Welcome to Tesla Motors Club
Discuss Tesla's Model S, Model 3, Model X, Model Y, Cybertruck, Roadster and More.
Register

FSD Beta Videos (and questions for FSD Beta drivers)

This site may earn commission on affiliate links.
Correct me if I'm wrong, but one year later and I think we are still seeing disengagements for the same reasons.

Yup, mostly true, except for #2 and #3 have been much improved:

Moves into different lane over complicated intersections
Difficulty turning into narrow roads with cars

I just like to point out things that did improve. A few posts ago, I mentioned that 10.2 is much better at unprotected lefts vs 10.1, over span of only 2 weeks. We can imagine that some of these other common disengagements get fixed one day, hopefully soon.
 
A few posts ago, I mentioned that 10.2 is much better at unprotected lefts vs 10.1, over span of only 2 weeks

Yet, it's still quite unreliable at performing this task correctly. It's not fixed by any stretch of the imagination. Just this morning, FSD charged down the area between the double yellow lines for no apparent reason, after I pushed it reluctantly to proceed with the unprotected left turn (there was no traffic but it had decided to halt in the eastbound traffic lanes). Hadn't done that before - FSD decided to ignore the double yellow lines after the merge lane this morning (usually it just swerves around wildly). No one was around to puzzle at the behavior. Eventually (after about 1-2 seconds) I disengaged so that we could get back in the westbound traffic lanes ( I don't know what it would have done if I had left it...presumably continued to drive between the double yellow lines until the traffic-light controlled intersection).

Hopefully didn't pick up any nails. There's a lot of debris there since no one drives in that area of the road. Video clip reported to Tesla. Maybe I'll write it up, but I already wrote up this intersection for them. 95 Safety Score this morning, I think mostly because TACC got left on after a disengagement and abruptly halted at a traffic light. Sigh. Maybe I'm not cut out for beta testing.
Screen Shot 2021-10-19 at 12.51.04 PM.png


So a long way to go. I expect it will be at least 6 months before this sort of issue is substantially better and turns can be made reliably, smoothly and correctly. May require a couple fundamental rewrites!
 
Last edited:
  • Like
Reactions: powertoold
Musk says the improvements are exponential. I'm just saying we may be way out on the negative axis where the change with time is slow, small slope. So it looks like not much is happening to get us to level 3 (time = 0) any time soon. Exponential doesn't necessarily mean what Musk implied, fast change with time.
Elon didn't specify the base of the exponent IIRC. The graph of an exponential function where the base of the exponent is <1 is a decreasing slope.

/math_nerd
 
Last edited:
...
I pulled out of my house yesterday, at the first light it stopped to make a right, light turned green, so it started to proceed into hitting a woman and her stroller. Had to disengage there.
It’s not even close to 500 miles unless you were on the interstate for 500 miles driving strait and even then it’s doubtful.

I feel like it is 10-15 miles on well marked roads with stoplights and 0-5 miles on other roads.
A lot of the disengagements depend on the presence of a third party. If there is (a woman with stroller) or if there is not.

If there are no disengagement-quality issues then FSD drives quite well, of course. Putting a distance to the disengagements is useful so long as we take into account the complexity of the drive.
 
Last edited:
Sigh. Maybe I'm not cut out for beta testing.


So a long way to go. I expect it will be at least 6 months before this sort of issue is substantially better and turns can be made reliably, smoothly and correctly.
Based upon, exactly, what? Please back this up with hard data.
May require a couple fundamental rewrites!
Sure, and China may annex Japan. Climate change may reverse itself without our intervention. Who knows?

These kinds of statements drive me nuts. All speculation with no basis.
 
Based upon, exactly, what? Please back this up with hard data.
It's my expectation (a guess!) based on my evaluation of FSD Beta so far. Overall it seems like it has a very long way to go. Bookmark my post and check back in 6 months to see how we're doing! I'd be thrilled if it were reliably approaching that unprotected left (the only difficult turn I routinely have to make), smoothly executing it, and entering the correct lane of westbound traffic, within the next six months, with no intervention. I just think it's unlikely (so far I've had to intervene every single time - pushing the accelerator to get advanced enough, and so far I'd have to steering disengage on every attempt if I wanted it to be smoothly and correctly executed - so about 0 for 6 so far).

What hard data could I possibly have to make such a claim? On the flip side, what hard data could anyone have to "prove" that it will be shorter?

Who knows?
The fundamental rewrite comment was obviously a little tongue in cheek. Wouldn't be surprised though!
 
  • Like
Reactions: Phlier
Based upon, exactly, what? Please back this up with hard data.

...
These kinds of statements drive me nuts. All speculation with no basis.
Speculating on faith, past history, and insider forecasting...? Elon does it too. "(something will improve) in (some amount of time) probably, (some longer amount of time) definitely"
 
Last edited:
  • Like
Reactions: AlanSubie4Life
So a long way to go. I expect it will be at least 6 months before this sort of issue is substantially better and turns can be made reliably, smoothly and correctly. May require a couple fundamental rewrites!
Just procedural code, won't need rewrites, as such. From AI day, looks like they are continuously trying to improve the number of simulations the optimization needs to run - so in that sense there is continuous improvements.

Yes, at some point may get into NN from procedural code ... that requires end-to-end training.

In this case - the problem is figuring out when it is "ok" to drive between 2 yellow lines (happens all the time when the traffic is backed up) vs when to strictly follow the rules.
 
The fundamental rewrite comment was obviously a little tongue in cheek. Wouldn't be surprised though!
You gotta remember that there are a lot of people that don't know that FSD is a fundamental re-write. For those that did know that little tidbit, the tongue-in-cheek is pretty obvious, and well-played, IMO.

I really hope they got it right with the current fundamental re-write, but certainly wouldn't be surprised if another ground-up (or two) attempt is needed.

As long as we're crytal-balling it, mine says that it'll show gradual improvement over the next year. But after that, we'll start to see very rapid improvements.

But level 4-5? I wouldn't even venture a guess at this point.
 
Just procedural code, won't need rewrites.
Unless it's a perception problem. It's really hard to say why it would drive across double yellow lines. Was that caused by an error in the procedural code (some escape path caused by a combination of circumstances in this simple turn case), or was it caused by a dropout or failure of perception somehow (did it lose track of the yellow line and forget about it, when it got too close to it)?

If it's a perception memory issue, do they need to redesign how they weight past perception and build their map of the environment (followed by extensive regression testing of course)?

To be clear, I have no idea. All extremely speculative and I have no idea why it fails the way it fails.
 
Last edited:
  • Like
Reactions: Phlier and Dan D.