FSD rewrite will go out on Oct 20 to limited beta

[Daniel in SD]

I love how someone disagreed with this post. Whoever they are, good luck waiting for anything that is "perfect"

You can click to see who it was. It was me.
First, Tesla has not presented that evidence needed to prove that Autopilot saves lives. We do not know the answer to that question.
Second, this is in no way analogous to vaccines. There is no reason to risk lives to test beta FSD. It is possible to test it safely, you just need a very alert and well trained drivers. Hopefully Tesla is monitoring the people testing it carefully to make sure that they are not increasing the risk of accidents.

 

[mikes_fsd]

First, Tesla has not presented that evidence needed to prove that Autopilot saves lives. We do not know the answer to that question.

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?

 

[helvio]

is Tesla Autopilot supposed to prevent strokes and heart attacks now?

Yes, please.

 

[mikes_fsd]

Yes, please.

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.

 

[Daniel in SD]

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?

We don't know if people are using Autopilot at the same rates under all conditions. Do people use Autopilot more or less when it is snowing? Do they use it more or less on the highway? Do they use it more or less at night?
The accident rate is higher when your headlights are on. Should you keep them off all the time? It's not so simple.
Also, Tesla doesn't even define what they consider a crash! (I assume it's airbag deployment but I have no idea)

 

[helvio]

We don't know if people are using Autopilot at the same rates under all conditions. Do people use Autopilot more or less when it is snowing? Do they use it more or less on the highway? Do they use it more or less at night?
The accident rate is higher when your headlights are on. Should you keep them off all the time? It's not so simple.
Also, Tesla doesn't even define what they consider a crash! (I assume it's airbag deployment but I have no idea)

That is not how statistics work lol

 

[Daniel in SD]

That is not how statistics work lol

Exactly! That's my point.
Sorry to sidetrack the beta FSD discussion. I should have just stuck with arguing against the idea that testing beta FSD requires human sacrifice.

 

[EVNow]

You can already set the mirrors to fold based on location. Mine fold the same time it fires my garage door opener, ends up being about a car length away from the edge of my driveway, most times.

Didn't know that. I'll have to check it out. But the idea of "tight parking mode" stands. The driving needs are obviously very different from normal road driving.

 

[helvio]

Exactly! That's my point.

I can partially agree with you on your Autopilot claim, there's more to be seen. But no AP and Safety Features on/off, can you agree that's out of question, meaning, improvement is clear?

 

[Daniel in SD]

I can partially agree with you on your Autopilot claim, there's more to be seen. But no AP and Safety Features on/off, can you agree that's out of question, meaning, improvement is clear?

Well I think automatic emergency braking, emergency lane departures assist, forward collision alert, etc. probably do reduce accidents significantly. On the other hand there could also be geographic effects as older Teslas may be concentrated in urban areas that have more accidents per mile. My point is, who knows? Tesla doesn't provide the data to determine that.

 

[MP3Mike]

Also, Tesla doesn't even define what they consider a crash! (I assume it's airbag deployment but I have no idea)

Actually they do:

Methodology:
We collect the exact amount of miles traveled by each vehicle with Autopilot active or in manual driving, and do so without identifying specific vehicles to protect privacy. We also receive a crash alert anytime there is a crash that is correlated to the exact vehicle state at the time. This is not from a sampled data set, but rather this is exact summations. To ensure our statistics are conservative, we count any crash in which Autopilot was deactivated within 5 seconds before a crash, and we count all crashes in which the crash alert indicated an airbag or other active restraint deployed. In practice, this correlates to nearly any crash at about 12 mph (20 kph) or above, depending on the crash forces generated. On the other hand, police-reported crashes from government databases are notoriously under-reported, by some estimates as much as 50%, in large part because most fender benders are not investigated. We also do not differentiate based on the type of crash or fault, and in fact, more than 35% of all Autopilot crashes occur when the Tesla vehicle is rear-ended by another vehicle. In this way, we are confident that the statistics we share unquestionably show the benefits of Autopilot.

 

[FloridaJohn]

So having watched a few of the City Streets beta videos, I think we are about to enter an interesting transition period:

First, its clear that 'City Streets' is running independently of the older NoA system, so Tesla are apparently running both the old "2D stack" (including the legacy NN) and the new "3D stack" (with the new NN that requires HW3). This is apparent by the hand-off between City Streets and NoA, and the lack of any changes in the behavior when the car is in the older NoA mode on freeways.

So, what happens when City Streets is more stable (later in 2021)? I'm guessing that at that point Tesla will re-work NoA into "NoA+" (my name), completely updating it to use the new 3D stack. This will allow the car to use the same sophisticated 3D stack for both City Streets and NoA that works seamlessly on all roads, and significantly enhance the smarts of NoA+ compared to NoA (better lane awareness, ramp negotiation etc etc).

Of course, NoA+ will requires HW3. So I'm guessing that at this point Tesla will branch the stacks/software. Those without FSD package will stay on the current NoA and HW2.x (which will become part of EAP in effect), while the combination of City Streets and NoA+ will become the FSD package.

Of course, they could package things differently (and probably will), but I'm still betting on the NoA+ re-write, as they will be anxious to retire the old 2D NN stack and move everything onto the new 3D stack (and thus focus development efforts only on the 3D stack).

(Note: 2D/3D names are not really accurate here, but you know what I mean .. the old non-integrated camera NN stack be the newer integrated NM stack.)

I think Smart Summon (and Reverse Summon) will also fall under that plan. Right now, only City Street driving is on the new NN; eventually all the automated driving will be running on the new NN.

