FSD Beta 10.69

[AlanSubie4Life]

Yep. For what it's worth, regarding V11 Elon said in a recent interview the team had to build larger NNs as several old NNs had reached their useful limit (level of incompetence).

Exactly what I was referring to.

So incremental, small changes per that Elon interview (which is what I was recalling).

We’ll see! Obviously something has to change, since the performance on city streets is clearly unacceptable for highways which are just much harder in those particular aspects (stopping and going specifically, which have proven to be the most challenging aspects, as evidenced by the vehicle telemetry that users have posted here).

 

[willow_hiller]

Did you also navigate to somewhere behind you? For regular routes, the turn lane display shows up more as navigation typically takes the "main" roads that are more likely to have better map data. But some people will still need to make turns at the smaller intersections too.

I tried again last night (too much snow to safely turn on FSD Beta anyway), and 8 of 19 intersections on a 2-way multilane road had turn lane display, so a little bit higher than my previous test's 35% but still only around 40%. Similar to what you noticed, a surprising amount of the correct display showed U-turns as well.

At least around here, there's a lot of two-way center turn lanes that turn into a dedicated left turn just before the traffic light, and FSD Beta 10.69.x has gotten confused about getting into the suicide lane too early for a left turn or believing the parking lane is the right-most lane for the upcoming right turn. Now with the turn display showing left | through | through;right, it makes some sense why FSD Beta gets confused.

Tried navigating to a destination behind myself, and it was closer to 50-50. That being said, it was a pretty clear pattern in my area. For suburban intersections where my road was controlled by a traffic light: 100% of them displayed lane directions. Intersections where I didn't have a light or stop sign (e.g. where the main road was uncontrolled but the intersecting roads were) did not display the lane directions. Also I wasn't able to test many urban intersections, since a lot of them on my route are single-lane, but it also did not display lane directions for complicated urban intersections where more than two roads joined.

Also an aside, I have a lot of poorly calibrated speed traps in my area, so I did this test on TACC as far as possible; and it seems like TACC is governed by FSD Beta on 10.69.25.1. While I was coasting down the road on TACC ignoring the GPS directions, my visualization kept highlighting cars in blue for upcoming turns, and slowing down, despite the fact that I was staying in my lane and not attempting the turn.

 

[aronth5]

As an avid cyclist I have always felt FSD doesn't move over far enough when passing cyclists. Where I live it is very common for cars to move partway into the opposite lane when passing. FSD doesn't move over enough, delays too long moving over and then moves back too quickly.
Will be interesting to see how Tesla handles the new law just passed in Massachusetts.

"On Monday, Governor Charlie Baker signed a bill into law requiring drivers to stay at least four feet away from “vulnerable users” of the road when passing them. " "In passing a vulnerable user, the operator of a motor 79 vehicle shall pass at a safe distance of not less than 3 feet when the motor vehicle is traveling at 30 miles per hour or less, with 1 additional foot of clearance for every 10 miles per hour that the vehicle is traveling above 30 miles per hour. "

https://malegislature.gov/Bills/192/S3162

 

[Edtreo]

As an avid cyclist I have always felt FSD doesn't move over far enough when passing cyclists. Where I live it is very common for cars to move partway into the opposite lane when passing. FSD doesn't move over enough, delays too long moving over and then moves back too quickly.
Will be interesting to see how Tesla handles the new law just passed in Massachusetts.

"On Monday, Governor Charlie Baker signed a bill into law requiring drivers to stay at least four feet away from “vulnerable users” of the road when passing them. " "In passing a vulnerable user, the operator of a motor 79 vehicle shall pass at a safe distance of not less than 3 feet when the motor vehicle is traveling at 30 miles per hour or less, with 1 additional foot of clearance for every 10 miles per hour that the vehicle is traveling above 30 miles per hour. "

https://malegislature.gov/Bills/192/S3162

4 feet is almost becoming universal. FSD reacts to the movement of objects in real time and slight movement is interpreted or misinterpreted as intent to move in certain direction. This is where NN using vision has a huge potential going forward.

I wonder if there will be an option in the future to train FSD by the driver for their driving style and specific use case - Adaptive FSD.

 

[EVNow]

Today FSDb would not turn because there was a person standing in the cornet talking over phone. It was not clear what she wanted to do … may be eventually FSD would have turned, but there was a car at the back so, had to take over and turn.

 

[nvx1977]

As an avid cyclist I have always felt FSD doesn't move over far enough when passing cyclists. Where I live it is very common for cars to move partway into the opposite lane when passing. FSD doesn't move over enough, delays too long moving over and then moves back too quickly.
Will be interesting to see how Tesla handles the new law just passed in Massachusetts.

