FSD Beta 10.69

[kabin]

Funny bit. @AlanSubie4Life made that comment about accel/brake/accel/brake. So, on the next major drive at speed (55 mph with the occasional stoplight and traffic) I went looking for it.

What I saw: brake/harder brake/brake/harder brake/etc. It... wasn't that noticeable, but it was there. A human would have been more consistent.

And, being the control theory guy that I am, I now make a half-A** guess that the observed behavior could be because the car has a less direct way of figuring the distance/relative speed of the car in front with vision vs. radar. Radar would be more accurate; vision has to gather an image and then, over multiple frames, work out how big/how small the image is getting over time; given the uncertainty, it would be no surprise that uncertainty would bleed over to the deceleration. Humans do a bit better job at this kind of thing (think: quarterbacks throwing footballs at wide receivers. But they miss fairly often.).

But was it a neck-snapper? Nope. Was it noticeable? Only when pointed out. Will it ever get fixed? (spoken in a Dorothy & Toto voice). Um. Probably yeah, after the fixers at Tesla get the more obvious bugs, like swing right on left turns, picking the wrong lanes, and all the other complaints.

Now, if this accel/decel/accel/decel stuff is more than the minor stuff I'm describing, then I'm wondering if there's something else at play. Tesla playing with different parameters on different cars? Camera calibration out of whack? Something broken somewhere? Dunno, but gotta wonder.

They tweaked the jerk limit so deceleration is a bit less dramatic now but it's still unnatural. Add to that I still get frequent bot-like decelerations at predictable locations much like Chuck Cook did with dual traffic lights in the field of view. It's a brief (short delta-t) deceleration minus the normal human gentle roll/fade on/off of the brake so it's about as bad as can be. Everyone in the car notices it.

 

[Supcom]

vision has to gather an image and then, over multiple frames, work out how big/how small the image is getting over time

Maybe there are other ways to determine range than just looking at apparent size growth? Since the car knows the position of the camera and its orientation as well as the orientation of the car, it should be able to gauge distance to nearby vehicles based on how far the bottom of the object is from the bottom of the image frame.

No doubt there are other, even better tricks to use.

 

[Tronguy]

Maybe there are other ways to determine range than just looking at apparent size growth? Since the car knows the position of the camera and its orientation as well as the orientation of the car, it should be able to gauge distance to nearby vehicles based on how far the bottom of the object is from the bottom of the image frame.

No doubt there are other, even better tricks to use.

No argument with any of that. But a radar is pretty simple: Send out a signal, wait for the reception, and it's 6.18 microseconds per radar range mile, exact. Two successive returns from the same object: Now you now distance and relative speed. Collect some more and it gets more accurate. Doesn't matter if the car is bouncing around, vibrating, and all that jazz: It Just Works.

Now, go for image processing. Car's bouncing, each camera shot full of pixels has to be converted to an image, images have to be stabilized, differences between images figured out, and so on.

Now, I happen to be an ex-RADAR techie and, while I haven't worked on them in nigh-on 35+ years, I've had the Navy training and hands-on experience with the Real Thing (tm); and, afterwards, for fun, took a college course on the subject. Not that I learned much new, mind you (easy A), but I have seen the math.

Image processing... isn't like that. If the Tesla guys say they can do speed control with images, well, I don't have that math, but a cursory look says that delta_X and delta_V has to be a more complicated analysis with vision than RADAR range and velocity. If Tesla says they can do it, well, throw enough hardware at a problem and Solutions Will Appear, so who am I to argue?

The obvious (at least, for short to medium baselines) would be stereo vision. For humans, that peters out around 30 m, and then we tend to use relative sizes and other tricks. And, since those are tricks, we can get confused. (Although any possible ancestors who got confused to easily aren't ancestors because they got eaten.. or fell out of trees while trying to get from point A to B. So, we have this genetic disposition to get it right, honed by zillions of years of trying, failing, and succeeding; mostly, only the ones who succeeded had progeny.)

We'll see. Isn't being on the cutting edge lots of fun?

 

[Supcom]

Doesn't matter if the car is bouncing around, vibrating, and all that jazz: It Just Works.

It's not quite so simple as that. There are potential issues with little things like multipath, mutliple targets, range ambiguities, range resolution and co-channel interference, to name a few complications. Not to mention the trade between detection threshold and false alarms.

