Google Using Self-Driving Car Data To Make Cars Smarter

cartechboy (2660665) writes "One thing Google has perfected is using massive data sets generated from users to improve user experience. Google's self-driving cars may be subject to the same cycle of improvement, as they have racked up considerable mileage on public roads, and each mile generates data that Google engineers can use to 'teach' vehicle. Meet Pricilla — a Google test driver on the self-driving car project as she does a video walk through of some of the improvements created so far. Some are fairly simplistic, for example: 'The car does move to avoid large obstacles." That said, the car can also detect a bicyclist signaling and stay clear — oddly, even when that cyclist changes his mind and zig zags a little." Google is now testing cars on the city streets of Mountain View.

Comments

Still waiting to see 3 things

[NewWorldDan]

1. Cop directing traffic
2. A more complicated construction zone with a badly marked detour
3. A snow storm

Things are coming along nicely, but I still imagine these are a decade away. Still, they should be common and affordable by the time I'm ready to plow through a farmer's market.

Re:Still waiting to see 3 things

[Missing.Matter]

A snow storm is the big one. Also, rain and dust can be a big problem as well. The thing is, when you R&D these systems in sunny California, silly things like "precipitation" seem to get forgotten. I remember seeing a presentation about the Google street view cars, and how when they deployed them to other regions, they had to institute lens cleaning procedures because they had pretty much forgotten it rains in other places in the world.

Re:Still waiting to see 3 things

[Missing.Matter]

I know what they use. I use the same sensors. Rain is a problem. Dust is a problem. Snow is a HUGE problem. Of the sensors on the car, the Velodyne HDL-64E on top is by far the most important, and provides the most critical data used for localization and detecting dynamic objects. With rain in particular laser beams can get refracted or reflected by the raindrops, which incredibly fuck up your distance measurements. The typical solution is to do a lot of filtering (i.e. take the median of two measurements) but this cuts down your effective frame rate, which already isn't that high to begin with.

Snow pretty much guarantees that "manual operation mode" is going to be a primary interface in autonomous cars for a long time to come, as we wait for not only sensors to get up to speed, but also machine learning in general.

Re:Still waiting to see 3 things

[American AC in Paris]

When the computer is good enough that you haven't had to do any driving in the past 3 months, how much are you really going to be paying attention when something goes wrong?

I'd suggest that once this is consumer-ready, the vast majority of "something goes wrong" scenarios where the car doesn't know what to do would fall into one of two categories:

• "I don't know what to do, therefore I will come to a complete stop (and pull over to the side of the road, assuming I can identify a safe path);" or
• "If I can't react adequately to this situation, there's very little chance that you, meatsack, would have done even half as well as I can manage right now."

These things'll never, ever be perfect. They will almost undoubtedly reach a point where they're at least an order of magnitude safer than humans, though. That'll be more than good enough for most people.

In the future...

[Ichijo]

...our children and grandchildren would wonder why we ever allowed humans to operate motor vehicles on public roads.

Fucking Cyclists are ruining the future.

[Anonymous Coward]

Fucking cyclists are going to have a free lunch with self driving cars.

Just plain old taking over the streets because they know the computers will give them right of way everytime.

Google needs to program the cars with some assholery in mind where scaring cyclists is as common as checking for the car's fuel.

Re:Fucking Cyclists are ruining the future.

[Ichijo]

What if there's no bike lane, or it's filled with debris, or the bicyclist needs to make a left turn?

Re:Fucking Cyclists are ruining the future.

[PvtVoid]

If grid lock, anger, douch baggery, and killing automated cars is you goal, then sure. Otherwise, we will just have the self righteous assholes going 15 MPH in a 25 MPH zone becasue they are too precious to use the bike lane.

And the car will have the simple common sense to wait ten or fifteen seconds until it's safe to go around the cyclist, who will inevitably catch up at the next traffic light anyway. This behavior seems to elude humans.

Not sure we need it

[nine-times]

These kinds of problems do need to be addressed, but I'm not sure they need to be *fixed* before you turn this into a product. For the snow-storm example, I don't think the car needs to be able to drive in the snow. It's much more important that the car is capable of detecting "this is a situation in which I can't operate safely," and refusing to try. It should be good enough if the car's AI can say, in effect, "Listen, human, I can't take responsibility for driving in this snow storm. If you're comfortable driving in it, go ahead and take manual control. Otherwise, we're staying right here." On the other hand, I could see an interesting application in providing some kind of intelligent 'driver assist' for bad weather conditions that helped the driver maintain traction.

Regarding details, I think the ideal would be for most road conditions, detours, and traffic issues to be kept up-to-date on a database that could allow for dynamic routing instead of the car relying completely on markers. It's not a complete solution, but again, it may be enough to pair a large database with some ability for the car to say, "I don't know what to do here, so I'm going to either give back manual control or pull over and wait."