FSD Beta 10.12

[johnmodely]

This version is still unusable in NYC. FSD simply can't figure out a path with 10+ cars in an intersection and dozens of pedestrians. Id love to see how Tesla will solve this one.

If they want robo taxies they will need them in the two major markets (ie LA/NYC). They are not even attempting to solve any dense urban areas right now (and all the youtubers/influencers are no where to be seen in these areas which is interesting as well).

Main issue is pedestrian logic, right now FSD yields to all pedestrians, which is an impossible logic to have in NYC. You have to assume you have the right of way and pedestrians will yield even when there are 20 of them hanging on the side of the street, otherwise you'll never move.

I get they have to be super cautious with pedestrians, but how do you maintain zero tolerance for pedestrian safety and still navigate in NYC?

 

[JHCCAZ]

...I get they have to be super cautious with pedestrians, but how do you maintain zero tolerance for pedestrian safety and still navigate in NYC?

Tesla's goal is to not be geo-fenced. But I don't think there's any way to avoid that it will have to be geo-responsive. Behavior that is required in one region can immobilize you, get you a ticket or start a fight in another region.

I think that as utonomous vehicles become commonplace, this will get easier. Traffic and pedestrians will adapt to the new situation, but we're certainly not there today.

 

[Reddy Kilowatt]

Earlier, both train crossing and pedestrian crossing ahead signs, neither visualized. Once closer, there was a vague visualization of something at the train signal location but no flashing lights. Stopped/slowed several cars back. Truck in front crept very slowly forward, perhaps without completely stopping (no clear brake lights). Movement to left was perceptible and indicated by gray line (no blue). FSD may have changed its mind about passing the cars as I disabled it. Two second, three second, and four second fsd frames with left view at beginning and right view at end.

View attachment 812534

 

[ItsNotAboutTheMoney]

This version is still unusable in NYC. FSD simply can't figure out a path with 10+ cars in an intersection and dozens of pedestrians. Id love to see how Tesla will solve this one.

If they want robo taxies they will need them in the two major markets (ie LA/NYC). They are not even attempting to solve any dense urban areas right now (and all the youtubers/influencers are no where to be seen in these areas which is interesting as well).

Main issue is pedestrian logic, right now FSD yields to all pedestrians, which is an impossible logic to have in NYC. You have to assume you have the right of way and pedestrians will yield even when there are 20 of them hanging on the side of the street, otherwise you'll never move.

I get they have to be super cautious with pedestrians, but how do you maintain zero tolerance for pedestrian safety and still navigate in NYC?

The current defensiveness is only necessary because the error rate is so high.

If they get the error rate low enough, they can assert their right of way and rather than anticipate assholes and dumbasses, react to them.

 

[nvx1977]

The truck certainly cannot be seen by the forward cameras because the car is turned slightly to the right.

+ @aronth5

But if the car is turned to the right, shouldn't the left repeater camera have visibility on the small truck? Might be worth mounting the gopro at the repeater, matching the camera angle and see what it sees (or just turn on the rear view with repeaters enabled).

 

[tmoz]

https://twitter.com/elonmusk/status/1533179948743151616?s=20&t=1QeJ7d_y0SOGUqT_duye7A
Start the 10.13 thread

 

[Supcom]

IMO this is the achilles heal of FSD. The reliance on the B-pillar camera and the fact the forward cameras cannot see 90 degrees left/right is going to diminish FSD when accidents get reported once you have owners who rely on FSD without carefully watching every turn.

The two things the car needs to do to resolve this is to recognize when it's view is not sufficient to make a turn decision and to angle the car during the creep to bring the road in view of the left repeater. I proved that the left repeater can get a perfect view of the road.

Currently, the car does angle itself, but only slightly. And based on more the ten tests of this one intersection, I can say that the car will attempt to enter the intersection if it does not see a car during it's current creep. What it does not appear to realize is that it does not have a clear view far enough down the lane it is about to enter.

I think it's quite doable, but that the car is not adequately trained for it yet.

