Hacker News new | ask | show | jobs
by bkandel 857 days ago
When I was in a Tesla Uber, I found watching the front screen that shows where cars are around you frightening. Cars would frequently move around and change position unexpectedly and in a way that had nothing to do with actual car movements, lane markings would be totally off, etc. If that screen is at all representative of what FSD "thinks" is going on, I'm not at all surprised that there are lots of crashes when relying on the system.
4 comments

YES. Thank you.

I rode in a Model Y last year and just could not believe the mistakes it was making. Disappearing and reappearing cars; the semi in front of us was apparently straddling the lane line for a few miles; somehow an early-2000s Dodge Ram was classified as a small sedan – the list goes on, and this was only a ~10-minute ride. I would be absolutely mortified if a product of mine ended up in front of a customer in that state.

I spent a decade working on commercial computer vision applications that, among other things, had to recognize and track cars. Those are exactly the sort of transient errors you'd expect to see in shipping products and they usually have heuristics to "smooth" over those sorts of problems.

That said, would I ever trust my life to a system like that? No.

I'm actually surprised they show it as raw as it looks. Doesn't inspire too much confidence, even though I bet the system must be reclassifying and changing way faster than it renders things on screen.
Don’t get me wrong, my background is also pretty CV-heavy, and I don’t expect perfection by any means.

But the display itself serves basically no purpose besides looking cool, and it just fails pretty badly at that. Also yeah, it made me maybe a little more nervous about being on the road with a Tesla than it should’ve…

We'd usually have something like that in our products as a developer/debug mode, not generally visible to customers.

If anything, if you've got self-driving on in a Tesla, you're not being nervous enough. :)

That’s not the perception for FSD, btw, that’s the output of a much older generation model that you’re seeing. But yeah, it’s pretty bad.
> If that screen is at all representative of what FSD "thinks" is going on, I'm not at all surprised that there are lots of crashes when relying on the system.

It has limited accuracy due to the poor processing power of the Intel Atom chips they used to put in the infotainment screen - they didn't even have anti-aliasing on the car models due to the poor processing power, they needed to make sure the rest of the infotainment experience remained snappy. (note this is not the same chip as what the FSD computer uses).

The "park assist" visuals[0] are closer to what the system actually knows about its surroundings from the cameras, but they're only enabled at low speeds for now (likely still due to processing power of the infotainment screen).

0: https://youtu.be/J5Tx3uEzz-g?si=AQRDcaukKQWrQ7-x

How on earth is it legal to sell a car with this sort of bug-riddled system attached?

That a car can disappear and reappear in the autopilot's knowledge of the world is really weird to me. I would have thought that each car would be tagged as an object that has physical attributes such as "can't teleport".

IT isn't what FSD sees FSD works with voxels, but they aren't rendering that out for the user, they're basically presenting the basic bare eseentials to entertain people and give people a little info.

To see what FSD is actually seeing see the new parking system that shows the voxels, the backend is like that but they cant afford the budget on the CPU to render that at high speed for visualization