Welcome to Tesla Motors Club
Discuss Tesla's Model S, Model 3, Model X, Model Y, Cybertruck, Roadster and More.
Register
This site may earn commission on affiliate links.
The next big milestone for FSD is 11. It is a significant upgrade and fundamental changes to several parts of the FSD stack including totally new way to train the perception NN.

From AI day and Lex Fridman interview we have a good sense of what might be included.

- Object permanence both temporal and spatial
- Moving from “bag of points” to objects in NN
- Creating a 3D vector representation of the environment all in NN
- Planner optimization using NN / Monte Carlo Tree Search (MCTS)
- Change from processed images to “photon count” / raw image
- Change from single image perception to surround video
- Merging of city, highway and parking lot stacks a.k.a. Single Stack

Lex Fridman Interview of Elon. Starting with FSD related topics.


Here is a detailed explanation of Beta 11 in "layman's language" by James Douma, interview done after Lex Podcast.


Here is the AI Day explanation by in 4 parts.


screenshot-teslamotorsclub.com-2022.01.26-21_30_17.png


Here is a useful blog post asking a few questions to Tesla about AI day. The useful part comes in comparison of Tesla's methods with Waymo and others (detailed papers linked).

 
Last edited:
You also somewhat provided a different heuristic of chilling and accelerating slower, but I'm sure plenty of people including you would complain that FSD Beta takes too long to speed up at green lights in the common case.
I’d be fine because this would be context dependent. My complaint on green lights is when I am first in line. It is painfully slow - not sure it is slower than when there is a lead vehicle (might be the same I guess).
Would people be happy if FSD Beta proactively left space for adjacent vehicles signaling to get into your lane?
I think that would be wise. It is again context dependent what the right behavior is - if speeds are closely matched it would be fine to yield to a non-stale turn signal.
it would seem like a totally normal situation of someone accelerating up to 45mph and FSD Beta should do the same to be reasonably fast.
The issue here was maintaining the buffer, which should be done in all cases. Note that at 5-10mph buffers can be quite short because the energy is so low. But at 20mph stopping distance is 4-16x as long (distance, not time which is 2x-4x of course). So once you push above 15mph you should be opening up substantial space to the typical 3-4 seconds. The point here being that you can be fairly aggressive to follow at low speeds (the initial roll), but at higher speed you should naturally open the space to keep things relaxed.

This can all be pretty formulaic and optimized, as far as I can tell, even without contextual clues to help with predictions.

In this case the car actually slowed down promptly (though not really aggressively enough at all which led to a disengagement) when brake lights came on, but the buffer was simply too small to work with.
 
Last edited:
I guess right scroll wheel then touchscreen is the easiest way to toggle it:
Interesting! Suggests they are really committed to this. We’ll see. I predict a lot of complaining, even though issues will “only” arise around interchanges and exits (places where very specific and very late lane choices are often desired).
 
I'm really hoping NOA options will return for FSD such as requiring manual lane change confirmation by the driver.

Ouch is that gone. I always leave it on. I trust is more on the street than I do on the freeway. It tends to have issues with cars next to it more on the freeway than the road. Besides I prefer to control which lane I’m in, especially when it thinks it has to get all the way over because an offramp is 3 miles away. I don’t want to get in the crawl lane at 3 miles.
 
Ouch is that gone. I always leave it on. I trust is more on the street than I do on the freeway. It tends to have issues with cars next to it more on the freeway than the road. Besides I prefer to control which lane I’m in, especially when it thinks it has to get all the way over because an offramp is 3 miles away. I don’t want to get in the crawl lane at 3 miles.
I'm hoping it's just a case of getting FSD code to a more basic and stable place before tacking on various options that could have unexpected and unintended effects on functionality.
 
V11.3.1 isn't perfect of course possibly some improved decision making and steering/driving smoothness. The latter is hard to say unless one's butt is in the seat. Still some late lane changes, solid line crossing, changing lanes in intersections, near rear end contact, ..

This is a bit of a mess. You can see it still looks and sounds stressful with the driver almost willing FSD do the right thing at the right time. I'm not digging how FSDb isn't able to keep a safe gap to the lead vehicle. The lead vehicle broke hard in case it needed to stop in front of the tracks.

 
  • Like
Reactions: scottf200
V11.3.1 isn't perfect of course possibly some improved decision making and steering/driving smoothness. The latter is hard to say unless one's butt is in the seat. Still some late lane changes, solid line crossing, changing lanes in intersections, near rear end contact, ..

This is a bit of a mess. You can see it still looks and sounds stressful with the driver almost willing FSD do the right thing at the right time. I'm not digging how FSDb isn't able to keep a safe gap to the lead vehicle. The lead vehicle broke hard in case it needed to stop in front of the tracks.

