FSD Beta 10.69

[AlanSubie4Life]

Is this 10.69.2 precisely calculating the median crossover region or lucky? (Yes, this is fisheye from the rear with lens distortion…)

View attachment 852284

No honks, no swerves, no disengagements!

It consistently does poorly with vehicle localization. Hopefully everyone testing is an expert, otherwise it’s just a matter of time before someone gets hit. This is one of the more dangerous FSD limitations.

This whole method of taking left turns is really novel to me. Seems hazardous unless the median is quite wide.

 

[Cheburashka]

Hopefully they’ll defeat your wheel weight soon; should be easy.

I'd like to see them try. I use a sock with some ball bearings in it, they bounce around a bit so the torque applied is uneven.

I highly doubt the company that can't figure out how to make reliable auto wipers will be able to defeat this.

Btw I only use this on straight freeways on long trips where I am paying attention.

 

[Cheburashka]

The smooth and stable turns on 10.69.2 constantly make me say "wow," whereas I'm constantly saying "what the fudge is it doing now" when it randomly brakes or decides to choose a dumb lane.

It also has no consideration of driving context. It'll happily accelerate into a pile of slowly moving cars ahead.

It feels a bit smoother but in terms of my personal disengage rate, nothing has changed.

I don't really do any ULPs on a regular basis so couldn't care less about how it handles those. Today on a green light it still hesitated to turn left for some reason. I always find myself either mashing the throttle to make it go faster or disengaging it when it does something utterly stupid.

 

[AlanSubie4Life]

I always find myself either mashing the throttle to make it go faster or disengaging it when it does something utterly stupid.

Same. Honestly don’t understand, unless people are just “seeing how it works” (a valid activity), how anyone can do otherwise. People gotta get places. Presumably that frustration is evident in my prior video. Like, when first at a light, with two lane changes to do with wide open road ahead, how can you not just take off, getting to the speed limit in the minimum possible time?

Even my wife is ok with me doing this (maybe not ok with actual minimum time…), and she hates FSD Beta.

 

[Daniel in SD]

Same. Honestly don’t understand, unless people are just “seeing how it works” (a valid activity), how anyone can do otherwise. People gotta get places. Presumably that frustration is evident in my prior video. Like, when first at a light, with two lane changes to do with wide open road ahead, how can you not just take off, getting to the speed limit in the minimum possible time?

That's a fairly aggressive maneuver, I can see why they don't do that. What I don't understand is the tailgating and aggressive braking (Autopilot does this too). It looks like it brakes way more aggressively than it accelerates.

 

[AlanSubie4Life]

That's a fairly aggressive maneuver, I can see why they don't do that. What I don't understand is the tailgating and aggressive braking (Autopilot does this too). It looks like it brakes way more aggressively than it accelerates.

It’s so f’ing stupid.

And it took 15 seconds to get to 50mph so I think there may be a middle ground here.

 

[D Good]

Apple uses LIDAR generated HD maps. Tesla does not and will never use HD maps, they're a crutch.

Ok. What Tesla has right now hasn’t improved in over 2 years.

I’m starting to think that pure vision can’t work in all circumstances. You absolutely cannot drive well without knowing what lane to be in.

Say for example, you have a right turn coming up in a mile. An unmarked turn lane just opened that is NOT for your turn (FSD even has trouble with arrow-marked turn lanes too though) A human can remember from past drives that that is not the lane to get in for their turn, or, can use basic logic that is simply not built into FSD yet that it’s not the lane for their turn.
Vision only FSD has no idea what to do so just gets in the lane. Now it’s stuck.
I’m telling y’all they are going to need better maps if they want a good experience with beta. If the goal is just for it to work (but also suck) then vision only could work.

We would all be having better drives right now (for the most part) if FSD used better maps. Period end of story. The anomaly is bad map data.

 

[Daniel in SD]

We would all be having better drives right now (for the most part) if FSD used better maps. Period end of story. The anomaly is bad map data.

Yeah, that was sarcasm. haha. HD maps, or any euphemism people want to use for them, are very useful. It's simply not realistic for a machine to have the superhuman intelligence necessary to anticipate everything as well as someone who's driven the road before.

 

[mtndrew1]

Tonight I did my test route which I’ve been doing since just after I joined the beta program.

