"Computers are famously devoid of common sense, and you can think of this pre-mapping as a way to bootstrap some common sense into the car," Urmson says.
Even so, if the world changes between the time the map was assembled and the time the test vehicle drives the route, it can lead to confusion. "There are things that right now are a challenge for us," Urmson says. "For instance, if most of the world stayed the same but the lanes are shifted—so the physical road didn't move but, for whatever reason, the department of transportation decided we should drive a half lane to the left—that would probably confuse the car today."
There are two main components to Google's efforts: "The first is reliability, which means having the car do the things we expect it to do over and over again; and the second is robustness, which is dealing with unusual situations and still being safe," Urmson says.
A large part of the ability to increase the system's reliability and robustness depends on developing new sensors that can see farther and provide a denser data set, according to Urmson. But the ability for a self-driving robot to deal with unexpected or unusual situations that get thrown at it are what make some question the self-driving car's apparent inevitability.
John Leonard, a Massachusetts Institute of Technology mechanical and ocean engineering professor who led that university's team to a fourth-place finish in the 2007 DARPA Urban Challenge, thinks that major technological hurdles in robot perception need to be overcome before self-driving cars can be deployed on a large-scale.
"I have tremendous admiration for my colleagues at Google," Leonard says. "The performance that they have achieved is amazing—for example, their ability to drive at highway speeds. However, because they are building the maps in advance and then having humans pick out stop signs and street lights and crosswalks and so forth, it's very different than turning a robot loose autonomously in the world with very little prior information."
Dealing with the extremes
Much of Leonard's work has been focused on Simultaneous Localization and Mapping (SLAM). Unlike the pre-mapping that Google's system requires, SLAM would allow a vehicle to drive through the world at the same time that it is mapping it. This holy grail of autonomous driving would greatly increase the self-driving car's ability to deal with dynamically changing information—but even SLAM would not be able to solve all problems.
More than 15 years ago the No Hands Across America team drove from Washington, D.C., to San Diego in an autonomous vehicle and made it 98.2 percent of the way without human intervention, Leonard says. What about that last 1.8 percent? "A key challenge even today is dealing with those unexpected moments," he adds. "To try to get to that 100 percent level of performance there's a common-sense reasoning—one of those elusive goals of artificial intelligence—that no amount of pre-mapping is going to prepare you for."
Even Google admits that they have no good way for dealing with these unexpected moments yet, which is why every test vehicle has two backup humans on board to monitor and take over when the car reacts strangely. "Our program is very much a research program at this point, and we haven't really addressed that issue yet," Urmson says. "If we were to take the people out of the cars today, they'd drive pretty well, and you probably wouldn't notice them on the road until something unexpected happened or some element of the [unreliability] of the system appeared."