It's something for the future addition list, but Beta will need to be able to handle intersection congestion at some point in the future. For now, I manually handle it as a driver, but we've got quite a few busy intersections that routinely fill up. Even though it's illegal to block an intersection in California, many people do it, and if Beta enters the intersection on a green light and then stops midway through the intersection because traffic stopped, it could end up blocking. Not sure how it's going to perceive that traffic ahead is slowing down and predict that it needs to stop at the line on a green light to avoid blocking the intersection.

Same for railroad crossings.
 
  • Like
Reactions: PVia and kabin
It's something for the future addition list, but Beta will need to be able to handle intersection congestion at some point in the future. For now, I manually handle it as a driver, but we've got quite a few busy intersections that routinely fill up. Even though it's illegal to block an intersection in California, many people do it, and if Beta enters the intersection on a green light and then stops midway through the intersection because traffic stopped, it could end up blocking. Not sure how it's going to perceive that traffic ahead is slowing down and predict that it needs to stop at the line on a green light to avoid blocking the intersection.

Same for railroad crossings.

Yep, it's too stressful for me and doesn't provide adequate reward. Little by little we should be able to use FSDb in more scenarios but this stuff is next level.
 
  • Like
Reactions: pilotSteve
It's a classic Tesla high risk, "hope we can get the software working soon!" kind of plan 😅. To be honest the more I think about it the more it seems like a bad plan.

They're trying to avoid the engineering work to get legacy AP working on HW4. That work is, in the long run, totally wasted time, so if this bet had worked we'd probably be talking about how this was a clever move.

But... it didn't work out (yet). And they look pretty dumb in hindsight. I imagine they probably had to make plans for production cutover to HW4 quite a while ago, at a point when it seemed reasonable that V11.x single stack would be ready in in time. People are not going to be happy getting a $90+k car without Autopilot.



I'm not sure. The official names for these features in the car are "Autosteer" (for basic AP), "Navigate on Autopilot", and "Full Self Driving (Beta)". They can make a build where the first two are implemented on top of what we talk about as FSD Beta V11.x "single stack" and just not change any of the names. That's why they've been testing single stack for over a year - to get to a point where they're confident swapping out the legacy AP code without an opt in beta program type thing.

If the AP stack isn't working today on HW4, I wonder if those cars have any active safety features working? I thought those were legal requirements. Maybe they work but not well?
Turns out just about everything I suggested here was wrong (happy to be wrong in this case and see V11.3.x go out to public testers, and hopefully me soon!)
 
Oops, took a left turn on a red light. Best viewed with 0.25x playback speed and volume off:

Oops! Did it see the green lights for the other road and think it could go? It's not a one way so a left on red would never be allowed here.

(Also, I would absolutely disengage there but each to their own... I guess that wasn't wildly dangerous since the car is basically treating it as an left turn at a stop sign)
 
I’m often in a straight lane when the car sees a green left allow and thinks it can go straight through the red line.

Likewise I see the opposite quite often with wanting to go on a red arrow because it sees the straight green even though I’m in the left lane.
 
Oops! Did it see the green lights for the other road and think it could go? It's not a one way so a left on red would never be allowed here.

(Also, I would absolutely disengage there but each to their own... I guess that wasn't wildly dangerous since the car is basically treating it as an left turn at a stop sign)
If you step through the turn frame by frame, you can see where one of the green signals on the right side that is for cross traffic gets visualized in blue just as the car makes the turn. An obvious error caused by the acute angle of the intersection.

You can clearly see two traffic lights in blue. The correct one is red, the incorrect one is green.
 
Not sure how it's going to perceive that traffic ahead is slowing down and predict that it needs to stop at the line on a green light to avoid blocking the intersection.

Same for railroad crossings.
One option is to keep more space to the car in front so it can slow down smoothly, avoid blocking intersection, railroad crossing etc.
Human vision is not just limited to the car in front but we can see through the windshield (debatable), see the traffic lot further due to curvatures on the road, see the brake lights of the car in front. Currently fsdb shows the turn signals and rear lights of the surrounding cars but it’s inaccurate.
 
That doesn't sound like the right fix. For me, the problem is that instead of following the correct map instructions to go straight through an intersection, my car will shift into the left-turn-only lane and then go straight through the intersection illegally from the left-turn-only lane. #4 above indicates that we'll just add a turn signal when we're making this illegal maneuver. Why not keep the car from moving to the left-turn-only lane and just stay in the correct lane and go straight through the intersection? Am I reading it wrong?
Yeah, my thought, too. “How about staying in the correct lane rather than being better at getting out of the wrong lane?”
 
  • Like
Reactions: pilotSteve