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:
Hmm… "We are almost at the point… that we can turn our attention to applying the code to highways" sounds like they haven't started or that it's not the top priority yet. There's definitely improvements to FSD Beta that would be quite nice for highway Autopilot:
That comment was kind of bizarre - even Omar loled at interventions being called rare.

It just sounds like Elon “buying time”. That “next month” on 11 isn’t happening … and Elon just wanted to say it by putting on as happy a face as he could.

At this rate, I wonder whether Tesla will introduce all important changes to 10.x and only keep “single stack” for 11. Nothing really prevents Tesla from putting all improvements into 10.x ….
 
Here’s a link: Elon: FSD Beta tweets

I posted this in that thread, too, but has anyone had FSD consistently navigate city intersections well? I’m not talking about stoplights but rather 4 way stops, 2 way stops, yields, etc? IME the car is so slow and unsure that the only way I’ll let it do it unaided is if there’s no traffic around. Otherwise I virtually always have to ’goose’ the accelerator. It seems like a little thing but is there’s more to it than it would seem and still means the car is failing.
My experience is similar to yours. It handles traffic lights and 4-way stops fairly well, but it's awfully slow to commit on 2-way and single stop signs turning onto country roads (especially the latter, often dangerously so). I'm actually quite surprised to see that that's apparently not the case for everybody - I had assumed this was due to the cameras not being able to see well enough. Maybe it has to do with mapping, or the lack thereof?
 
  • Like
Reactions: sleepydoc
My experience is similar to yours. It handles traffic lights and 4-way stops fairly well, but it's awfully slow to commit on 2-way and single stop signs turning onto country roads (especially the latter, often dangerously so). I'm actually quite surprised to see that that's apparently not the case for everybody - I had assumed this was due to the cameras not being able to see well enough. Maybe it has to do with mapping, or the lack thereof?
First of all - this has been discussed to death. Early releases had a lot of hesitation on unprotected right. It has become much better now.

As to fast country roads, there is not much discussion probably because most of us are in cities/suburbs and don't use country roads much.

Besides, in my county if there are rural intersections like those - they are either not busy or if busy, they have traffic controls. Afterall busy intersections like those would be dangerous for human drivers too ....

ps : Can you time it and tell me how many seconds it takes to turn on an unprotected right ? It used to take ~ 10 seconds to go from stop to right turn. Now it takes 5 seconds or less - humans driving properly will take 2 or 3 seconds. So we are talking about 2 extra seconds.
 
The best right turn I have is where there is No Stop sign leaving my neighborhood. FSD drives up very quickly to the edge of the crossing road (no creeping) and makes a decision to turn right or left immediately. Quicker than I would. First few times I disengaged until I learned FSD's behavior.

At the opposite end of this road about 1/4 mile away there is a Stop sign. FSD creeps to the crossing road and then takes several seconds to decide when to turn. The behavior couldn't be more different.
 
The best right turn I have is where there is No Stop sign leaving my neighborhood. FSD drives up very quickly to the edge of the crossing road (no creeping) and makes a decision to turn right or left immediately. Quicker than I would. First few times I disengaged until I learned FSD's behavior.

At the opposite end of this road about 1/4 mile away there is a Stop sign. FSD creeps to the crossing road and then takes several seconds to decide when to turn. The behavior couldn't be more different.
But, is the behavior at the first turn safe? I.e. if there is a car coming, does it wait ?

Of course, the other issue is right of way rules when there are no signs and whether there are significant differences among states.
 
Last edited:
But, is the behavior at the first turn safe? I.e. if there is a car coming, does it wait ?

Ofcourse, the other issue is right of way rules when there are no signs and whether there are significant differences among states.
Yes the first turn is safe 95% of the time. The remaining 5% IMO wouldn't result in an accident but would cause the approaching vehicle to brake. FSD should wait on those.

What I don't like is that FSD stops perpendicular to the road. A human would turn slightly left or right making the turn quicker and smoother and would reduce the chance of an accident. It also sometimes causes the car to cross the center line since there simply isn't enough turning space in the road which wouldn't be a problem If the car turned slightly when stopping. I suspect the behavior is so the front cameras have maximum visibility left and right. So many problems are caused by the lack of a forward camera. For example a 2nd camera in the front fender location
 
  • Informative
Reactions: EVNow
If the car turned slightly when stopping. I suspect the behavior is so the front cameras have maximum visibility left and right. So many problems are caused by the lack of a forward camera. For example a 2nd camera in the front fender location
Yes I've assumed the perpendicular positioning is to optimize the front-facing camera size view (something the B-pillar cameras cannot help with as they are too far back). My feeling has always been that the car should have cameras mounted roughly at headlight position angled to give forward and side views from as close to the front of the car as possible. Although this would probably end up scaring humans as the car would be able to see up/down the cross street way before a human driver :)
 