10.69.2 tied for its best performance so far (see link below). The numbers are improving and this is a very challenging route but the rate of improvement is nowhere near Musk’s lofty proclamations. Anyway, on to the next version.


FSD Tracking Spreadsheet

 

[Jeff N]

I’ve taken 10.69.2 on a few of my regular routes around San Francisco and am generally seeing better behavior. It correctly perceived and drove through several areas that it used to consistently get wrong. It seems to have somewhat better speed control when making sharp turns and when driving down narrower winding streets with lots of parked cars. It also correctly drove straight through an area where it used to dive into a left-only turn lane. It still messed up a couple of times crossing through complex intersections. I also noticed it now handling pedestrians and bikes better than it used to. There were a couple of times when drivers were getting out of cars parked on the street or bicycles were passing along next to the car when I thought it would freak out but it was relaxed and unperturbed.

 

[MrTemple]

When they release FSD beta later this year to all FSD purchasers the cumulative miles increase even faster! If that were actually a measure of capability why didn't they do that two years ago?

Uh… because it wasn’t nearly as capable.

This is kind of my point.

I think you’re using a very different definition of capability than I find in the dictionary.

FSD is WAAAAY ahead of where it was even a year ago. Way more capable. And it’s getting more and more competent and able to handle more and more driving situations. So good it is safer to loosen the stringent requirements of who can join the Beta test program. This is the definition of capable.

I don’t know why you’re trying to argue which is a better metric.

I just said this EXTREMELY easy to understand graph (showing not progress, but sudden and dramatically accelerating jump in miles driven) is illustrative of the pace of the progress.

Which it is.

Sure, while it’s easy to understand what the graph is showing, I can see where you can’t make the connection the rest of us can make.

To understand exactly WHY what it’s showing is a good illustrator of the progress, you need to understand the context I wrote in the message you quoted.

 

[MrTemple]

I'd like to see them try. I use a sock with some ball bearings in it, they bounce around a bit so the torque applied is uneven.

I highly doubt the company that can't figure out how to make reliable auto wipers will be able to defeat this.

Btw I only use this on straight freeways on long trips where I am paying attention.

Doing this, having the physical anti-safety device in the car, and bragging on the internet achieves the impressive quadruple-self-own:

1. Increases the chance you kill yourself and/or others.

2. Provides your insurance company investigator evidence to deny your claim.

3. Gives any competent lawyer ammunition to sue you and your family blind.

4. Hands your local District Attorney an open and shut negligent homicide case.

Uh… congratulations?

 

[2101Guy]

I'd like to see them try. I use a sock with some ball bearings in it, they bounce around a bit so the torque applied is uneven.

I highly doubt the company that can't figure out how to make reliable auto wipers will be able to defeat this.

Btw I only use this on straight freeways on long trips where I am paying attention.

This one is on the way for me to test with the yoke. Seller claims it will work and if it doesn’t I can return it. Will report back

https://www.ebay.com/itm/234610491237

 

[boonedocks]

I guess I’ll have to take Chuck’s approach and start posting videos and tweeting and causing a .


My daily commute within 1 mile of my home has had the same errors on the way to work and on the way home. Both places require me to disengage BEFORE I get there if there is any traffic of any kind. I try it separately at times when traffic is light to see if there are any marked changed but even with 10.69.2 it does the same wrong lane every single time.


On the way to work, the red arrow, the right turn on to McEver always stays in the right most “right only in to a CVS” and at the last second hard brakes and turns on the blinker and tries to get back in to the straight lane.


On the way home, the green arrow, after making the left turn on to Browns Bridge Rd it selects the correct lane and then turns on the right blinker and moves in to the lane that ends in 200 feet and then turns the left blinker on and has to get BACK in to the straight lane as that lane ended.


The system has no retention of previous mistakes and corrections made nor does it use the internal map database to realize the lanes ending or turn only position.


So as of 10.69.2 I still have disengagements less than a mile from my home in both directions and for different reasons. There are 4 other very simple points on my commute where I still have to “take over” as it is actually very unsafe for other drivers not knowing FSDB intentions and direction.

 

[2101Guy]

Doing this, having the physical anti-safety device in the car, and bragging on the internet achieves the impressive quadruple-self-own:

1. Increases the chance you kill yourself and/or others.

2. Provides your insurance company investigator evidence to deny your claim.

3. Gives any competent lawyer ammunition to sue you and your family blind.

4. Hands your local District Attorney an open and shut negligent homicide case.

Uh… congratulations?