BTW, doppler radars do not need multiple returns to determine velocity of the target.

 

[Mardak]

I thing I noticed today about "random turn signals" is that a lot of them are just unnecessary as opposed to incorrect that requires rerouting or a disengagement if not cancelled.

For example, there were multiple instances where I was in the left-most lane, and map data knows there's a dedicated left-turn lane for the next intersection while navigation knows it wants to go straight, so FSD Beta relying too much on map data wants to switch right to avoid being forced to turn left not realizing just a few seconds later, it would have seen the dedicated left-turn lane appear. This desire to switch out of the lane also causes unnecessary slowdowns that one could notice if carefully watching the adjacent visualized vehicles turn blue as FSD Beta is trying to find a gap to switch.

A different kind of random turn signal is for forking lanes near a curve or intersection (including some driveways). FSD Beta knows it doesn't actually need to change lanes (and doesn't highlight adjacent vehicles in blue), but most people probably would wonder why the car is signalling to stay in its lane. However, this is extra confusing for the driver as I'm pretty sure there isn't any distinction between this "stay in lane" vs "switch lane" signal where the latter you need to be fast to cancel to prevent an unnecessary lane change.

I suppose generally, FSD Beta might actually work better in areas that have a lower concentration of multi-lane roads where FSD Beta just wouldn't have the option of getting confused about which lane to pick.

 

[bhakan]

Better application for FSD beta is wide well-marked suburban streets (parkway-like roads) with dedicated turn lanes and no changes in lane numbers (no merges or forks). Ideally travel in just a straight line with no turns and no hills. Light traffic with all green lights hopefully. This will result in fairly good performance, though you may have to use signals to cancel unnecessary lane changes.

The key is to avoid turning, stopping, and starting. And other vehicles. I have found the performance to be quite good in this domain.

Elon should release two separate tier robotaxis when that year comes , one running in the burbs and the other in densely populated city blocks. Like the game Rush hour the robotaxis will be stuck on the city streets.

 

[EVNow]

Looks like you’re averaging almost 6 miles per disengagement! Huzzah.

We’re supposed to be chasing nines by now. We should be in the thousands of miles per disengagement neighborhood if FSD is to ever amount to anything resembling the original breathless promises.

At this rate of improvement you’re tracking to be at 30 miles per disengagement in another year and a half. I wonder if that will result in “the greatest increase in asset value in history.”

Do you have a Google link to a common sample route of yours? I’m curious to see what it takes to get six continuous miles of FSDb performance without intervention.

I've been saying this for a long time. For human equivalent driving Tesla still needs to improve some 1000x.

A lot of my disengagements go away if Tesla works properly at roundabouts, which account for ~ 60%.

No links to my routes though - they are my usual routes that start or end at my home. I don't do special runs to test FSDb - I just use it in all of my trips.

 

[black knight 3]

After five years of waiting for FSD to materialize and 18 months of watching it stumble around city streets I’ve determined that what I truly want is something like supercruise. I want to zone out in numbing traffic for two hours per day.

Ever since Tesla focused on FSD, highway autopilot development has slowed to a crawl. At least for me the system has gone backwards on the highway. When radar disappeared autopilot’s performance in stop and go traffic became very uncomfortable. Accel-brake-ACCEL-BRAKE over and over.

Maybe single stack will be a leap forward. Maybe it will be another regression. But I can say one thing, 18 months of FSDb has given me no confidence in Tesla’s solution for autonomy.

I really bought into the hype (literally!) and that was foolish of me.

You're not alone so don't feel bad. we all bought into the hype unfortunately...

 

[aronth5]

That is…surprising. I guess it depends on what one is used to?

Or perhaps your car just behaves differently?

 

[aronth5]

You're not alone so don't feel bad. we all bought into the hype unfortunately...

Not all. My ask has always been L3 on restricted access highways which is a reasonable goal. I never believed any hoopla over driverless FSD with the current hardware and software approach.

 

[sleepydoc]

Not all. My ask has always been L3 on restricted access highways which is a reasonable goal. I never believed any hoopla over driverless FSD with the current hardware and software approach.

The sad part is, Tesla could easily have achieved (has achieved?) this long ago but has chosen pursue the loftier goal of FSD everywhere rather than at least give owners somethign more useful in the mean time.

 

[AlanSubie4Life]

Or perhaps your car just behaves differently?

