Dutchy Ron schreef op 12 oktober 2018 09:00:
Artikel WSJ
To achieve the dream of autonomous vehicles and robots, it’s going to take much more than computer vision and artificial intelligence. Cars, drones, delivery bots, even our vacuum cleaners and robot chefs are going to need something that our ancestors developed millions of years ago: a sense of place.
“I definitely don’t think people understand how reliant autonomous cars are on the fidelity of the map,” says Mary Cummings, a professor of mechanical, electrical and computer engineering at Duke University. “If the map is wrong then the car is going to do something wrong.”
It turns out that, whether it’s Waymo’s self-driving cars or the many auto manufacturers relying on tech from Intel Corp.’s Mobileye, so-called “autonomous” vehicles are cheating, in a way. This is also true of models that are already commercially available, such as Cadillacs with Super Cruise.
Rather than perceiving the world and deciding on the fly what to do next, these autonomous and semi-autonomous vehicles are comparing their glimpses of the world with a map stored in memory. The incredibly detailed maps they rely on are what engineers call a “world model” of the environment. The model contains things that don’t change very often, from the edges of roads and lanes to the placement of stop signs, signals, crosswalks and other infrastructure.
That self-driving cars—and eventually, all other forms of autonomous robots—require such a map has big implications for who will need to partner with whom in the autonomous driving space. It implies a great deal of collaboration, or at least licensing, because the amount of data and engineering required to build these maps is so gargantuan. It means that, at least for the foreseeable future, no matter how sophisticated a company’s self-driving technology, it must engage in a massive effort or else partner with someone capable of making an ultra-detailed map of every road on which they might drive—companies like Ushr, TomTom and Here.
In urban environments where global positioning systems can be inaccurate, a vehicle must navigate by landmarks, says Sam Abuelsamid, a senior analyst with Navigant research who specializes in mobility. Once a vehicle is navigating using lidar—a 3-D laser view of the environment—along with cameras and possibly radar, it can cross-reference certain buildings, lamp posts or street markings, to identify its stretch of the road down to the centimeter, he adds.
When the car knows precisely where it is, it can follow predetermined routes in its memory, simplifying the driving process. When the Super Cruise system is activated in a Cadillac, the car stays in its lane by following a route that has been determined ahead of time, says Christopher Thibodeau, senior vice president and general manager at Ushr. Ushr makes the Super Cruise system’s maps—terabytes of map data boiled down to a few hundred megabytes of relevant route information.