FSD Beta 10.12

[Bladerskb]

See- here's a great example of a guy who admits he didn't even watch the video and just made up some stuff he imagined it might have said instead.

You didn't answer the question and yes i have watched the video when it first came out.
I have watched all Tesla Tech presentation videos/podcast unlike you.
Now answer the question.

1. Did the guy also mention that they use a radar that is almost a decade old and has multiple orders of magnitude less resolution than radars others use for autonomous driving?
2. Did they talk about how two radars are not alike and that there are better radars which they are not using?
3. Or did they just talk about how radars suck?

 

[Knightshade]

You didn't answer the question and yes i have watched the video when it first came out.
I have watched all Tesla Tech presentation videos/podcast unlike you.

Weird, since I actually know what's in them while you previously wrote

Did the guy also mention that they use a radar that is almost a decade old and has multiple orders of magnitude less resolution than radars others use for autonomous driving?

Did they talk about how two radars are not alike and that there are better radars which they are not using?

Or did they just talk about how radars suck?

So you watched it but somehow have no idea what was in it?

If you're gonna keep trolling the thread with nonsense can you at least try and make it consistent nonsense?

Now answer the question.

1. Did the guy also mention that they use a radar that is almost a decade old and has multiple orders of magnitude less resolution than radars others use for autonomous driving?
2. Did they talk about how two radars are not alike and that there are better radars which they are not using?
3. Or did they just talk about how radars suck?

Even weirder- you just did it again!


You claim you watched it, but keep demanding people explain to you what it said.


Is the problem that you DID watch it but didn't understand the words?

Which words, specifically, were giving you trouble?

 

[Bladerskb]

Even weirder- you just did it again!

You claim you watched it, but keep demanding people explain to you what it said.

Is the problem that you DID watch it but didn't understand the words?

Which words, specifically, were giving you trouble?

So you can't actually answer the 3 questions. Gotcha.

 

[Jmolloy3]

Sure would be nice if there was less pissing and more constructive dialog.

 

[Knightshade]

So you can't actually answer the 3 questions. Gotcha.

Of course I can. Since unlike you I actually watched the presentations.


The thing is- if you had watched the videos you claim you did, you wouldn't be asking the questions. Since they're answered by watching them.


Your current postings are like coming in here claiming you watched Empire Strikes Back then asking if anyone can tell you who Lukes father is.

Gotcha indeed


And the best way to answer your question would be to tell you to go actually watch the videos.

 

[dramsey]

Saw a phantom blinker on a rendered car today. Might explain why the beta always tends to veer right at that intersection - it must think the left lane is actually a left turn lane when actually the right lane is a right-turn-only lane. The car in front of me did not have their left blinker on at any point and they indeed went straight.

IMHO the single biggest problem the current FSD has is being in the wrong lane at the wrong time. I was thinking about this and would guess that the Google maps data the car uses to plot its route doesn't include lane info, like "this road has three lanes, with a temporary right turn lane branching off at the next intersection." If this is correct then the car must discern the lanes on its own, every time.

One of the most common reasons I have to disengage is my car finding itself in a dedicated turn lane and having to either give up or make a very aggressive maneuver to get back in the correct lane. This behavior is distinct from the "turning right on a street and immediately moving into the leftmost lane even though there's another right turn coming up at the next intersection a block away" behavior.

Prior to the FSD beta, my car's visualizations would clearly show turn arrows painted on the road to designate dedicated turn lanes. The FSD visualization never shows these arrows, and I'd guess FSD doesn't "know" a lane is a turn lane until it sees something else like a curb or turning outer lane line...

 

[Knightshade]

IMHO the single biggest problem the current FSD has is being in the wrong lane at the wrong time. I was thinking about this and would guess that the Google maps data the car uses to plot its route doesn't include lane info

Tesla does not use google maps for routing.

Prior to the FSD beta, my car's visualizations would clearly show turn arrows painted on the road to designate dedicated turn lanes. The FSD visualization never shows these arrows, and I'd guess FSD doesn't "know" a lane is a turn lane until it sees something else like a curb or turning outer lane line...

Weird- my FSDBeta absolutely still shows arrows visible in lanes. It also shows those "this lane is ending, merge THAT way" arrows-- but still waits to the last minute to do anything about it.


