Monday, November 10, 2014

Another side to the driverless car discussion

For years now, there have been two basic narratives when it came autonomous cars. The first is what I've called the ddulite version: driverless cars are just around the corner and they are about to change our lives in strange and wonderful ways if we can just keep the regulators out of the way. The second version is more skeptical: while lower levels of autonomy are coming online every day, the truly driverless car still faces daunting technological challenges and, even if those are met, these cars may not not have the often-promised impact. (You can probably guess which side I took.)

If you follow this story through the New York Times or the Economist, you are overwhelmingly likely to get the first version, You may not even know that this bright future is contested. If, on the other hand, you talk to the engineers in the field (and I've talked to or exchanged emails with quite a few recently), you are far more likely to get the second.

This recent Slate article by Lee Gomes is one of the very few to take the second approach.
For starters, the Google car was able to do so much more than its predecessors in large part because the company had the resources to do something no other robotic car research project ever could: develop an ingenious but extremely expensive mapping system. These maps contain the exact three-dimensional location of streetlights, stop signs, crosswalks, lane markings, and every other crucial aspect of a roadway.

That might not seem like such a tough job for the company that gave us Google Earth and Google Maps. But the maps necessary for the Google car are an order of magnitude more complicated. In fact, when I first wrote about the car for MIT Technology Review, Google admitted to me that the process it currently uses to make the maps are too inefficient to work in the country as a whole.

To create them, a dedicated vehicle outfitted with a bank of sensors first makes repeated passes scanning the roadway to be mapped. The data is then downloaded, with every square foot of the landscape pored over by both humans and computers to make sure that all-important real-world objects have been captured. This complete map gets loaded into the car's memory before a journey, and because it knows from the map about the location of many stationary objects, its computer—essentially a generic PC running Ubuntu Linux—can devote more of its energies to tracking moving objects, like other cars.

But the maps have problems, starting with the fact that the car can’t travel a single inch without one. Since maps are one of the engineering foundations of the Google car, before the company's vision for ubiquitous self-driving cars can be realized, all 4 million miles of U.S. public roads will be need to be mapped, plus driveways, off-road trails, and everywhere else you'd ever want to take the car. So far, only a few thousand miles of road have gotten the treatment, most of them around the company's headquarters in Mountain View, California.  The company frequently says that its car has driven more than 700,000 miles safely, but those are the same few thousand mapped miles, driven over and over again.


Noting that the Google car might not be able to handle an unmapped traffic light might sound like a cynical game of "gotcha." But MIT roboticist John Leonard says it goes to the heart of why the Google car project is so daunting. "While the probability of a single driver encountering a newly installed traffic light is very low, the probability of at least one driver encountering one on a given day is very high," Leonard says. The list of these "rare" events is practically endless, said Leonard, who does not expect a full self-driving car in his lifetime (he’s 49).

The Google car will need a computer that can deal with anything the world throws at it.
The mapping system isn’t the only problem. The Google car doesn’t know much about parking: It can’t currently find a space in a supermarket lot or multilevel garage. It can't consistently handle coned-off road construction sites, and its video cameras can sometimes be blinded by the sun when trying to detect the color of a traffic signal. Because it can't tell the difference between a big rock and a crumbled-up piece of newspaper, it will try to drive around both if it encounters either sitting in the middle of the road. (Google specifically confirmed these present shortcomings to me for the MIT Technology Review article.) Can the car currently "see" another vehicle's turn signals or brake lights? Can it tell the difference between the flashing lights on top of a tow truck and those on top of an ambulance? If it's driving past a school playground, and a ball rolls out into the street, will it know to be on special alert? (Google declined to respond to these additional questions when I posed them.)

Computer scientists have various names for the ability to synthesize and respond to this barrage of unpredictable information: "generalized intelligence,” "situational awareness,” "everyday common sense." It's been the dream of artificial intelligence researchers since the advent of computers. And it remains just that. "None of this reasoning will be inside computers anytime soon," says Raj Rajkumar, director of autonomous driving research at Carnegie-Mellon University, former home of both the current and prior directors of Google's car project. Rajkumar adds that the Detroit carmakers with whom he collaborates on autonomous vehicles believe that the prospect of a fully self-driving car arriving anytime soon is "pure science fiction."


  1. Mark:

    I tend to be on your side on this particular issue, but I do have a question here. Is zero accidents really the appropriate comparison point? After all, human drivers run people over all the time.

    1. Andrew,

      Of course, zero accidents is not appropriate. I'd even go further and say that driverless cars are probably already safe enough. The insurance industry, the legal system and society as a whole can handle eighteen-wheelers and teenagers driving SUVs; The occasional malfunction in a slow moving Prius is not that big a deal.

      My take-away from this article is that the functionality is more limited than we've been told. Some of these limitations are serious but should be solvable (parking). Others are potentially technology-killers.

      Unless Google can largely eliminate the time and labor intensive mapping procedure described here, I don't see how their approach to autonomous driving can be viable on a large scale. Right now, I see the incremental (and less hyped) approaches of companies like GM, BMW and Nissan as more likely to pan out.