Yes I've assumed the perpendicular positioning is to optimize the front-facing camera size view (something the B-pillar cameras cannot help with as they are too far back). My feeling has always been that the car should have cameras mounted roughly at headlight position angled to give forward and side views from as close to the front of the car as possible. Although this would probably end up scaring humans as the car would be able to see up/down the cross street way before a human driver :)
If the side camera were mounted on the A pillar then their view would be similar to a driver leaning forward. They would also be closer to the front cameras which might be an advantage.
 
First of all - this has been discussed to death. Early releases had a lot of hesitation on unprotected right. It has become much better now.

As to fast country roads, there is not much discussion probably because most of us are in cities/suburbs and don't use country roads much.

Besides, in my county if there are rural intersections like those - they are either not busy or if busy, they have traffic controls. Afterall busy intersections like those would be dangerous for human drivers too ....

ps : Can you time it and tell me how many seconds it takes to turn on an unprotected right ? It used to take ~ 10 seconds to go from stop to right turn. Now it takes 5 seconds or less - humans driving properly will take 2 or 3 seconds. So we are talking about 2 extra seconds.
I timed three different instances of unprotected right turns onto country roads, and each one was about 4-6 seconds so you're correct on that aspect. However, my bigger issue isn't so much the time to execute but HOW it executes. By the time the car is ready to commit, it's already sticking out into the road. Visibility is often poor where I live (lots of winding, hilly roads). If FSD is sticking out and hesitating as described, and then a car suddenly appears barreling towards me at 50+ MPH, the car either slams the brakes and remains stuck partially in the road, or it hits warp speed and slams the accelerator (usually it's the former). The first scenario is unsafe; the second is quite uncomfortable. Both are nerve-wracking. And I can't help but feel that this is due to the car not being able to see well enough WITHOUT having to stick out into the road in such a manner. Not sure how that gets resolved without repositioning the cameras for a better view.

Don't get me wrong - FSD has improved quite a bit in a lot of areas, including this one. But Elon's recent comments about disengagements left me floored. We've still got a ways to go, and from my POV, there are some problems that can't be solved without drastic hardware changes that aren't likely to happen on any cars on the road at the moment. I think Tesla's going to nail this, eventually...but whatever version of "delivered" FSD my M3 eventually gets, I think it's going to be far inferior both to the initial vision, and to what newer Teslas with newer hardware are capable of. Maybe they'll continue giving free upgrades for those that purchased, but at some point I imagine they'll call it "complete" in that all features basically work, even if not all that well (or as well as newer cars with newer hardware).

Just my two cents, and I hope I'm wrong (it certainly wouldn't be the first time :) ).
 
By the time the car is ready to commit, it's already sticking out into the road.
This is my experience as well.

This morning there was no traffic so I just let the car do its thing. Turning from a side street (30 MPH with a stop sign) onto a county road (50 MPH, no stop sign) with clear lines of sight. The car stops then inches forward several times to the point that the nose is about 3 feet into the lane of traffic, then stops. Normally at this point there's a car coming so I just hit the accelerator to complete the turn and get out of the way. Today I happened to look down at the screen and I caught a glimpse of an alert that said something to the effect of "waiting for high speed cross traffic. Press accelerator to continue." First, the road was empty - there was no traffic at all. Second, if it can't continue without my intervention then it's a fail no matter how you cut it.

I'll try to repeat this and see if I can get a picture of the message.
 
This is my experience as well.

This morning there was no traffic so I just let the car do its thing. Turning from a side street (30 MPH with a stop sign) onto a county road (50 MPH, no stop sign) with clear lines of sight. The car stops then inches forward several times to the point that the nose is about 3 feet into the lane of traffic, then stops. Normally at this point there's a car coming so I just hit the accelerator to complete the turn and get out of the way. Today I happened to look down at the screen and I caught a glimpse of an alert that said something to the effect of "waiting for high speed cross traffic. Press accelerator to continue." First, the road was empty - there was no traffic at all. Second, if it can't continue without my intervention then it's a fail no matter how you cut it.


This happened at nearly all T-stops, even one with a light, in 10.8.x... 10.10 is better, in that it does eventually go on its own without driver intervention- but it still creeps into what would be oncoming traffic (if there were any) even with clear visibility and nobody coming either way before proceeding.
 
