FSD Beta v11.x

aronth5 · Apr 21, 2023

11.3.6 seems to be settling down which is consistent with releases that seem to improve over a couple of weeks. No idea why this appears to happen. I am back to the point with 11.3.6 that the majority of drives are disengagement free. Overall now 11.3.6 is an improvement for me.

Today my only "No Right on Red" sign was actually recognized and the car didn't turn. First time ever that has happened. I saw posts in Reddit that stated the same.

Yesterday I drove 150 miles on controlled access highways and had 1 disengagement which was so I could verbally provide feedback when the car wanted to go to the passing lane when I was approaching an exit. While I still don't like some of the zip merges I am finding that if I just wait the car is merging "successfully" so I'm waiting a lot longer before considering disengaging.

Supcom · Apr 21, 2023

aronth5 said:
11.3.6 seems to be settling down which is consistent with releases that seem to improve over a couple of weeks. No idea why this appears to happen. I am back to the point with 11.3.6 that the majority of drives are disengagement free. Overall now 11.3.6 is an improvement for me.

Today my only "No Right on Red" sign was actually recognized and the car didn't turn. First time ever that has happened. I saw posts in Reddit that stated the same.

Yesterday I drove 150 miles on controlled access highways and had 1 disengagement which was so I could verbally provide feedback when the car wanted to go to the passing lane when I was approaching an exit. While I still don't like some of the zip merges I am finding that if I just wait the car is merging "successfully" so I'm waiting a lot longer before considering disengaging.

It seems that the car is heeding No Turn on Red signs. If so, Tesla should add something to the visualization to indicate this. As timid as the car has been when it does not have the right of way, it's hard to say whether this is intentional.

Mardak · Apr 21, 2023

This double left turn seems to be causing trouble for 11.3.6:

Seems like something map data will need to handle as this Google Street View perspective from the right-most of three 1-way lanes shows the white car turning from the left-most lane into the middle/right of a really wide lane that soon forks to a left-turn-only lane and straight lane. The tricky visual aspect is that this overpass/bridge is a fixed width for 4 total driving lanes but it happens to be split into 1 oncoming and 2 destination lanes for the double left turn, but there's no visible dashed white lane separators until past the white car.

It would help if there were guiding stripes painted for the double left or maybe even clearer / earlier lane separators, but these still would only help when close enough to see them, so medium-term planning to get into the appropriate lane would still need to rely on pretty precise lane details that indicates there's initially 2 lanes with the left one immediately splitting.

The current behavior seems to be it really wants to be in the outer turn lane because it believes the inner turn lane will be forced to make a left turn, but this also is problematic in that during the outer left turn, it tries to switch to where people in the inner turn lane would drive (people don't usually make this 2-step U-turn.) Similarly, if I force it into the inner turn lane, it tries really hard before the turn to get out, but when actually making the inner left turn, it stays really far left making people think we're taking the next left turn and is slow / unable to get into the straight lane because it thinks there's 3 lanes instead of the current lane forking.

What's worse is at least OpenStreetMap's data incorrectly indicates there's 4 total lanes here (with no distinction of how many in each direction).

EVNow · Apr 21, 2023

Ramphex said:
no more jerking off the steering yoke and no more watching me via cabin camera to make sure I’m jerking off the steering yoke.

So - basically no driver monitoring ? Even with L3 - there may still be some driver monitoring. You are out of luck unless ....

How well do you know NHTSA Director ?

aronth5 · Apr 21, 2023

Ramphex said:
It’ll be good enough for me whenever it meets a certain set of my personal requirements, regardless of the % of completion. Those requirements are simple: no more jerking off the steering yoke and no more watching me via cabin camera to make sure I’m jerking off the steering yoke.

At least for FSD I'm very glad I don't have a steering yoke. At least with the steering wheel I just let the wheel slide thru my hand(s). Super easy and I'm always in control. The yoke for FSD at least would seem to be a nightmare based on all the videos I've watched. No wonder some posters view the yoke jerking as a major problem.

YNoVA · Apr 21, 2023

JulienW said:
This is an extraordinary claim that defies all Tesla release scheduling logic. Extraordinary claims that sound unrealistic require proof. Unless you car was somehow "mistaken" for a qualified employee you all but certainly don't have 11.4.

