Hacker News Reader

Comment by revelation | Hacker News Reader

The first will be allowing the car to stray from routes that it has been specifically trained to drive.

I think the author is hugely overstating the importance of this.

I agree 100%. My commute is over the same route, every day, some 260 days a year. The value of freeing me up to do something else for that hour each day can't be understated.

I live in NYC and the best use-case for this is taxis. Taxi do not take the same route twice. Traffic is a nightmare during different seasonal, temporal, or accident events. So like a Garmin device it should be able to "remap" the trip.

In our lifetimes this will be possible but maybe not in next 1-3 years. (I hope I am wrong)

The car already doesn't need to be on planned routes, it remaps the trip as needed. That was done back during the DARPA Challenges.

They're just limiting the car here for the media.

Google has covered large parts of the (populated) world with Street View. Pictures [1] indicate that the cars were equipped with laser sensors. So Google probably has enough data to have the cars drive new routes.

But thats not the point. Most of peoples daily driving routine is the same routes over and over, commuting to and from work.

Going even further: I'd imagine that you could train the car on a specific route by driving it manually. Such information could be crowdsourced, much as Google already does with its traffic data on Maps.

[1]: http://cdn.physorg.com/newman/gfx/news/hires/streetviewca.jp...

I believed that as well, but now I'm not so sure. The density of data gathered by the Velodyne on the self driving cars is far higher than the SICK lasers in that picture. Also, the scene may have changed in the interim.

Yes, but we're still talking cars that measure metres by metres, and streets with lots of extra room. As you mention, the scene may have changed - so trying to preplan the whole route to the inch will only hurt you. What is needed is basic data above GPS, and I think the Street View lasers can be that.

I'm sorry I was unclear - I meant that the scene could have changed since the Street View team was last through.

The Google Maps data is surprisingly bad - we saw several presentations about the Car, and they highlighted how awful it really is.

I don't think that they're preplanning. I think that they're analyzing the scene for differences to simplify and speed calculation. Having good a priori data is really powerful.

You might be surprised. It's a bit more than having to just program in new GPS coordinates. Imagine the car drives the route once collecting tons of data with the laser range finder and other sensors on the car. Now they let the car sit for a few days while they process all the data. Finally the car can drive the route. There's still some pretty hard challenges to solve in making those offline algorithms realtime. Maybe some of it is just waiting a couple years for faster processors, but I'm guessing they might have some pretty intense engineering challenges ahead.

I don't think he meant doing it was an easy task, but rather that a large majority of the time when you take your car you go on a route you've already taken in the past, and if the car can self train its routes when in manual mode then you only need to drive it once / the first few times to go to work/groceries/school/...

Reply on news.ycombinator.com