FSD rewrite will go out on Oct 20 to limited beta

[Mo City]

I like the approach of this Zeb Hallock guy in North Carolina. Very good video posted a few hours ago, especially the first 7.5 mins.

At ~6:25, FSD makes a left turn at a red light with a yellow arrow.

Only complaint is I would like to see at least a couple of failures in a 10 min video (unless there were none ).

 

[turnem]

I like the approach of this Zeb Hallock guy in North Carolina. Very good video posted a few hours ago, especially the first 7.5 mins.

At ~6:25, FSD makes a left turn at a red light with a yellow arrow.

Only complaint is I would like to see at least a couple of failures in a 10 min video (unless there were none ).

Yea - I'd like to see some serious stress testing to understand where it fails. We've seen great documentation of what it does well and we all know there are going to be areas that it struggles with so I'd like to see what those are.

In particular... with the non-beta software my M3 will often misread lane lines when there is a bend in the road (almost always a left hand bend) and the lane lines break in the bend. It almost always wants to pull LEFT very abruptly as if it's seeking out that left lane marker as its grounding point.

I'm curious if issues like this are resolved or not with the FSD Beta software.

 

[Mo City]

A couple of shorties from James Locke

 

[Mardak]

Quite impressive past-the-roundabout prediction even with Autopilot cameras positioned slightly higher than what's used to record Tesla FSD BETA roundabout testing (plus unprotected left) @ 2:08. There's a traffic light barely visible for a few seconds past the trees and central mound vegetation, and the visualization shows a predicted angled roundabout exit while the road is obscured as engineered. Maybe it's also using the visible building's wall and lined trees to infer the angle of the road?

 

[AlanSubie4Life]

Only complaint is I would like to see at least a couple of failures

Failure #1: It stopped at the yield sign on entry to the roundabout, which would result in a deduction on a driver's test if there were no traffic coming. Not only is it unsafe (someone behind you can run into you because they do not expect you to stop), but this can also cause massive congestion at roundabouts. Roundabouts tend to be extremely bad for traffic flow at heavy times due to inadequate interleaving (roundabouts in the UK are often a nightmare of congestion, in my experience). Stopping on entry doesn't help this problem. You're only supposed to yield, not stop!

Failure #2 (minor): It also failed to signal right to exit the roundabout (the left signal on entry, since he wasn't taking the first exit is correct, as explained), which is what you are supposed to do on exiting a roundabout, right as you pass the exit prior to your exit.

 

[willow_hiller]

FSD beta avoiding debris at the edge of a lane:

First time I've seen it do that; and it even seems to have shown up on the visualization for a frame or two as a blue dot.

 

[GZDongles]

Quite impressive past-the-roundabout prediction even with Autopilot cameras positioned slightly higher than what's used to record Tesla FSD BETA roundabout testing (plus unprotected left) @ 2:08. There's a traffic light barely visible for a few seconds past the trees and central mound vegetation, and the visualization shows a predicted angled roundabout exit while the road is obscured as engineered. Maybe it's also using the visible building's wall and lined trees to infer the angle of the road?

View attachment 605180

Doesn't it know where the exit is from the map? (Hoping the 40 pages of arguing about the definition of HD in HD maps is behind us )

 

[mikes_fsd]

Hoping the 40 pages of arguing about the definition of HD in HD maps is behind us

Speaking of this...
There was another 10+ pages of arguing that the landing page for Autopilot should not be considered "accurate" because it is not linked to from within Tesla's main site.

But Tesla keeps updating the Autopilot landing page...
Today....

One year ago:

 

[MP3Mike]

Brandon is back with an hour long video: (I haven't watched it yet.)

 

[Daniel in SD]

Brandon is back with an hour long video: (I haven't watched it yet.)

I watched a little of it. It seems like a major concern of his, and mine, is how the car deals with cross traffic at unprotected intersections. The issue is that if the view is obstructed the car can start to cross quickly without the driver being able to verify that the path is clear. It seems like a very difficult maneuver to monitor as one must be aware of traffic in both directions and be very quick to stop the car if it starts moving when it shouldn't. I wonder if an interface where the driver is required to keep their foot on the brake and release it when it is safe to go would be a good idea?

 

[turnem]

I watched a little of it. It seems like a major concern of his, and mine, is how the car deals with cross traffic at unprotected intersections. The issue is that if the view is obstructed the car can start to cross quickly without the driver being able to verify that the path is clear. It seems like a very difficult maneuver to monitor as one must be aware of traffic in both directions and be very quick to stop the car if it starts moving when it shouldn't. I wonder if an interface where the driver is required to keep their foot on the brake and release it when it is safe to go would be a good idea?

I agree that this area still needs work - if for no other reason than to give the driver confidence that it is indeed all-clear.

Later on in the video he shows how the car DOES sometimes creep up to peak around the corner (and it notifies you that it's doing so in the UI). This may be enough to make drivers more comfortable. But it needs to do it consistently.