Nah I hacked my Tesla to make it think it's an employee vehicle

Twiglett · Apr 21, 2023

YNoVA said:
Nah I hacked my Tesla to make it think it's an employee vehicle

wow, what a coincidence so did I, did you get the secret extra super speed acceleration or the extra 50 mile range ?
Of course I can't prove it either...

Without proof we can claim anything to be true.

KLam · Apr 21, 2023

aronth5 said:
At least for FSD I'm very glad I don't have a steering yoke. At least with the steering wheel I just let the wheel slide thru my hand(s). Super easy and I'm always in control. The yoke for FSD at least would seem to be a nightmare based on all the videos I've watched. No wonder some posters view the yoke jerking as a major problem.

True, why jerking the steering wheel especially at slow speed turns and residential streets? After more than one year of coding and NN tuning? Tesla needs to put someone who knows how to drive on coding the drive mechanics module!! This is not a racing game implementation!

Therealm80 · Apr 21, 2023

I agree with the poster a few posts ago about everything calming down. I drove about 50 miles today and only had one disengagement intervention and a couple Accelerator pedal pushes. everything was fairly smooth. Even did a couple roundabouts pretty well. The disengagement was before the roundabout. It was accelerating pretty quick and a car was entering the roundabout to my left, and it caught it almost too late and slammed on the brakes and I disengaged at that point, but Other than that things are a lot better. Looking forward to 11.4.

JB47394 · Apr 21, 2023

KLam said:
True, why jerking the steering wheel especially at slow speed turns and residential streets? After more than one year of coding and NN tuning? Tesla needs to put someone who knows how to drive on coding the drive mechanics module!! This is not a racing game implementation!

I'm as annoyed as the next guy about this stuff, but let's remember that nobody has ever done this before. This is a research project in pretty much every way. The hardware (multiple sets), the driving software, the planning software, the training infrastructure, the simulations, analysis tools, version control, safety requirements, performance requirements, visualizations, different driving environments for weather, road types, road configurations - the list of items that have to be considered and resolved must be absolutely endless. I think the only thing that was generally solved when they started was identifying objects in images and drawing boxes around them.

Criminy. They even got yanked into Twitter for a while there.

I believe that it is all the gubbage surrounding the project that is slowing down progress the most. I wonder how much time they get to spend on the actual driving software, algorithms and neural networks. Any other engineers here know that sensation from their own professional life?

KLam · Apr 21, 2023

JB47394 said:
I'm as annoyed as the next guy about this stuff, but let's remember that nobody has ever done this before. This is a research project in pretty much every way. The hardware (multiple sets), the driving software, the planning software, the training infrastructure, the simulations, analysis tools, version control, safety requirements, performance requirements, visualizations, different driving environments for weather, road types, road configurations - the list of items that have to be considered and resolved must be absolutely endless. I think the only thing that was generally solved when they started was identifying objects in images and drawing boxes around them.

Criminy. They even got yanked into Twitter for a while there.

I believe that it is all the gubbage surrounding the project that is slowing down progress the most. I wonder how much time they get to spend on the actual driving software, algorithms and neural networks. Any other engineers here know that sensation from their own professional life?

Actually, it's a very simple requirement. Jerky steering wheel indicates "out of control". It indicates that you do not know what you are doing. That's why all the wife/girl friend/passengers hates it. Look at the Waymo drive clips. Do you see them jerking the steering wheel?

Therealm80 · Apr 21, 2023

IM IMPRESSED! Just drove in a complete downpour and FSD didn’t even blink kept going on the freeway I couldn’t see anything outside and it never disengaged worked very well.

JB47394 · Apr 21, 2023

KLam said:
Look at the Waymo drive clips. Do you see them jerking the steering wheel?

Talking about Waymo and FSDb in the same breath is a bit silly. Waymo is a commercial taxi service that operates at low speeds in geofenced areas using very expensive hardware. FSDb is a consumer system that operates at all speeds everywhere using consumer hardware. Tesla is tackling a vastly more difficult problem, and that's why they can't spend the time to fine tune everything just yet. I don't know how long you've been using FSDb, but the wheel jerkiness has been massively reduced. When the FSDb team has the time, I'm sure they'll get the whole thing running buttery smooth - including steering at very low speeds.

