FSD Beta 10.69

[AlanSubie4Life]

When critics are grasping at straws you know you are winning

Huh? There's not any competition here, really. Just trying to get over the hurdles to arrive at a useful driver assistance feature.

These are the basic failures we've seen for a while now. It's the really hard part to get right! They're not surprising at all - and also it's not grasping at straws.

The basic perception of vehicles and the go/no-go decisions seem pretty good now - a huge accomplishment over where things were before they started to try to solve this turn. And the complex dealing with people turning left, etc., also appears to be well handled. How reliably is this handled? Way too small a sample to know - we can say that part seems good with all of the recent attempts, but it is just such a small number of attempts, less than 100 so far, it is hard to know.

But when will it have the ability to communicate clearly (both with potential audio cues or preferably, just body language) to the driver what it plans to do?
Does it have the speed to complete the turns successfully? When will they increase crossing tempo? Why have they not already done so?
When will it be able to build confidence of the driver by logically taking available gaps, rather than missing them frequently?
When will it come remotely close to being good enough to reliably complete turns in a timely manner without intervention?

Those are the questions going forward. We'll continue to plateau at around this 90% success level until some of these items are closed, is my prediction. Nearly to the first 9, though! Then we can begin marching.

Where will we be in a year, I wonder?

To me it seems like the main issues right now are the tuning of jerk and acceleration to clearly communicate what is going on and to make maneuvers fast enough. Hard to know how reliable the perception is - I'm assuming it's ok, but I don't really know.

In any case my FSD investment continues to pay off, with my super secret inside knowledge of its workings, allowing me to take (fair) advantage of people who have not used it themselves. Paid off 1% so far.

 

[AlanSubie4Life]

think currently it is just too slow to figure out whats happening. It can't be assertive when its uncertain.

Are you say it's slow to determine a plan based on the perception latency? That seems unlikely - seems like that should occur in a superhuman fashion.

Regarding the uncertainty & slowness - I'm not talking about the creeping behavior (though that needs work clearly on the jerk & acceleration profile). I doubt it is lacking certainty after it has made the go/no-go decision. I certainly hope it's not uncertain (it could be, as I said, we'd have to have a much larger sample size to determine whether there is enormous danger here)! It just need to go more assertively when it has made that decision. And strangely, it is not. Just very, very slow to cross. Done the analysis many times here and it's clear there is margin for 20-30% faster crossing speed with no impact on user comfort (in fact it would help, and Chuck has commented on this many, many times).

It's weird it is not done since it seems like it should improve safety. If it really is a certainty issue, that is troubling, because that would imply it goes slowly because Tesla knows the car can proceed on a collision course with an oncoming vehicle which is clearly visible, in some circumstances. We just haven't seen it yet. If that can happen, lower crossing speeds would be safer since it would allow slightly more time for driver disengagement. But it would still be exceedingly dangerous.

 

[kabin]

Chuck with the jitters means another beer for me from @Daniel in SD !!! (Combining this with the earlier 6 attempts from 10.69.3 - which we counted as 6/6 even though it was actually 5/6 - you’ll recall my magnanimity.)

This was the version that had the better cross-traffic speed estimation (wasn't it?), and that seems to have led to some regression in its ability to shoot gaps.

As discussed before, it's essential that it be more aggressive in crossing the road. It will not be comfortable at all until it drives more assertively.

Agreed, 10.69.3.1's UPL decision making seemed a little off. Gotta wonder if the improved cross speed estimation was the result from filtering as that would add some lag time.

The Memorial drive near miss could also possibly be related to a laggy cross traffic velocity/angle/range estimation.

 

[2101Guy]

Huh? There's not any competition here, really. Just trying to get over the hurdles to arrive at a useful driver assistance feature.

These are the basic failures we've seen for a while now. It's the really hard part to get right! They're not surprising at all - and also it's not grasping at straws.

The basic perception of vehicles and the go/no-go decisions seem pretty good now - a huge accomplishment over where things were before they started to try to solve this turn. And the complex dealing with people turning left, etc., also appears to be well handled. How reliably is this handled? Way too small a sample to know - we can say that part seems good with all of the recent attempts, but it is just such a small number of attempts, less than 100 so far, it is hard to know.

But when will it have the ability to communicate clearly (both with potential audio cues or preferably, just body language) to the driver what it plans to do?
Does it have the speed to complete the turns successfully? When will they increase crossing tempo? Why have they not already done so?
When will it be able to build confidence of the driver by logically taking available gaps, rather than missing them frequently?
When will it come remotely close to being good enough to reliably complete turns in a timely manner without intervention?

Those are the questions going forward. We'll continue to plateau at around this 90% success level until some of these items are closed, is my prediction. Nearly to the first 9, though! Then we can begin marching.

Where will we be in a year, I wonder?

To me it seems like the main issues right now are the tuning of jerk and acceleration to clearly communicate what is going on and to make maneuvers fast enough. Hard to know how reliable the perception is - I'm assuming it's ok, but I don't really know.

In any case my FSD investment continues to pay off, with my super secret inside knowledge of its workings, allowing me to take (fair) advantage of people who have not used it themselves. Paid off 1% so far.