 

[JulienW]

https://twitter.com/elonmusk/status/1533179948743151616?s=20&t=1QeJ7d_y0SOGUqT_duye7A
Start the 10.13 thread

Welcome to yesterday.

FSD Beta 10.13

teslamotorsclub.com

 

[Supcom]

+ @aronth5

But if the car is turned to the right, shouldn't the left repeater camera have visibility on the small truck? Might be worth mounting the gopro at the repeater, matching the camera angle and see what it sees (or just turn on the rear view with repeaters enabled).

I was able to access the dashcam for the left repeater and for where the car creeps on FSD, the road is not in the field of view. The car must be angled a little more to the right.

 

[Oil Freedom]

A few observations for where I drive:

Better:

• Confidence and smoothness making turns.
• Handling of crossing traffic and vehicles turning out of path (less unnecessary slowing).
• Merging with traffic to/from freeway ramps.
• Lane selection (still makes some mistakes, but correction is improved, and also fewer random turn indicators for bike lanes).
• Less jerkiness navigating in parking lots, but still unpleasant experience.
• Recognizing and showing cars with turn indicators (new feature but still work in progress based on intermittent success / failure).
• More detailed representation of other vehicles (bonus: also works in garage).
• Spot for one more icon is a plus! (now want another)

Worse:

• The Waze for Tesla web page doesn't work (again) because this version broke the location info (again). Hope that's fixed (again) soon! SMH

Status quo (figured these would be fixed by now!):

• Still doesn't read the No Right Turn on Red signs so have to hit the brakes.
• Still takes curves too widely on the freeway and surface streets (to the point other drivers in the neighboring lane would get nervous, so I take over). It's weird behavior since the car shows that it's not centered in its lane, so it's intentionally going wide and not correcting itself.

 

[Oil Freedom]

Now that we had the wide, release, here is the final chart. Took 52 days from last major release.

The question is - if this is the level of change we can see after nearly 2 months, how can Tesla reach human level driving in 6 more months i.e. 3 more releases?

View attachment 811808

It's a fair point, but I suspect Tesla spent a lot of this time and their engineering resources not just improving the driving but moving the solution closer to single stack.

 

[jabloomf1230]

Worse:

• The Waze for Tesla web page doesn't work (again) because this version broke the location info (again). Hope that's fixed soon! SMH

It works for me. Did you get the new location permission prompt from Chromium when you ran TeslaWaze? It asks to either allow once, allow forever or prevent?

The bad news is that TeslaWaze crashes the browser every ten minutes or so. It appears to occur when new map data loads.

 

[EVNow]

It's a fair point, but I suspect Tesla spent a lot of this time and their engineering resources not just improving the driving but moving the solution closer to single stack.

To me its definitely one or two steps backward. What they have done is made some platform level changes. I don't think it is showing up in real world testing ... to me atleast.

In the drive I'm just coming back from I had more disengagements than any other drive this year ...

 

[Oil Freedom]

It works for me. Did you get the new location permission prompt from Chromium when you ran TeslaWaze? It asks to either allow once, allow forever or prevent?

The bad news is that TeslaWaze crashes the browser every ten minutes or so. It appears to occur when new map data loads.

It only read something like "Location data unavailable." without any prompting for permission like it normally does. It kept crashing for me on the prior version, but at least it worked some of the time. I'll keep trying...

 

[Oil Freedom]

To me its definitely one or two steps backward. What they have done is made some platform level changes. I don't think it is showing up in real world testing ... to me atleast.

In the drive I'm just coming back from I had more disengagements than any other drive this year ...

Exactly... I think they're making more foundational changes to better position themselves for the longer term. The current architecture is probably holding things back.

 

[drtimhill]

To me its definitely one or two steps backward. What they have done is made some platform level changes. I don't think it is showing up in real world testing ... to me atleast.

In the drive I'm just coming back from I had more disengagements than any other drive this year ...

It's mixed for me .. I've seen lots of stuff handled better, but today the car suddenly swerved to the left mid-way through an intersection ½ way into the oncoming traffic lane, with no apparent cause. I've noticed some "phantom cars" showing up on the display, which seem to be related to the car seeing reflections in the side windows of other cars as real, and wondering if this triggered this swerve. Either way, I reported to Tesla with an email also.

