Hacker News new | ask | show | jobs
by ACCount37 267 days ago
Yeah! Just add more sensors! We're only 992 more sensors away from full self-driving! It totally works that way!

The debris? The very visible piece of debris? The piece of debris that a third party camera inside the car did in fact see? Adding 2 radars and 5 LIDARs would totally solve that!

For fuck's sake, I am tired of this worn out argument. The bottleneck of self-driving isn't sensors. It was never sensors. The bottleneck of self-drivng always was, and still is: AI.

Every time a self-driving car crashes due to a self-driving fault, you pull the blackbox, and what do you see? The sensors received all the data they needed to make the right call. The system had enough time to make the right call. The system did not make the right call. The issue is always AI.

4 comments

You want the AI to take the camera's uncertainty about a road-colored object and do an emergency maneuver? You don't want to instead add a camera that sees metal and concrete like night and day?
It’s a lot easier to make an AI that highly reliably identifies dangerous road debris if it can see the appearance and the 3D shape of it. There’s a fair bit of debris out there that just looks really weird because it’s the mangled and broken version of something else. There are a lot of ways to mangle and break things, so the training data is sparser than you’d ideally like.
The problem can be sensors, even for humans. When a human's vision gets bad enough, they lose their license.
We had sensors that can beat "a human whose vision got bad enough to get the license revoked" used as far back as in the 2004 DARPA competition.

That got us some of the way towards self-driving, but not all the way. AI was the main bottleneck back then. 20 years later, it still is.

We don't have a bottleneck anymore. We have Waymo. They seemed to have solved whatever the issue was...I wonder what the main difference between the Waymo system and the Tesla system is?
The "main difference" is that Waymo wouldn't even try to drive coast to coast.

Because it's geofenced to shit. Restricted entirely to a few select, fully pre-mapped areas. They only recently started trying to add more freeways to the fence.

You're right, they wouldn't try, but I don't think there's any evidence for the idea that Waymo couldn't pull this trip off now from a technical POV. Even if they're pre-mapping, the vehicles still have to react to what's actually around them.
Adding more sensors (such as LIDAR) can certainly make it easier for the "AI" (Computer Vision) to detect & identify the object.