 

[Cheburashka]

Yeah that video is pretty rough to be honest. It needs a lot more work. I seriously doubt we will have any sort of public rollout next month.

 

[Bill Price]

I was waiting for someone to say that.

90% of us were waiting for someone to say that!

 

[powertoold]

I seriously doubt we will have any sort of public rollout next month.

The fact that we're evening thinking in the timeframe of months is incredible. Tesla's come a long way.

 

[mhan00]

I watched a little of it. It seems like a major concern of his, and mine, is how the car deals with cross traffic at unprotected intersections. The issue is that if the view is obstructed the car can start to cross quickly without the driver being able to verify that the path is clear. It seems like a very difficult maneuver to monitor as one must be aware of traffic in both directions and be very quick to stop the car if it starts moving when it shouldn't. I wonder if an interface where the driver is required to keep their foot on the brake and release it when it is safe to go would be a good idea?

The car is also struggling with the weird ass “roundabouts” he has there, which are, as far as I can tell, just four way intersections with a concrete circle in the middle for some reason. He said his car handled those in .10 and .11, but here in this vid of .12 the car was routing through them and he had to take over. I’m wondering if it’s related to a vid I saw back in the original beta release (.10) where there was a circle painted in the middle of the intersection that the beta thought was a traffic circle and tried to drive around?

 

[turnem]

The car is also struggling with the weird ass “roundabouts” he has there, which are, as far as I can tell, just four way intersections with a concrete circle in the middle for some reason. He said his car handled those in .10 and .11, but here in this vid of .12 the car was routing through them and he had to take over. I’m wondering if it’s related to a vid I saw back in the original beta release (.10) where there was a circle painted in the middle of the intersection that the beta thought was a traffic circle and tried to drive around?

I had this same thought. I remember seeing that same video.

Elon said there would be regressions in some instances and this is clearly one of them. We have a few of these odd roundabouts in the Charlotte, NC area as well. I think the point is honestly just to slow people down but it seems like a speed hump would've been a LOT cheaper to build!

I still believe that we are likely to see a public release of this software within the next 90 days. They need more cars on the road with this software in order to REALLY push it forward.

The Beta FSD appears to already work better than the non-beta version while in the city.

 

[Daniel in SD]

The car is also struggling with the weird ass “roundabouts” he has there, which are, as far as I can tell, just four way intersections with a concrete circle in the middle for some reason. He said his car handled those in .10 and .11, but here in this vid of .12 the car was routing through them and he had to take over. I’m wondering if it’s related to a vid I saw back in the original beta release (.10) where there was a circle painted in the middle of the intersection that the beta thought was a traffic circle and tried to drive around?

Ha, I saw that video too, that's a good theory. The display certainly showed it routing right through the circle.
I wonder if it's a mapping issue? Which maps does Tesla use?
The roundabouts show up on Openstreetmaps but not Google Maps. 900 24th Street Sacramento is the location.

 

[Daniel in SD]

I still believe that we are likely to see a public release of this software within the next 90 days. They need more cars on the road with this software in order to REALLY push it forward.

Not so sure about that. This single drive produced enough disengagement data to keep them busy for a long time! Maybe progress will be quick but right now they seem to have no shortage of data.

 

[heltok]

Quite impressive past-the-roundabout prediction even with Autopilot cameras positioned slightly higher than what's used to record Tesla FSD BETA roundabout testing (plus unprotected left) @ 2:08. There's a traffic light barely visible for a few seconds past the trees and central mound vegetation, and the visualization shows a predicted angled roundabout exit while the road is obscured as engineered. Maybe it's also using the visible building's wall and lined trees to infer the angle of the road?

View attachment 605180

Impressive. But also remember that the car camera is a lot higher up than the camera recording the video. When you sit higher you see the world very differently. Also the sign maybe has some radar signature that can help the neural network. This new 4D is awesome, it will start to predict things that the human labelers could not label from the current camera images by using a few pixels and a few radar detections here and there filling the blanks.

With more data, better tuning of the neural network, longer training the system will become even more impressive! What a time to be alive!

 

[Ludalicious]

I watched a little of it. It seems like a major concern of his, and mine, is how the car deals with cross traffic at unprotected intersections. The issue is that if the view is obstructed the car can start to cross quickly without the driver being able to verify that the path is clear. It seems like a very difficult maneuver to monitor as one must be aware of traffic in both directions and be very quick to stop the car if it starts moving when it shouldn't. I wonder if an interface where the driver is required to keep their foot on the brake and release it when it is safe to go would be a good idea?

Have seen the issue come up in a few video's. Is it possible that Tesla has hidden a camera in the new front headlights? I am under the assumption that there is a giant blind-spot for vehicles crossing in front. I am sure there isn't but makes you wonder if there is.