Wiki MASTER THREAD: Actual FSD Beta downloads and experiences

[EVNow]

We've definitely seen cases where it's clear phantom braking is not warranted, like shadows of trees. Or when there's really nothing at all. So these need to be eliminated.

Ofcourse. We are just saying there needs to be better definition of phantom braking.

 

[EVNow]

Here's a phantom brake that happened to me. A bike was traveling on the sidewalk in the direction of the red arrow. I could see that there wasn't an issue because I knew the bike would soon turn to follow the sidewalk parallel to the road. FSD, however, assumed we were on a collision course and slammed on the brakes.

I've seen slight slowdowns whenever a car turns into the oncoming lane from the left. Of course we all assume those cars are going to turn properly and not hit us - FSD assumes there is a certain non-zero probability of it hitting us and slows down. They need to adjust cut off probability - after all the probability of someone turning wrong and hitting us is quite low.

In your case, the code (NN ?) that tracks the motion of the bike and tries to estimate where it is going needs to know the bike path and how its turning. Needs to be map aware ...

 

[thewishmaster]

Still cannot tell why it keeps stopping at this intersection
Green light, with or without a lead car, FSD beta just pulls up to the line and stops. I paid close attention to the screen and there is no stop sign or anything unusual visualized. Maybe hard to see traffic lights? Though I vaguely recall seeing them visualized just fine. There's always too much traffic around to see whether beta will treat it as a stop sign or sit forever but if the conditions line up I'll give it a go.

That didn’t last long. After fixing itself in the middle of 10.4, the last 2 drives with 10.5 have had a 100% failure rate here. This morning I thought maybe the windshield is dusty and AP was misinterpreting something, but I washed the car and happened to pass the same intersection at night and it’s back to stopping

Also noticed a bunch of false positives when passing cars parked 90 degrees, facing the street - FSD came to a stop both times when the street was fairly empty and there were just 2 cars spaced about a block apart on the right side of the street. Full stop, then proceeds once it realizes there is no danger. But it also behaves weird when passing a whole row of cars parked the same way - it stays further left on such unmarked roads and phantom brakes more often than on other types of streets. I think this is also what makes it hesitate so much at 4-way stop signs - takes too long to figure out that the cars waiting on the cross street are not moving.

Finally, encountered a blinking yellow arrow that I’ve never seen before - T junction where the road directly ahead is one-way in the opposite direction and you can only go left or right. FSD was rightly confused and I took over.

 

[mxnym]

Just got FSD for the first time today! I have a 98 score. A couple questions - I live in the roundabout capital of the US, Carmel IN. We have over 100 of them. The car often activates turn signals on FSD when going through roundabouts - not good. It will activate the right turn signal when entering and then left turn signal midway through….just weird and confusing for other cars. Does anyone else notice this on roundabouts/rotaries?

I see this most of the time when I'm not taking the first exit. When I am taking the first exit, it's about 50/50 whether there will be any signal at all. Part of the problem is lack of consistency in both regulation and understanding. For instance, see some excerpts from this IndyStar article where drug charges were thrown out in March of 2020 because of an illegal traffic stop (for not signaling on a roundabout in Indiana):

“Why would a roundabout turn be different than any other turn,” asks Carmel Mayor James Brainard, leader of the so-called “roundabout capital” of the country. “It just doesn't make logical sense that because it’s a circle (as) opposed to a square, you would treat it differently.”

Brainard, who read the court’s decision, told IndyStar signaling on a roundabout should eventually be required and result in tickets for people who fail to do so.

So what is common sense turning in a roundabout? Before reading up, I'd argue signal right when taking the first exit and signal left when passing more than one exit. That may be based on previous input bias, though. I'm pretty sure I was instructed to signal when passing exits and when exiting decades ago, but it's a lot easier to treat the exit as a straight if passing the exit is a left (except when taking the first exit as a right turn, where a signal benefits others who are overly cautious about entering, to which point I agree Tesla's behavior is dangerous, but that doesn't mean regulation doesn't call for it somewhere), then again there are also potentially lane changes, especially in much larger traffic circles, plus there are roundabouts that are designed to NOT have lanes changed in regardless of the number of directions in which they have more than one lane.

The officer believed the motorist committed a traffic infraction because the officer had "always (been) instructed that you needed to signal for exiting a roundabout,” the court said.