"On Monday, Governor Charlie Baker signed a bill into law requiring drivers to stay at least four feet away from “vulnerable users” of the road when passing them. " "In passing a vulnerable user, the operator of a motor 79 vehicle shall pass at a safe distance of not less than 3 feet when the motor vehicle is traveling at 30 miles per hour or less, with 1 additional foot of clearance for every 10 miles per hour that the vehicle is traveling above 30 miles per hour. "

https://malegislature.gov/Bills/192/S3162

FSDb doesn't move over enough for pedestrians walking in the neighborhood either.

I remember a few years back when Tesla introduced the lane departure correction feature. Every time I manually moved over for a bike, the car would want to swerve back. I disabled that feature really quick.

I've been hit 2x by cars while biking in MA. decided to just train indoors. not worth my life to assume any law would remove the disdain and/or carelessness some drivers have for VRUs.

 

[MindBlend]

Has anyone else noticed the new build 10.69.25.1 works WAY better on highway stack than 10.689.25. It doesn't drift over nearly as far when the lines widen for merge lanes. Its more confident in exit ramps / on ramps. It almost feels like they might be pushing more of the FSD logic into the EAP stack..

Anyone else???

 

[powertoold]

Has anyone else noticed the new build 10.69.25.1 works WAY better on highway stack than 10.689.25. It doesn't drift over nearly as far when the lines widen for merge lanes. Its more confident in exit ramps / on ramps. It almost feels like they might be pushing more of the FSD logic into the EAP stack..

Anyone else???

Yes, I agree. 25.1 is superior to .25 in almost every way.

25.1 is really great... Especially when driving with little traffic and manually using the stalk to change lanes ahead of time. (Referring to city streets)

Last night, I felt giddy after it drove me home with no disengagements. It was the smoothest, most confident I've experienced with fsdb. I started "seeing" the robotaxi future of Tesla's fsd program.

 

[archae86]

Yes, I agree. 25.1 is superior to .25 in almost every way.

Thanks for the nudge--you motivated me to go the garage and touch software. It took unusually long in the "checking for updates" state, then moved directly into downloading.

 

[aronth5]

Yes, I agree. 25.1 is superior to .25 in almost every way.

25.1 is really great... Especially when driving with little traffic and manually using the stalk to change lanes ahead of time. (Referring to city streets)

Last night, I felt giddy after it drove me home with no disengagements. It was the smoothest, most confident I've experienced with fsdb. I started "seeing" the robotaxi future of Tesla's fsd program.

25.1 for me seems like a typical point release. Some improvements along with some regressions. Takes awhile to determine whether it's a net improvement or not. Just once I'd like to try a new release with significant improvement and say WOW. Still waiting.

• Regressions (for me)
  • More phantom braking
  • Tried to make a right instead of a left hand turn at street lights I've driven over 100 times successfully.
  • Tried to turn into a cemetary immediately after a left hand turn. Has done this before but was recently fixed, so broken again.
  • Failed a sharp curve and I had to take over to avoid hitting a sidewalk curb.
  • Failed to take a left hand turn because 2 cars were waiting to come out of the street. No matter which direction I approach this intersection it messes up. Thinks it a 4 way stop when it's only a 2 way stop. Plus it's up an incline which totally screws FSD up. Certainly a mapping error but this intersection is now worse than it used to be on .25
• Improvements (for me)
  • Haven't really seen any yet.

 

[MindBlend]

25.1 for me seems like a typical point release. Some improvements along with some regressions. Takes awhile to determine whether it's a net improvement or not. Just once I'd like to try a new release with significant improvement and say WOW. Still waiting.

• Regressions (for me)
  • More phantom braking
  • Tried to make a right instead of a left hand turn at street lights I've driven over 100 times successfully.
  • Tried to turn into a cemetary immediately after a left hand turn. Has done this before but was recently fixed, so broken again.
  • Failed a sharp curve and I had to take over to avoid hitting a sidewalk curb.
  • Failed to take a left hand turn because 2 cars were waiting to come out of the street. No matter which direction I approach this intersection it messes up. Thinks it a 4 way stop when it's only a 2 way stop. Plus it's up an incline which totally screws FSD up. Certainly a mapping error but this intersection is now worse than it used to be on .25
• Improvements (for me)
  • Haven't really seen any yet.

Have you tried the highway stack? It seems to be a drastic improvement. I agree, the city stack isn't much changed or worse.

 

[momo3605]

Have you tried the highway stack? It seems to be a drastic improvement. I agree, the city stack isn't much changed or worse.

Why would highway stack change? Autopilot hasn’t been updated in a while. V11s new highway stack is around the corner. Seems pointless to change autopilot in a X.X.X.1 point release

 

[powertoold]

Why would highway stack change? Autopilot hasn’t been updated in a while.

