Hacker News Reader

Comment by dml2135 | Hacker News Reader

Great summary. This will be an interesting experiment that I think will ultimately fail.

On the basis of what evidence? Seems quite pessimistic given that waymo has been stepping up its activity in sf considerably over the last year.

Just curious more than anything, has there been any model/ai/robot-car that has even come close to responding to the kinds of nuances described above? It seems like so far we are just celebrating that the cars that can stay in a lane, not hit things, which is a huge I get it, but I don't see how sophisticated CV heuristics could even come close to approximating the kinds of edge cases that exist here, both in the social sense of communicating with other drivers, and responding to things that should not be the case (no lines on a road, a double parked car), but in practice are the case everyday in cities.

There are videos of Waymo successfully negotiating a Costco parking lot. Here is a recent video of some scenarios in SF that are close to the nuances described above: https://www.youtube.com/watch?v=2CVInKMz9cA. Notice how there are activity icons indicating pedestrians running, truck stopped with a door open and so on. They can understand a scene pretty well.

Behavior prediction is a large research focus now according to Waymo, so they are thinking about complex situations like these. But a lot of them (like no lane lines, construction, 2 way roads becoming 1 way) can just be encoded in HD maps and distributed to the whole fleet.

Yes, I routinely drive next to these driverless things and they are negotiating plenty of annoying & complicated stuff. I think I must live in a peak location because they are everywhere in my neighborhood in SF, which has lots of fast cars and pedestrians and bikers and skateboarders doing all sorts of crazy things.

Like these actually exist, I guess people on the East coast don't realize it yet?

https://m.youtube.com/watch?v=TFEvkmvIjVo Clearly not perfect in how it handles this as it doesn’t really back up or go onto the sidewalk but these Chinese self-driving cars are handling quite difficult scenarios.

It’s like there’s a phase transition from theoretical ideal driving to ignoring street markings and navigating based on how much space there is around you and intentions of surrounding vehicles. Looks difficult but not impossible.

There are similar videos showing Teslas handling double parked vehicles and oncoming traffic too.

My old neighborhood in sf was crawling with these things negotiating all kinds of traffic conditions. It doesn’t seem like that much of a stretch to me having driven past and alongside on the order of dozens of these things.

Reply on news.ycombinator.com