FSD rewrite will go out on Oct 20 to limited beta

[Mardak]

"very detailed" map is a map that have many details that are not needed for straight navigation. These details are precisely located on the map. The details could be anything from pothole, traffic control devices, crosswalks/stop lines and other road features

In terms of history, OpenStreetMap has been tagging more information that needed for simple vehicle navigation (as maps are used for more than just driving). At least from this OSM wiki page on footways/crosswalks Tag:highway=footway - OpenStreetMap Wiki The article history shows an earliest edit from 2012.

I would think Tesla map data partially derived from OSM had all this data, and Autopilot team figured, "well, there's all this data that we readily have available for some areas…" This could be used in vehicle for driving and visualization as well as assisting in collecting data for training.

These "very detailed" maps have been around for a long time. Whereas 3D HD lidar maps are relatively new.

 

[Daniel in SD]

As for testing with "well trained drivers", what does that prove? FSD is supposed to work with regular drivers, not trained ones.

FSD is supposed to work with no driver at all!
The whole stated purpose of beta FSD is to train the system to operate autonomously twice as safely as a human driver. I believe that can only be done safely with carefully monitored and well-trained safety drivers. I think it's possible that Tesla's current approach is safe though. I'm skeptical that releasing it to everyone before it is safer than a human would be safe unless they have some way of ensuring that everyone is testing it safely (lifetime bans for abuse? I'm sure they'll come up with something).

 

[emmz0r]

Well, to be honest, we have access to sources that know, for instance: Karpathy knows.
This year alone, in 2 public engagements, he said, we do not use HD Maps. The problem is that we have "experts" here that say otherwise.

They might not be using centimeter level maps but they are using something more than normal maps for sure.
Otherwise roundabout wouldn't be displayed as a perfect circle way beyond what the camera could see etc.
There is nothing wrong in using them either. You wouldn't want the car to act like it saw something for the first time every single time and drive like a confused tourist.

 

[verygreen]

These "very detailed" maps have been around for a long time. Whereas 3D HD lidar maps are relatively new.

yup. 3D lidar scans are relatively new and very detailed maps existed for a relatively long time and were accuulating more and more details along the way.

 

[mikes_fsd]

They might not be using centimeter level maps but they are using something more than normal maps for sure.
Otherwise roundabout wouldn't be displayed as a perfect circle way beyond what the camera could see etc.

a roundabout on a map is a normal map since ~ 2015 (actually earlier).
Today, even the most basic of Google Maps has traffic lights noted at intersections. i.e. that is not HD mapping as far as autonomous driving is concerned. but "experts" on here say otherwise.

 

[emmz0r]

a roundabout on a map is a normal map since ~ 2015 (actually earlier).
Today, even the most basic of Google Maps has traffic lights noted at intersections. i.e. that is not HD mapping as far as autonomous driving is concerned. but "experts" on here say otherwise.

Ah I didn't know it had roundabout as vectors as well. Nice though. I also know that google maps have lanes. I agree it's not HD , and frankly I don't think it's needed since vision will do the rest.

 

[mikes_fsd]

I also know that google maps have lanes.

Most mapping solutions today have lanes (no these are not lane line markings recorded) these are "centerline" lane delineation - it started with highways and now has slowly grown to other road types.
Each lanes centerline is recorded in the vector tiles. For 4 lanes, you would have 4 lines marked on the center of the lane. Tesla gathered this data back in 2015 as well when they first changed out the mapping engine.

 

[EVNow]

yup. 3D lidar scans are relatively new and very detailed maps existed for a relatively long time and were accuulating more and more details along the way.

Aren't the 3D lidar scans, integrated with very detailed maps what the industry calls HD maps ?

 

[verygreen]

Aren't the 3D lidar scans, integrated with very detailed maps what the industry calls HD maps ?

no. mobileye calls their maps HD maps and there's no lidar #D scans. commai calls their maps hdmaps and they are not produced with lidar either.

mapbox calls their container hdmaps and Tesla uses their technology....

 

[MP3Mike]

Through an apartment parking lot:

And now going into a round-about with it no longer stopping first: (at least on this one.)

 

[GZDongles]

View attachment 602858

I mean, are we going to be re-defining what crash rate means as well??
Or is Tesla Autopilot supposed to prevent strokes and heart attacks now?

I think the data is encouraging, but the point he is making that autopilot usage could coincide with the least risky type of driving anyway, so it's not an apples to apples comparison. A true comparison would be autopilot vs. non autopilot on the same type of road (e.g. a non autopilot urban highway drive vs an autopilot urban highway drive). I won't claim that the driver assistance features don't reduce the total accident rate, because of course they do. However, the difference probably isn't as big as these statistics indicate. Moreover, they include a lot of really old cars with no safety features like emergency braking or blind spot warning that are available or standard on many cars across the price spectrum, so I'm sure the average new car has a lower accident rate than the national average as well.

I agree with @Daniel in SD that the statistics as presented don't actually tell us much.

 

[Mardak]

no. mobileye calls their maps HD maps and there's no lidar 3D scans. commai calls their maps hdmaps and they are not produced with lidar either.

Here's comma ai's blog HD Maps for the Masses

"The data in OSM or Google Maps was primarily optimized for assisting humans, resulting in an over simplified and abstract representation of the world. Self-driving cars have generated an entirely new set of requirements. HD maps used for self-driving need to be a complete centimeter-accurate depiction of the real world, containing all things relevant to road navigation such as lanes, curvature and road boundaries."