To answer your question of where will we be in a year? Not sure but I can tell you with absolute certainty where we won’t be: Robo taxi service that was promised for end of 2019.

 

[mhan00]

DT had some of the best roundabout behavior I’ve seen from beta. Excellent drive up to that point, and then a bunch of weird slowdowns requiring foot on go pedal interventions to keep from impeding drivers behind the car, lol.

 

[Mardak]

Anybody else noticing 10.69.3.1 being quite courteous for pedestrians at crosswalks? The visualization turns blue for the dog walker probably over 5 seconds before they would even enter the road if continuing straight, but they end up turning:

These people with a stroller turn blue at 0:30 and cross the visualized red road edge line at 0:37:

 

[EVNow]

Done the analysis many times here and it's clear there is margin for 20-30% faster crossing speed with no impact on user comfort (in fact it would help, and Chuck has commented on this many, many times).

I think they are trying to optimize for various mechanical comfort items .... but not the most important one, fear of collision.

It might be perfectly safe the way they are doing some things .... but as test drivers we are not convinced they are being safe enough. Tesla will have to take that into account.

Good example is the creeping wall. It might be perfectly safe - but we want some buffer between the lane in front and how far the car would creep. Or, when a car is coming hurtling down from the left, we want to get out of the way fast. When approaching a roundabout, if a car is coming from the left, we want to slow down well in advance to communicate to the test driver that the car will stop and to the vehicle from the left that they will indeed get right of way.

Huh? There's not any competition here, really. Just trying to get over the hurdles to arrive at a useful driver assistance feature.

These are the basic failures we've seen for a while now. It's the really hard part to get right! They're not surprising at all - and also it's not grasping at straws.

Thus the at the end.

I'm just happy they have come a long way compared to last year.

 

[powertoold]

It'd be a big letdown if the only difference between 69.3.1 and 11 was:

enableFSDStackOnHighway = true

 

[willow_hiller]

It'd be a big letdown if the only difference between 69.3.1 and 11 was:

enableFSDStackOnHighway = true

I hope we see better performance overall, just due to no longer running multiple processes on top of each other.

If you disengage FSDb with the steering wheel, it instantly kicks into TACC with the original AP visualization. This makes me think that they may be running the original AP stack in the background of FSDb all the time.

 

[EVNow]

It'd be a big letdown if the only difference between 69.3.1 and 11 was:

enableFSDStackOnHighway = true

Well, V11 will definitely be different on highways. Totally different networks. Hopefully better ... but I expect better in some ways and worse in some others.

 

[powertoold]

I hope we see better performance overall, just due to no longer running multiple processes on top of each other.

If you disengage FSDb with the steering wheel, it instantly kicks into TACC with the original AP visualization. This makes me think that they may be running the original AP stack in the background of FSDb all the time.

Hope so, but I doubt it. The time to switch one set of NNs for another on the compute is dependent on the cache / ssd loading speed, which I’d imagine is less than a second. This is just my speculation of course.

 

[kabin]

10.69.3.1 looks smoother but also slower to make decisions. Overall it looks like a step in the right direction but they need to find a way to cut decision making time by at least a half from the current 5 to 7 secs. I get the feeling all available real time h/w processing is tapped out.

And those 30-50ft early slow/stops at stop signs with subsequent slow progression to the cross walk are confusing to pedestrians and drivers. Not sure if that's something the team is using to hide processing/decision-making time.

 

[DanCar]

It'd be a big letdown if the only difference between 69.3.1 and 11 was:

enableFSDStackOnHighway = true

Still exciting to see how the new highway software works. When will pothole detection work?

 

[powertoold]

When will pothole detection work?

Once fsd beta stops doing this:

 

[jeremymc7]

10.69.3.1 looks smoother but also slower to make decisions. Overall it looks like a step in the right direction but they need to find a way to cut decision making time by at least a half from the current 5 to 7 secs. I get the feeling all available real time h/w processing is tapped out.

Accuracy is of course a solid choice. But I already “feel” like the builds are getting slower. It also “feels” like the processor is just not able to keep up; but that be right either.

The car even currently seems to take forever in situations where it appears to have clear visibility and potential threats. It needs to get to a point where others are much less annoyed by beta.

 

[thewishmaster]

Once fsd beta stops doing this:

What a mega awkward interaction lol

 

[Stefanv1993]

Does anyone know if this release resets the strikes?

 

[jeremymc7]

Does anyone know if this release resets the strikes?

Good question. I haven’t seen that happen in a while; for whatever reason.

I only have one currently; at an intersection with a red light oddly enough.

Isn’t it currently five strikes; when it was three previously?

 

[TS.42]

DT had some of the best roundabout behavior I’ve seen from beta. Excellent drive up to that point, and then a bunch of weird slowdowns requiring foot on go pedal interventions to keep from impeding drivers behind the car, lol.

No offense but even my wife said “He declares every update the best ever”. Hmmm, not wrong.

 

[Sigma4Life]

No offense but even my wife said “He declares every update the best ever”. Hmmm, not wrong.

Shouldn’t every update be better than the previous one? We are not making progress if that’s not the case.