WRONG on all 4 points. Why? For starters. You have zero evidence to support #1. Lets start there.
For 1-4? As indicated in every T&A, document, form, manual, instruction..you name it. The DRIVER of the vehicle is always 100% accountable for anything that happens with the vehicle. Whether the car is on AP, FSD, FSD(b), FSDB, manual mode...whatever. The DRIVER is always responsible. Under the law of every state in the USA (and probably every other country as well).

If a person runs over a pedestrian in a tesla and kills them? The driver of the Tesla is 100% accountable. There is no getting out of a claim under any circumstance with an excuse of "the tesla was on FSD at the time of the crash". With or without a weight. Also, the law cares not one bit about whether the car was on FSD, AP, or whether or not some thing was or was not on the steering wheel. The driver will get charged with manslaughter or murder (depending on state) and will be sued. (as they should) REGARDLESS.

 

[sleepydoc]

Same. Honestly don’t understand, unless people are just “seeing how it works” (a valid activity), how anyone can do otherwise. People gotta get places. Presumably that frustration is evident in my prior video. Like, when first at a light, with two lane changes to do with wide open road ahead, how can you not just take off, getting to the speed limit in the minimum possible time?

Even my wife is ok with me doing this (maybe not ok with actual minimum time…), and she hates FSD Beta.

If it accelerated faster others would complain that it's too aggressive. You're complaining about style. There's nothing wrong with how it accelerates, you just don't like it.

 

[sleepydoc]

Yeah, that was sarcasm. haha. HD maps, or any euphemism people want to use for them, are very useful. It's simply not realistic for a machine to have the superhuman intelligence necessary to anticipate everything as well as someone who's driven the road before.

The intrinsic problem with relying on HD maps is that as soon as they're out of date the system fails. Roads are continually being updated and changed all over and you have to spend an inordinate amount of time updating the maps to keep the system working.

 

[AlanSubie4Life]

If it accelerated faster others would complain that it's too aggressive. You're complaining about style. There's nothing wrong with how it accelerates, you just don't like it.

15 seconds 0-50 is too slow.

Analysis above thought that was intentional to allow lane changing.

I think it is possible for it to be much more assertive without frightening anyone. You see videos with it speeding up much faster all the time and no one seems bothered.

Why does it brake so hard?

Nothing to do with style. You can see this created safety issues - the video is right there.

 

[Daniel in SD]

Uh… because it wasn’t nearly as capable.

This is kind of my point.

I think you’re using a very different definition of capability than I find in the dictionary.

FSD is WAAAAY ahead of where it was even a year ago. Way more capable. And it’s getting more and more competent and able to handle more and more driving situations. So good it is safer to loosen the stringent requirements of who can join the Beta test program. This is the definition of capable.

I don’t know why you’re trying to argue which is a better metric.

I just said this EXTREMELY easy to understand graph (showing not progress, but sudden and dramatically accelerating jump in miles driven) is illustrative of the pace of the progress.

Which it is.

Sure, while it’s easy to understand what the graph is showing, I can see where you can’t make the connection the rest of us can make.

To understand exactly WHY what it’s showing is a good illustrator of the progress, you need to understand the context I wrote in the message you quoted.

View attachment 852364

All the graph shows is that Tesla released FSD beta to more people between Oct 2021 and Dec 2021 and then again in May. I think the correlation of capability to safety is very weak for a supervised system and in fact I think it will go negative at some point due to automation complacency. For all we know safety is mostly a function of the capability of the driver monitoring system and that's why Tesla felt comfortable releasing it to more people.

Here's a graph for Autopilot:

I feel like increased reliability has been more significant than increased capability but I guess it's all sort of semantics. It's capable of driving further between errors.
After Tesla releases FSD beta to all paying customers what metric can we use to measure capability?

 

[boonedocks]

The intrinsic problem with relying on HD maps is that as soon as they're out of date the system fails. Roads are continually being updated and changed all over and you have to spend an inordinate amount of time updating the maps to keep the system working.

Until FSDB has the ability to "remember and recall" there needs to some map reliance My posts a few up is a perfect example of that.