That said I agree it often makes the wrong choice about what lane to be in, and this is very likely from bad or incomplete map data.

But the question is- should the fix be "have perfect map data everywhere" or "make the system smart enough to not rely on map data"

Both are actually pretty hard to do


One example I run into often since it's a turn the car has to make on the way to work.... the road it's on is 2 lanes each way... divided (though there are intersections). At the next light it will need to get into a right turn lane (that only exists near the intersection) and turn right.

But-- a little bit before that.... a DIFFERENT temporary/short right turn lane appears to turn into a business park.

every time, the beta dives into that lane, since it "knows" it has an upcoming "temp right turn lane" to get into. Except that lane doesn't go to the light. It goes to a little island that ends the lane right after the turn into the business park.

So every time I need to manually intervene to correct it.

 

[sleepydoc]

IMHO the single biggest problem the current FSD has is being in the wrong lane at the wrong time. I was thinking about this and would guess that the Google maps data the car uses to plot its route doesn't include lane info, like "this road has three lanes, with a temporary right turn lane branching off at the next intersection." If this is correct then the car must discern the lanes on its own, every time.

One of the most common reasons I have to disengage is my car finding itself in a dedicated turn lane and having to either give up or make a very aggressive maneuver to get back in the correct lane. This behavior is distinct from the "turning right on a street and immediately moving into the leftmost lane even though there's another right turn coming up at the next intersection a block away" behavior.

Prior to the FSD beta, my car's visualizations would clearly show turn arrows painted on the road to designate dedicated turn lanes. The FSD visualization never shows these arrows, and I'd guess FSD doesn't "know" a lane is a turn lane until it sees something else like a curb or turning outer lane line...

Lane selection has been the achilles heel of the last 2 versions. Like you, it accounts for the overwhelming majority of my disengagements followed by difficulties with yield signs. They’re supposedly working in this and there were some marginal improvements from v10.11 to 10.12 but it’s clearly still an issue.

 

[dramsey]

Tesla does not use google maps for routing.

Really? Who do they use? Isn't the in-car nav database from Garmin?

Weird- my FSDBeta absolutely still shows arrows visible in lanes. It also shows those "this lane is ending, merge THAT way" arrows-- but still waits to the last minute to do anything about it.

Mine absolutely does not. There are many of these arrows in my neighborhood. The main road leading into my area is a two-lane road with very clear markings. Turn lanes branch off to side streets in either direction-- a right turn lane for streets to the right and a center turn lane (bordered in yellow) for streets to the left. Each has a large arrow that was clearly rendered in the pre-FSD visualizations, but isn't now. I'll try to pay attention to arrows in other areas to see if they show up there.

Oh, yeah: the prior visualizations also showed the bike lane markings, i.e. the little bike symbol, and FSD does not.

One example I run into often since it's a turn the car has to make on the way to work.... the road it's on is 2 lanes each way... divided (though there are intersections). At the next light it will need to get into a right turn lane (that only exists near the intersection) and turn right.

But-- a little bit before that.... a DIFFERENT temporary/short right turn lane appears to turn into a business park.

every time, the beta dives into that lane, since it "knows" it has an upcoming "temp right turn lane" to get into. Except that lane doesn't go to the light. It goes to a little island that ends the lane right after the turn into the business park.

So every time I need to manually intervene to correct it.

I've had the same thing happen many times. This implies that the map data doesn't have the turn lane info and that the car is figuring it out itself since it doesn't know how far from the turn the turn lane will start.

Man, FSD is indeed hard.

 

[Knightshade]

Really? Who do they use? Isn't the in-car nav database from Garmin?

That has been a matter of considerable debate.... with some evidence they get at least some stuff from TomTom, and more evidence they get a fair bit from Mapbox and OSM.

For routing they use Valhalla

Valhalla multi-purpose routing engine - help, consulting, instant download of Tilepacks - Interline Technologies

Valhalla is a multi-purpose routing engine. It's open-source software, originally created by Mapzen. It can be used to compute turn-by-turn directions, time/distance matrixes, optimized routes, access isochrones, and location-to-map matches. Interline Technologies provides help, custom...

www.interline.io

 

[Dewg]

During navigation, my car does show lanes when there will be a maneuver. For example, if there is a right turn ahead, at the top of the navigation panel, it will show the lanes (for example 4 lanes), with the lane the car needs to be in highlighted in black (in this case the #4 lane).

