FSD Beta 10.69

[Watts_Up]

As with school zone signs, there appears to be no consistent format for speed bump signs. Not that FSD couldn't deal with a hundred, or thousand, variations. But it does add yet another complication.

I believed that FSD determines local speed using both posted signs and GPS?

If so, could GPS mapping of posted speed bump, school zones, or signs with time of the day restrictions... could be used by FSD?

 

[Jeff N]

As with school zone signs, there appears to be no consistent format for speed bump signs. Not that FSD couldn't deal with a hundred, or thousand, variations. But it does add yet another complication.

FSD seems to already read the standard speed bump markings on the road itself. Of course, that doesn’t work if the road is covered with snow but then lots of other aspects of FSD are problematic on snow-covered roads.

 

[Twiglett]

There's a tar strip in the middle of a briefly wide lane on my way to work. (Think very long bus stop.) Given the angle of the sun in the morning, the tar looks lighter than the asphalt. FSD treats that lighter tar strip like a lane marker. Thus it swerves into the wrong "lane" and has to swerve again when the road narrows a hundred yards later.

But in the afternoon the tar looks darker so FSD ignores it.

Forcing a disconnect by holding the steering wheel to reject the first swerve, then hitting the report button, gives Tesla the training data to reject tar strips in more lighting conditions.

Doing it every day gives them a bunch of similarly geolocated disconnects. (That's a signature that can be mined out of a sea of big data.) Doing it every day also provides subtle variations in the captured video data with respect to lighting, weather, other vehicles, etc.

I applaud your efforts sir
I take a different view and presume that they don't use multiple hits from the same source as a weight when establishing a capability issue. Rather I hope they are grouping similar issues into something they can correct. Something like that seems to have happened with the tar stripe being detected as pedestrians, pretty sure it wasn't just me reporting that
But that was when we got updates every two weeks and the pool was smaller.
However, as we are in TMC, we specialize in attempting to decode the hidden Tesla methodologies

 

[Bladerskb]

The only real hurdle is accurately recreating the world in vectorspace at decent framerate.

And they’ve demonstrated by now that HW3 can do that. It’s not quite there yet, but that just takes a bunch more NN targeting, training, and tuning.

After that there is a TON of work still to do to add the heuristics behind key information detection and decision-making, but this is essentially a mountain of relatively trivial work.

And by trivial, I don’t mean it doesn’t take a lot of smarts and won’t have mistake/decision setbacks. But the implementation process for these is trivial.

Once you can detect a “no left turns during school days while mercury is in retrograde” sign, the decision and implementation of what the car needs to do on that is trivially easy from an architectural point of view.

It just takes a LOT of work. Not a lot of compute thankfully. (The vectorspace modeling is what takes the real compute.)

Anybody thinking FSD Beta’s current driving ability is a bad sign on where it will be on the exact path it is currently on (hardware, architectural approach, etc) is just wildly, and I mean WILDLY, misunderstanding its development process.

This isn’t like testing Mk1 of a spacecraft and finding its faults to go back and tweak designs. This is more like the tank hop tests where the nose cone is not planned to be put on for many months.

Driving policy is multiple orders of magnitude harder than perception. I hate to break it to you but the person who is misunderstanding the development process is you.

 

[Dewg]

I applaud your efforts sir
I take a different view and presume that they don't use multiple hits from the same source as a weight when establishing a capability issue. Rather I hope they are grouping similar issues into something they can correct. Something like that seems to have happened with the tar stripe being detected as pedestrians, pretty sure it wasn't just me reporting that
But that was when we got updates every two weeks and the pool was smaller.
However, as we are in TMC, we specialize in attempting to decode the hidden Tesla methodologies

Tar seams used to be much worse, and sometimes confused for other objects. They are working on it - all the reports are helping, and you can see it in recent release notes:

"Improved the precision of VRU detections by 44.9%, dramatically reducing spurious false positive pedestrians and bicycles (especially around tar seams, skid marks, and rain drops)."

 

[Twiglett]

Tar seams used to be much worse, and sometimes confused for other objects. They are working on it - all the reports are helping, and you can see it in recent release notes:

"Improved the precision of VRU detections by 44.9%, dramatically reducing spurious false positive pedestrians and bicycles (especially around tar seams, skid marks, and rain drops)."

that's exactly what I meant. The difference was indeed dramatic.
In that case it was extremely difficult to spot the momentary appearance of the pedestrian on the display, it took two of us in the car because the driver was busy with managing the blaring alerts
But my point was really that it took many different people reporting that same issue.

