FSD Beta 10.69

[JohnB007]

I'm REALLY hoping that now this FSD can handle other (easier but more common) ULTs more gracefully/successfully.

 

[MP3Mike]

But he did say, for the record, that they were gonna solve Chuck’s turn 100% (on 10.69).

But you have to remember, that their definition of solved isn't necessarily the same as yours. They could consider waiting for a 10 second gap as acceptable. (Where you expect a maximum of a 5 second gap to be used.)

 

[AlanSubie4Life]

I'm REALLY hoping that now this FSD can handle other (easier but more common) ULTs more gracefully/successfully.

I have extremely low expectations for mine! It is rarely successful at completing this properly and employs a remarkable array of weird solutions. We may all need to start YouTube channels for each tricky one. (Mine is not tricky, to be clear.)

 

[AlanSubie4Life]

But you have to remember, that their definition of solved isn't necessarily the same as yours. They could consider waiting for a 10 second gap as acceptable. (Where you expect a maximum of a 5 second gap to be used.)

Sure, though a 10-second gap is clearly not a solution, and unacceptable for any reasonable driver (and also mysterious because it would imply a 400m visual range, which would be big, if true, and also they go on 5-6 second gaps, and in the past have gone on gaps shorter than prior gaps). The gaps are all very mysterious.

There have also been failures which don’t involve the wait time (see yesterday with bad posing, and a near collision when using an 8-second gap).

Also, 100% is a weird goal, as previously discussed. It makes no sense. How many 9s?

 

[EVNow]

But he did say, for the record, that they were gonna solve Chuck’s turn 100% (on 10.69).

Did he say they will 100% solve chuck’s turn or solve Chuck’s turn 100% ?

 

[EVNow]

What is? It’s very important from a demonstration-of-capability standpoint to be able to safely do this every time. It’s an easy left turn but with substantial requirements on perception. If it can’t do this I am not sure what is the point (since this would also likely mean it could not turn right reliably, either, or perform merges into traffic from a stationary position, or correctly position itself on the road, or see reliably over 200m!)


It’s definitely a waste of my time, though, lol.

Its all about disengagement rates. First get easy ULTs reliably - that will improve the disengagement rate not this edge case.

 

[Mardak]

Will 10.69 include "single stack"? I ask since 10.69 will apparently include some big architectural changes according to Elon

"Single stack" has been part of FSD Beta firmwares although so far only for shadow mode data collection. It seems like having single stack FSD Beta replace current Navigate on Autopilot will happen with the major version bump to FSD Beta 11 (and continue ongoing improvements with 11.x presumably until it's ready for no-safety-score-wide-release).

I believe the "big architectural changes" have already happened as part of 10.13, and while there could still be more improvements since 10.13 went to employees, it seems like Elon Musk wanted to highlight the importance of those existing changes since 10.12 instead of "the usual" next version bump.

 

[AlanSubie4Life]

First get easy ULTs reliably

While not the easiest, this is not a hard left turn by any stretch.

this edge case.

Not an edge case, at all. Very very common, nothing unusual about it.

What do you think makes it an edge case?

Requires a base level of capabilities. Doing it very very (very very) reliably is a BFD because it says that certain basic capabilities are sufficient to be safe.

 

[EVNow]

Not an edge case, at all. Very very common, nothing unusual about it.

What do you think makes it an edge case?

What % of drives in US involve such ULT ?
Compared to say one or two lane ULT at 35 mph ?

 

[Supcom]

What % of drives in US involve such ULT ?
Compared to say one or two lane ULT at 35 mph ?

I have a similar ULT leaving my neighborhood about 25% of the time. Another 25% is a URT that has very high occlusion. I can point out similar ULTs all over my town and the adjacent town to me. Just pick any multilane divided artery in a city and you should have no problem finding similar intersections.

 

[LowlyOilBurner]

I doubt most people who currently own FSD paid $12k. (I know I didn't.)

I did. And I paid $6k two years ago for FSD on the car I traded in.

 

[AlanSubie4Life]

What % of drives in US involve such ULT ?
Compared to say one or two lane ULT at 35 mph ?

There are plenty of such situations, as has been mentioned.

I don't see how a 6-lane ULT is significantly different than a 4-lane ULT, really. Arguably, you can mostly ignore traffic in the furthest lane so you just have one additional lane to deal with. And 4-lane ULTs are very common! I do one nearly every day (see above), with traffic traveling at 45-70mph (there's a wide range, adds to the challenge, since the posted 45mph speed limit can't be legally enforced due to the California speed trap law (posted limit is lower than a traffic study would find is the norm, so no traffic study has been done, since people don't want a higher posted limit)!).

How are the requirements on FSD different for a 4-lane ULT (2 lanes each direction) vs. a 6-lane (3 lanes each direction) ULT, both with traffic traveling at up to 60mph?

 

[EVNow]

There are plenty of such situations, as has been mentioned.

