FSD beta on Laurel or Coldwater Canyon?

[Beachybumm]

Know this is a longshot, but is there someone with the FSD beta willing to go up and down either Laurel or Coldwater Canyon in LA to see how it performs? In the current fsd iteration, it's extremely dangerous. My car goes over the double line into oncoming traffic regularly. It reads stop signs on side roads that are angled toward the main road and brakes hard. It constantly forces disengagement due to the number of tight turns. And when you get to the intersection of Coldwater and Mulholland, the system freaks out because the road lines are offset.

I'd be super impressed if fsd could navigate this terrain, as it's one of the most difficult in LA.

 

[4SUPER9]

I would be interested in knowing this too. OTOH, I do my best to get on Coldwater by myself whenever I can, as it is just too fun to drive.

 

[Beachybumm]

No takers on this, apparently. I saw one today where a beta tester did Lombard Street in SF and the M3 wasn't even close to being able to do it. I assume the same for Coldwater and Laurel.

 

[Misterbee]

Perhaps not germane but, I have taken my regular autopilot M3 on Mulholland, and it navigates it beautifully as long as you set it at the speed limit.

 

[Spacep0d]

Perhaps not germane but, I have taken my regular autopilot M3 on Mulholland, and it navigates it beautifully as long as you set it at the speed limit.

Nice to know and I think it is germane.

 

[Beachybumm]

Perhaps not germane but, I have taken my regular autopilot M3 on Mulholland, and it navigates it beautifully as long as you set it at the speed limit.

I haven't tried it on Mulholland, but the current AP absolutely does not work well on Coldwater or Laurel. If you engaged it and let the car do everything, you'd be in a head on collision within seconds. Despite having tons of room on the non-oncoming traffic side, the car still chooses to cross the double line and risk a head on collision.

 

[BobbyKings]

Is this what you are discussing about? Watching from a distance it seems to be "trainable" I guess....

 

[Beachybumm]

Is this what you are discussing about? Watching from a distance it seems to be "trainable" I guess....

View attachment 604382

Yes, this is Coldwater. In particular at the intersection of Mulholland and Coldwater the current AP freaks out and start alerting and slamming on the brakes. But the entire stretch of Coldwater Canyon is a nightmare for AP. Back and forth turns throughout, some pretty tight. Cliff walls on each side of the curve, so if taken too fast, you'll hit the side of a mountain.

I'd also be interested in seeing how the beta performs in the Hollywood Hills. Almost every road is unmarked, and with cars parked on the side of the road, it's usually only one lane shares by traffic coming in both directions. As a human driver, it's not unusual to either backup or pull all the way over to let incoming traffic go by. Wonder if the beta knows to do that?

 

[Misterbee]

Mulholland is twisty, but much less steep than Coldwater or Laurel Canyon. I am guessing that the current software is a bit challenged when you add twisty plus steep.

 

[Beachybumm]

Mulholland is twisty, but much less steep than Coldwater or Laurel Canyon. I am guessing that the current software is a bit challenged when you add twisty plus steep.

To say the least. That's why I was hoping a beta tester would do the route. I assume the neural network would automatically correct and fix this software if the driver has to constantly disengage.

 

[TMYToday]

 

[Spacep0d]

Know this is a longshot, but is there someone with the FSD beta willing to go up and down either Laurel or Coldwater Canyon in LA to see how it performs? In the current fsd iteration, it's extremely dangerous. My car goes over the double line into oncoming traffic regularly. It reads stop signs on side roads that are angled toward the main road and brakes hard. It constantly forces disengagement due to the number of tight turns. And when you get to the intersection of Coldwater and Mulholland, the system freaks out because the road lines are offset.

I'd be super impressed if fsd could navigate this terrain, as it's one of the most difficult in LA.

All of Mulholland would be a doozy of a test, especially 'The Snake'.

 

[EVNow]

To say the least. That's why I was hoping a beta tester would do the route. I assume the neural network would automatically correct and fix this software if the driver has to constantly disengage.

Not automatically. They have to manually add that as training data (AFAIK).

As a human driver, it's not unusual to either backup or pull all the way over to let incoming traffic go by. Wonder if the beta knows to do that?

I was hoping that drive had some cars coming in the opposite direction. In what I watched, there wasn't anything. There are bridges which are single lane - we have talked about this situation in the past.

 

[gaswalla]

That video is awesome! Love how it stops for the pedestrian that jumps off the sidewalk

 

[TMYToday]

Is this what you are discussing about? Watching from a distance it seems to be "trainable" I guess....

View attachment 604382

Speaking of.. just saw this on my feed.

21:45 for the infamous triangle intersection.

 

[Mardak]

Speaking of.. just saw this on my feed.

21:45 for the infamous triangle intersection.

Wow. Very impressive taking many 15mph curves with no problem, and that sharp right onto Mulholland looked too easy for FSD beta. Separately, there was a neat mis-prediction where visually, it's not hard to see why the neural network got it wrong:

The left-side curbs before and after the intersection plus landscaping perfectly line up briefly leading to the neural network predicting a 3-way intersection. But getting closer to actually see the road on the left quickly switches to a 4-way intersection.

 

[EVNow]

Wow. Very impressive taking many 15mph curves with no problem, and that sharp right onto Mulholland looked too easy for FSD beta. Separately, there was a neat mis-prediction where visually, it's not hard to see why the neural network got it wrong:

View attachment 613184
View attachment 613186

The left-side curbs before and after the intersection plus landscaping perfectly line up briefly leading to the neural network predicting a 3-way intersection. But getting closer to actually see the road on the left quickly switches to a 4-way intersection.

They have to be using detailed maps to do something like this. This can't be just done visually. Infact to drive "like humans" you need atleast "human like" memory of the roads.

 

[gluu]

Wow. Very impressive taking many 15mph curves with no problem, and that sharp right onto Mulholland looked too easy for FSD beta. Separately, there was a neat mis-prediction where visually, it's not hard to see why the neural network got it wrong:

View attachment 613184
View attachment 613186

The left-side curbs before and after the intersection plus landscaping perfectly line up briefly leading to the neural network predicting a 3-way intersection. But getting closer to actually see the road on the left quickly switches to a 4-way intersection.

Good Eye. I think this is also a cause of phantom breaking in Autopilot, non-FSD. Images can spuriously give you the wrong impression, which is why the video training likely helps a great deal.

 

[gluu]

They have to be using detailed maps to do something like this. This can't be just done visually. Infact to drive "like humans" you need atleast "human like" memory of the roads.

I think it's ok to mispredict the intersection when you're a ways back, once you get closer you would hope that no misprediction would occur. There's no penalty for mispredicting it early on.