And HERE blog about Mobileye HD mapping HERE and Mobileye: crowd-sourced HD mapping for autonomous cars

"autonomous cars need to be able to precisely position themselves on the road, making high-definition mapping an imperative part of the autonomous-driving-puzzle"

I suppose the main difference between Tesla's map usage and non-lidar-generated HD maps from comma ai/Mobileye is the precision on lane lines for localization. Where Tesla's maps might have information about "this road is approximately here and has 4 lanes" vs HD maps have "this lane line is exactly here" x5.

 

[azred]

But seriously, even just taking Tesla fleet into account.
4.59 million miles between crashes on Tesla's with AP engaged.
1.79 million mile between crashes on Tesla's with no AP and no safety features engaged.
----
That still makes Autopilot 2.56 times safer than a Tesla without Autopilot.

I think I probably have at least 2.56 times accident-free drives thanks to autopilot. But that surely would not be true if I wasn’t watching both of my Teslas closely while on autopilot. The sudden decelerations near certain overpasses, the swerving around in freeway merge lanes and the desire to hunt for the correct lane when on slight curves that don’t have perfect lane markings are every day issues that could make that 2.56 go to zero. I can’t imagine trusting my cars to safely execute intersections anytime soon given it has been nearly four years of AP2 (upgraded recently to AP3) in much simpler driving situations.

Robotaxis in a year or so? What a joke!

 

[mikes_fsd]

I suppose the main difference between Tesla's map usage and non-lidar-generated HD maps from comma ai/Mobileye is the precision on lane lines for localization.

No, the main difference is what is your base foundation.
If your approach is map-first, you need:

"autonomous cars need to be able to precisely position themselves on the road, making high-definition mapping an imperative part of the autonomous-driving-puzzle"

If your approach is vision-first, you do not need those maps.

 

[mikes_fsd]

Robotaxis in a year or so? What a joke!

You're in Chandler AZ the holy land of robotaxi's why do you even own a Tesla at all, just hop on to one of those "true robotaxi's" and sell your Tesla.

 

[drtimhill]

I think I probably have at least 2.56 times accident-free drives thanks to autopilot. But that surely would not be true if I wasn’t watching both of my Teslas closely while on autopilot. The sudden decelerations near certain overpasses, the swerving around in freeway merge lanes and the desire to hunt for the correct lane when on slight curves that don’t have perfect lane markings are every day issues that could make that 2.56 go to zero. I can’t imagine trusting my cars to safely execute intersections anytime soon given it has been nearly four years of AP2 (upgraded recently to AP3) in much simpler driving situations

Its pretty clear even just watching the beta videos that Tesla have significantly advanced the visual discrimination of the AI system. Like it or not, Tesla have focused mostly on FSD for the last 18 months, so yes NoA has taken a back seat, but that was (imho) because FSD took priority, not because they could not improve NoA. At some point they will upgrade NoA to use this same NN stack, and I think we will see a vastly improved NoA experience at that time (and it will be completely integrated with City Streets), with many of these kinks worked out (that was, after all, the whole goal of the V11 re-write of the NN stack).

 

[Mardak]

No, the main difference is what is your base foundation.

Yup, totally agree that Tesla's overall approach doesn't need maps. I was just highlighting the difference between the precision available in Tesla's maps when they are available vs the precision from comma ai/Mobileye maps. Clearly there is Tesla map data used for at least navigation, and that same map data sometimes has additional metadata at precision insufficient for centimeter-accurate localization (as Autopilot doesn't need it).

 

[verygreen]

Here's comma ai's blog HD Maps for the Masses

"The data in OSM or Google Maps was primarily optimized for assisting humans, resulting in an over simplified and abstract representation of the world. Self-driving cars have generated an entirely new set of requirements. HD maps used for self-driving need to be a complete centimeter-accurate depiction of the real world, containing all things relevant to road navigation such as lanes, curvature and road boundaries."

And HERE blog about Mobileye HD mapping HERE and Mobileye: crowd-sourced HD mapping for autonomous cars

"autonomous cars need to be able to precisely position themselves on the road, making high-definition mapping an imperative part of the autonomous-driving-puzzle"

I suppose the main difference between Tesla's map usage and non-lidar-generated HD maps from comma ai/Mobileye is the precision on lane lines for localization. Where Tesla's maps might have information about "this road is approximately here and has 4 lanes" vs HD maps have "this lane line is exactly here" x5.

That's one way at looking at it. Tesla roads are of course not "approximately" here, they are pretty exactly positioned.They record lanes, but the exact detail besides the count is not definitely known.

 

[Redbrick]

As I replied to someone else, in this context we need to proceed as even beta autopilot saves lives
The parallel is to vaccines - we require guinea pig humans to be infected & tested on to ultimately save lives

Uh...okay...you go first...tell let us how that works out for you...(...perhaps the wacky-est idea I've read on this forum..wait, is this reddit?)

 

[Soda Popinski]

no. mobileye calls their maps HD maps and there's no lidar #D scans. commai calls their maps hdmaps and they are not produced with lidar either.

mapbox calls their container hdmaps and Tesla uses their technology....

I don't know why I'm wading into this semantics argument, but doesn't MobilEye generate their HD maps with VIDAR? To me that's functionally equivalent to LIDAR since it gets cm level accuracy.