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:
Still stuck on the obsolete 11.3.4 for what seems like forever! C'mon, Elon! Show some love towards those of use that have been providing good data on 11.3.4!

Seriously, from the videoas I've seen, it looks like th emajor change for 11.3.6 was merging the USS and non-USS branches. That's a good thing, but there doesn't seem to be any change in the issues with uncontrolled intersections. Still waiting to see something on highway merge behavior, but I suspect it has not changed. Maybe 11.4...

Oh, and I hope we aren't going to spend the year on another major redesign to accommodate this new diffusion technique. I was hoping that we might see some genuinely new capabilities.
 
If you are going by what he posted 2+ years back ... you are missing out a lot.

He has made a lot of changes including even showing the brake/accelerator because people claimed he was pushing the accelerator when he was not. Also see his commentary in some of the videos - he calls out all the issues.
Omar would have been a lot better off if he stayed away from exaggerated YouTube titles and Twitter posts and just let his videos speak for themselves. I give him a lot of credit for the time and effort those videos take and his Cruise and Waymo comparison drives. He is a good resource to the Tesla community over-all.
 
  • Like
Reactions: GSP and FSDtester#1
I think this is kind of an ultimate turing test for FSDb. Ideally other drivers, especially none tesla geeks, shouldn't be able to tell a FSD driven tesla from a manually driven tesla.
What I'd like to see are drives from outside / behind where we have to guess whether the car is on FSDb or not. The front car should engage / disengage randomly. Can the following car get it right most of the time ? If there are no stop sign / creeps, it would be difficult.
 
On city streets I’d expect it to start to get over a half mile before and on highways at the one mile mark.


I had a fun/frustrating disengagement on 11.3.3 last week related to getting over for a turn, but in this case I think it was confounded by it not wanting to use the suicide lane?

I have a pic of the intersection below but not dashcam footage.... in red is the outline of a box truck that got into the left turn lane before it technically became one... and my car in blue in the main lane.... which refused to get behind him because the lines were dotted yet and instead sped up- but not enough to get in front of him because there was a light coming up (which was red).... so I ended up stuck in the straight lane, first at the light, with now multiple cars in the turn lane I had to wait for, and people behind me in the straight lane honking in annoyance.


badleft1.jpg
 
I had a fun/frustrating disengagement on 11.3.3 last week related to getting over for a turn, but in this case I think it was confounded by it not wanting to use the suicide lane?

I have a pic of the intersection below but not dashcam footage.... in red is the outline of a box truck that got into the left turn lane before it technically became one... and my car in blue in the main lane.... which refused to get behind him because the lines were dotted yet and instead sped up- but not enough to get in front of him because there was a light coming up (which was red).... so I ended up stuck in the straight lane, first at the light, with now multiple cars in the turn lane I had to wait for, and people behind me in the straight lane honking in annoyance.


View attachment 926649
FSDb does not always make the pass or follow decision correctly. On highways, I find FSDb making the very human, yet wasteful decision to pass another vehicle just before exiting the highway. Often the passing maneuver was calculated correctly and FSDb is just able to complete the pass in time to exit. This is something that humans often do that uses additional energy to save no more than a couple seconds over just sitting behind the slower vehicle until the exit is reached.

It seems that FSDb calculates that two seconds saved is worth the expenditure, and the risk that the other car might speed up during the maneuver. And, since FSDb treats the maximum set speed as inviolable, it cannot decide to momentarily exceed the set speed to ensure it can complete the maneuver in time.

Your example above seems to fit this pattern. Perhaps FSDb did not account for the deceleration profile required to make the turn when it calculated the distance required to pass the truck. Or, perhaps the truck sped up a little. In any event, curious that FSDb does not recognize the issue soon afterward and simply slow down to get behind the truck.
 

I watched this guy's videos 3 times. He did not have disengagements.
I took the same route to the airport last week. My car moved from the lane going to freeway 5 to the wider shoulder. Disengaged. Maybe his setting is different? Or camera?

My disengagement happened at the shoulder at 8:31 in that video.
 
Last edited:
  • Like
Reactions: jebinc
Took my first drive on V11.3.6 this morning. Previously had 11.3.4. No noticeable changes in the updated version. Still having problems where the car pulls into a turn-only lane instead of remaining in the right lane to continue forward. Fairly smooth acceleration and deceleration with these versions, though, much less "slamming on the brakes" than I had in 10.69.

I'm satisfied with the ongoing improvements, and understand I have to closely supervise my car instead of texting or reading a book while driving.

I did drive for several miles without actually touching the yoke, by remaining perfectly still with my eyes on the road ahead. I like that a lot!
How frequent are the nags with this newest update?
 
I did about 180+ miles today on FSD Beta 11, most of it on highways. I got to say I was impressed with the highway driving. There were no unnecessary lane changes like NOA used to do. It handles highway transitions, on ramp and off ramp really well. It lane changed early to follow a route. I was never worried it might miss an exit like I used to be with NOA. It also handled lane changes safely, it saw when cars were coming fast behind me in the adjacent lane and waited for them to pass and them executed a smooth lane change when there was a clear gap. Now granted, conditions were ideal (sunny day, not a ton of traffic) but I had no interventions in 100+ highway miles.

