Hacker News Reader

Comment by ocdtrekkie | Hacker News Reader

I have yet to hear Google has solved the main reason their Self-Driving Cars aren't practical: That every road they drive on must be precisely mapped ahead of time, at a much higher fidelity level than Google Maps/Street View is normally done nationwide.

Tesla and other self-driving developers are working without the luxury of a worldwide mapping database, and while not perfect, their products more or less could work on any road, because they're based on input from the sensors. But Google cars are dependent on input from Google servers, which means, where that data is lacking, the cars can't drive.

If they're looking to run some sort of business to start making money, I could maybe see like an Uber-like service across very small geographic areas they have mapped. But the narrow fence around the area you can practically drive with one makes it useless to sell as a consumer vehicle.

View on HN

Source? I thought Google's cars used many sensors as well and weren't just a "follow this high-fidelity map and hope it's correct!" technology.

In fact, Google has the most expensive set of sensors on any self-driving prototype out there, AFAIK. They spend about $150k per car on sensors. But almost all the technology they've shown is about detecting people, dogs, other obstacles, on a largely unchanging setting.

As of late 2014[0], Google engineers admitted the mapping required for their cars to work wasn't feasible at a nationwide scale, and that if they didn't know to look for say, a stoplight, on their maps, they'd run right through it. (Looking for arbitrary red lights is probably a hard problem too.)

That's two years ago, yes, but they've announced nothing, technologically, in this field in this time (please comment if you have an additional source relating to the mapping of the roadway for a Google car), despite tons of PR about the number of miles they've driven or what cities they were driving in, with the same technology they've already had. Also bear in mind, almost all of Google's senior talent in self-driving cars has left Google X. So if they were having trouble with this problem in 2014, their newer, less experienced people now, probably are having more difficulty in 2016.

[0]http://www.slate.com/articles/technology/technology/2014/10/...

I highly doubt this is still the case considering auto manufacturers far behind the curve on this tech [1] are doing things like recognizing stop signs and reading speed limit signs using on-board cameras. Lane keeping also uses onboard sensors. As far as I know, the most important thing high-resolution maps are needed for is decision-making for things more than a few hundred feet ahead (ex. when to get in the right lane to exit or turn). It also helps with driving in adverse conditions where it's difficult to use normal sensors to see road markings. There are plenty of players working on this, though, so I don't think it's a huge obstacle. For example, [2].

[1] https://en.wikipedia.org/wiki/Traffic_sign_recognition [2] http://blog.caranddriver.com/how-high-definition-maps-are-pl...

If I'm not mistaken, this is the presentation that discusses mapping the roads before autonomous travel. They compare the maps to what the car sees as it drives. https://www.youtube.com/watch?v=YXylqtEQ0tk

But there are a number of reasons the need pre-mapping are not a huge issue. A) When autonomous cars start to be introduced people will still know how to drive, and can pilot the car through unmapped areas. By the time everyone has an autonomous car, and no one knows how to drive the world will be mapped. B) Google has already shown with Streetview that they can create detailed, up-to-date maps of most roads. C) This problem will likely be solved in time.

Premapping is still required as of today, but they do collect data to try and correct those maps: https://medium.com/waymo/building-maps-for-a-self-driving-ca...

My impression was the Google uses both approaches, so I don't think they are at any disadvantage here.

I also don't think precisely mapping each road is going to be a difficult problem if you enlist a huge network of cars contributing data. The problem is highly parallelizable.

It's difficult to map with a car that is reliant on the map to function. As far as I know, they need a human-driven map car ahead of time on any route their self-driving cars will run on. Tesla could maybe solve this problem just from using their sensors on cars that aren't self-driving, but again, their cars aren't reliant on an authoritative map to begin with.

Reply on news.ycombinator.com