 

[turnem]

I don't have the FSD Beta but on 2020.40.8, I've noticed this behavior too. My car stopped well short of the stop line and then lunged forward twice to reach the stop line.

I've seen this behavior as well. It will get your attention REAL QUICK! LOL!

I do have to say one thing though. I'm really glad that I bought my Tesla when I did so that I can clearly see the progression and then (hopefully) get to experience this huge leap forward into FSD Beta territory. Exciting times for sure!

 

[MP3Mike]

I personally think Tesla should make a much bigger/more visible reference to this statistic:

in fact, more than 35% of all Autopilot crashes occur when the Tesla vehicle is rear-ended by another vehicle.

 

[helvio]

I do have to say one thing though. I'm really glad that I bought my Tesla when I did so that I can clearly see the progression and then (hopefully) get to experience this huge leap forward into FSD Beta territory. Exciting times for sure!

That makes 2 of us, I got mine in late August and couldn't be happier with the timing!

 

[mhan00]

Actually they do:

are you expecting people on the internet to read before they render judgements? Lol.

 

[Daniel in SD]

are you expecting people on the internet to read before they render judgements? Lol.

They did not use to provide that information. And as you can see my assumption was actually right!

 

[verygreen]

Can you tell us more about the improvements that you see? Can you, for instance, give us two clips of video, one in a region with these maps and one in a region without, and then point out where the maps are assisting the vision?

not cilps, but see this thread for some examples of map-assisted display https://twitter.com/greentheonly/status/1320604581534015488

It's not 100% conclusive of course, but there's another subthread that dives into some actual map data:
https://twitter.com/greentheonly/status/1320959889582264321 (scroll up for the frames)

Using these you can form your own opinion.

Can you define what is a "very detailed map"? What is the source of these maps? Can we examine these maps? (If not "we" (the general public), can "you" (a gifted Tesla hacker)?) What is the source of these maps?

"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 to outlines of roadside objects and relevant supportign data to locate them (Tesla used radar-outlined "furniture in the past but stopped). The important part is all these extra details are described in very accurate manner (position wise, so no, waymo/google maps "there's a stopped car somewhere 0.1 miles from here" does not count) and are actually not needed for the basic task of navigation.

Tesla does not make their maps public, they use a fork of valhalla/mapbox that you can see an old version of in their github repo. I have a copy of the maps but I am not about to go dive into the binary formats like that at this time because I don't find it important. I can give you a dump of the map data (and I certainly did before when there were people interested in taking things apart, see the now abandoned ADAS tile maps thread for an example: Tesla Autopilot maps ) if you want to dive in and there are other ways to confirm certain data classes stored there anyway. They source their map data from openstreet maps and some other sources that are less clear (tomtom is specifically mentioned in metadata), they also augment them with what their fleet collects. We know they store stop signs/traffic lights on the map because it's outlined in the documentation (and you can independently verify behavior of mapped/not mapped ones by the amount of advance warning you get).

Finally I am not a "gifted tesla hacker" or a bunch of those other things people claim I am. I am mostly a random curious guy on the internet (and I am far removed from anything automotive or anything ML/NN) that can also publish my findings somewhat freely because what I find is not important enough for Tesla to make me sign NDAs (and there are people that they did silence with NDAs so we know they do it) but on the other hand I realize not everybody is in position to perform the kind of digging I do for a variety of reasons. Don't take my word for gospel or anything like that, I've been wrong before and I am going to be wrong in the future too. I try to clearly separate what I see as facts and what I think that means for the reason that you can then take the facts and perform your own thinking. If you come to different conclusions given the facts - that's all fine with me.

 

[Mardak]

If someone wants to take a closer look although this won't exactly prove or disprove map usage as Tesla pulls in multiple sources to generate their maps. Here's the OpenStreetMap data with crosswalk line data circled (some intersections use a 4-line square shown as a single circle while other intersections use single lines for each crosswalk at the intersection):

I've also highlighted in blue/green the route taken in this video:

There are crosswalks at all of the larger intersections but only some of them are mapped in OSM. Notably down 19th street in the middle, N/O/P streets have unmapped crosswalks as well as along P street. I didn't notice much difference in the visualizations.

 

[drtimhill]

You can click to see who it was. It was me.
First, Tesla has not presented that evidence needed to prove that Autopilot saves lives. We do not know the answer to that question.
Second, this is in no way analogous to vaccines. There is no reason to risk lives to test beta FSD. It is possible to test it safely, you just need a very alert and well trained drivers. Hopefully Tesla is monitoring the people testing it carefully to make sure that they are not increasing the risk of accidents.

Actually it is like testing a vaccine. Do you think Tesla did no testing before the beta release? Drugs go through various stages of testing, from labs, to animals (no ethical comment here, just stating), to limited human trials, to full human trials. Right now Tesla is at the equivalent of limited human trial (and yes, I'm sure they did not test earlier with animal drivers!).

As for saving lives, everyone (including Tesla, I am sure) hopes that a mature AP system will indeed save lives (and, as I have said before, human drivers are a pretty low bar to start with). But this is very different from drug testing, since the goal is diffuse; its much more like public health testing, which can only be done retrospectively.

As for testing with "well trained drivers", what does that prove? FSD is supposed to work with regular drivers, not trained ones. I dont see any way you can extrapolate from results with "well trained drivers" to what would happen with regular untrained drivers. So, ultimately, you have to test with untrained drivers, but you do so gradually so that you can back-off quickly if there are issues. Which is exactly what Tesla is doing.