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:
Someone should do the math on this. To me it seems like there is a hidden category here
I believe we're missing out on True Negatives from the first 2 improvement numbers (recall & precision) and assuming the before/after comparison did not change the Total number of examples (e.g., some of the 80k added clips were included in the baseline).

Recall = True Positive / (True Positive + False Negative)
Precision = True Positive / (True Positive + False Positive)
Error = (False Positive + False Negative) / (Total = TP + TN + FP + FN)

This is very much dependent on the number of each of the 4 possible outcomes in Tesla's mix of examples in their training vs evaluation set.
 
  • Like
Reactions: AlanSubie4Life
In Bangalore, I was scared to cross the street. You see old ladies walk right through the moving traffic and it just kind of flows around them.

I guess it works most of the time but you'd better not hesitate. The drivers are extrapolating the pedestrian path for planning, but they certainly aren't considering "multiple VRU futures" even though I sure was!

This is how I described the "method" earlier.

Looks easier than the one near my home (in India) ;) Basic idea is - go slow and occupy the space first. That gives you the right of way (atleast for a few feet). Also, first few feet to the side are reserved for vehicles driving the wrong way.

Actually the best way to think about it is - its like a slightly faster parking lot.

BTW, I should say ... after all these years, I feel scared to drive in India. Somehow I wasn't scared to cross the road .... probably in the "genes" ;)
 
-Improved lane guidance module to feed in long range routing "hints" to the network for which lanes ego needs to be in to reach its destination. Also significantly improved per-lane routing type autolabeler. These changes combined resolved 64% of all interventions caused by bad routing type.
I think now (finally !) Tesla is working on planner issues specifically - and probably in general issues that result in interventions. Thats a good thing. Hopefully they can start addressing the most annoying items on a weekly basis.

AFAIK, this is the first time they have actually mentioned reducing interventions by a certain %.
 
FSD Beta 11.3.x only really started going wider the last few weeks of March, but including highway driving seems to have already increased the monthly FSD Beta mileage rate by 25% now up to 16M miles/month. The wide release to 400k vehicles including those who did not proactively opt-in had a smaller boost of around 5% bump in January compared to October's previous high. Here's a graph that I estimated from their 2023Q1 report:

fsd beta 23q1 cumulative miles.png
 
I ran into this on version 10, but not 11. I have a two lane road with a series of low rolling hills, and FSDb would brake pretty hard at the crest of each hill. A lead car or oncoming traffic would be enough for the car to know that the world didn't suddenly vanish beyond the crest.
I have to amend this a bit as 11 is still giving me hesitation and sometimes just plain braking when reaching the crest of small hills. I had one instance of dropping from 40 to 33 mph.
 
  • Like
Reactions: VanFriscia
I have to amend this a bit as 11 is still giving me hesitation and sometimes just plain braking when reaching the crest of small hills. I had one instance of dropping from 40 to 33 mph.
Yes, I've experienced this too. More like a slight hesitation while cresting a rise as the car peers over the peak, usually only 2-3 mph.
(@JHCCAZ - traveling River Rd. between Swan and Sabino Canyon)
 
Yes, I've experienced this too. More like a slight hesitation while cresting a rise as the car peers over the peak, usually only 2-3 mph.
(@JHCCAZ - traveling River Rd. between Swan and Sabino Canyon)
I would consider this as a plus for safety. We know the car's forward camera is situated high on the windshield. If the camera can't see what is ahead over the hill, I am sure the driver, unless tall with his head touching the roof, can't see also.
 
  • Like
Reactions: heapmaster
In Bangalore, I was scared to cross the street. You see old ladies walk right through the moving traffic and it just kind of flows around them.

I guess it works most of the time but you'd better not hesitate. The drivers are extrapolating the pedestrian path for planning, but they certainly aren't considering "multiple VRU futures" even though I sure was!
Try to avoid the holes in the ground as well…. 💩
 
  • Funny
Reactions: JHCCAZ
I would consider this as a plus for safety. We know the car's forward camera is situated high on the windshield. If the camera can't see what is ahead over the hill, I am sure the driver, unless tall with his head touching the roof, can't see also.
I consider this a safety hazard. MX suddenly slows from low 60’s-low 50’s with no traffic or other reason, maybe the turn signal starts and the car lunges to the right (OK that’s better than to the left), or maybe it doesn’t.
So far this is happening to me on county and state hiways without a lot of traffic. It worked well on I-5 in traffic today, then on WA525 had erratic moves.
I am using it a lot, but that means disengaging it a lot -for safety.