One Thing That Makes Google's Self-Driving Cars Work
theatlantic.com
theatlantic.com
If you wish a less irritating title, then, I'd put it at: "Google-scale machine learning and data collection is the thing that makes Google's self-driving cars work"
1. They map the streets to inches, right down to the height of the curb.
2. They record prior trips in such detail, data-mine it appropriately, and then hand-massage the results.
So far they've done this in little more than the Mountain View area. Having done the hard work in that limited region to that much detail, they now can derive how to automate the process. I imagine their recent high-detail-mapping "phone" will be a key part of this, tracking minute aspects of how someone travels from A to B, and merging that info with many other identical & overlapping trips so the cars can make the trip darn near blind, with a little help from monitoring immediacies (like actual position, pedestrian detection, weather, etc).
edit: Although there's a distinction between tech that can be previewed (Tango, Street View) and those which is developed in relative secrecy.
http://www.newyorker.com/reporting/2013/11/25/131125fa_fact_...
They also hinted that the cars drive over the speed limit to match prevailing traffic.
Who knows though, I was reading about a transmission that used mapping data to anticipate shifts. I guess it was in the new Rolls Royce Wraith.
Easier to automate would be trains, large container ships, tube trains, planes or buses (fixed route, designated stopping zones).
Granted planes and ships travel in more dangerous regions, but I think they should be flown 99% of the time by computer and pilots only taking over in an emergency.
I found it hard to believe anyone act that way. Relying on car-to-car communication would create an enormous chicken and egg problem.
I suspect the author overstated the level of disbelief here.
1. sudden road changes, like the appearance of potholes
2. avoiding puddles which can cause one side of your car to hydroplane
3. taking a 'rain line' when the road is wet. (Traffic tends to put slight ruts in the road. They're enough to make slight puddles. If you drive slightly off-center, your tires will be on the high spots and less likely to hydroplane, that's that 'rain line'.)
4. avoiding icy spots
It's much easier for the on-board processing to start with a baseline and then figure out how the world differs than to start from scratch from the sensors.
There are other issues, too, such as I've been behind gravel trucks that were constantly dropping gravel, and other vehicles with a clearly unstable load that I get away from. I also give wide berth to cars that don't appear to be driven in a stable manner.
The LIDAR system the car relies on doesn't work in rain, snow, or fog. Also, it would have a lot of trouble with the massive changes to roads and driving patterns caused by seasonal snow accumulation in the mid-west and north east.
It's relatively easy to create a first-time-through map of something. You simply record everything you see (or thought you saw). On revisiting, there are now questions of:
• Was that here the first time?
• Did it move? Was it mismeasured?
• What about the stuff that's recorded that we don't see?
Early cartogrophers (as well as researchers in conceptual fields) often mistakenly create or record features where none actually exist (or at least no subsequent evidence is found). Often the ghosts of these features haunt maps (or textbooks) for centuries.
Essentially, you need a map, ultimately, which has a sense for the permanence of any highway feature, and to not fall prey to map-territory relation assertion fallacies: http://en.wikipedia.org/wiki/Map-territory_relation
The map is not the territory.
Obviously the Google car's data collection is more extreme and precise than that which happens in your head, but that seems like it's mostly just because they don't yet entirely know what they need.
And this, in the long run, will be part of the advantage of these cars once they get the process more automated. I expect that when you can finally buy one of these cars they'll have a (perhaps opt-out) system where terrain data is collected and merged into Google's own system to provide better data for future drives for not just you, but everyone else as well.
And that'll be one of the reasons it'll be safer. If also a whole bunch more creepy.
Like the author says, the idea really does have a strong whiff of Googleyness. Bind-bogglingly ambitious, technically extremely impressive, and fairly creepy.
one can also imagine that to renew car registration one would have in addition to evidence of insurance to also provide evidence of updated subscription to the Google road database
Or, more likely, if you haven't kept your google car data subscription up to date (or can't agree with their latest privacy policy change), your self-driving vehicle suddenly becomes useless?