Hacker News Reader

Comment by samcheng | Hacker News Reader

I was just thinking about how the Google Self-Driving Car effort has become more aggressive recently. On a five-mile drive through Mountain View this afternoon (on El Camino and a residential neighborhood) I saw FOUR of the little bubble cars, and one of the Lexuses.

Of course, on each of the cars I saw closely, the operator had their hand on the joystick...

View on HN

Their hands probably always have to be on the joystick, following it's motion. That way if anything goes wrong, they can take over faster. That doesn't mean the car isn't in auto-drive.

They probably realized that the incremental approach to self driving cars is getting to a very reasonable point, and if they want their big-bang approach to be feasible, they need to put the pedal to the metal, as they say.

Maybe they're testing convoy algorithms now. I can easily imagine autonomous vehicles forming convoys, using some short-distance communication to synchronize and thus optimize their movements.

I actually studied exactly that! The optimal calculations for headway distance get a bit involved, but I coded up a modest algorithm and simple simulator. https://github.com/dvbuntu/oddities

Gotta rack up those miles, man. They're all on the same stretches of road they've been driving for years, but Google sells that "miles driven" stat hard.

Obviously you do need to drive in other areas, particularly to get different weather conditions. However, even driving in the same area different incidents are encountered everyday - it is not a waste of training.

Google's cars have:

0 miles in snowstorms

0 miles in historic city centers with small mixed-use (pedestrian, bike and car) streets without lane markings

0 miles in mixed-used residential streets (basically, it's both a footpath and a street, usually with kids playing on it, and also used for cars)

etc.

They mastered only highway driving and driving on perfectly segregated roads.

That's the easy part that VW and Sony's cars had already working in 2008.

i think self driving cars will be a reality, but probably in no less than 50 years.

it's 2016 and we can't even keep websites up without armies of people tending to them 24/7. i don't know what reality everyone who is enthusiastic about this technology is living in.

A lot of "AI"/machine learning/etc. examples show impressive results on relatively narrow and well-controlled use cases. A lot of people look at those and extrapolate to much broader and less controlled environments with lots of messy corner cases and assume it's just a bit more incremental work to get there. Whereas, in fact, getting from essentially a demo to mostly just works in messy real world environments can easily be 90% of the work.

hint: Google hired from CMU's DARPA team

If that is the case, the vehicles might have a better chance at operating under winter conditions, but it likely still won't help them in the streets of Mumbai, Amsterdam, or Rome.

That, and actually increasing the dataset...

Reply on news.ycombinator.com