FSD Beta 10.13

[jebinc]

Bold prediction

Yes, very bullish!

 

[tmoz]

I just got Version 11!
But I also have HW1.
So I think it will mainly be finding where they put my cheese again.

 

[Electroman]

Just like every release, be prepared to get disappointed with this upcoming release, and then come to the realization after a few days it is actually marginally improved.

 

[Todd Burch]

what good is your life being safe 50.1% of the time....... I hope you don't want to turn around and come home the same way.

The flaw is in your thinking that remarkable technical achievements come along like the flip of a switch on a single day.

 

[jebinc]

 

[Silicon Desert]

So? Even if “just” 50.1% of the time, that would be a remarkable achievement. No other autonomous driving company can even attempt such a challenging maneuver.

People in these FSD beta threads have far too little appreciation for how challenging this problem is.

yup this ^^^

 

[GranularHail]

While we wait for 10.13.??
Just another example of how FSD performs differently from car to car. FSD for me works best on rural roads. Almost always a zero disengagement drive. Hard braking for the most part is gone and while FSD makes alot of mistakes there are very few places where I won't enable FSD. For now I'll give FSD a "C+" grade compared to when I first got it last October when it was at best a "D". For me like others lane selection is the biggest problem.

Same - rural roads are great for me except for phantom blinkers approaching intersections where the road splits into turn lanes and a straight lane.

The areas I find entirely un-usable are multi-lane roads, where my car (on FSD's "Chill" profile) can't go a half mile without changing lanes. The lane changing does often correlate to specific locations, but not to traffic conditions. It'll move towards slow cars, away from slow cars, out of the way of a car behind me, into the way of a car behind me, or change lanes when no cars are around, all with equal likelihood.

 

[S4WRXTTCS]

Just like every release, be prepared to get disappointed with this upcoming release, and then come to the realization after a few days it is actually marginally improved.

This release is different than the others though.

It has the clear purpose of doing a lot better on Chucks unprotect left.

Heck it's not even really about how we feel, but Chuck.

Chuck is the Judge, Jury, and Executioner of 10.13

 

[JHCCAZ]

This release is different than the others though.

It has the clear purpose of doing a lot better on Chucks unprotect left.

Heck it's not even really about how we feel, but Chuck.

Chuck is the Judge, Jury, and Executioner of 10.13

Some people predict it will continue to fail too often on this famous UPL, others that it will do well because of the specific training - but may then be "over-fit" and not a general Improvement.

I'm excited to see if the improvements (however great or small they may be) will be generalized to a widening list of UPL and UPR cases:

• That particular turn in different time of day, lighting, weather, or seasons of the year
• Similar turns nearby, maybe with better visibility
• Similar turns traffic-wise, but in different-looking environments
• UPRs onto the same or similar roads
• Oncoming-traffic UPLs (Chuck has one or two standard test-cases for these also, and then of course there's an infinite variety of these in everyday driving)

So at this point, I personally care less about how flawlessly Chuck's specific UPL is handled than I do about how generalized the improvements prove to be. To me this is a fascinating question hanging over the whole training approach.

And I'll say once again that it would be fine, probably even better, if the car were to convert that particular UPL into a right turn plus U-turn at the light.

 

[Electroman]

Some people predict it will continue to fail too often on this famous UPL, others that it will do well because of the specific training - but may then be "over-fit" and not a general Improvement.

I'm excited to see if the improvements (however great or small they may be) will be generalized to a widening list of UPL and UPR cases:

• That particular turn in different time of day, lighting, weather, or seasons of the year
• Similar turns nearby, maybe with better visibility
• Similar turns traffic-wise, but in different-looking environments
• UPRs onto the same or similar roads
• Oncoming-traffic UPLs (Chuck has one or two standard test-cases for these also, and then of course there's an infinite variety of these in everyday driving)

So at this point, I personally care less about how flawlessly Chuck's specific UPL is handled than I do about how generalized the improvements prove to be. To me this is a fascinating question hanging over the whole training approach.

And I'll say once again that it would be fine, probably even better, if the car were to convert that particular UPL into a right turn plus U-turn at the light.

Ha! Tesla engineers are testing on Chuck's UPL.

https://twitter.com/x/status/1549881782211907586

 

[HappyPanda]

And I'll say once again that it would be fine, probably even better, if the car were to convert that particular UPL into a right turn plus U-turn at the light.

I have a busy street similar to Chuck's near my house, but the median blocks you from taking a left turn directly to the opposite side. Instead, there are dedicated U-turn spots, so you have to turn right (going into the left-most lane) and U-turn. As far as I know, FSD can't do this currently, but it would be nice to see! I did try to let it do a U-turn while already on the road when there wasn't traffic, but it just ended up making a left instead.

 

[S4WRXTTCS]

And I'll say once again that it would be fine, probably even better, if the car were to convert that particular UPL into a right turn plus U-turn at the light.

There is no disagreement that it would be better to avoid a UPL and especially one like that.

But, and its actually a But^2.

The navigation isn't setup to avoid difficult situations. In fact I couldn't use FSD Beta to go to work because the navigation wants to take me to an UPL over multiple lanes of high speed traffic. So instead of follow the navigation I avoid that road, and I stay on the main road to get to the light.

Elon is pretty obsessed with the whole non-geofenced thing and routing for ease smells a lot like geofencing. In fact I'd call myself an L4 driver because I geofence myself quite a bit. Like you think I'm going anywhere near a chick-fil-a at noon?

 

[JHCCAZ]

There is no disagreement that it would be better to avoid a UPL and especially one like that.

But, and its actually a But^2.