City driving was less good. On a one lane rural road, it stopped too late for a stop sign. It detected the stop sign but stopped almost a full car length after the sign, in the middle of the intersection (there was no traffic). Several times, FSD Beta moved into the right turn lane only when I needed to go straight. In one instance, it was able to jerk back into the straight lane at the last minute when it realized it was in the wrong lane for my route. Another instance, it basically stopped in the middle of the right turn lane and I had to take over.
 
I did about 180+ miles today on FSD Beta 11, most of it on highways. I got to say I was impressed with the highway driving. There were no unnecessary lane changes like NOA used to do. It handles highway transitions, on ramp and off ramp really well. It lane changed early to follow a route. I was never worried it might miss an exit like I used to be with NOA. It also handled lane changes safely, it saw when cars were coming fast behind me in the adjacent lane and waited for them to pass and them executed a smooth lane change when there was a clear gap. Now granted, conditions were ideal (sunny day, not a ton of traffic) but I had no interventions in 100+ highway miles.

City driving was less good. On a one lane rural road, it stopped too late for a stop sign. It detected the stop sign but stopped almost a full car length after the sign, in the middle of the intersection (there was no traffic). Several times, FSD Beta moved into the right turn lane only when I needed to go straight. In one instance, it was able to jerk back into the straight lane at the last minute when it realized it was in the wrong lane for my route. Another instance, it basically stopped in the middle of the right turn lane and I had to take over.
I agree the highway driving is an improvement over the previous software build. The city driving is so treacherous that I'm probably going to stop using it.
 
which refused to get behind him because the lines were dotted yet and instead sped up
There's no question that the planner is feeble-minded. The car just moves along and assumes that the opportunities for lane changes will appear. They have a long way to go on planning.

Like many here, I moved from 11.3.3. to 11.3.6 this morning. I had a highway and secondary road trip today of about an hour total. It's a regular trip for me. I thought it was a pretty good drive, but that might partly be due to a growing familiarity with FSDb's quirks. Some of it may be changes to route planning data that gets updated behind the scenes. Some of it may be the driving environment (following vs being lead car).

The one thing I was most curious about was the way that FSDb 11.3.3 refused to re-engage on narrow, but well-paved and well-marked road that I take. I repeated my drive along it and still can't engage on the road unless the roadway widens out. This time I waited until going over a small bridge that had wider shoulders, and the FSDb icon appeared.

Oh, and others may already know this, but it's apparently not possible to cancel a lane change on the highway when needing to take an exit to stay on route. I had one which I had to cancel over and over again as the exit approached. When the car got very close to it, the car started the lane change without any opportunity to cancel, so I cancelled FSDb and stayed in my lane.
 
  • Like
Reactions: jebinc
Is it just me or is 11.3.6 incredibly touchy on when it can be activated off of the highway? It seems the grey wheel icon is gone more than it is on the screen anymore. Tired of hearing the rejection beeps telling me autosteer isn’t ready.
 
Is it just me or is 11.3.6 incredibly touchy on when it can be activated off of the highway? It seems the grey wheel icon is gone more than it is on the screen anymore. Tired of hearing the rejection beeps telling me autosteer isn’t ready.
It's not just you. Others have posted about it in this thread. If you look up one post, you'll see me describing my encounter with it. I hit it in 11.3.3 for a narrow two-lane road with perfect markings and road surface. The problem is still there in 11.3.6, and my testing says that I cannot engage due to road/lane width. Chuck Cook also hit it for unmarked roads when he moved too close to the shoulder.
 
  • Like
Reactions: SK360
I did about 180+ miles today on FSD Beta 11, most of it on highways. I got to say I was impressed with the highway driving. There were no unnecessary lane changes like NOA used to do. It handles highway transitions, on ramp and off ramp really well. It lane changed early to follow a route. I was never worried it might miss an exit like I used to be with NOA. It also handled lane changes safely, it saw when cars were coming fast behind me in the adjacent lane and waited for them to pass and them executed a smooth lane change when there was a clear gap. Now granted, conditions were ideal (sunny day, not a ton of traffic) but I had no interventions in 100+ highway miles.

City driving was less good. On a one lane rural road, it stopped too late for a stop sign. It detected the stop sign but stopped almost a full car length after the sign, in the middle of the intersection (there was no traffic). Several times, FSD Beta moved into the right turn lane only when I needed to go straight. In one instance, it was able to jerk back into the straight lane at the last minute when it realized it was in the wrong lane for my route. Another instance, it basically stopped in the middle of the right turn lane and I had to take over.
As I keep saying controlled access highways aren't far from L3 if Tesla wanted to push that as a pre robotaxi phase. It would probably take a competitor to get their first and take business away from Tesla for it to ever happen. I think a lot of people who won't pay $15k and only have HW3 would pay something for this level of FSD.
  • Invoke L3 only after merging
  • Enable minimal lane changes (initial phase)
  • Cap speed at 70mph
  • Average setting
  • Need new software to enable graceful user handoff for edge cases
  • Enable video.
  • No nags for phone usage.
  • Don't allow for snow
  • Doesn't appear this would need HW4