Welcome to Tesla Motors Club
Discuss Tesla's Model S, Model 3, Model X, Model Y, Cybertruck, Roadster and More.
Register

FSD Beta Videos (and questions for FSD Beta drivers)

This site may earn commission on affiliate links.
3:20 - car is confused about the lanes as it’s moving to a left turn and starts swerving a bit. Disengagement.
Hopefully this is a case that the new Lanes network should help improve in the near future. This large intersection where navigation still thinks the turn is 100ft away probably confused FSD in thinking the immediate left turn lane is only for the sharp left turn for this 6-way intersection. It looks like the behavior here is to go straight to turn left at the "next" intersection, which is the correct behavior of other back-to-back intersections.

milwaukee & grand.jpg
 
  • Helpful
Reactions: mhan00
Has nothing to do with 360 view. Being able to see something at a position and being able to see it well enough to judge the speed of an approaching car at a distance are two different things. There are positions that maximize the resolution of the given cameras for a given FOV (for example taking advantage of the narrow camera).
This looks like it should be right in the middle of the narrow camera so there's no need to advance towards the oncoming cars.
Someone is probably going to argue that the the view could be obscured but the trees but that's fine. After it looks clear from this angle you start the turn as the other cars pass and if a previously occluded car comes in to view you stop. That is reassuring behavior to the oncoming drivers and gives the Tesla driver more time to intervene if the system makes a mistake.
1629143712967.png
 
Has nothing to do with 360 view. Being able to see something at a position and being able to see it well enough to judge the speed of an approaching car at a distance are two different things. There are positions that maximize the resolution of the given cameras for a given FOV (for example taking advantage of the narrow camera).
thanks for pointing out to the fanboys that FSD does not have 360 view
 
To be clear, are you suggesting that the car should turn the wheels in preparation for the turn, but then be prepared to quickly turn them back if it sees itself about to be rear-ended? You know I'm not at FSD hater at all, but this proposition seems dubious to me.

Nope, not saying that at all. I'm just pointing out that some best practice for humans may not be as important or necessary for a robot.

I agree that best practice is to not turn the wheels until it's necessary, but I also don't think turning the wheels beforehand is *that* big a deal. There are all sorts of risks when you get rear-ended in different situations and environments.
 
  • Love
Reactions: Electroman
Nope, not saying that at all. I'm just pointing out that some best practice for humans may not be as important or necessary for a robot.

I agree that best practice is to not turn the wheels until it's necessary, but I also don't think turning the wheels beforehand is *that* big a deal. There are all sorts of risks when you get rear-ended in different situations and environments.
My guess is that some drivers turn their wheels prior to stopping in preparation for turning and other times don't And I suspect don't even realize it when they do its just how they drive. Like many other questions the answer is "it depends" on the turn and situation. Frankly whether FSD does or doesn't turn the wheels in advance has pretty much zero impact on how well it FSD performs. Time to move on.
 
My guess is that some drivers turn their wheels prior to stopping in preparation for turning and other times don't And I suspect don't even realize it when they do its just how they drive. Like many other questions the answer is "it depends" on the turn and situation. Frankly whether FSD does or doesn't turn the wheels in advance has pretty much zero impact on how well it FSD performs. Time to move on.
Probably around the same number of drivers turn their wheels as those who perform California stops or signal while (not before) turning. Just because some drivers have poor habits does not excuse the car, it should have perfect habits - yet the FSD Beta does indeed have those bad habits too. Rather than driving like a poor to average driver Tesla should be striving for better.
 
The car does eventually make some successful lefts through the cross traffic. I'd much rather have the car do the loop than do what it did in version 8.x, which was to get stuck midway or not commit.

I think it's very impressive that V9 can assess uncertainty and make "real-world" decisions based on that.
Except at the end of that loop comes... another loop. It does not seem to want to make that right turn, followed by a U-turn in the next signaled intersection. Chuck went nowhere with 9.1 and 9.2 other than in circles.
 
FSD Beta 9.2 almost crashed into this stop sign at a kinda-6-way intersection. It's a sharp right turn from Shannon onto Kennedy with limited visibility leading up to it, but maybe the driveway off of Deer Park also confused things as that is also a potential right turn. There was another vehicle that waited and fortunately not directly behind as the driver slammed on the brakes but probably was wondering why someone would drive straight towards the stop sign and trees.

View attachment 697382
Did the graphics change on 9.2?
 
The biggest improvement they could make right now is to not turn the wheels and move towards the road when there's oncoming traffic. It should maintain a straight trajectory until it's ready to begin the left. That would give the driver much more time to intervene and make it much less of a white knuckle experience. The way it behaves right now is very unsettling to oncoming drivers.
Totally agree. I always cringe when I see drivers do that. Tesla needs to stop that. I've noticed Waymo doing it too.

It's funny, most people are surprised when I point that out so they were probably never taught the correct behavior when they learned to drive.
 
I hate to bring up this wheel turning issue again, but check out how inhumanly fast fsd beta can turn the wheel. See it turning left for no reason, and once it's cleared to proceed, it instantly spins back to turn right:


So what if FSD Beta can turn the wheel super fast? FSD Beta should not be making unnecessary wheel turns in the first place.
 
