FSD Beta Videos (and questions for FSD Beta drivers)

[DDotJ]

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.

That's good practice for non-FSD human drivers as well. That way if someone rear ends your car you aren't thrust into oncoming traffic. I've always been taught to keep the wheel straight until the turn is clear to avoid this scenario, instead of winding up the wheel while still waiting for the turn. Would be nice if FSD would do the same.

 

[AlanSubie4Life]

minor preferential issues

Are you talking about the angle of the car and the turning of the wheels here, prior to the left turn?

 

[Daniel in SD]

Yes, but once it's able to do difficult lefts at ~100%, these minor preferential issues shouldn't be difficult to iron out.

That's really not a preferential issue, it's necessary to be a better than average driver. Watch people making the left turns, they don't start the turn until it's clear. Several testers have gotten negative reactions from oncoming drivers.
It would also reduce the number of unnecessary disengagements which would mean less data for Tesla to analyze and faster progress towards 99.999% (or whatever human success rate is).

 

[powertoold]

That's really not a preferential issue, it's necessary to be a better than average driver. Watch people making the left turns, they don't start the turn until it's clear. Several testers have gotten negative reactions from oncoming drivers.

I agree with you. It should act more natural and in a way that the driver can easily anticipate its behavior.

I just say it's preferential because it's difficult to define an "average" driver, and yes, I have seen many people start to turn when it's not clear yet.

My point with that prior post was that once the more difficult problem of lefts is "solved," then Tesla can easily iron out the "preferential" / more minor issues.

I always go back to Brandone's and Chuck's issues with the car weaving within lanes in earlier versions. They eventually fixed that, and it should be the similar with these turns, eventually.

 

[powertoold]

Looks like they expanded the testers:

 

[powertoold]

That way if someone rear ends your car you aren't thrust into oncoming traffic. I've always been taught to keep the wheel straight until the turn is clear to avoid this scenario, instead of winding up the wheel while still waiting for the turn.

But have we considered that fsd has a 360 degree view at all times and a 100ms or better reaction time?

 

[diplomat33]

Another update from Elon:

https://twitter.com/x/status/1426934693195304965

Elon seems to be saying that FSD Beta might skip from V9.4 straight to V10. Also "significant architectural changes" sounds like another major rewrite. I wonder what that will be?

 

[aronth5]

The thing is, it could do 100 unprotected lefts in a row perfectly and that wouldn't come close to proving that the problem has been "conquered". When it's capable of doing that people should keep that in mind.
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.

It's not whether the wheels are turned or not its whether the upcoming cars see the FSD car continuing to creep into the lane when the car should be stopped. It's bad creeping behavior that is the root cause for the ongoing cars freaking out. I will typically turn my wheels a bit so when the opening is their I can immediately make the unprotected left turn. That extra time saved where I don't have to turn the wheel is much more likely to avoid an accident then the very slight chance someone will rear end me.

 

[Daniel in SD]

It's not whether the wheels are turned or not its whether the upcoming cars see the FSD car continuing to creep into the lane when the car should be stopped. It's bad creeping behavior that is the root cause for the ongoing cars freaking out. I will typically turn my wheels a bit so when the opening is their I can immediately make the unprotected left turn. That extra time saved where I don't have to turn the wheel is much more likely to avoid an accident then the very slight chance someone will rear end me.

Yes I agree it’s the creeping or even lunging that’s the bigger issue.
I disagree on the wheel turning. There’s plenty of time to start moving and turning as the last oncoming car is passing.
This behavior would also be better for FSD as it would alert the driver that the car thinks it’s clear to turn.

 

[stopcrazypp]

It's not whether the wheels are turned or not its whether the upcoming cars see the FSD car continuing to creep into the lane when the car should be stopped. It's bad creeping behavior that is the root cause for the ongoing cars freaking out. I will typically turn my wheels a bit so when the opening is their I can immediately make the unprotected left turn. That extra time saved where I don't have to turn the wheel is much more likely to avoid an accident then the very slight chance someone will rear end me.

I've talked about it in another post a while back, but I suspect the angle of the car is necessary for the cameras to get the proper view (both the lanes and also the one the car is targeting).

 

[BMW740iL]

I've talked about it in another post a while back, but I suspect the angle of the car is necessary for the cameras to get the proper view (both the lanes and also the one the car is targeting).

hahaha, but but but fsd has a 360 degree view at all times

 

[evdude69]

Another update from Elon:

https://twitter.com/x/status/1426934693195304965

Elon seems to be saying that FSD Beta might skip from V9.4 straight to V10. Also "significant architectural changes" sounds like another major rewrite. I wonder what that will be?

every update is exciting We are getting there

 

[Terminator857]

Is anybody counting all the crashes and humans-killed-by-FSD that Tesla does not have to answer for because good people like Chuck Cook and Trevor are taking all the responsibility?

