FSD Beta 10.69

[AlanSubie4Life]

But, if Tesla has ideas about becoming L4/L5, at some point, they really need to be able to pass the guy doing 45 on a 75 mph road. Certainly not required for L2 or even L3.

I added this above, but I don’t think L4 vehicles will do it except (perhaps) in very extreme circumstances.

If you’re L4, there is no rush. And it’s just too difficult.

 

[Supcom]

Thinking about this critically, this is actually a really good case study in human perception and context clues.

The first (and obvious) context clue in determining which light is ours is one being slightly less visible than the other. This context is probably what most perception networks should attempt to perceive first:

View attachment 850437

But failing that, and taking this edge case to the extreme, what happens when the lights appear as if they are equal in size? Without distance or shadows, it's literally impossible to know which light is ours in this case:

View attachment 850436

Oddly enough, I think the next best context clue that we all pick up on (but don't consciously realize we consider) is actually the orientation of the wire on which the lights are. Here's the lights plus the wire as they appeared in the video:

View attachment 850435

And here's the wire if you were to view it from the perspective of a driver in the perpendicular lane:

View attachment 850434

So we instinctually know that because the lights are designed to face two adjacent angles in the 4-way intersection, the light that faces the direction that the wire opens must belong to the other lane, and can be ignored.

I regularly see traffic lights that are wrongly pivoted in ways that make them difficult to see. And your examples using wire angles become difficult to discern at night. A traffic light should never be visible to the wrong direction of travel. The example in the video should be corrected by adding longer shades to the lamps.

 

[Daniel in SD]

Thinking about this critically, this is actually a really good case study in human perception and context clues.

The first (and obvious) context clue in determining which light is ours is one being slightly less visible than the other. This context is probably what most perception networks should attempt to perceive first:

View attachment 850437

But failing that, and taking this edge case to the extreme, what happens when the lights appear as if they are equal in size? Without distance or shadows, it's literally impossible to know which light is ours in this case:

View attachment 850436

Oddly enough, I think the next best context clue that we all pick up on (but don't consciously realize we consider) is actually the orientation of the wire on which the lights are. Here's the lights plus the wire as they appeared in the video:

View attachment 850435

And here's the wire if you were to view it from the perspective of a driver in the perpendicular lane:

View attachment 850434

So we instinctually know that because the lights are designed to face two adjacent angles in the 4-way intersection, the light that faces the direction that the wire opens must belong to the other lane, and can be ignored.

I think I know because one light is facing me and the other is facing side traffic, hence why they're painted yellow to make it easy to see. I guess everyone's brain works differently!

 

[EVNow]

I’ll probably post a few more. My commitment to you all is to keep production value low.

And my commitment is to comment without watching.

Just like we all do for linked stories on twitter.

 

[EVNow]

If I came upon it in the evening or night I could easily get it wrong.

I am confident you would not.

I'm confident someone driving there for the first time would get confused.

Currently Beta drives every inch like its never driven there before. A lot of "instinct" and "you would know" are just memorized responses. You see drivers new to the area struggling all the time ... and new drivers too. Combine that and you know how FSD Beta drives

 

[Bladerskb]

I'm confident someone driving there for the first time would get confused.

Lol no. Proven by the fact that i was the one who posted that video

I watched it for the first time without knowing beforehand that it tried to run that red light.
I also watched it with-out audio and even I had the "oh *sugar* its about to run a RED light!" response just by watching it for the first time.
I have never been to New York (i hope to go soon though )

This was 100% obvious, plus I have seen "worse".

Currently Beta drives every inch like its never driven there before. A lot of "instinct" and "you would know" are just memorized responses. You see drivers new to the area struggling all the time ... and new drivers too. Combine that and you know how FSD Beta drives

Funny someone once said that HD map provides foresight.

 

[Daniel in SD]

Is there some sort of orientation blindness condition I've never heard of?
Here are four traffic lights, which one is facing you?

I think some people are trying to imagine themselves as artificial NN's instead of the product of a billion years of evolution...

 

[EVNow]

Lol no. Proven by the fact that i was the one who posted that video

If that is the kind of "proof" you can provide - you have already lost the argument.

 

[AlanSubie4Life]

Currently Beta drives every inch like its never driven there before. A lot of "instinct" and "you would know" are just memorized responses. You see drivers new to the area struggling all the time ... and new drivers too. Combine that and you know how FSD Beta drives

I agree. This seems very bad. What do you think will be done to resolve this?

I'm confident someone driving there for the first time would get confused.

I’m really quite confident they would not. I don’t understand what is confusing about these traffic lights. Their meaning seems extremely clear!

 

[willow_hiller]

Also, in light of this video in NY, I've realized that counting disengagements is a relatively useless metric of progress for FSD Beta. It's way too circumstantial and way too dependent on a test driver's chosen routes.

If this guy chooses a route with an obstacle that FSD Beta categorically cannot handle, he'll have a disengagement every day. And likewise if someone just happens to have a test route that doesn't pass any edge-cases, they might never have a disengagement.

The better metric is the percentage of routes (or road miles) that can be driven disengagement free. At what point can 50% of the roads in your area be driven disengagement free, 80%, 90%, 99%, etc. And then jointly, what percent of roads in the US can be driven disregard free for any given firmware version. It moves the local variability out of the error term for the metric and turns it into the metric itself.

 

[FSDtester#1]

"Took the long City Streets route to AutoZone and was super impressed with the results. Everything felt smoother!! Went through school zones, round about, construction, etc."

I didn't know Tesla drivers went to Autozone....