Well, that doesn't match what I thought sounded like common sense, does it? The article also links to this INDOT page which has only this excerpt regarding turn signals:

• Exit:
  
  • Signal and exit to the right
  • Left turns are completed by circling around the center island and then making a right turn to exit

I can see how that would make sense, too, but if the education has changed over the decades and there isn't even any regulation in Indiana, who knows what other states' regulations hold.

 

[GranularHail]

Does this happen at night for you as well? When I drive at night, phantom braking is much reduced (practically nonexistent).

Same here. My first 10.5 drive was at night, so I had a very positive first impression, but daytime driving is a whole different experience

 

[Twiglett]

wonderfully odd thing today.
We have a longish driveway that's about 90ish feet long. FSD has always marked the sides in red as it should, but has always marked the end of the driveway in red as well, so I could never FSD out of the driveway.
I noticed this morning that my driveway showed like a regular T junction onto the road, so engaged FSD on the driveway.
The car couldn't cope
It wheel started making tiny movements left and right and it would make the smallest of movements forward then stop.
I sat there watching it do that for a minute or so before disengaging and retrying, same story.
That seemed to be a warning for the day so I drove manually

 

[EVNow]

Moved from the tweets thread.

Definitely worth a rewatch, now that I've had firsthand experience with the beta! Is it your understanding that the NN is completely in charge of the red light (mis)behavior seen in the video? Also wondering if perhaps the lane guides within the intersection confused the network and made it think it was stopping behind the line? Still a mistake that no human driver would ever make.

Looks like a route planner problem.

Sure, and this explains most of the failures in the video. But FSD will absolutely need to be able to dynamically read road signs and markings in order to reach any sort of city-street L4, or even halfway useful L2. I wonder where Tesla is at internally with this.

Nobody does that. Even most humans now rely on car nav or memory. Good 2-D map is needed for FSD. They will slowly add other signs ... but that will only be to override maps - like it happens with speed limit now.

A totally different approach they could take would be to make a monolithic GAN-style network that uses unsupervised unlabeled learning to simply predict what a human would do in any given situation, with an adversarial network able to take FSD-predicted paths and actual human-driven paths and determine which is which, then train the FSD network to generate paths that fool the discriminator. With this approach it could learn all the road signs implicitly, with no human labeling or annotation, and it might even be more robust than feature-engineered supervised learning.

End-to-end training is the holy grail of the industry. Everyone talks about it but nobody does it. I think the training compute needs are beyond what is practical currently. May be in 10 years ...

 

[tivoboy]

So I managed to capture on video what the AP system is seeing, and it sort of explains why it's doing what it is,, which is,
Come to a stop at this STOP LIGHT, and then start to launch into the intersection - at which point I brake and disenage.
then, I re-engage FSD again still sitting at the same stop light, and FSD starts to launch into the intersection again. Cars are coming from RIGHT and LEFT at speed mind you.

I've seen this on 10.4 and 10.5, so it's clearly not fixed.

I have NO idea why the cameras and system are seeing these RED LIGHTS as having a flash (one of them only) but watching it in the UI, it's not a constant rate, but at this light it probably saw the light go NOT red (so it's NOT seeing green till is actually green) and interpreting the NOT red of this one light in the middle as time to go into the intersection.

I can tell you I watched the stop light for nearly the whole time, it never flickered or blinked, it wasn't in the sun or shade. It was CONSTANT red and the car thought it was appropriate to go right through it. If I let the car do what it wanted and intended to do, it would have gone into the intersection and I would have been t-boned on the drivers side.

I have a video which I will send to tesla FSD beta, but wow. Here are four screenshots from the video all within about 3 seconds

 

[spokey]

So I've had the beta since the first rollout in mid Oct or so. But I don't often go out after dark due to poor eyesight. Tonight I let Sonya take us home and I must say that as some other have noted it does seem to work better at night. I think I had maybe one slight phantom break but other than that it performed flawlessly.

Got a nag question though. I think I've seen others speak about a chime if a nag is triggered. I don't get that and it would be helpful as I sometimes miss them at first as I"m looking at the road and not the screen. Is there one and if so how do you turn it on.

