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:
Well, there has been some discussion. As long as FSD is a driver supervised feature ("beta" or not, and whether in the title or description of the feature), most people agree that you are liable for mistakes and at-fault accidents. Map errors? Google Maps occasionally call for an illegal or unavailable maneuver; it's up to you to mind the one-way sign they missed or to not drive on the bridge that was washed out.

There is somewhat less agreement on what happens when Tesla moves to and unsupervised self-driving capability, i.e. not just advanced driver assistance. Some people say that Tesla will :never take liability, but I think this is quite wrong. See my post in a different thread:
Even if you are allowed to remove your hands from wheel and not pay attention for 30 seconds, that would shift liability in my mind. If 30 seconds of inattention is not allowed, it is certainly driver assistance at best and not FSD, arguably not safer or more relaxing for the driver. I have not been in an accident for many years, so safer will be hard to prove, but after 4 years of playing prepaid-beta-tester, I am more stressed while using FSD. Only Elon could figure out how to get fools like me, to pay to develop their software and teach the AI about the world.
 
That was a very categorical statement. How do you substantiate this position?
You mean the cold fart trusters are Fantasy?
I was ready to fly. This is the most comical of the lies Elon told, any engineer would see the impracticality, and safety issues, even if for track use only. still laughing. runner up lie would be that your car would remember wheel positions to be able to smart summon out of a parking garage, yes he said that
 
  • Like
Reactions: pilotSteve
Yup, I don't think Tesla will ever provide a "free" level 3 or higher service. It will all be charged per mile (or subscription to Tesla insurance) to account for their liability.
There always needs to be insurance, someone will pay for it and that someone will be us, whether it's indirectly or directly. I suspect that's also behind the move towards making FSD a monthly fee. That's a better way to fund the ongoing development, maintenance and support costs of such a system. From a PR perspective, people will focus more on the availability than the monthly fees.
Drove 11.4.1 to and from work for the first time. Roughly 45 minutes and 30 miles each way. Morning started rough. Couldn't make a protected left turn it has done well previously. I had to intervene after some failed jerky moves. I would add that the sun was just up and particularly blinding so that might explain the behavior. Then, a few minutes later, no sun involved, while going along a four lane divided road (US 1 in FL) FSD decided to move to the right in a turn lane into a plaza. No idea why. Intervened. The rest of the drive was quite nice.

Drive home: absolutely great not a single intervention. A couple of slow downs (more than I would have) when cars from the opposite direction turned left and crossed my path but otherwise very smooth, very comfortable drive. Most impressive since I started on FSD in November of 2021.
I've commented in the past that previous FSD versions seem to improve somewhat over time. It's unclear whether it's expectations tempering over time or something else but I've noticed several specific, identifiable items that improve so I think it's more than simply perception. It may be better map data, improvements in the neural nets or other data pushed to the car that FSD uses in decision making.
 
  • Like
Reactions: pilotSteve and GSP
I've noticed a few dedicated right turn lanes, like below, 11.3.6 is starting the turn and then coming to a near stop mid turn (red mark shows where the car starts slowing down), even with a green light, and given the lane is dedicated and not merging for some 100+ feet after the turn. The last two versions handled these intersections/turns properly, without slowing down.

View attachment 938178
I've never had FSD consistently handle such right turns well. Most of them involve yields but it will frequently stop before the crosswalk on green lights, regardless of the presence of pedestrians. If the light is red it will always stop, even when there's no traffic at all.
 
  • Like
Reactions: Dewg
I think it may have just stopped raining in California.

Last time I drove in a misty rain associated with a warm front from the last couple rounds of rain, which had more subtropical tap characteristics, it was awful. It could not handle it. Tiny drops it simply could not understand.

But then 11.3.6 came out, which caused it to stop raining.

We’ll see. It does seem like a problem they should be able to solve. Unlike some other problems.
I've noticed some improvements but still far from perfect. My guess is that you end up with some streaking at the end of the wiper arms/top of the window where the cameras are. The system isn't patient enough to let the streaks dry so it wipes again and renews the streaks. Lather, rinse, repeat.
 
