Hacker News new | ask | show | jobs
by Retric 6 hours ago
> The garage door I can certainly believe. Its AI brain just won’t be trained to look for that sort of thing.

Why not? That’s likely to come up at minimum thousands of times per day, and likely vastly more as the system improves.
1 comments
Reason077 6 hours ago
I get the impression that it kind of comprehends the world as a horizontal plane. It’s focused on objects on the surface and isn’t looking for hazards coming from above. Could that be improved? Sure! Is it a priority for them? Maybe not…
link
Retric 6 hours ago
At minimum it needs to pay attention to railroad crossing bars and bridges.

They show a 2D representation to the driver because that’s good enough for drivers, but I wouldn’t assume that represents how the system operates internally at every stage. Even navigating requires the concept of bridges crossing roads without intersecting them.

link
Reason077 4 hours ago
It is indeed known to have issues at railway crossings:

https://news.ycombinator.com/item?id=47783427

AFAIK, there's no issue with bridges that cross roads, however. It will just ignore a road that's above or below the one you're travelling on that doesn't intersect with yours. Just like a human driver would do.

link
Retric 4 hours ago
> It is indeed known to have issues at railway crossings:

At scale you can “have issues” and still work 99% of the time. If it’s completely incapable of handling railroad crossings that’s relatively easy to verify.

> AFAIK, there's no issue with bridges that cross roads, however. It will just ignore a road that's above or below the one you're travelling on that doesn't intersect with yours. Just like a human driver would do.

As a user of the system that’s worked it all out it’s not an issue for you, but getting that behavior from raw sensor data is non trivial.

link