I also see some markings on the road, such as arrows, but these are just visualizations and not indicative of logic the system will act on. I don't think the car sees the merging arrows painted on the street and acts upon that visualization. It still watches lane markers, and when a lane ends, it moves over when required - which usually is at the last moment. That can be frustrating for some people who are expecting different. This could be helped with more detailed mapping data, but it seems Tesla is going for less mapping data. If that's the case, then they need to definitely start picking up more signs, both on poles and painted on the road to help navigate. I do understand why Tesla is going for less - it's been their position that humans don't have HD maps in our heads, we drive by reading road signs, following flows of traffic, and reading painted cues in lanes.

 

[jabloomf1230]

I've never seen it suggest or offer more than one route

Nor have I. Maybe it's being tested in a limited USA region.

 

[Harriscott]

Humans also improve their lane selection, speed, etc. as they drive the same route more than once. This is particularly helpful in heavy traffic when painted cues are often obscured and last-moment adjustments are difficult. So mapping of some sort remains useful.

 

[JulienW]

Nor have I. Maybe it's being tested in a limited USA region.

That is because it is NOT enabled and no one in the US has it. Greentheonly (a hacker) found it looking through the code and "forced flagged" it on so he could try it.

 

[DrChaos]

IMHO the single biggest problem the current FSD has is being in the wrong lane at the wrong time. I was thinking about this and would guess that the Google maps data the car uses to plot its route doesn't include lane info, like "this road has three lanes, with a temporary right turn lane branching off at the next intersection." If this is correct then the car must discern the lanes on its own, every time.

One of the most common reasons I have to disengage is my car finding itself in a dedicated turn lane and having to either give up or make a very aggressive maneuver to get back in the correct lane. This behavior is distinct from the "turning right on a street and immediately moving into the leftmost lane even though there's another right turn coming up at the next intersection a block away" behavior.

Prior to the FSD beta, my car's visualizations would clearly show turn arrows painted on the road to designate dedicated turn lanes. The FSD visualization never shows these arrows, and I'd guess FSD doesn't "know" a lane is a turn lane until it sees something else like a curb or turning outer lane line...

This sounds like a misguided insistence on never using maps or avoiding buying good maps. I don't think the autonomy engineers would choose that decision.

I think functioning somewhat acceptably in areas where maps are unavailable (rural areas) is needed, but in the common areas they need maps because the driving computers aren't at general human intelligence and won't ever be. Humans use semantic information from diverse sources---and their own memory---to decide on the proper behaviors. A very important one is watching what other cars do (because that leads to less surprising driving for others), and another is remembering what cars did when they saw this intersection before. Others are complex things like "city buses only 4-8 pm weekdays". Those routes are definitely drivable but shouldn't be.

This is one area where an Apple Car might have a strong advantage: Apple Maps is now excellent. Good maps make up for trying to figure everything out from AI on vision.

This is not an impossible to solve problem either---they could collect driving path data from their large fleet of human drivers with AP/FSD off and build up this lane information on their own over months, at least for the well travelled areas. But the well travelled areas are of course the high volume ones people experience all the time.

If it were me I would build this dataset, then train the driving policy machine learning nets with half the examples of the map data dropped out and half with mapped data included, perhaps even with multi-task learning, like training nets where many layers are shared but a few layers are specific to mapped case vs map-less case. Not a big deal with modern nnet packages, I've personally done this in a different problem.

The mapping data could even be used to govern the optimization target for the mapless situations, such that producing a solution which stays close to pre-mapped routes is preferred over one which contradicts those routes.

 

[DrChaos]

Really? Who do they use? Isn't the in-car nav database from Garmin?

Mine absolutely does not. There are many of these arrows in my neighborhood. The main road leading into my area is a two-lane road with very clear markings. Turn lanes branch off to side streets in either direction-- a right turn lane for streets to the right and a center turn lane (bordered in yellow) for streets to the left. Each has a large arrow that was clearly rendered in the pre-FSD visualizations, but isn't now. I'll try to pay attention to arrows in other areas to see if they show up there.

Oh, yeah: the prior visualizations also showed the bike lane markings, i.e. the little bike symbol, and FSD does not.

I've had the same thing happen many times. This implies that the map data doesn't have the turn lane info and that the car is figuring it out itself since it doesn't know how far from the turn the turn lane will start.

