How will Tesla demo FSD?

[gluu]

I'm not convinced at all that this shows a major step towards FSD at all. The Waymo system is way ahead of what was shown here.

There was very little traffic on the local roads, there was no cross traffic at any stop that the car had to wait for, nor were there any pedestrians or bicycles. The killer (literally) is going to be all the corner cases and they didn't show anything close to one.

Tesla is doing this with cameras only. Waymo is geofenced, so it might look great in these types of demos but doesn't have a chance in hell of scaling up to the entire world. Tesla does.

 

[bak_phy]

Tesla is doing this with cameras only. Waymo is geofenced, so it might look great in these types of demos but doesn't have a chance in hell of scaling up to the entire world. Tesla does.

Yep. Although the demo route was preset there were obviously many chances for a corner situation to develop. If a significant number of the audience members took a demo trip no doubt some of them would have encountered one. Tesla would have to have been fairly sure that the car could handle such a situation.
Given the way that Tesla claims the learning is done there's no reason to think that there was anything special or reprogrammed about the route other than that perhaps it had been checked for roadwork, debris, and perhaps a tricky merge.

 

[electronblue]

Right now all you can do is trust that the hardware software approach is correct. (or not )

That was the big news from yesterday right?

Tesla has locked their hardware suite for Tesla Network launch. This is it. No additional sensors. Not even a nose camera for parking lots?

They are leasing out cars that will ”return” to the Network according to them as we speak...

That is bold.

 

[DrDabbles]

I guess you are one of those who say, glass is 1/100th empty?

For me it works, and it is true I am guessing for over 90% of Tesla drivers. I will take that.. There is always for someone somewhere it doesn't work.

No, I'm the type of person that says, "Hey cool. @Electroman has a glass. My glass has holes in it, so it's still a glass but it doesn't hold water as well as his".

The fact is, NoAP works fine if you live in an area of the world where highways are extremely simple and the land is flat. That's great. But if you're going to say NoAP works 90%, then I'm going to produce stark counter examples. Two weeks ago, when I got my latest update, I took the car out for a ride on a route that has traditionally had trouble. That route is less than 200 miles, and it's go everything from Interstate to state highway, multi-lane roads to twisting back roads. I can easily reproduce the same exact type of error, in the same exact place, every single time I drive past it. Not almost every time, I'm talking 100% rate.

In the cases where I don't force NoAP to do the right thing by tapping the accelerator or taking the wheel, it will come to a nearly full stop or swerve dangerously. Again, not some of the times I drive through these scenarios, but 100% of the time. If I was do these tests when there is traffic, it would create a very dangerous situation for anybody around me. That's not FSD. That's nowhere near FSD.

Now, I'm willing to accept that Tesla could produce a system that covers 100% of your needs by the end of 2019. I'm even willing to accept that they could produce a system that eliminates all of the scenarios that fail for me by the end of 2019. But to imagine that they will have trained their NN to the point where it can be considered Level 4 or Level 5 by the end of 2019 is something I can not do. As Elon notes, the real world is too messy. There are too many things that they will not have trained the network to understand, and that fact alone means they will not be able to provide full autonomy.