Does not sound like it! (He said he noticed it.)

Humans do a bit better job at this kind of thing (think: quarterbacks throwing footballs at wide receivers. But they miss fairly often.).

If humans have laser-like focus on this task they do an extraordinary job of this with near 100% success. Of course sudden stops are sometimes still required due to timing. But even then the acceleration vector magnitude monotonically increases to its peak, and then monotonically decreases to zero. No one wants to test the initial bite.

 

[aronth5]

Does not sound like it! (He said he noticed it.)

But only when he paid close attention. Sounds like your car's behavior is more pronounced.

 

[JulienW]

The sad part is, Tesla could easily have achieved (has achieved?) this long ago but has chosen pursue the loftier goal of FSD everywhere rather than at least give owners somethign more useful in the mean time.

The current AP stack is in no way even close to being capable of Full Self Driving and Tesla hasn't even released the FSD highway Stack yet. As long as the highway Stack is taking Tesla might have a better chance of L3 on city streets than on highways.

 

[AlanSubie4Life]

But only when he paid close attention. Sounds like your car's behavior is more pronounced.

Hard to say, but I doubt it. Vector magnitude is highly dependent on conditions and I live in a very hilly area with high speed roads.

Also highly dependent on attention.

 

[kabin]

I think that has been normal operation for a long time. It treats you engaging FSD as a signal that it is OK to proceed through the light/stop sign. (Chuck has run into that on his UPLT, he has to enable FSD far enough back from the stop sign or it just goes right through.)

I tried it again this morning on one of my more frequently taken left turns and I couldn't duplicate it. It's curious for sure.

 

[black knight 3]

Not all. My ask has always been L3 on restricted access highways which is a reasonable goal. I never believed any hoopla over driverless FSD with the current hardware and software approach.

I get your point on that and I've owned 3 tesla's since 2018. However, it's questionable whether they have an advantage over their competition with regards to L3 on restricted highways if that's your only concern. And the never ending phantom braking is always present although much improved. For me personally, I would not have purchased my S and bought Mercedes if my criteria were only limited to L3 and restricted highways. I may eventually go that route as it looks like FSD is going to take a long time to be useful in my opinion.

 

[AlanSubie4Life]

I never believed any hoopla over driverless FSD with the current hardware and software approach.

I was always doubtful about this FSD, including L3 at low speed on freeways. (Tesla taking liability? Seems unlikely!)

I got FSD to get HW3 (sadly for $3k, not $2k) to ensure I would not be shut out of the promise of high performance active safety features (the “car that won’t crash”), realized over 3-5 years. Unfortunately these remain largely unrealized, and the current latency of FSD is disappointing and suggests this promise may not be realized. It’s also apparently not a focus for Tesla.

The dream remains unfulfilled. At least the HW3 was value realized, though I am not clear on exactly what that is getting me right now.

 

[Twiglett]

I was always doubtful about this FSD, including L3 at low speed on freeways. (Tesla taking liability? Seems unlikely!)

I got FSD to get HW3 (sadly for $3k, not $2k) to ensure I would not be shut out of the promise of high performance active safety features (the “car that won’t crash”), realized over 3-5 years. Unfortunately these remain largely unrealized, and the current latency of FSD is disappointing and suggests this promise may not be realized. It’s also apparently not a focus for Tesla.

The dream remains unfulfilled. At least the HW3 was value realized, though I am not clear on exactly what that is getting me right now.

I did the same - got FSD for $2k, no way I'd pay $15 for it.
That is the most frustrating part for me. Instead of finishing up on highway autopilot and making it shine and maybe even remove the beta tag, they instead regressed the whole fleet on a goal that is years away with every part unfinished.
Heck, they even disabled the also unfinished auto park - they seem to have an odd definition of "progress"

 

[AlanSubie4Life]

I did the same - got FSD for $2k, no way I'd pay $15 for it.
That is the most frustrating part for me. Instead of finishing up on highway autopilot and making it shine and maybe even remove the beta tag, they instead regressed the whole fleet on a goal that is years away with every part unfinished.
Heck, they even disabled the also unfinished auto park - they seem to have an odd definition of "progress"

It’s all going to be in v11; that is why it is taking so long. Tesla is always learning and this time will be delivering a finished product immediately. They now know the huge downside of over-the-air updates. 2023: The End of Beta