Zero killed by FSD and zero major crashes. There is a long line of people wanting to test FSD and willing to accept responsiblity so there is a lot of good people eager to help test. Any other questions?

 

[Daniel in SD]

I've talked about it in another post a while back, but I suspect the angle of the car is necessary for the cameras to get the proper view (both the lanes and also the one the car is targeting).

Has anyone quantified this "blind spot" because it's certainly not intuitive.
It really doesn't look like there's any blind spot relevant to left turns. There's a ton of overlap between B pillar camera and wide front camera.

source:

https://twitter.com/x/status/1425915165074235392

Also related to "blind spots", I still haven't seen anyone explain how b pillar camera position is any different than driving a car with a longer hood (which humans can do).

 

[AlanSubie4Life]

Has anyone quantified this "blind spot" because it's certainly not intuitive.
It really doesn't look like there's any blind spot relevant to left turns. There's a ton of overlap between B pillar camera and wide front camera.

View attachment 697201
source:

https://twitter.com/x/status/1425915165074235392

Also related to "blind spots", I still haven't seen anyone explain how b pillar camera position is any different than driving a car with a longer hood (which humans can do).

I don’t think there is any blind spot relevant to the unprotected left with oncoming traffic discussion here. I’m not sure why the angle of the car would matter much unless the middle of the FOV is somehow easier to deal with for the NN’s (which would be bad!!!)

In the context of those pictures, @verygreen ’s contention is that indeed the car has to poke forward into the intersection, into traffic, slightly, to get the visibility around certain corners. It’s true that humans might have to do this (though slightly less since they can lean forward) too, but the contention is that is an unnecessary risk even if it exists to some extent for humans, given sensors could be located in the front corners of the vehicle and eliminate this problem entirely.

Left turns with occluded visibility are dangerous for humans, and they’re also dangerous for FSD to the extent it has to creep into the lanes of traffic to get sufficient visibility (and there is the question of how to program it to “know” that it can see - something humans can obviously figure out, but might be more difficult to program - though it seems possible based on perception of how much road can be seen). Of course, this issue of “knowing” it can see is something that would have to be solved with cameras in the corner of the car as well.

 

[MP3Mike]

The other thing I've not seen is behavior at railroad crossings when trains are present. If anyone has seen any of the beta testers doing this, please link to it.

Here you go, FSD 9.2 stopping for train crossing lights:

Sadly he took it out of FSD and backed up, because he was worried about rocks being thrown by the train, but it seemed to do good.

 

[mhan00]

Here you go, FSD 9.2 stopping for train crossing lights:

Sadly he took it out of FSD and backed up, because he was worried about rocks being thrown by the train, but it seemed to do good.

3:20 - car is confused about the lanes as it’s moving to a left turn and starts swerving a bit. Disengagement.

7:40 - the car did a great job at an unprotected left.

 

[linux-works]

Zero killed by FSD and zero major crashes. There is a long line of people wanting to test FSD and willing to accept responsiblity so there is a lot of good people eager to help test. Any other questions?

do the by-standers, who could also be affected, have any say?

hmmm?

real roads, beta tested by 'users'.

yeah. right.

this is not an easy issue. if they crash, they dont just hurt themselves.

in society, you think of things like this. well, some of us do...

 

[JHCCAZ]

But have we considered that fsd has a 360 degree view at all times and a 100ms or better reaction time?

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.

First, I don't accept that turning the wheels while waiting for the left-turn maneuver saves any time at all. (In fact I'll throw in another related point, which is that Tesla in FSD, Autopark or Summon should not be spinning the steering wheel & scrubbing the tires when the car is at a standstill.) IMO there is nothing important to be gained by pointing the tires in anticipation of the turn, much less in pointing the car itself which is disturbing behavior, and much less in nosing the carforward or even intruding into the oncoming lane before the real go-commitment.

Second, regarding quick reaction time, that is no doubt true in some aspects - but I have to say that the obvious indecisiveness of traffic-gap judgment and go-commitment, evident in v8.x and v9.x, indicate that there is a latency pipeline with these turn decisions that's presently much longer than with human drivers. I hope that is a temporary situation, maybe having to do with NN perception handing off control to the hand-programmed path module, and will be re-architected in v10.

 

[stopcrazypp]

hahaha, but but but fsd has a 360 degree view at all times

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).