 

[MrTemple]

Driving policy is multiple orders of magnitude harder than perception. I hate to break it to you but the person who is misunderstanding the development process is you.

No. Read again where you misunderstood the point.

The architecural challenges of implementing driving policy are essentially nil.

That does NOT mean (as I stated) that this is fast or easy to get right.

It’s just the challenges are COMPLETELY outside the hardware and software architectural realm.

Once you know what you want the car to do for an edge case in driving policy, the implementation is trivial.

Thus there is no limitation other than time it takes to learn from real-world edge case to define the requirements and logic-out the heuristic decisions.

This is the nose cone that has not been put on yet. It’s not that it’s not working now, it’s that we are not there in the schedule yet where it is expected.

And let’s not forget that knocking down edge cases in heuristic driver policy pushes the car further and further beyond the human drivers on various points of the driver-quality bell curve.

Of course during this training and learning phase the AI is going to make different mistakes than a poor human driver. So the bell curves will be slightly different (unless you measure something like serious accidents per km).

But we will gradually and surprisingly soon get to a point where the computer has better driver policy than 99% of humans. And the nines will keep coming.

 

[KLam]

My model S has the capability of remembering the locations for raising/changing my suspension height. I wonder if it can be used to remember pothole locations and driving speed on my local route in FSD.

 

[tivoboy]

My model S has the capability of remembering the locations for raising/changing my suspension height. I wonder if it can be used to remember pothole locations and driving speed on my local route in FSD.

So, there is a third party company that already does this and I think Elon said at one point that they were going to introduce some form of this feature - but most likely NOT from the company that already does this for some OEM’

It’s apparently not AVOIDANCE which would be easier and faster to implement but rather a dynamic REACTION to the road impairment like a pothole or other

 

[uscbucsfan]

My model S has the capability of remembering the locations for raising/changing my suspension height. I wonder if it can be used to remember pothole locations and driving speed on my local route in FSD.

Tesla Model S And Model X Now Recognize Potholes On The Road, And Will Soon Be Able To Avoid Them

Tesla has added a feature to Model S and Model X that allows them to recognize road conditions, detect potholes, speed bumps and other imperfections on the road, and adjust the car suspension accordingly. The idea for the future is to combine this features within the Autopilot so that the system...

www.torquenews.com

 

[AlanSubie4Life]

Ok, it's a bet!
I wonder if the randomness is to achieve the stated goal of achieving zero intervention drives. If you do the same thing every time then you may fail every time but if you do a random thing every time then sometimes you will succeed.

Are you sure you don’t want to wait for 10.69.1 results before finalizing this? You might reconsider!

 

[Daniel in SD]

Are you sure you don’t want to wait for 10.69.1 results before finalizing this? You might reconsider!

Nah, I'm confident they'll fix all known issues for 10.69.2.
They're still testing Chuck's UPL with a sample size large enough that if there are no failures then 90% success should be well within the confidence interval.

 

[Supcom]

Once you know what you want the car to do for an edge case in driving policy, the implementation is trivial.

My many years developing systems has taught me that anyone who claims something is trivial either does not understand the problem or is not responsible for the implementation.

 

[Silicon Desert]

I could be a minority, but I do prefer to keep FSD on, and just scroll down the max speed for these occasions, rather than turning FSD off all together. When max speed changes again, it will pick it up automatically, so it's just one-off action. I can live with it, until they fix it.

We both might be in a minority. I don't know what you replied to since I ignored that person long ago due to snippiness, but I do exactly what you specified above for FSD. I keep it on and scroll down the max speed when desired. However, it could be done better.

 

[Silicon Desert]

He's got to stay on schedule to justify the price hike on September 5th

not really. He could change the dates on both if he wants. Oh and lets hope it does not happen for you folks. Break a trend and stay on time

 

[AlanSubie4Life]

Nah, I'm confident they'll fix all known issues for 10.69.2.
They're still testing Chuck's UPL with a sample size large enough that if there are no failures then 90% success should be well within the confidence interval.