  • Like
Reactions: VanFriscia
And I can't help but feel that this is due to the car not being able to see well enough WITHOUT having to stick out into the road in such a manner. Not sure how that gets resolved without repositioning the cameras for a better view.
I think there has been a lot of discussion previously on this.

My take is - this is a planning issue. This is a small % of total cases and Tesla hasn't paid attention. We'll know if it gets better with more robust planning that is coming in 11 (or earlier) with NN based MCTS.
 
I'm unclear how planning fixes not being able to see around a corner where the B pillar cam is blocked?

Chuck Cook, among others, have done videos on this- showing the feed from the B-pillar position vs the feed you'd get with a side-cam near the front turn lights and it's a pretty dramatic difference at such intersections-- to where you can see just fine from the stop line with the more forward mounted cam, but have to creep near or sometimes into cross traffic to get a good enough view with the B-pillar location.


Now- if they plan to just stay L2, they can solve that by just making it a situation that requires the human to confirm it's safe to go (as it already does now at some T-intersections I've seen)... but if they want to go L3 or above they need to actually be able to SEE coming cross traffic with the cameras without being IN the traffic.
 
I'm unclear how planning fixes not being able to see around a corner where the B pillar cam is blocked?

Chuck Cook, among others, have done videos on this- showing the feed from the B-pillar position vs the feed you'd get with a side-cam near the front turn lights and it's a pretty dramatic difference at such intersections-- to where you can see just fine from the stop line with the more forward mounted cam, but have to creep near or sometimes into cross traffic to get a good enough view with the B-pillar location.


Now- if they plan to just stay L2, they can solve that by just making it a situation that requires the human to confirm it's safe to go (as it already does now at some T-intersections I've seen)... but if they want to go L3 or above they need to actually be able to SEE coming cross traffic with the cameras without being IN the traffic.
Better planning will reduce lag and time it takes to make up the mind and move. Yes, it can't produce perception that is not there.

Unprotected right turns / left turns on to busy high speed roads should be avoided. Like Waymo does. They are a hazard, for AVs and for humans. Shows those cities / counties are not doing their job.
 
Better planning will reduce lag and time it takes to make up the mind and move. Yes, it can't produce perception that is not there.

Unprotected right turns / left turns on to busy high speed roads should be avoided. Like Waymo does. They are a hazard, for AVs and for humans. Shows those cities / counties are not doing their job.
Disagree with this - the alternative is putting stoplights at every such intersection which is a huge expense and impacts traffic flow on the larger street. It all depends on traffic patterns. There are 2 such intersections leading out of our development so I can't avoid them but in the 10 years we've live there accidents have not been a problem so it can't really be argued that they're a hazard.

Regardless, such intersections are common throughout the U.S. so any FSD software needs to be able to manage them.
 
  • Like
Reactions: VanFriscia
Regardless, such intersections are common throughout the U.S. so any FSD software needs to be able to manage them.
Oh yes - bad infrastructure is common throughout US. Question is will AVs develop alternate strategies to deal with them ...

BTW, that there are (apparently) no accidents on those two spots don't make those spots safe, as such. There is a lot of luck involved - because people who are driving should also not slow down to make accidents happen. Most people will slow down when they see someone turning into the lane ahead.
 
  • Like
Reactions: VanFriscia
Oh yes - bad infrastructure is common throughout US. Question is will AVs develop alternate strategies to deal with them ...

BTW, that there are (apparently) no accidents on those two spots don't make those spots safe, as such. There is a lot of luck involved - because people who are driving should also not slow down to make accidents happen. Most people will slow down when they see someone turning into the lane ahead.
True - the presence of accidents is good evidence of an unsafe intersection but the converse is not necessarily true. Safety is not binary, it's degrees of risk and we see this all the time in street design. Sight lines and traffic patterns play a big role, too. An intersection that is safe for one level of traffic can be horribly unsafe for a different level.

In this case you can see cars coming at least 1000 feet in advance and they can see you. If you see a car inching out 1000 feet ahead of you you assume that it will turn out, accelerate and have plenty of time to come up to speed long before you get to the intersection, even if it's a Prius. In this case the car inches out and leaves its nose in the road while hesitating. 1000 feet is 12-15 seconds at 45-50 MPH but the car takes longer than that, creating a safety issue.
 
In this case the car inches out and leaves its nose in the road while hesitating. 1000 feet is 12-15 seconds at 45-50 MPH but the car takes longer than that, creating a safety issue.
I think this can be corrected by better planner and faster perception.

Issues that can’t be corrected are where you can’t really see much, unless you lean forward/creep into the intersection.