 

[EVNow]

I have a video which I will send to tesla FSD beta, but wow. Here are four screenshots from the video all within about 3 seconds

For some reason the NN probably thinks the probability of the red being lit is close to 50 ? So - as the probability varies, it goes on and off ? Normally you would expect the probability to be 90% or so - and thus a few percentage variation over time doesn't affect the classification.

 

[Gigaron]

That isn't what the release notes say. It is if you are put in AP jail three times:

Hmmm, I wonder if this includes forced disengagement for exceeding 80 mph. So far I have one because I needed to accelerate out of an accident situation on the freeway.

Meanwhile, I have to take over plenty of times when driving in the Bay Area.

280 south to 380 west - car wants to be in the backed-up far right lane while the next lane is clear, but it can’t handle where one lane becomes 3 with one exiting.

380 west to 101 north - car wants to get over to the right lane full of traffic instead of sticking to an open lane that also will exit. Then tries to kill me merging onto 101.

101 north to Grand Ave - 1 exit-only and another lane which make a 90 degree turn at the bottom of a hill. Car again likes the far right which also is an on-ramp lane and is messy.

Merges onto a freeway are problematic and turn signals never early enough.

 

[nvx1977]

Hmmm, I wonder if this includes forced disengagement for exceeding 80 mph. So far I have one because I needed to accelerate out of an accident situation on the freeway.

That forced disengagement does not count as a strikeout. Proven by a few people around here already.

 

[Ben W]

Looks like a route planner problem.

Can you elaborate? The projected path was straight ahead the entire time. Stepping frame-by-frame, the car slowed from 40mph to 25mph by the time it entered the intersection (while the light was still yellow, barely), then braked further to 15mph within the intersection before I overrode it. On the visualizer, the cross-traffic lane guides in the intersection kept flickering between dashed and solid, so maybe it was confused about whether those may have been the correct stopping points for the red light? Or possibly it couldn't accurately gauge the distance to the stopped cars far ahead? (Unlikely; they don't even show up on the visualizer.) I'm not sure it's knowable whether the car would have stopped completely in the intersection had I not overrode it, or whether it would have continued creeping through. My instinct in the moment was not to find out! But I would be very curious to know why it slowed down in the intersection, and why there isn't a higher-level system monitoring for obviously incorrect behavior like this and overriding it (or at least alerting the driver to override). Stopping in intersections for no apparent reason has been a big problem with both 10.4 and 10.5, in my experience.

Nobody does that. Even most humans now rely on car nav or memory. Good 2-D map is needed for FSD. They will slowly add other signs ... but that will only be to override maps - like it happens with speed limit now.

Strong disagree (re "nobody does that"). A good human driver will always look for and obey signs like No Right Turn on Red, especially at non-standard intersections (e.g. not 90°, on a hill, poor cross-visibility, etc.). "Relying on memory" begs the question. And how does a car nav tell a driver not to turn right on red? Ideally the maps would be as up-to-date as possible, but they will always be out of sync somewhere, and the level of driving reliability needed for L4 is much higher than the realistic reliability of maps. Unless of course, the maps are continuously and automatically updated by FSD systems trained to recognize them. But that again begs the question.

End-to-end training is the holy grail of the industry. Everyone talks about it but nobody does it. I think the training compute needs are beyond what is practical currently. May be in 10 years ...

I agree that FSD may not have the compute capability (inference) for true city-street L4 until HW4 at the earliest, perhaps HW5. Eventually it may be retroactively possible to achieve sufficient reliability with HW3 levels of computation, due to advances in ML efficiency, but that will probably be several years after it first becomes possible with HW4/HW5, and more compute will always be better/more reliable.

A problem with the current feature-engineered approach is that it's quite brittle; too much essential and subtle information is discarded during the categorizing and labeling. It may be the only workable approach at HW3 levels of compute, given the current state of ML knowledge, but it still may not get us where we need to be. Perhaps Dojo can leapfrog this and create a system that's even more training-heavy and inference-light, to allow a more end-to-end approach capable of running on HW3? One can hope!

 

[thewishmaster]

Tiny traffic circle thing confused 10.5 to the point that it stopped halfway into the intersection. This was a pretty small one in the middle of an otherwise normal intersection, with stop signs for the crossing street and no other traffic control for the road I was on. The route planner curved like it wanted to take a right turn. Seemed like it didn’t understand how it could go straight per the nav route.

