FSD Beta 10.69

[aronth5]

These videos are straight up spam. Why post them here? He's also using a nanny defeat device. That should not be promoted.

Pretty simple drive so difficult to assess how well FSD did.

 

[Knightshade]

I've had mine do roundabouts without issue- but they tend to be simple single-lane ones here, and they're pretty uncommon so it's only been a few times

 

[sleepydoc]

Lots of roundabouts in my suburb here.

They will get to roundabouts one of these days - like they tackled unprotected left turns.

It's odd that they haven't done better on roundabouts. They're nothing but unprotected right turns which are actually easier than left turns. Regardless, they need to handle the same logic for the dedicated right turn lanes with yield signs that are incredibly common at intersections.

 

[2101Guy]

Yes, sure but the headline is click bait and claims “50% improvement” over 69.2.2 without anything to back it up.

Oh, no doubt. In fact, I wouldnt be suprised if he is still on 69.2.2

 

[2101Guy]

It's odd that they haven't done better on roundabouts. They're nothing but unprotected right turns which are actually easier than left turns. Regardless, they need to handle the same logic for the dedicated right turn lanes with yield signs that are incredibly common at intersections.

because there apparently arent any roundabouts between the NYC to LA route that was promised to happen 5 YEARS ago

 

[2101Guy]

it always goes like this:
- Day 1 of the new release: "Man, this version is SO much smoother and better. I think its so close to being almost there. This is truly groundbreaking!'
- Day 2 (same driver): "Wow. This thing stops mid intersection. and doesnt do roundabouts at all. And the pb is worse. I keep having to hover my foot over the accelerator".

EVERY TIME.

 

[Ramphex]

because there apparently arent any roundabouts between the NYC to LA route that was promised to happen 5 YEARS ago

They’ll add a navigational toggle to “avoid roundabouts” before they teach FSD how to navigate them

 

[jabloomf1230]

It's odd that they haven't done better on roundabouts. They're nothing but unprotected right turns which are actually easier than left turns. Regardless, they need to handle the same logic for the dedicated right turn lanes with yield signs that are incredibly common at intersections.

Single lane roundabouts meet your criteria. But multilane roundabouts may require either a lane change or "cutting off" the traffic in the outer loop. It's not trivial behavior, especially considering that a significant fraction of US drivers don't understand how to navigate a roundabout.

 

[willow_hiller]

My biggest problem with the roundabouts in my area is that it does treat them like an unprotected right. Complete stop, creeping for visibility, etc. Granted, many drivers in my area do the same thing, but it's controlled with a yield sign for a reason, not a stop sign.

I imagine Tesla might be waiting until they're able to improve the confidence of the vision system before fully implementing roundabouts. Because to my mind, the optimal way to navigate them is to reduce speed slightly, find a gap, and then accelerate to match the speed of the vehicles in the circle. And all of that will require pretty precise and rapid trajectory estimation without coming to a complete stop, unless necessary.

 

[Twiglett]

I just played "spot the difference" between the 2020.69.2.2 and 69.2.3 release notes
Can't find a single difference

 

[willow_hiller]

Installations seem to have slowed down for the night. I have a theory as to why:

 

[powertoold]

2.3 disengages itself, counts as "zero" disengagements:

 

[aronth5]

It's odd that they haven't done better on roundabouts. They're nothing but unprotected right turns which are actually easier than left turns. Regardless, they need to handle the same logic for the dedicated right turn lanes with yield signs that are incredibly common at intersections.

Roundabouts are definitely different then UPL's since FSD has to decide whether to merge with traffic, stop or enter the roundabout with minimal slow down depending on the roundabout design. Plus FSD has to decide which exit to take and negotiate around the roundabout which is very different then a UPL turn.

The large single lane roundabout I take frequently with no cars coming should be entered between 15-20mph. 10.69.2.2 is the first release where FSD does not come to a complete stop. Now it slows down to 2 mph which at least helps a little with drivers behind me.

 

[bhakan]

Following is from my recent drive. I don't have 69.2.3 yet.

1. I was on FSDb and making a left turn onto Chapman Ave ( Google map for context). There was a big garbage truck or something parked on the right lane at the corner and my car turned to left lane. A pedestrian crossed the road from the right as he could not see my car because of the blocked truck and Tesla slowed and stopped even tho it was not a crosswalk. The pedestrian just fell down on his own as he was surprised. I was shaken even tho my car stopped and disengaged the AP. This could be any car that could be there when the guy was crossing and fell down. See the dashcam link below as I don't have a video recorder in the car. I sent the snapshot to Tesla.




Untitled

2. I still have cabin camera issues that Tesla should fix as it does not appear to be software issue. But my SC wants me to work with FSD team. oh well. Since I have two strikes due to cabin camera issues while driving at night if I have to drive in the evening as @Dewg recommended I am driving with bright light shining on my face from my screens. Not happy.

 

[willow_hiller]

2.3 disengages itself, counts as "zero" disengagements:

I wonder if that's unintended given the design of the yoke. A quick rotation with all of the weight on one side generates enough torque to disengage. I'd be curious to see if that happens on a 3/Y wheel.

 

[sleepydoc]

The difference is with roundabouts without cars coming the car should enter the roundabout with minimal slow down depending on the roundabout design. Plus FSD has to decide which exit to take and negotiate around the roundabout which UPL's don't have to do.

The large single lane roundabout I take should be entered between 15-20mph with no cars coming 10.69.2.2 is the first release where FSD does not come to a complete stop. Now it slows down to 2 mph which at least helps a little with drivers behind me from honking.

Yeah, I'm guessing FSD still needs more time to analyze traffic and assess whether to go or wait. With an UPL you always need to stop so you will have more time to make this assessment. With a UPR/yield sign you need to do it on the fly and be prepared to proceed directly so if the algorithm is still too slow it needs to stop.

 

[Dewg]

I wonder if that's unintended given the design of the yoke. A quick rotation with all of the weight on one side generates enough torque to disengage. I'd be curious to see if that happens on a 3/Y wheel.

That was my thought - it seems the wheel jerks a bit at the moment of disengagement, and the tone sounds like a torque disengagement.

 

[powertoold]

https://twitter.com/x/status/1577783872896409601

 

[AlanSubie4Life]

2.3 disengages itself, counts as "zero" disengagements:

Whompy wheel and alignment from the massive curb strike which caused no problems a while back.

 

[GranularHail]

I got 10.69.2.3 today, which I didn't expect because I was in the last batch to get 10.69.2.2, so it sure seems like it could be random who gets it when.

I went out for a quick drive, and didn't notice any obvious difference vs. 10.69.2.2. I'll get a much better feel for it after a full commute tomorrow where I know what 10.69.2.2 was good at and not good at.