Hacker News Reader

Comment by argonaut | Hacker News Reader

"Fundamental architectural failure" is a very bold claim to make given that absolutely no other approaches have worked as well as Google's for urban driving.

View on HN

There's a serious problem at the heart of "throw a bunch of data at it" strategy - driving has an extremely long tail that's relatively fat. You can hit the "average case" relatively quickly (you can assume that Google's done that for the SoCal environment), but being able to extend that dataset to other, less likely scenarios, like "heavy downpour in SF" or "freak snowstorm" or "sudden construction due to water main break" is really hard.

None of what you said has anything to do with my point that the pre-mapping and then LIDAR-localization approach taken by Google is the only approach shown to work.

You're making a point about end-to-end trained ML systems; if anything, that's an indictment of Tesla's approach, not Google's more traditional sensor-based-approach.

How do you know that? There hasn't been a competition since 2007, and that was before Google was involved.

* That we know of. Tesla et al. have not publicized urban driving videos/examples, so I'm assuming they're not capable of it as well as Google, right now.

https://www.wired.com/2015/04/delphi-autonomous-car-cross-co...

That article has 0 information about Delphi's approach. In fact, the Delphi car has LIDAR too. For all we know they could be taking Google/Uber's approach.

I used to work on this - the approach is very different. That's a pretty hollow statement on the internet though, so take it as you will.

An interesting aside - it seems like most of Google's team left this year. Chris Urmson left, Anthony Levandowski left, and many of their engineers are gone as well (many went to Otto). It'll be interesting to see what happens now that they've lost all of the original leadership (Sebastian Thrun left long ago).

What is the approach, then?

At a very high level, instead of comparing everything against a map, we try to figure out what it is. For example, the way Google detects traffic lights is using geometry to infer where a traffic light will be. Their maps are so good that they know where the car is, and where the traffic light is, and just use math to figure out where in the image there should be a traffic light. Then they just look at the pixels and figure out which color is lit - red, yellow, green. The way we did it is constantly scan for traffic lights (you can do lots of sanity checks, for instance using GPS are we near an intersection?), which means figuring out whether or not a traffic light is in the image, and then once we detect one figure out what color is the light.

Google's approach is simpler but is highly dependent on the maps. The other approach is easier to scale, but is obviously harder and likely takes more work to get the same results. Something I think a lot of people gloss over is that these maps take up huge amounts of data - not something you can stream over a cell network and probably not even high speed WiFi (maybe you'd download a map for your trip the night before). And also, as I understand it, the maps require regular maintenance.

The real question is how will this scale? And honestly I don't know. My theory is the tech is so hard and the rewards are so great that infrastructure will change to make it easier. By installing smart traffic lights, having cars talk to each other, and adding various road marks specifically for self-driving cars you can drastically reduce the complexity of the problem.

Reply on news.ycombinator.com