Hacker News Reader

Comment by udfalkso | Hacker News Reader

Waymo’s tech requires the area to be meticulously mapped correct? Is this why the test is limited to a small geography?

Roads in Chandler are very regular (nice, wide grid), and the weather is clear and sunny the vast majority of the days. The particular geography helps isolate the number of factors the car needs to consider besides "just" driving.

From a NYT article[0],

>"The company did not say whether it was testing the driverless cars in environments considered challenging for autonomous vehicles, like bridges or tunnels, or more difficult conditions, like driving at night or in rain and snow — usually not a big concern in the dry Phoenix climate."

[0] https://www.nytimes.com/2017/11/07/technology/waymo-autonomo...

So crazy to see this video shot in the neighborhood I lived in the summers as a kid. Chandler's got a lot of good things going for it as eggpy mentioned -- well marked roads, wide, very little rain. On the downside, the entire Phoenix area has some terrible drivers, which I suppose is good training for the Waymo cars. It will be interesting to see if they park the cars during monsoon season; conditions can get pretty treacherous on the first rainy days of the season -- roads get very slick, drainage isn't always the best so you can end up with large puddles of water on the road that cause hydroplaning, etc.

Everywhere has terrible drivers - that's the point!

The Ars Technica article says this is still the case:

> To aid with navigation, Waymo has built high-resolution three-dimensional maps of its service area. Self-driving software in each car can compare the objects identified by sensors to objects on the map, allowing it to quickly distinguish stationary objects like trees and buildings from mobile objects like cars and pedestrians.

> As Waymo expands its map and acquires more vehicles, it will also expand its service area. Before too long, Waymo expects to offer service across the entire Phoenix metropolitan area. Eventually, Waymo will extend service to other metropolitan areas using the same incremental approach.

https://arstechnica.com/cars/2017/11/fully-driverless-cars-a...

This article from Dec 2016 seems to suggest that's still the case [0]. However, it also looks like those maps can be built by the cars themselves, which should simplify expansion.

[0]: https://medium.com/waymo/building-maps-for-a-self-driving-ca...

How much of a problem is this mapping?

Huge part is up-front-work, and then?

How often do streets change?

Not significantly more difficult than what goes into google maps, you just need a modified sensor suite, if I understand correctly. (Disclosure: googler, but I have nothing to do with maps or cars)

That's pretty terrifying that it can't read street signs and is dependent on mapping. Here in D.C., there are all sorts of fine-grained road rules, such as streets that are one-way only certain times of day, streets that reverse direction at certain times of day, turns that are legal or illegal depending on time of day, etc. That's not including random road closures.

These databases apparently don't get updated that often. Recently, D.C. moved a major highway on-ramp. It used to be that you could get on 395-S heading west on H street, by taking a left onto the on-ramp. They replaced it with an underground entrance ramp on eastbound Mass. Ave., such that westbound traffic on H or Mass. Ave. could no longer use the on-ramp, and had to go several blocks south to use a different one. This was a widely-publicized move planned long in advance. And at least Apple Maps tried to steer me down the old route for awhile.

Similar thing in our neighborhood just outside the Annapolis city limits. They closed our street to car traffic with barriers, so that you can get in an out but can't drive along it for more than one cross street. Whatever database Uber uses still hasn't figured this out--I'll get Uber drivers waiting for me on the other side of one of the barriers.

It can read street signs, even hand-held STOP signs in the hands of construction workers. And the cars are continuously mapping their surroundings. But they ALSO use a shared, extremely high quality map database for first-level path planning.

edit: https://www.theatlantic.com/technology/archive/2017/08/insid...

And it can even see and understand the hand gestures (e.g. turn signs) of cyclists.

It's the cyclists that don't do hand gestures that are the problem. Which is - where I live - the vast majority of them.

I bet it's still safer for the cyclists than the average human-driven car, though.

Time will tell.

Plus can even understand cyclist that can keep their bike stopped but no feet on the ground.

>That's pretty terrifying that it can't read street signs

Not sure how you arrived at that. There's nothing in the article or in the comment you were replying to that supports it.

>it can't read street signs and is dependent on mapping.

Says who? The article didn't say anything about that.

There's little external information about how map dependent their implementation is.

They certainly have meticulous maps available in areas they operate, but it is easy to imagine other reasons for that. Like they could be using the map to separately process their sensor data, producing a reliable diff and evaluating the sensor only scene based on that.

Perhaps, late at night, some autonomous vehicles will drive unmapped areas at about 25MPH, gathering mapping data. You don't need much mapping data if you drive slow.

Probably a solved problem in the eyes of Google..

Reply on news.ycombinator.com