KLam · Apr 21, 2023

JB47394 said:
Talking about Waymo and FSDb in the same breath is a bit silly. Waymo is a commercial taxi service that operates at low speeds in geofenced areas using very expensive hardware. FSDb is a consumer system that operates at all speeds everywhere using consumer hardware. Tesla is tackling a vastly more difficult problem, and that's why they can't spend the time to fine tune everything just yet. I don't know how long you've been using FSDb, but the wheel jerkiness has been massively reduced. When the FSDb team has the time, I'm sure they'll get the whole thing running buttery smooth - including steering at very low speeds.

I thought I was talking about low speeds in residential areas. Something which do not need a lot of Beta driving setups and feedbacks. Certainly lower effort than UPL turns. Just imagine taking a DMV driving test, if you jerk your steering wheel driving out of the DMV office. Your examiner probably will tell you to head back immediately.

thewishmaster · Apr 21, 2023

Observed what seems like new, or at least newly mentioned in the UI "Merging out of the rightmost lane" behavior today; it was a stretch of freeway with no onramps / offramps on the right side. The nav actually needed the car to stay in the right lane due to me needing to exit at the next exit.

It did end up merging back to the right lane eventually, but not after 2 attempts to get out of the right lane. First one I wasn't paying attention to the screen and disengaged when I felt the car suddenly start to leave the lane (the blinker activated late), and the second time I let it merge to see how soon it would merge back.

Also it seems like the "sound / vibrate" feature to warn before changing lanes doesn't work for FSDb because I have both enabled

aronth5 · Apr 21, 2023

KLam said:
Actually, it's a very simple requirement. Jerky steering wheel indicates "out of control". It indicates that you do not know what you are doing. That's why all the wife/girl friend/passengers hates it. Look at the Waymo drive clips. Do you see them jerking the steering wheel?

Simple requirements don't necessarily mean simple solutions. The jerky wheel drives me a bit nuts as well but I wouldn't presume to tell the developers it's an easy fix.

EVNow · Apr 21, 2023

KLam said:
I thought I was talking about low speeds in residential areas

Two things ...
- knowing exactly where everything is makes it easier, duh.
- search and you will see someone complaining about jerky steering by Waymo. You just don't feel it from the back seat - just like my kids don't feel the jerky steering by FSDb

jebinc · Apr 21, 2023

Ramphex said:
I think Alan will approve it before Jeb does.

It’ll be good enough for me whenever it meets a certain set of my personal requirements, regardless of the % of completion. Those requirements are simple: no more jerking off the steering yoke and no more watching me via cabin camera to make sure I’m jerking off the steering yoke.

My biggest gripe with the software is not its crap driving but the constant nagging along with the crap driving. Adding those two ingredients is what creates the Faux Self Driving: Junk.

https://twitter.com/x/status/1649285600472862720

jebinc · Apr 21, 2023

Therealm80 said:
I agree with the poster a few posts ago about everything calming down. I drove about 50 miles today and only had one disengagement intervention and a couple Accelerator pedal pushes. everything was fairly smooth. Even did a couple roundabouts pretty well. The disengagement was before the roundabout. It was accelerating pretty quick and a car was entering the roundabout to my left, and it caught it almost too late and slammed on the brakes and I disengaged at that point, but Other than that things are a lot better. Looking forward to 11.4.

Seriously, how is this even possible? The on board software does not evolve within the car. Clearly a placebo effect, or driver learning/acceptance curve.

If there any proof of how this could even happen (“settling down”)???

jabloomf1230 · Apr 21, 2023

jebinc said:
Seriously, how is this even possible? The on board software does not evolve within the car. Clearly a placebo effect, or driver learning/acceptance curve.

If there any proof of how this could even happen (“settling down”)???

It's probably due more to the various factors (traffic, lighting, nav route, etc ) that affect FSDb averaging out as the driver undertakes additional trips.

FSD Beta v11.x

Long Time Follower

Active Member

Active Member

Well-Known Member

Long Time Follower

Member

Single pedal driver

Member

Member

Active Member

Member

Member

Active Member

Member

Active Member

Long Time Follower

Well-Known Member

Well-Known Member

Well-Known Member

Minister of Silly Walks

Similar threads