FSD discussion

[EVNow]

As an investor . . . I like the progress compared to "original" FSD that the cars come with.

I don't see much discussion about FSD as an investor.

I mean - people like Daily Tesla (Rob) and Dave Lee et al talk about FSD - but never really say much about the end game.

Ostensibly Elon's goal is "robotaxi". To get to Robotaxi, Tesla needs to get to about 10,000 miles per disengagement. I get this number from a couple of different places - Waymo's current disengagement rate and number of miles per any kind of small accident for humans.

But from all indications, FSDb disengagement rate is about 10 miles.

Home

FSD Community Tracker

www.teslafsdtracker.com

From DirtyTesla ...

My own calculation ... (I've watched and tabulated all my recordings) ...

So FSDb needs to improve about 1,000 times before Robotaxi quality can be achieved. I see no path to that for next several years atleast ....

 

[Bet TSLA]

So FSDb needs to improve about 1,000 times before Robotaxi quality can be achieved. I see no path to that for next several years atleast ....

I do. As neural nets replace algorithmic code, the executing programs get faster and smaller. So, effectively, the AI can think faster on more minimal hardware. As Dojo takes over training, iteration becomes faster, so newly found (march of 9s) bugs can get cleaned up faster.

So... progress speeds up by an order of magnitude or so. And as the most common reasons for disengagements disappear, their frequency goes down much faster than linear. So, virtuous cycle.

Is this likely? I don't know.

 

[EVNow]

I do. As neural nets replace algorithmic code, the executing programs get faster and smaller. So, effectively, the AI can think faster on more minimal hardware. As Dojo takes over training, iteration becomes faster, so newly found (march of 9s) bugs can get cleaned up faster.

So... progress speeds up by an order of magnitude or so. And as the most common reasons for disengagements disappear, their frequency goes down much faster than linear. So, virtuous cycle.

Is this likely? I don't know.

That sounds kind of generic.

The question really is around the planner, I think. Nobody has used end-to-end NN for planner. So, lot of it will be normal code.

1,000x better is a tall order in the next 3 years.

 

[Buckminster]

https://twitter.com/x/status/1595682322707267584

 

[JusRelax]

That sounds kind of generic.

The question really is around the planner, I think. Nobody has used end-to-end NN for planner. So, lot of it will be normal code.

1,000x better is a tall order in the next 3 years.

This youtube video has a discussion on AI for autonomous vehicles between employees of Tesla, Wayve, Waymo, and NVIDIA. One of the major topics that is discusssed is the possibility of end-to-end NN. Unfortunately, I watched the video a while ago, so don't remember exactly where in the video it was discussed, nor is there a timestamp for it.

Note that Kate Park was one of the engineers that presented at this year's AI Day (Unfortunately, she played more of a facilitator role in this discussion).

 

[Gray Man]

https://twitter.com/x/status/1595682322707267584

Ok, so I'm a new owner, just picked up on 11/11. I've requested FSDb, but as I understand it this announcement really means nothing for me since I'm on 2022.40.4.1. I'm SooL until v11 comes out, right?

 

[bkp_duke]

Ok, so I'm a new owner, just picked up on 11/11. I've requested FSDb, but as I understand it this announcement really means nothing for me since I'm on 2022.40.4.1. I'm SooL until v11 comes out, right?

Probably, I believe they need to merge the code changes from the FSD beta version into that version first.

 

[lafrisbee]

<--- Looking around the room
Where the hell is every... oh? maybe they are still out "driving"

 

[Joe F]

First ever FSDb drive with 10.69.3.1 last night in a less than 3 mile round trip. Hoping the several interventions were maybe due to this area never have had any FSD enabled traffic on them. I know, beta, but two interventions were so bad it felt like the camera were blind.

After following earlier reports, viewing YouTube videos showing major improvements with each new release, I was not expecting the car to head for an island or a cement base light pole in a parking lot. Will have to retrace the route in daylight to see how it does.

Perhaps my local area needs more FSD traffic as it was not what I was expecting, even though I knew to expect it to do the wrong thing at the wrong time, it was not what I believe others have experienced given the comments that I’ve seen that this release is so much better than the last.

 

[jerry33]

First ever FSDb drive with 10.69.3.1 last night in a less than 3 mile round trip. Hoping the several interventions were maybe due to this area never have had any FSD enabled traffic on them. I know, beta, but two interventions were so bad it felt like the camera were blind.

After following earlier reports, viewing YouTube videos showing major improvements with each new release, I was not expecting the car to head for an island or a cement base light pole in a parking lot. Will have to retrace the route in daylight to see how it does.

Perhaps my local area needs more FSD traffic as it was not what I was expecting, even though I knew to expect it to do the wrong thing at the wrong time, it was not what I believe others have experienced given the comments that I’ve seen that this release is so much better than the last.

Try recalibrating the cameras.

 

[EVNow]

One of the major topics that is discussed is the possibility of end-to-end NN.

There are long videos of MobilEye discussing various planner ideas. Check it out.

In general - when there are various road rules you have to follow (which, BTW change between countries and over time) - NN for planner becomes impractical. It will require way more computing resources to train and possibly in the car.

 

[EVNow]

First ever FSDb drive with 10.69.3.1 last night in a less than 3 mile round trip. Hoping the several interventions were maybe due to this area never have had any FSD enabled traffic on them. I know, beta, but two interventions were so bad it felt like the camera were blind.

After following earlier reports, viewing YouTube videos showing major improvements with each new release, I was not expecting the car to head for an island or a cement base light pole in a parking lot. Will have to retrace the route in daylight to see how it does.

