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

Elon: FSD Beta tweets

This site may earn commission on affiliate links.
#941
viet658 said:
Try this spreadsheet. Should tell you the miles you need to get the score you want.

docs.google.com

Tesla Safety Score - Tracking & Planning v10

Safety Score This is set to read only. To make your own sheet go to the file menu, 'Make a Copy' and open in Google Sheets (not in Excel). No share requests please. There are hidden columns with some of the working formulas so be careful what you change or move. <a href="https://teslamotorsclu...
docs.google.com docs.google.com
Click to expand...

Thanks for posting my sheet. I've gotten great feedback indicating it has helped many folks.
Spreadsheet link:
docs.google.com

Tesla Safety Score - Tracking & Planning v10

Safety Score This is set to read only. To make your own sheet go to the file menu, 'Make a Copy' and open in Google Sheets (not in Excel). No share requests please. There are hidden columns with some of the working formulas so be careful what you change or move. <a href="https://teslamotorsclu...
docs.google.com docs.google.com
1634307708008.png
 
  • Like
Reactions: edseloh, boonedocks and viet658
#946
  • Like
Reactions: qdeathstar and dcesq06
#947
EVNow said:
Single stack is a distraction. What a stupid trivial reason to want single stack, too. I first want FSD beta to stop being jerky and stop the FCWs.
Click to expand...
Agreed but I believe Elon said that summons, etc. will greatly improve with single stack. To the extent that's true, then I view it as more than simply a distraction.

Clearly, getting rid of the jerkiness and FCWs (although I haven't had a FCW in several months) is a priority. But I view that as a long-term work-in-progress.
 
#949
dcesq06 said:
I believe Elon said that summons, etc. will greatly improve with single stack.
Click to expand...
I just did a bunch of testing in a large plaza parking lot with multiple sections/buildings, and I believe single stack with parking lot training data will improve driving behavior on odd city streets such as those with sharp angles and back-to-back intersections. The curbs can also be irregular in parking lots, so it'll help in general curb detection on city streets such as intersections that have a little bit jutting out for a safe place for pedestrians to wait (where currently 10.2 seems to have trouble predicting and visualizing very short and sudden changes to the curb).

A single stack should also bring general improved driving behavior to highway driving such as 360 video understanding of surroundings to smoothly change lanes with large trucks nearby, handling construction, and generally places that currently warn "unsupported NoA maneuver" such as high curvature and short interchanges.

Improved highway is nice, but generally fixing corner cases for city streets is the more important aspect of single stack.
 
  • Like
Reactions: JupiterMan, dcesq06 and powertoold
#950
Mardak said:
Improved highway is nice, but generally fixing corner cases for city streets is the more important aspect of single stack.
Click to expand...
I think it’s a myth that somethings get automagically fixed with single stack. This particular thing - corner cases for city driving - how would it get fixed when added with freeway NOA stack ?

Single stack will have lots of regressions that need reoptimization to solve.
 
Last edited:
  • Like
Reactions: Skione65, daktari, powertoold and 2 others
#951
EVNow said:
I think it’s a myth that somethings get automagically fixed with single stack. This particular thing - corner cases for city driving - how would it get fixed when added with freeway NOA stack ?

Single stack will have lots of regressions that need reoptimization to solve.
Click to expand...
Yeah I don’t really see the big deal. Especially if it regresses interstate autopilot
 
  • Like
Reactions: nvx1977
#952
EVNow said:
corner cases for city driving - how would it get fixed when added with freeway NOA stack ?
Click to expand...
Sure, it's not "free" but it seems like FSD Beta neural networks have been trained on mostly city streets data so far, and now that they're more confident they have the correct architecture (hopefully no more massive "rewrites"), Tesla has been applying the approach to more predictions. For example, they likely started 360 video with training intersections and moving objects, and they've been adding static objects like cones and traffic lights. So keep expanding that to include parking lots and more.

The point of a single stack is likely to avoid custom code / networks for handling parking lots, i.e. the same neural network makes predictions for city streets and summon. The reason why FSD Beta somewhat works in parking lots already is that things look a bit like driving on regular streets. So similarly getting summon / parking lot working well on the single stack should improve situations where a city street looks somewhat like a parking lot.

Yes there are likely to be regressions where a parking lot behavior might not be appropriate on city streets, but that's the point of automated testing and fleet testing to then collect more training data (or worst case realize another rewrite is necessary to understand the differences).
 