The navigation isn't setup to avoid difficult situations. In fact I couldn't use FSD Beta to go to work because the navigation wants to take me to an UPL over multiple lanes of high speed traffic. So instead of follow the navigation I avoid that road, and I stay on the main road to get to the light.

Elon is pretty obsessed with the whole non-geofenced thing and routing for ease smells a lot like geofencing. In fact I'd call myself an L4 driver because I geofence myself quite a bit. Like you think I'm going anywhere near a chick-fil-a at noon?

I couldn't agree more that the navigation needs to adjust for safer / more sensible choices to avoid hazardous situations. This is simply part of smart defensive driving and I don't see it as and any admission of failure or incapability.

It's obviously a matter of degree - if the car simply refused all UPLs, that wouldn't be very good either.

Regarding navigation with City Streets, I would be in favor of a "guided autopilot" mode that allows the car to drive autonomously, but waits for your input to execute your choice of turns. I have several ideas for how this could work smoothly.

 

[JHCCAZ]

I have a busy street similar to Chuck's near my house, but the median blocks you from taking a left turn directly to the opposite side. Instead, there are dedicated U-turn spots, so you have to turn right (going into the left-most lane) and U-turn. As far as I know, FSD can't do this currently, but it would be nice to see! I did try to let it do a U-turn while already on the road when there wasn't traffic, but it just ended up making a left instead.

Right, the consensus is it simply isn't programmed to do U-turns yet. Clearly that needs to come anyway, and then the navigation logic for right turn plus U-turn can follow.

 

[BBTX]

It can do some sorts of U-turns here in TX, when there's a dedicated U-turn lane carved out (physical curb barrier) on a highway feeder road underneath a highway bridge (as shown in Texas U-turn - Wikipedia ). It has mixed success, but it does usually/mostly work. There are two key problems I run into when Nav+FSDb sends me through one of these, both kinda intermittent depending on location + conditions:

1. They're often ~2+ cars wide through most of the turn shape, which is to allow large commercial trucks to use them with their wider turning needs, but there's no lane markings inside them and they're definitely intended to be used as if it's a single lane. At least some of the time, FSDb seems to assume there's two lanes here due to the width, and I've seen others report it trying to pass stopped cars inside these lanes (hasn't happened to me yet, but seems very plausible), when it should be queuing up behind them.
2. For some reason, at least on a few that I commonly use near my house, it sometimes is very hesitant on entry. I've seen the same turn sometimes go flawlessly and proceed on its own, but I've also see it slow down to a near-stop and proceed very cautiously on turn-in, and I've seen it come to a complete stop there as well (both of which were fixable by pushing the go-pedal).

 

[Callahan27]

While we wait for 10.13.??
Just another example of how FSD performs differently from car to car. FSD for me works best on rural roads. Almost always a zero disengagement drive. Hard braking for the most part is gone and while FSD makes alot of mistakes there are very few places where I won't enable FSD. For now I'll give FSD a "C+" grade compared to when I first got it last October when it was at best a "D". For me like others lane selection is the biggest problem.

My car is the same way. With the most recent update FSD is much more usable for me just about everywhere. Sure it struggles with things here and there but most normal driving it does fine. I look forward to 10.13

 

[Supcom]

It can do some sorts of U-turns here in TX, when there's a dedicated U-turn lane carved out (physical curb barrier) on a highway feeder road underneath a highway bridge (as shown in Texas U-turn - Wikipedia ). It has mixed success, but it does usually/mostly work. There are two key problems I run into when Nav+FSDb sends me through one of these, both kinda intermittent depending on location + conditions:

1. They're often ~2+ cars wide through most of the turn shape, which is to allow large commercial trucks to use them with their wider turning needs, but there's no lane markings inside them and they're definitely intended to be used as if it's a single lane. At least some of the time, FSDb seems to assume there's two lanes here due to the width, and I've seen others report it trying to pass stopped cars inside these lanes (hasn't happened to me yet, but seems very plausible), when it should be queuing up behind them.
2. For some reason, at least on a few that I commonly use near my house, it sometimes is very hesitant on entry. I've seen the same turn sometimes go flawlessly and proceed on its own, but I've also see it slow down to a near-stop and proceed very cautiously on turn-in, and I've seen it come to a complete stop there as well (both of which were fixable by pushing the go-pedal).

I have the same experience as you on the Texas U-turns. Generally, good, though the car tends to stop on entry. Likely keying off of the traffic lights at the intersection to the right.

 

[JohnB007]

I would grade the current version a solid B. Yes, it rarely gets me home by itself completely but it is on 90% of the drive. I do experience my share of PBs too. Still, I find this far more usable than not having any assist features at all. Much better than the BMW system I had prior to Tesla.

It is so nice that it recognizes stop signs, red lights and performs simple right turns. Reliable too. So much more relaxing doing road trips (especially new areas) and my wife tells me she can clearly feel it. When I'm on the section where I know the Beta can handle it, it is always on.

Looking forward to experiencing 10.13. Hope it gets to B+ in my scale at least!

 

[rjack1282]

I wouldn't consider "smart routing" to be geofencing as a lot of human drivers (including me) do this from time to time to avoid difficult maneuvers that may not be necessary to get from point A to point B. I'm sure we've all seen Waze routing want you to take a side street and then a UPL to avoid a traffic light where the left turn may be backed up. I generally ignore that routing and opt to wait a couple cycles at the light for example if I know the UPL is crossing several lanes on a busy street. I may be stuck at that UPL longer than just waiting a couple cycles at the traffic light....not to mention the protected turn at the light is much safer.

 

[Bill Foster]

Am I missing something? Teslafi is only showing one install so far?