Perhaps my local area needs more FSD traffic as it was not what I was expecting, even though I knew to expect it to do the wrong thing at the wrong time, it was not what I believe others have experienced given the comments that I’ve seen that this release is so much better than the last.

No - you don't need more traffic in your area.

It takes quite a bit of time to get used to how FSD drives. You tube will always make it much smoother and easier (like watching climbing Mt Everest is obviously easier than actually climbing Mt Everest).

Keep very vigilant, figure out where FSDb drives ok and where it needs help. You will do fine ...

BTW, see FSD sub-forum for lots of info on this.

AI, Autopilot, & Autonomous/FSD

Discussion about AI, Tesla Bot, Tesla Autopilot (AP), the promise of Full Self Driving (FSD), as well as other Autonomous Vehicles.

teslamotorsclub.com

 

[navguy12]

Try recalibrating the cameras.

All that does is eliminate the “fine calibration“ system bias that takes a few hundred (cumulative) kms of straight roads after the system has done a “coarse calibration“ to get it up and running.

Best analogy I can come up with: old fashioned inertial navigation units could “coarse align” in about 120 seconds and then one could set them from “align” to “navigate“ and go fly…but if one kept the units in “align” for more than 120 seconds, say more like 20 minutes, the accuracy of the units after a 10 hour sortie would be vastly superior to the two minute alignment example.

My mobile repair chap showed me how the various cameras (although letting me use NOA/Autosteer) were in various stages of complete calibration; in my situation, the pax side “b” pillar camera was still at 70% calibration whilst the rest of the suite was over 95% calibration.

 

[Discoducky]

First ever FSDb drive with 10.69.3.1 last night in a less than 3 mile round trip. Hoping the several interventions were maybe due to this area never have had any FSD enabled traffic on them. I know, beta, but two interventions were so bad it felt like the camera were blind.

After following earlier reports, viewing YouTube videos showing major improvements with each new release, I was not expecting the car to head for an island or a cement base light pole in a parking lot. Will have to retrace the route in daylight to see how it does.

Perhaps my local area needs more FSD traffic as it was not what I was expecting, even though I knew to expect it to do the wrong thing at the wrong time, it was not what I believe others have experienced given the comments that I’ve seen that this release is so much better than the last.

Parking lots isn't good right now, wait for "actually good smart summon".

Enable FSD on a straight road as you get used to FSD deficiencies.

Great to see more people posting about their drives!

 

[Discoducky]

My still limited impressions of 10.69.3.1

• Handles roundabouts better,
  • single lane issues are much less and don't require much throttle intervention,
  • two lane issues are better but still require intervention
    • FSD does not choose the correct lane often
    • FSD changes lanes during the roundabout, sometimes twice
• Handles target car turning left/right out of lane MUCH better
  • Almost perfectly actually as FSD proceeds immediately after the target car has left the lane. It is not accelerating fast enough for me and I'd actually like it to proceed just prior to the target car exiting the lane.
• Not enough data yet for the following issues
  • Acute/obtuse turns and NO creep wall
  • Unnecessary slowdowns
  • Late lane changes
  • Late blinker use
  • No blinker for merges
  • Missed turns due to late lane changes
  • Using dedicated lanes incorrectly
  • Not identifying partial medians correctly (i.e. Too strict lane usage)

PSA and as a reminder: FSD is best for public roads. Private roads and parking lots are not a great experience currently. "Actual smart summon" should address this in a future build.

 

[bkp_duke]

Parking lots isn't good right now, wait for "actually good smart summon".

Enable FSD on a straight road as you get used to FSD deficiencies.

Great to see more people posting about their drives!

Agreed. My wife was super-excited about getting FSD beta until I told her "under no circumstances do you use this in a parking lot . . . ever".

 

[Chunky Jr.]

I think FSD is making great progress, but I think we're still years away from robotaxi level of quality.

 

[Discoducky]

I think FSD is making great progress, but I think we're still years away from robotaxi level of quality.

Why exactly?

 

[Chunky Jr.]

Why exactly?

There are many reports (YouTube, here on TMC) of serious interventions that are needed even with the latest software. Some people have great experiences, and some are having not so great experiences. In order to get reliability needed for robotaxi, they'd need months of virtually no interventions in order to shake out the last sets of bugs. Look how long they have spent testing the semi or other car models before shipping. It's a long way between "hey, this works pretty well most of the time!" to a fully autonomous robotaxi

Basically I think the rate of improvements is not high enough to be close to shipping and it will take many more releases, perhaps a dozen or more, before it is close to being out of beta.

 

[Joe F]

Advice on not using FSDb in parking lots noted. In my case I enabled it on approach to an exit, when it promptly veered towards a light pole. On exit, it made a hard right into a dedicated right turn lane, yet I was turning left, two lanes over, as indicated by the route. Was kind of expecting it to screw up based on comments I read earlier.

Will try the route again, but it at one point it wanted to route a few blocks away, making 3 turns when going straight and making one left turn was needed. I had to disengage twice at parking lot entrances as it appeared to want to u-turn and put me back on the long way around.

That left leads to a tight circle, and not only was the entry way too fast, the map showed it wanted to route the wrong way, going against what would have been oncoming traffic. I don’t know what it would have done as I disengaged and took over.

Perhaps due to the circle being connected to a shopping center, it warrants the same caution as a parking lot. Will still try again in daylight.

Only other comment is I felt it did an “unsafe follow“ behind a fedex truck, much closer than I would have been. I thought it was funny as the safety score went away with this release, and my thought was it didn’t want to ding itself for an unsafe follow.

I applaud all those who used earlier releases, and thank them for their bravery.