I guess since I extracted a beer from you already I should just go along and go with this bet on 10.69.1 as well (do you want to bet on one at a time?). Or we can stick with just 10.69.2, but that gives me a chance to bail…. Conditions:

I believe that FSD Beta 10.69.2 will have less than or equal to 90% success on Chuck’s UPLs described below:

1) Attempts with no traffic at all on the visualization in the relevant directions during the decision phase don't count.
2) Car has to go at the first opportunity when a decent human driver would go, when there is a 5-second opening or greater. No honking allowed from other drivers (failure for that).
3) The car cannot cause any other car to slow down or react at any point in the maneuver (by creeping too far, moving forward at the wrong time, etc.) Including if it stops halfway through in the median - it cannot result in any car slowing or yielding to it (for good cause), or anything like that. Prudent defensive driving by other drivers when the vehicle is stationary and in the correct spot is allowed; we can’t help other drivers being courteous or cautious.
4) The car can't do anything incorrect, like stopping in the first lanes of traffic (even if there is no one coming) to wait for traffic from the other direction. No stuttering, sticking partially into lanes, and if median is used, must properly sit in the median with correct buffer zones (no lane position changes performed by passing traffic), etc.
5) Interventions including accelerator application count as failures.
6) All unprotected left attempts count.
7) Must be greater than 10 attempts. Roof-mounted video from Chuck of car surroundings must be provided (there were a few dark videos done last time which we ignored).

 

[Silicon Desert]

I guess since I extracted a beer from you already I should just go along and go with this bet on 10.69.1 as well (do you want to bet on one at a time?). Or we can stick with just 10.69.2, but that gives me a chance to bail…. Conditions:

I believe that FSD Beta 10.69.2 will have less than or equal to 90% success described below:

Hey there. Always like your posts. You might want to bet Daniel a cheaper lite beer just in case. Here is why I say that. I got back from the bay area yesterday and able to try the .69 on the older MX. I was absolutely in shock at the improvement from the previous 10.12.12 from June 4. Wife and I came home from the Atlantis casino buffet out of the parking lot, on surface street, onto highway 395, later merging on I80, then off to the surface streets and 7 miles on surface street to home. Total of about 17 miles in medium mid-day traffic. I never took over once and would call it a successful drive for the way I drive. For the first time, it handled the two round-abouts near my home. Never did that before.

Now to your point. On one of the sections of Pyramid highway this morning, I rate that it violated 3 of your bullet points on a long section of the road where they are repairing and widening the road (partial lane closure and manual redirection. So in summary, really impressive improvements, yet some occasions where it has an "Alzheimer's moment" as you have probably seen in some online videos. For me and my limited driving of about 60 miles so far, I would say it is at about 90% or a little better success (time and mileage wise). It will be interesting to see what happens on my drive out to the Fallon Naval Air Station tomorrow. I expect it might be close to perfect since it is mostly easy interstate travel. I just hope it doesn't run over any expensive Navy jets.

 

[AlanSubie4Life]

You might want to bet Daniel a cheaper lite beer just in case

I don’t know how to quit while I am ahead and I feel sorry for my friend Daniel, both due to the beer, and also he does not have FSD.

Also, I trust the inexplicable inconsistency of FSD to deliver me the victory (and also the 5-second gap rule).

I am not surprised at all that 10.69 is a lot better. Glad to hear it.

 

[Silicon Desert]

Oh gosh, I made a typo. I said my previous FSD version from June 4 was 10.12.12. It was actually FSD 10.12.2 (2022.12.3.20)

 

[mmyers300]

So after watching hours of 10.69 FSD videos I kept hearing that it is overall better. I'm trusting the Beta Beta testers, but if I watched the hours of video without their commentary, I'm not entirely sure I would as confidently agree. But still can't wait to get this version!
Anyway, here is a summary of the comments and a compilation link:

• More affirmative, more confident, more polished.
• Overall it is Smoother - less jerky turns, stays better in its lanes in roundabouts and just about everywhere
• Not as much hard braking for pedestrians and speeds back up sooner
• Slowdowns for speedbumps don’t seem to have changed
• Stops too long at stop signs. Of course, it used to roll through stops
• Since Chuck Cook is literally mentioned in the release notes for unprotected lefts, I watched many of his videos and he is happy
• It looks like phantom braking is reduced as Dirty tesla saw no phantom braking in areas that used to have many per mile.