An autonomous car has been built, and has driven 140,000 miles without accident.
Apparently mass production is intended within ten years.
To me — at least, to me wearing my software engineering hat — I see parallels with projects that use unrealistic test data, then blow up catastrophically in the real world.
We developers often use inadequate test harnesses, data sets that are too small, and clean test inputs. We get a false sense of confidence in our code, and this is only deflated by thorough exercise during betas. At a higher level, software companies seem to have a tendency to target our own insular demographics: we build startups that target 25–35 year olds living in the Bay Area, for example, and wonder why they fail to get broad appeal.
The equivalent for the autonomous car?
I doubt very much that Google’s autonomous car has been tested in a Northwest blizzard, dragging its thin all-weather tires through 8 inches of snow. In the Bay Area, that’s not a problem: you might have to scrape a windshield free of ice one day each year. Outside the Bay Area, one has to cope with snowplows, whiteouts, drifts, spray, cars sliding across the road in front of you, and frequent loss of traction… then sucking mud, laddered gravel roads, and dust storms (with the consequent problems with both visibility and sensor clogging).
The parking camera on my truck, which is at chest height on a tall man, usually ends up completely iced over and useless by the end of a journey. The front of my truck looked like the inside of an old freezer after today’s 3-hour drive in a blizzard. How will the sensor array on a small autonomous car fare?
Color me skeptical.