FSD Beta Videos (and questions for FSD Beta drivers)

[EVNow]

Chuck's stance (which I agree with) is that it's fine for the car to be center-biased when there's no one around, but move over as as soon as an oncoming car is detected, and move over when approaching an intersection where you have to stop; otherwise you are blocking others trying to turn onto your road.

Personally, I would like to see FSD Beta just stick to the side. In some of my rides, the car will go back and forth because it will center itself, then move over when there is a car, then center itself when the car passes, then move over again when the next car comes etc... It would be smoother if it just stayed to one side.

I think you guys have more unmarked roads than we have here. Most of the unmarked roads here are local roads with little traffic but a lot of parked cars.

So, the default logic changes. Better to stay towards the center if you have a lot of parked cars - otherwise you have to wave in and out.

So, I’d say the logic should be to stay towards the edge if there are no parked cars. If a car comes from the other direction, always move to the side after passing any close by parked cars.

BUT, we are just talking hard coded rules - they have to translate this to costs associated with various paths such that the optimum path results in the behavior we want. Not that simple.

 

[EVNow]

I wonder how Tesla AI learn to drive on the right side of the road when there is no way for the FSD Beta testers to input their preference, cause any wiggle of the steering wheel will disengage. To allow the FSD Beta to drive on the middle of the unmarked road will only encourage the neural net to learn and do the wrong thing.

That is not how learning works. You just report when it doesn’t drive properly.

 

[KLam]

That is not how learning works. You just report when it doesn’t drive properly.

Really? You mean that I have to go out and drive and disengage before I even get to my street corner every day to beat it to the head (neural net) of the AI machine? While you have all these people putting on miles and miles of driving without disengaging when it center on the unmarked road? I can tell that to my kid on the first day of driving and they learn it right a way. As is now FSD beta will fail a DMV driving test just by getting out of the DMV parking lot. How is that safe driving.

 

[EVNow]

Really? You mean that I have to go out and drive and disengage before I even get to my street corner every day to beat it to the head (neural net) of the AI machine? While you have all these people putting on miles and miles of driving without disengaging when it center on the unmarked road? I can tell that to my kid on the first day of driving and they learn it right a way. As is now FSD beta will fail a DMV driving test just by getting out of the DMV parking lot. How is that safe driving.

I think you should listen to the AI day to understand how the whole thing works.

What we are talking about here is planning and that is not done in NN.

BTW, kids have the benefit of a billion years of evolution

 

[KLam]

I think you should listen to the AI day to understand how the whole thing works.

What we are talking about here is planning and that is not done in NN.

BTW, kids have the benefit of a billion years of evolution

OK. Its not too good a planning then. If they can plan to stop when seeing a dog on the road, could they just do some planning on where to put the car on the road first with no dogs?

 

[EVNow]

OK. Its not too good a planning then. If they can plan to stop when seeing a dog on the road, could they just do some planning on where to put the car on the road first with no dogs?

I don’t think it’s as simple as that. If you try to actually come up with rules for all the scenarios, you will see how difficult it is.

We have discussed this a lot over the months and you can search ….

 

[black knight 3]

OK. Its not too good a planning then. If they can plan to stop when seeing a dog on the road, could they just do some planning on where to put the car on the road first with no dogs?

Perhaps it's not enough understanding of what training and inference is.

Here's a link to Andrej Karpathy at Tesla that explains training for Neural networks..

A Recipe for Training Neural Networks

Musings of a Computer Scientist.

karpathy.github.io

 

[KLam]

Perhaps it's not enough understanding of what training and inference is.

Here's a link to Andrej Karpathy at Tesla that explains training for Neural networks..

A Recipe for Training Neural Networks

Musings of a Computer Scientist.

karpathy.github.io

Thank you for your link to the article/explanations.

I tried to read it but I have to admit that I don't understand anything what he said.

But, I sense that there is something wrong in the Tesla implementation, when till these days, in AP that it cannot handle the merging entry lane. (i.e. still try to center on the slightly wider diminishing merge lane)

 

[EVNow]

But, I sense that there is something wrong in the Tesla implementation,

Self-Driving is though - infact the toughest artificial intelligence problem.

So, its not that there is something wrong with implementation - more like nobody has "solved" self-driving yet.

 

[diplomat33]

But, I sense that there is something wrong in the Tesla implementation, when till these days, in AP that it cannot handle the merging entry lane. (i.e. still try to center on the slightly wider diminishing merge lane)

AFAIK, regular AP is not using the FSD Beta code. Regular AP is basically just dumb lane centering and cruise control. It is not designed to handle merging entry lanes. So it is not surprising that it cannot handle it.

 

[KLam]

AFAIK, regular AP is not using the FSD Beta code. Regular AP is basically just dumb lane centering and cruise control. It is not designed to handle merging entry lanes. So it is not surprising that it cannot handle it.

Then, I hope they are not using the current AP dumb lane centering code on the FSD Beta code. That said, some how I seemed to have seen similar problem in the FSD Beta test drives on city streets.

Actually, the Tesla engineers should have already tested their City driving code on the Freeways already! Isn't it true that Freeways are less complex than city streets?

 

[diplomat33]

Then, I hope they are not using the current AP dumb lane centering code on the FSD Beta code.

No. FSD Beta is basically completely new code.

Isn't it true that Freeways are less complex than city streets?

Yes. But highway driving is faster so accidents can be more dangerous.

 

[mhan00]

Couple of bad things here. First, the car clearly got confused by a sharp right turn and was very close to turning too far and running into a parked car. The driver had to take over to prevent an accident at the very last second (always keep your hands on the wheel, kids).

Next was the car trying to run a red light immediately after making a turn onto the road with the light. I’ve had something similar happen to me, but with stop signs. The car stops at a stop sign, turns left onto a road that immediately has another stop sign it needs to honor, but it ignored it and I had to hit the brake. That was on an earlier version and it hasn’t done it since I got an update, but clearly still an issue here with the red light.

 

[EVNow]

Just for fun …

https://twitter.com/x/status/1505737984267665412

 

[Terminator857]

FSD fails to detect blockade at end of video:

https://twitter.com/x/status/1512463149869543435

 

[JulienW]

FSD fails to detect blockade at end of video:

https://twitter.com/x/status/1512463149869543435

Of course the Tesla apologist in me says it is in a parking lot and FSD is not meant to be used there.

BUT I also know (from actually experience) that it will do the same on a street.

Number 1 rule: Don't hit things.

 

[EVNow]

FSD fails to detect blockade at end of video:

https://twitter.com/x/status/1512463149869543435

Those are super difficult to detect.

I remember Elon posting about them (low occlusion) once - and it was supposed to have been included in one of the releases ?

https://twitter.com/x/status/1447581205407232000

 

[Terminator857]

Those are super difficult to detect.

Super easy to detect.

 

[Knightshade]

Super easy to detect.

Where can I buy the self driving software you wrote?

 

[Terminator857]

Where can I buy the self driving software you wrote?

Detection of objects is super easy and a well defined process. Collect samples, train on them. wahla, it works. There are standard APIs in cloud networks that work very well. Standard APIs in other formats for object detection also.