 

[Oil Freedom]

It's mixed for me .. I've seen lots of stuff handled better, but today the car suddenly swerved to the left mid-way through an intersection ½ way into the oncoming traffic lane, with no apparent cause. I've noticed some "phantom cars" showing up on the display, which seem to be related to the car seeing reflections in the side windows of other cars as real, and wondering if this triggered this swerve. Either way, I reported to Tesla with an email also.

Your post reminded me: While driving on a surface street and approaching a parked truck (18 wheeler) that was off to the right and only slightly in my way/path, FSD Beta seemed at first like it was going to drive around it (There was plenty of space, and I've seen it successfully complete similar maneuvers before.), but this time it suddenly veered to the left like it was going to make a hard left turn instead.

Obviously, I had to quickly take over and correct. And, there were cars coming from the other direction, but they weren't very close quite YET. I'm sure they thought I had spilled a very hot coffee into my lap or something!

 

[tinkererguy]

I did some limited testing yesterday on normal work route and the new visualizations are cool but I had to do the same interventions/disengagements I was doing with 10.11.2 so no real difference I can tell. I did not experience any red hands of death. Speed bump visualizations are still not back and it fails to recognize and slow at about the same percentage as before.

Yeah, testing 10.12.2 today, speed bump (or speed hump) visualizations (and slow down) are definitely not back yet, they used to render just fine at this same clearly marked location, I think it was two #FSDBeta releases ago.

 

[AlanSubie4Life]

but assuming the car can accurately estimate the distance

Yeah, not clear how good/consistent their distance estimation is. Still, it should be able to do a lot better than it is doing. Inconsistent distance estimation might lead to less smooth slowing but even that could probably be smoothed.

It seems to do a little better at traffic lights (though it is far from good), which is odd.

I take this back. A bit more driving today, and I noted that it behaves apparently exactly the same for traffic lights. Just less of an issue typically, since it’s reasonable to expect to wait a bit at traffic lights, so minimizing the run-up time to the light is not as important as it is for a stop sign. But definitely an issue when coming up to red lights where you will be turning right, etc.

And I guess I don’t understand the people saying it is more aggressive. It may be that on occasion it will mildly gun it in a turn, where it did not before, but certainly no alacrity on the starts from a light. Maybe people are just commenting on that subtle change. I feel in aggressive mode you should be doing 50mph in 5-6 seconds (don’t take exact numbers here literally - I would have to time it, this is an estimate of “reasonable”) from a light. Maybe at some point they’ll tune this. But right now it is frustratingly slow. Maybe a bit cautious on following traffic from a light, too. Overall perhaps not that bad but should start moving more quickly when traffic in front starts moving and then open up the space to a comfortable distance. Rather than creating a ton of space right away. Not sure I’ll ever figure out their algorithm since typically I am accelerator overriding - just tried not doing it a couple times today to observe.

 

[aronth5]

The two things the car needs to do to resolve this is to recognize when it's view is not sufficient to make a turn decision and to angle the car during the creep to bring the road in view of the left repeater. I proved that the left repeater can get a perfect view of the road.

Currently, the car does angle itself, but only slightly. And based on more the ten tests of this one intersection, I can say that the car will attempt to enter the intersection if it does not see a car during it's current creep. What it does not appear to realize is that it does not have a clear view far enough down the lane it is about to enter.

I think it's quite doable, but that the car is not adequately trained for it yet.

I was able to access the dashcam for the left repeater and for where the car creeps on FSD, the road is not in the field of view. The car must be angled a little more to the right.

Is the "left repeater" the same as the "B-Pillar"? The problem comes when the obstruction is so bad FSD needs to creep the car into the intersecting road so far it's impeding on coming traffic. I have several T-intersections where this occurs and in that situation you're basically screwed. The human driver leans way forward gaining 2+ feet which is significant. FSD cannot safely do the same. Turning further to the right helps but doesn't solve the inherent problem.