…..

How are the requirements on FSD different for a 4-lane ULT (2 lanes each direction) vs. a 6-lane (3 lanes each direction) ULT, both with traffic traveling at up to 60mph?

You are not answering the question.

simple 35 mph ULT (say out of neighborhood or at traffic lights) are a LOT more common. They need to be dealt first. My guess is they cause 100x more disengagements than Chuck’s ULTs.

 

[AlanSubie4Life]

You are not answering the question.

simple 35 mph ULT (say out of neighborhood or at traffic lights) are a LOT more common. They need to be dealt first. My guess is they cause 100x more disengagements than Chuck’s ULTs.

Sure I would certainly agree those are more common. All I was saying is that Chuck’s turn is not an edge case and is very common - not that it represents the majority of turns or anything like that.

Still: What differences are there in capabilities required in the ULTs you suggest (let’s say no traffic light - I think we can agree that such turns are extremely common) vs. Chuck’s?

I just don’t really see much difference in what is required.

 

[Supcom]

Sure I would certainly agree those are more common. All I was saying is that Chuck’s turn is not an edge case and is very common - not that it represents the majority of turns or anything like that.

Still: What differences are there in capabilities required in the ULTs you suggest (let’s say no traffic light - I think we can agree that such turns are extremely common) vs. Chuck’s?

I just don’t really see much difference in what is required.

The ULTs I referred to earlier in my area are onto roads with 45-50 mph speed limits. We like to drive fast in Texas! I can also easily find 70-75 mph divided and undivided highways near me with at grade intersections. These might be interesting test cases for ULTs. Without single stack, it'll be FSD beta turning onto a road that will immediately enable NOA. That might be very interesting...

 

[AlanSubie4Life]

The ULTs I referred to earlier in my area are onto roads with 45-50 mph speed limits. We like to drive fast in Texas! I

Yeah, in California too.

Anyway my point, if it was not clear, was that the speed limit doesn’t really matter much, anyway. If the speed limit is 25mph, you definitely want to be able to detect someone coming at 70mph and clear the intersection prior to their arrival. This sort of situation is quite common and disaster is usually averted. (Some would argue that this is one of the primary advantages of autonomous vehicles in fact - they are always vigilant and watching for such events.)

I just fail to see what fundamental differences exist (in terms of what is required from the vehicle) between Chuck’s turn (which is common) and much more common turns.

Fundamentally, you always need visibility (measured in time) that exceeds the time that is required to complete the exposed maneuver.

 

[Supcom]

Yeah, in California too.

Anyway my point, if it was not clear, was that the speed limit doesn’t really matter much, anyway. If the speed limit is 25mph, you definitely want to be able to detect someone coming at 70mph and clear the intersection prior to their arrival. This sort of situation is quite common and disaster is usually averted. (Some would argue that this is one of the primary advantages of autonomous vehicles in fact - they are always vigilant and watching for such events.)

I just fail to see what fundamental differences exist (in terms of what is required from the vehicle) between Chuck’s turn (which is common) and much more common turns.

Fundamentally, you always need visibility (measured in time) that exceeds the time that is required to complete the exposed maneuver.

Well, its possible that someone could be coming at 100+ mph! That could be a bad result, regardless of who's at the wheel.

 

[Silicon Desert]

But you have to remember, that their definition of solved isn't necessarily the same as yours. They could consider waiting for a 10 second gap as acceptable. (Where you expect a maximum of a 5 second gap to be used.)

YES, I love your logic. I was thinking something similar. Hopefully not to get off the topic, but what is considered an "unprotected left turn", or actually what IS a "protected left turn?"

 

[AlanSubie4Life]

Well, its possible that someone could be coming at 100+ mph! That could be a bad result, regardless of who's at the wheel.

Obviously at SOME point, a collision would be unavoidable (for example recent horrific LA accident was unavoidable for most of those people). But the idea is to be WAY down the distribution (it’s quite common for someone to go 60mph in a 35mph zone, but much less common for them to be going 80mph) . And also, in a simple low traffic density situation, many humans would likely detect someone coming at 100mph in time, assuming they were actually visible.

The same fundamental idea applies though - there’s a big benefit to completing exposed maneuvers QUICKLY, since it reduces risk due to misjudgment of oncoming vehicle speed. And it lowers the requirement on required visual range (measured in time).

It’s something we don’t yet see happening in the apparent testing of 10.69; the vehicle is quite slow and lacks assertiveness which increases risk.

 

[Supcom]

YES, I love your logic. I was thinking something similar. Hopefully not to get off the topic, but what is considered an "unprotected left turn", or actually what IS a "protected left turn?"

An unprotected left is a left turn where you do not have right of way and must wait until it is safe to proceed. You typically have a stop sign, as opposed to a traffic light.

A protected left is one where there is a traffic light that gives you right of way to make the turn (green arrow).