I second the fact that 11.3.6 has improved since I first got it. This is not just an “It feels better” analysis. A spot that has always had an awkward lane change through an intersection before a turn has been fixed, even though when I first got 11.3.6 it had not been fixed.
They definitely have an ability to do something without an official update.
How could it have gotten better since you got it? The only way it could get better is through OTA updates, no?
 
Last edited:
That was a very categorical statement. How do you substantiate this position?
The car is years and likely hw revisions away from driverless operation much less the ability to just go out and be a taxi, unlock doors for people, monitor the car, charging logistics, etc.

The biggest joke people fell for was that Tesla was keeping leases for robotaxi, it was never true...you could literally follow them selling leases by the vin from the start.

It's the same people that believe their car can actually be a profit making taxi services that believe that the Roadster will fly, the yoke from Nov 2021 will get an OTA update to be a horn, or that Teslas patents were actually open.

I love my cars and Elon has done great things, but he lies or, in the very least, pushes out bad info all of the time.
 
Driven/ridden 11.3.6 for a week now. Substantially better than pre 11 versions for sure. Up to 1/2 hour between interventions for me. Could have gone further but eventually take over when FSD is inconveniencing other road users too much by slowing/stopping where not necessary or when it wants to drive 25 mph over cobblestones and 5" protruding man holes where no reasonable person would go more than 5 mph.
In earlier versions I have experienced incorrect lane planning as well as occasionally turning into oncoming traffic. None of that happened in 11.3.6.

The unnecessary slowing and stopping appears to be meant as an abundance of caution. Other than that, the hard parts of driving are almost there. Once basic safe driving is reasonably ready and cautionary slowing can be gradually removed, the team will portably start adding the trivial features.

Those include recognizing most road signs, treating rail road crossings correctly, slowing for or circumventing road imperfections, going all the way inside a destination parking, navigate parking garages and so on. As others have said, sharing road hazards and detours between FSD cars in real time al'a Waze would be super helpful. Not sure if that's trivial or not.
 
  • Like
Reactions: FSDtester#1
The car is years and likely hw revisions away from driverless operation much less the ability to just go out and be a taxi, unlock doors for people, monitor the car,
Why do you think so? What is wrong with the current door locks and monitoring system? I would be more worried about the pedals and steering wheel.
charging logistics, etc.
With current charging logistics, the car could drive back to its handler for charging when there is not enough range for the next order. Like Uber, customer order would include destination.
The biggest joke people fell for was that Tesla was keeping leases for robotaxi, it was never true...you could literally follow them selling leases by the vin from the start.
Speed of machine learning is hard to predict. Many, probably including Elon, believed Robotaxi would be ready in time. Also don't underestimate the effect stretch goals have on your organization. Engineers work best under pressure. compare WWII, moon landing inside 1960s etc. I hope you don't think the moon landing was fake too.
he lies or, in the very least, pushes out bad info all of the time.
one knows others by knowing oneself
 
  • Like
Reactions: Yelobird
Ok.

Except even AMONG the "AV and HD map industry at large" there are multiple, different, definitions of what an "HD map" is.

This cites cm scale accuracy for example.

This cites 10 cm accuracy

This cites "between 10 and 20 cm"

I suspect if I invest more than 30 seconds into looking I'll find at least 2 more different definitions so it appears you can get at least 5 different answers even after moving the goalposts as far from my original statement as you did.

But by all means, if you believe you know of a SINGLE "correct" and complete definition of HD maps that everyone in the industry agrees on, please provide it, with a source.

What you're arguing here is fundamentally misguided. The assertion that a High Definition (HD) map needs to be exactly 10 cm accurate in order to qualify as an HD map is simply not being made. The metrics provided merely represent the specifications of their current technology - it's you who's attempting to enforce this standard.

To draw parallels, it's akin to suggesting that iPhones aren't smartphones simply because Samsung Galaxies perform at a faster pace, or because the Pixel brand boasts superior camera quality. This isn't a fair assessment of what constitutes a smartphone. The varying specifications don't change the fundamental definition of the device, but rather contribute to its unique identity.

Similarly, by this logic, you might argue that AMD Radeon graphics cards aren't truly graphic cards because they don't match the speed of the RTX series. This line of reasoning is clearly flawed.

