbo]

Can confirm the EU maps are still woefully behind.

Tile: gcpsp-dev
Bounds: 0.00000000, 0.00000000 - 0.00000000, 0.00000000
Version String: eu20170306: 2017-06-30 10:06
Number of Sub-Tiles: 56
Number of Splines: 1004

Tile: gcpsp-live
Bounds: 0.00000000, 0.00000000 - 0.00000000, 0.00000000
Version String: eu20170306: 2017-07-02 14:45
Number of Sub-Tiles: 56
Number of Splines: 1004

You know you are not a priority when the DEV data is older than the LIVE data

 

[avoigt]

View attachment 268092

just deleted the post. Its old news! My bad.

 

[robertvg]

You know you are not a priority when the DEV data is older than the LIVE data

Well at least you have tiles from July, in NL they date from March 2017

 

[AZ Desert Driver]

Is there any discussion on better route planning? My local routes pick roads I know have faster paths, less construction - and the route planned is sometimes just weird. Google on my Mac picks better routes. Not talking about "aBetterRoutePlanner" that optimizes toward superchargers. Just a simple comparison from Tesla to Google.
Do fancy new base maps come with fancy new routes?

 

[verygreen]

Please refresh our memories @verygreen. I can't seem to remember any references to stereo vision, like @DamianXVI I've always thought this was a mere (theoretical) possibility given the separation between the two cameras. Have you seen anything concrete in the code implying stereo vision?

@jimmy_d laid it out here: HW2.5 capabilities