Also ran into the same left turn as before where it was occluded by buses, but this time no buses and it behaved the same way - 1) missed the turn lane completely and 2) basically did not even attempt the left turn once in the turn lane and moving through the green light. It worked much better before so definitely a regression somewhere.

 

[JulienW]

First FSD Beta road trip and for me what I found. When driving at highway speeds (65MPH) on the FSD BETA Stack (NOT Vision AP Stack) on straight and wide 2 lane highways with minimal traffic the Phantom Braking is BEYOND horrendous. I could not go more than ½ a mile without having 1 or 2 events. It would happen with no shadows or anything to see but straight wide open highway. It would sometimes slow to 50MPH. Found the ONLY way to dive at highway speed on the Beta Stack was to keep the accelerator pedal pressed to the correct point. This of course leads to "Cruse Control Will Not Brake" warnings and TOTALY defeats using it but is the only way it will work. If this is a preview of Beta Stack integration Tesla is NO where near close.

So strange that the FSD Beta Stack is so completely overwhelmed by FB at highway speeds but the AP Vision Stack isn't. Seems like FSD Beta is going in the WRONG direction for FB at highway speeds on all Stack integration.

Funny (or sad) Anecdote:
Had just SC at the Metter SC and was crossing I16. A car was waiting on me to pass to turn left. Tesla PB and the other car took that as an opportunity to turn left in front of me. Tesla then accelerates and I had to hit brakes to avoided T-Boning. Reengaged FSD and first thing it does is PB brake again. So PB when nothing is there and accelerate when a car is in the path. Kinda backwards there Tesla.

 

[aronth5]

Tiny traffic circle thing confused 10.5 to the point that it stopped halfway into the intersection. This was a pretty small one in the middle of an otherwise normal intersection, with stop signs for the crossing street and no other traffic control for the road I was on. The route planner curved like it wanted to take a right turn. Seemed like it didn’t understand how it could go straight per the nav route.

Also ran into the same left turn as before where it was occluded by buses, but this time no buses and it behaved the same way - 1) missed the turn lane completely and 2) basically did not even attempt the left turn once in the turn lane and moving through the green light. It worked much better before so definitely a regression somewhere.

Here is a round about where my car stops 100% of the time entering the roundabout and then struggles trying to go around. Mostly just fails so I almost always take over since cars behind either honk or try and pass since my car "wanders" around. Pathetic effort actually.

 

[EVNow]

Tiny traffic circle thing confused 10.5 to the point that it stopped halfway into the intersection. This was a pretty small one in the middle of an otherwise normal intersection, with stop signs for the crossing street and no other traffic control for the road I was on. The route planner curved like it wanted to take a right turn. Seemed like it didn’t understand how it could go straight per the nav route.

Is the round about in TomTom map ? It gets confusing to FSD if it’s not in the map.

Without looking at the map, there is a way to tell. On the Nav steps, FSD will show any roundabouts.

 

[tivoboy]

For some reason the NN probably thinks the probability of the red being lit is close to 50 ? So - as the probability varies, it goes on and off ? Normally you would expect the probability to be 90% or so - and thus a few percentage variation over time doesn't affect the classification.

Don't you think the vision is actually trying to verify and quantify the RED light on state in real time?

 

[EVNow]

Don't you think the vision is actually trying to verify and quantify the RED light on state in real time?

Right - that is why the probability changes over time (say 30 times a second or whatever is their refresh frequency). If the probability is around 55%, may be sometime it falls to 45% - and will be thought of as Off. This can make it look like a on-off red light.

Pretty rare, I guess.

 

[tivoboy]

Right - that is why the probability changes over time (say 30 times a second or whatever is their refresh frequency). If the probability is around 55%, may be sometime it falls to 45% - and will be thought of as Off. This can make it look like a on-off red light.

Pretty rare, I guess.

Yeah, I guess it would be.. what is odd is that there was more than ONE RED light in this intersection and just as the NN will confer with each other/s to come up with concenssus, I wonder why it wouldn't take ALL the visible lights into consideration to validate RED vs. NOT red. The other lights in this intersection were of course all still RED to the human eye and they were ALSO red in the UI (at least what was displayed on the Tesla screen)