 

[LowlyOilBurner]

Brutal. Happens to everyone. Watch that regen; that’s how you know!

Note the car applied power going into the turn lane. Then maxed out regen (with some brakes I think but would need a bigger screen; the nose of the car dives). There was no way to predict it would need to stop, I guess.

So, I’m thinking some of this behavior is part of the 2022.20.X branch. The regen/friction brake thing doesn’t work well at all on my car running 2022.20.8. In fact, I can’t even toggle the button to enable it, it always defaults to “off”. The slow acceleration from a stop sign is also a thing on .20.8. The auto high beams are pitiful as well. There are a few other things, but this is just a theory I’m tossing out there.

 

[ChaoticNeutral]

Also, in light of this video in NYC, I've realized that counting disengagements is a relatively useless metric of progress for FSD Beta. It's way too circumstantial and way too dependent on a test driver's chosen routes.

If this guy chooses a route with an obstacle that FSD Beta categorically cannot handle, he'll have a disengagement every day. And likewise if someone just happens to have a test route that doesn't pass any edge-cases, they might never have a disengagement.

The better metric is the percentage of routes (or road miles) that can be driven disengagement free. At what point can 50% of the roads in your area be driven disengagement free, 80%, 90%, 99%, etc. And then jointly, what percent of roads in the US can be driven disregard free for any given firmware version. It moves the local variability out of the error term for the metric and turns it into the metric itself.

Tesla/Google chooses the routes though, the car is only going to follow what Navigation says regardless of whether it’s able to handle the obstacle.

 

[willow_hiller]

Tesla/Google chooses the routes though, the car is only going to follow what Navigation says regardless of whether it’s able to handle the obstacle.

If you watch that video from NY, he specifically says he's disengaging in order to place the car in a position where it will need to attempt a left-turn that it has previously failed to do. So many of the FSD Beta test drivers seem to have their favorite routes and road features to test.

 

[jeremymc7]

I regularly see traffic lights that are wrongly pivoted in ways that make them difficult to see. And your examples using wire angles become difficult to discern at night. A traffic light should never be visible to the wrong direction of travel. The example in the video should be corrected by adding longer shades to the lamps.

Ahh but how to fix the same direction of traffic w/ a dedicated turn lane with light and separate straight (and right turn) lanes with their own dedicate light.

When the turn and straight lights are next to each other beta often gets confused and does the opposite; thus running a red.

I guess there is no validation between I'm in the left turn lane, my left turn blinker is on, so I should only be looking at the left arrow light. Instead the left arrow is red and the green straight next to it tells the car it should run the left red arrow.

Ugh.

 

[Supcom]

If you watch that video from NY, he specifically says he's disengaging in order to place the car in a position where it will need to attempt a left-turn that it has previously failed to do. So many of the FSD Beta test drivers seem to have their favorite routes and road features to test.

It's fair game to go through some gyrations to cause the car to create a route through an area that you want to test. The car needs to be able to handle any legal routing. I put my car in locations so it tests certain roads/intersections like a lot of testers. Edge cases need to be submitted to Tesla so that they can be (hopefully) fixed.

That said, the fact that the car fails certain (many) scenarios is to be expected at this stage of development.

 

[jeremymc7]

Is there some sort of orientation blindness condition I've never heard of?
Here are four traffic lights, which one is facing you?
View attachment 850452View attachment 850453View attachment 850454View attachment 850455

I think some people are trying to imagine themselves as artificial NN's instead of the product of a billion years of evolution...

There are instances of these situations at say 5-way intersections where both beta and human drivers often get confused. The angles of some of the lights are so close to each other depending on where you are in the lanes and which street. Especially since there are few/no arrow lights. Further complicated by the no turn on red in many directions due to crossing over another street or two to get to your desired street.

Is that the turn light or the straight light?

 

[Supcom]

Ahh but how to fix the same direction of traffic w/ a dedicated turn lane with light and separate straight (and right turn) lanes with their own dedicate light.

When the turn and straight lights are next to each other beta often gets confused and does the opposite; thus running a red.

I guess there is no validation between I'm in the left turn lane, my left turn blinker is on, so I should only be looking at the left arrow light. Instead the left arrow is red and the green straight next to it tells the car it should run the left red arrow.

Ugh.

Tesla obviously needs to work on these cases. How to fix it is Tesla's problem.

 

[Supcom]

There are instances of these situations at say 5-way intersections where both beta and human drivers often get confused. The angles of some of the lights are so close to each other depending on where you are in the lanes and which street. Especially since there are few/no arrow lights. Further complicated by the no turn on red in many directions due to crossing over another street or two to get to your desired street.

Is that the turn light or the straight light?

View attachment 850493

From the signage, I'd conclude that there is no straight light because all lanes turn left or right. Probably no distinction between left and right at this intersection since the center lane is permitted to go left or right. Probably also why the lights I see are not arrows.

 

[willow_hiller]

I guess there is no validation between I'm in the left turn lane, my left turn blinker is on, so I should only be looking at the left arrow light. Instead the left arrow is red and the green straight next to it tells the car it should run the left red arrow.

I came across this monstrosity the other day:

Five lanes. Left turn lane with an arrow shaped light, three forward lanes signaled by regular lights, and a single right-turn lane with two regular (non-arrow) lights. I guess the city ran out of right-arrow lights and just decided to go with two regular lights and a "Right turn signal" sign.

I was in the fourth lane from the left, and my green-light chime sounded when the right-turn signals turned green (while my lane's signal stayed red). And then it chimed again when my proper light turned green.