The point is, the definition of an HD map for autonomous driving is not contingent on its precision or the technology used to create it - be it Lidar, camera or radar. Whether it's 1 cm, 5 cm, 10 cm, or 20 cm accurate, it doesn't matter. The crux of the matter is the ability for a car to localize in it - that is the fundamental and sole criterion for an HD map. The rest are simply varying attributes, contributing to the unique character of the map, not defining it.
 
Last edited:
I've noticed some improvements but still far from perfect. My guess is that you end up with some streaking at the end of the wiper arms/top of the window where the cameras are. The system isn't patient enough to let the streaks dry so it wipes again and renews the streaks. Lather, rinse, repeat.
No, the issue in the case of mist is zero wiping. I don’t care about extra wipes.
 
Why do you think so? What is wrong with the current door locks and monitoring system? I would be more worried about the pedals and steering wheel.

With current charging logistics, the car could drive back to its handler for charging when there is not enough range for the next order. Like Uber, customer order would include destination.

Speed of machine learning is hard to predict. Many, probably including Elon, believed Robotaxi would be ready in time. Also don't underestimate the effect stretch goals have on your organization. Engineers work best under pressure. compare WWII, moon landing inside 1960s etc. I hope you don't think the moon landing was fake too.

one knows others by knowing oneself
FSD has gotten better in the 2 years and 50,000 miles that I've been a tester. I think it was a worthy investment for a level 2 ADAS and I do enjoy it, but it's been very very slow progress. Every update is 1 step forward, 2 steps back. The whole argument of AI making exponential improvements has yet to come to fruition. Tesla hasn't made much progress with school zones, school buses, yield signs, etc. Much less "edge" cases.

I don't believe that the current camera placement will ever be good enough without supervision, because there are too many cases on turns where the cameras cannot see and the car has to extend itself dangerously into traffic.

For your moon landing comment, that's an absurd accusation that's obviously made because you have blind hope in Robotaxi coming true. I've been called a fanboy for liking FSD Beta much more than other options like Ford's Blue Cruise, but it's still not to the point where there's even a possibility this car could just drive without me being there to help it through situations or to save it from wrecking.

Your last line is just as ridiculous as the moon landing...Believing FSD is even a few years away from completion, much less becoming a robotaxi is much more of a leap than believing it won't. There are thousands of testers that see the capability or lack thereof every day...that's verifiable...the idea that it will be a robotaxi is just Elon talking, which isn't. The burden of proof that robotaxi will be reality is on you, because I can show you why it won't by just getting in a car and turning on FSD Beta.
 
FSD has gotten better in the 2 years and 50,000 miles that I've been a tester. I think it was a worthy investment for a level 2 ADAS and I do enjoy it, but it's been very very slow progress. Every update is 1 step forward, 2 steps back. The whole argument of AI making exponential improvements has yet to come to fruition.
Seems like we have on the order of similar amounts of FSD Beta experience, almost. Mine is 2 years 25k miles.

I agree it has been slow progress and think of the iterations as building and testing a series of channels for machine learning, in a quest to find the most effective one. This doesn't necessarily lead to improved performance. Once they find/if they find a good enough set up I think it can improve exponentially. I am definitely not sure either that it will happen in the near future, but I think I am starting to see a take off. Neither de we know how the brain works, nor does anyone know every detail of what's happening inside of a neural network, so how could we predict when one will catch up with the other? Who would have predicted Chat GPT would have surpassed most human writing and reasoning already?
Tesla hasn't made much progress with school zones, school buses, yield signs, etc. Much less "edge" cases.
I believe that they deliberately assign nearly 100% of the autopilot team to developing the fundamentals of setting up the best environment for machine learning. This means no resources remain for these features. They will be added later.
 
Last edited:
Even if you are allowed to remove your hands from wheel and not pay attention for 30 seconds, that would shift liability in my mind. If 30 seconds of inattention is not allowed, it is certainly driver assistance at best and not FSD, arguably not safer or more relaxing for the driver. ...
That is NOT a possible option in the Full Self Driving Levels. Until its is SAE L3 or above you are required to pay attention all the time, hands off or hands on is irrelevant. You are in full control of the car.

SEA.png
 
  • Like
Reactions: sleepydoc