It does and has changed with every fsdb point release for me. I am 100% sure of this (not about every single point release, but that AP has been updated / improved with fsdb point releases).

They've been working on the car's behavior for merging and dealing with merging cars (among other things like slowing earlier with stopped cars ahead on highway).

 

[momo3605]

It does and has changed with every fsdb point release for me. I am 100% sure of this (not about every single point release, but that AP has been updated / improved with fsdb point releases).

They've been working on the car's behavior for merging and dealing with merging cars (among other things like slowing earlier with stopped cars ahead on highway).

Do you have any evidence that it has changed. Release notes are pretty detailed and it’s not mentioned in there. Early beta YouTubers have never mentioned it, and I personally have not noticed any difference to autopilot behavior in a while.

I’m guessing it’s just placebo.

 

[powertoold]

Do you have any evidence that it has changed. Release notes are pretty detailed and it’s not mentioned in there. Early beta YouTubers have never mentioned it, and I personally have not noticed any difference to autopilot behavior in a while.

I’m guessing it’s just placebo.

It's 100% not placebo. I'm not an evidence collector, so we can corroborate our experiences.

What I've noticed since 69.2:

1) On curves, the car slightly biases toward the curve in a more natural and confidence instilling way (25.1)
2) When merging onto highway with ~5mph traffic, it picks a spot more naturally rather than just ramming towards the front of the slow cars (~since 69.2.2 or so). I am 100% confident of this change, although it could be a result of fsdb behavior change (as NoA doesn't usually kick in until the car is in the highway lane).
3) Merging onto highway in general is much improved, as it is better at picking a spot
4) On certain point releases, it no longer jumps toward lead cars in stop and go traffic, the stop / accel behavior is much more natural
5) In general, each point release has subtle changes in how it deals with cars merging onto highway as well

 

[old pilot]

It's 100% not placebo. I'm not an evidence collector, so we can corroborate our experiences.

What I've noticed since 69.2:

1) On curves, the car slightly biases toward the curve in a more natural and confidence instilling way (25.1)
2) When merging onto highway with ~5mph traffic, it picks a spot more naturally rather than just ramming towards the front of the slow cars (~since 69.2.2 or so). I am 100% confident of this change, although it could be a result of fsdb behavior change (as NoA doesn't usually kick in until the car is in the highway lane).
3) Merging onto highway in general is much improved, as it is better at picking a spot
4) On certain point releases, it no longer jumps toward lead cars in stop and go traffic, the stop / accel behavior is much more natural
5) In general, each point release has subtle changes in how it deals with cars merging onto highway as well

I don't know about the rest of your points but I disagree with the assertion "(as NoA doesn't usually kick in until the car is in the highway lane)." What I have observed is that shortly after entering an on ramp for a highway my selected speed limit (+10 mph) shows on the display ie. speed limit of 70mph shows 80mph. as max. speed.

 

[powertoold]

I don't know about the rest of your points but I disagree with the assertion "(as NoA doesn't usually kick in until the car is in the highway lane)." What I have observed is that shortly after entering an on ramp for a highway my selected speed limit (+10 mph) shows on the display ie. speed limit of 70mph shows 80mph. as max. speed.

But I think you're still seeing the Fsdb visualization right? The fsd visuals are on until the car is fully within the first highway lane.

 

[old pilot]

But I think you're still seeing the Fsdb visualization right? The fsd visuals are on until the car is fully within the first highway lane.

I don't know, I'm sorry, I'll have to look for that next time. The only thing I can say with certainty is the max speed does change.

 

[aronth5]

Have you tried the highway stack? It seems to be a drastic improvement. I agree, the city stack isn't much changed or worse.

Not yet but I have a 150 mile highway drive tomorrow so I'll know after that.

 

[scottf200]

Why would highway stack change? Autopilot hasn’t been updated in a while.

It does and has changed with every fsdb point release for me. I am 100% sure of this (not about every single point release, but that AP has been updated / improved with fsdb point releases).

They've been working on the car's behavior for merging and dealing with merging cars (among other things like slowing earlier with stopped cars ahead on highway).

NoA/AP:
I have used NoA part of AP ever since it came out and it has improved considerably over time. It is way better at making better & sooner merging to get it the correct lanes on exits and merges better on entries. It is very reasonable now ... before not so much.

FSDbeta:
It is puzzling to me that it can get the lanes wrong on turns screws up the below almost 1/2 the time. Today in the same lighting conditions this intersection was handled two ways. It has done both a few times before today too. It does something similar on the way to a 10am appt today but on a right hand turn. It missed the turn because it screwed up and got into the left most lane on purpose about 1/4 mile ahead of the right turn. The regular nav had the directions exactly correct.

Green below is what it should always do and has done maybe 50% of the time. Yellow is what it does the rest of the time. Did both in same lighting today on 2 drives.