I hate to bring up this wheel turning issue again, but check out how inhumanly fast fsd beta can turn the wheel. See it turning left for no reason, and once it's cleared to proceed, it instantly spins back to turn right:

Driver and FSD both had a problem with passing the cyclist over the double yellow at a blind hill. Though they did it anyway. A good swerve if they met an oncoming car and the cyclist would have been hit.

Around.png

Around2.png



No screen UI visualization of the giant nut in front of the car. I'd be confused too.
Nut.png
 
Probably around the same number of drivers turn their wheels as those who perform California stops or signal while (not before) turning. Just because some drivers have poor habits does not excuse the car, it should have perfect habits - yet the FSD Beta does indeed have those bad habits too. Rather than driving like a poor to average driver Tesla should be striving for better.
Actually FSD just needs to drive as well as I do:) Got my license in 1968 and have a perfect driving record. No accidents not even a minor scrape and have never gotten a moving violation. Not bad for driving in the greater Boston area. So if FSD wants to turn its wheels early is fine with me since that is how I drive!
 
This looks like it should be right in the middle of the narrow camera so there's no need to advance towards the oncoming cars.
Someone is probably going to argue that the the view could be obscured but the trees but that's fine. After it looks clear from this angle you start the turn as the other cars pass and if a previously occluded car comes in to view you stop. That is reassuring behavior to the oncoming drivers and gives the Tesla driver more time to intervene if the system makes a mistake.
View attachment 697288
I was going to respond to your other comment, but I can respond to this one also. It's not that the cameras are affected by a blind spot per se (other than occlusions by trees), just that the position better utilized the cameras.

Here's a top down view to make my point more clear. The blue lines show the front cameras (wide, main, narrow). The green shows the side b-pillar camera. Length of the lines are based on the range of the camera. This is based on the visualization on the Tesla AP page which may not necessarily be correct (it seems to be assuming the published specs for the FOV is based on horizontal not diagonal), but should work for a rough idea.

This is the view for the position in your screenshot. You can see about half of the narrow camera's FOV is wasted on covering the traffic going the same direction instead of oncoming traffic. It also gets a worse view of the road it is turning into on the B-pillar cams. The red lines show the occlusion by the trees. The shaded red area shows the additional coverage of the second case for occluded areas.
view.jpg

This is the view moving the car more forward and angled in preparation for the turn (I looked at previous drone footage of V8.2 to get the position). I measure from the drawing that it gets about 50 ft more of leftmost lane, 100 ft more for half view of leftmost lane, 160 ft more of center lane, 260 ft more for half view of center lane, 200 ft more for rightmost lane (to edge of given range for narrow camera). And for the road it is turning to, it gets basically full coverage to the limit of the 80m (260 ft) range.
view2.jpg


So the position is it setting up at simply gives significantly better coverage of both roads. Sure, there is a risk of rear ending moving the car into oncoming traffic, but that risk is much smaller than being hit by oncoming traffic just attempting to make the turn. Rear ending risk is much reduced for this type of intersection already anyways given the left turner is pulled to the side, and no longer in a lane of through traffic. I agree it should be tweaked a bit not to go so close to the line (so as to not freak oncoming traffic and also the driver), but don't think it's as big a deal to have the car turned a bit.
 
IN
I was going to respond to your other comment, but I can respond to this one also. It's not that the cameras are affected by a blind spot per se (other than occlusions by trees), just that the position better utilized the cameras.

Here's a top down view to make my point more clear. The blue lines show the front cameras (wide, main, narrow). The green shows the side b-pillar camera. Length of the lines are based on the range of the camera. This is based on the visualization on the Tesla AP page which may not necessarily be correct (it seems to be assuming the published specs for the FOV is based on horizontal not diagonal), but should work for a rough idea.

This is the view for the position in your screenshot. You can see about half of the narrow camera's FOV is wasted on covering the traffic going the same direction instead of oncoming traffic. It also gets a worse view of the road it is turning into on the B-pillar cams. The red lines show the occlusion by the trees. The shaded red area shows the additional coverage of the second case for occluded areas.
View attachment 697596
This is the view moving the car more forward and angled in preparation for the turn (I looked at previous drone footage of V8.2 to get the position). I measure from the drawing that it gets about 50 ft more of leftmost lane, 100 ft more for half view of leftmost lane, 160 ft more of center lane, 260 ft more for half view of center lane, 200 ft more for rightmost lane (to edge of given range for narrow camera). And for the road it is turning to, it gets basically full coverage to the limit of the 80m (260 ft) range.
View attachment 697597

So the position is it setting up at simply gives significantly better coverage of both roads. Sure, there is a risk of rear ending moving the car into oncoming traffic, but that risk is much smaller than being hit by oncoming traffic just attempting to make the turn. Rear ending risk is much reduced for this type of intersection already anyways given the left turner is pulled to the side, and no longer in a lane of through traffic. I agree it should be tweaked a bit not to go so close to the line (so as to not freak oncoming traffic and also the driver), but don't think it's as big a deal to have the car turned a bit.
Interesting analysis of the situation for the car turning too far in to the lane. Now do the analysis for the same situation where Chuck Cook had to disengage when fsd goes full Bowfinger trying to run three lanes of incoming traffic.