  • Like
Reactions: powertoold and pilotSteve
#953
Mardak said:
The point of a single stack is likely to avoid custom code / networks for handling parking lots, i.e. the same neural network makes predictions for city streets and summon.
Click to expand...
So, when you say single stack - what you are saying is add parking lot to current fsd beta ? That takes a lot of effort. So will adding highways.

This is the basic idea of single stack. This is all just my understanding.

Currently, the video / images are fed into two separate CNN. The old one which is used by summon, freeway NOA etc. New one used by city FSD beta. So, the CNN separately classify the images and identify various objects. A single stack would merge these two CNNs to form a single CNN. The CNN will be fed video/images which will identify all the objects - and the planners for city & highway can use this perception to plan and do the actual driving. This will save a lot of cycles on the computer and be more efficient.

But to accomplish this - essentially they have train the new network with all the highway training set as well (assuming the new CNN is similar to the one used for FSD Beta) - then the entire network needs to be optimized. As you can see this is a lot of work. That is why I've been skeptical of a single stack anytime soon - and I think we'll probably see a lot of regressions and wonder why we wanted single stack in the first place :oops:
 
#954
EVNow said:
A single stack would merge these two CNNs to form a single CNN
Click to expand...
Ah, yeah I believe single stack is more of throwing out the old network as the current FSD Beta network will expand in capabilities. Similar to how the neural network is taking over various "Software 1.0" responsibility, it can take over "old Software 2.0" responsibility -- namely parking lot and highway driving, which in some sense were "incremental hacks" being replaced by a "doing it right" solution. Yes this happens to also free up some compute in not needing to run the old network.

FSD Beta already drives on interstate-speed (posted 65mph) "city street" highways that can have traffic lights and intersections, so it doesn't seem that big of a stretch to officially have it handle "easier" controlled-access highways like the interstates. There are likely things that are special to interstates but also similar enough to other highways that including interstate training data will improve the FSD Beta network overall.

I totally agree this isn't a trivial task, but at AI Day, Tesla said it took them less than 3 months to collect data to replace radar. It's been 3 months since Elon Musk hinted at "one stack to rule them all," and clearly it's taken longer than he hoped for.
https://twitter.com/x/status/1415782123298902019
 
  • Like
Reactions: powertoold
#955
Mardak said:
I totally agree this isn't a trivial task, but at AI Day, Tesla said it took them less than 3 months to collect data to replace radar.
Click to expand...
The results are less than stellar with a lot of regression around FCW.

Do we really want single stack before they solve FCW issues ? Not pretty on the freeway ...

ps : I think it makes a lot of sense to first solve known FSD Beta issues rather than take on more features, just so people can get pretty visuals on highways.

pps : I feel people are attracted "single stack" because it sounds very nerdy, not for any practical benefits :D
 
  • Like
Reactions: Skione65 and ahoen117
#960
powertoold said:
Once they can validate that there are no or insignificant regressions with single stack. They'll deploy it.
Click to expand...

jabloomf1230 said:
The highway NOA works great. Not perfect, but it does get you from point A to point B. Will single stack break that? Stay tuned.
Click to expand...

Yeah it would be good for them to get rid of the current regressions in highway NOA with FSD Beta before they move forward. Need to get on par with radar, have speed limit max set to 90mph, etc. And maybe learn to anticipate traffic a bit. Being able to look more than 100-200 meters ahead would be a nice touch.
 
  • Like
Reactions: Cheburashka and fo_Sean
You must log in or register to reply here.

Similar threads

B
FSD 11.4.9, First time use, subscribed to get v12 Beta
Replies
12
Views
711
AI, Autopilot, & Autonomous/FSD
Tethlaaaa
Tethlaaaa
Joey D
  • Article
FSD 12 perspective from a previous skeptic
Replies
76
Views
4K
AI, Autopilot, & Autonomous/FSD
Saxman
Saxman
D
Path to FSD Beta wide release by end of year
Replies
10
Views
2K
AI, Autopilot, & Autonomous/FSD
BBTX
B
T
FSD Beta 10.5 - will autopilot on freeway suffer?
Replies
12
Views
2K
AI, Autopilot, & Autonomous/FSD
Goose66
G
TMC Staff
  • Article
Blog FSD Beta 10 Rolls Out to Testers
Replies
30
Views
6K
AI, Autopilot, & Autonomous/FSD
USCGDAV
USCGDAV
Top
Back
Blog
New
Forum list
Marketplace
Menu