Man, FSD is indeed hard.

It wouldn't be that hard if they crowdsourced route paths from human drivers off AP. There would be lots of paths which end up having turned on the intersection, and the majority of them would have followed the correct procedure. The algorithm "do what the median human (with score 90+) has been previously measured to do in this intersection for those humans who are going in the same route we are" is a good heuristic.

People do this too. Humans might have made the similar mistake, but they'd only do it once or twice. And in significant traffic where humans can see the people turning ahead would also see the vast majority of people avoiding the first exit line vs the second, knowing that most people will be turning onto the main road and they probably want to do that too.

 

[jabloomf1230]

That is because it is NOT enabled and no one in the US has it. Greentheonly (a hacker) found it looking through the code and "forced flagged" it on so he could try it.

View attachment 823212

I posted the original Twitter thread. We have no idea who else has it. The flag could be enabled for either Tesla employees or other people who are testing the option. greentheonly is often tipped off on hidden features like this.

 

[sleepydoc]

It wouldn't be that hard if they crowdsourced route paths from human drivers off AP. There would be lots of paths which end up having turned on the intersection, and the majority of them would have followed the correct procedure. The algorithm "do what the median human (with score 90+) has been previously measured to do in this intersection for those humans who are going in the same route we are" is a good heuristic.

People do this too. Humans might have made the similar mistake, but they'd only do it once or twice. And in significant traffic where humans can see the people turning ahead would also see the vast majority of people avoiding the first exit line vs the second, knowing that most people will be turning onto the main road and they probably want to do that too.

No one knows for sure where Tesla gets its mapping data; seems to be from a combination of sources including google and tom tom. Regardless, I've wondered why they need to - they have an army of cars equipped with GPS and cameras roaming the streets every day and it seems like they could crowdsource their maps from their own cars. It would likely give more rapid information on road changes as well.

Mine absolutely does not. There are many of these arrows in my neighborhood. The main road leading into my area is a two-lane road with very clear markings. Turn lanes branch off to side streets in either direction-- a right turn lane for streets to the right and a center turn lane (bordered in yellow) for streets to the left. Each has a large arrow that was clearly rendered in the pre-FSD visualizations, but isn't now. I'll try to pay attention to arrows in other areas to see if they show up there.

Oh, yeah: the prior visualizations also showed the bike lane markings, i.e. the little bike symbol, and FSD does not.

During navigation, my car does show lanes when there will be a maneuver. For example, if there is a right turn ahead, at the top of the navigation panel, it will show the lanes (for example 4 lanes), with the lane the car needs to be in highlighted in black (in this case the #4 lane).

I also see some markings on the road, such as arrows, but these are just visualizations and not indicative of logic the system will act on. I don't think the car sees the merging arrows painted on the street and acts upon that visualization. It still watches lane markers, and when a lane ends, it moves over when required - which usually is at the last moment. That can be frustrating for some people who are expecting different. This could be helped with more detailed mapping data, but it seems Tesla is going for less mapping data. If that's the case, then they need to definitely start picking up more signs, both on poles and painted on the road to help navigate. I do understand why Tesla is going for less - it's been their position that humans don't have HD maps in our heads, we drive by reading road signs, following flows of traffic, and reading painted cues in lanes.

Something to remember is that the screen visualization is not necessarily a complete representation of all the data that the car is sensing and acting on. Interestingly, if you try to use plain AP off-highway, it may show the lanes with arrows but it completely ignores them. The AP logic is such that it picks the path that is straight ahead (or continuing on the trajectory the car was on,) regardless of the type of lane it is. FSD applies much more analysis to the task, even if it doesn't always get it right.

 

[tmoz]

So bringing back EAP seems that Musk acknowledges that FSD is not right around the corner anymore.
Perhaps his experience with the Single Stack made him change his mind.
I wonder what that does to the FSD roadmap - continue with separate stacks longer?

 

[Knightshade]

So bringing back EAP seems that Musk acknowledges that FSD is not right around the corner anymore.

As mentioned elsewhere this is more a Rorschach test than anything else.

To some it means what you suggest.

To others it means exactly the opposite-- he's so confident FSD will be ready for wide release soon that they can now sell it as its own thing again (when previously they had to move all the EAP features into it for it to have any value as a package)