Why automated vehicles must become social
Human drivers deploy more than a physical model of the road, other vehicles, and physical environment. They also rely on a behavioral model of the drivers, bicyclists, and pedestrians around them, sometimes without being entirely conscious of it. That guy in the grey Toyota has been edging close to the right lane, and likely wants to get over to exit soon. That bicyclist is an older woman carrying a heavy load on her back rack, so may wobble. That pedestrian is more attentive to his iPod than to the traffic lights. The communication runs two ways; we ease back a bit to let that grey Toyota driver know we understand his intent, even before he signals. That kind of social interaction is necessary, especially when negotiating urban streets and intersections. Or when an automated vehicle picks up and lets off passengers at a downtown bar.
It would be easy but mistaken to think that automated vehicles could get away without that kind of social interaction. They don’t need to drive like humans. But they will be operating in an environment with humans. Perhaps automated trucks can get away without social behavior, if they use only controlled access highways, start and stop at special terminals, and call for backup when encountering a situation that is too complex. A more general purpose automated vehicle has to do more, including behave sensibly when people around it are not. (What’s that coked-up bar hopper doing on my hood?!)
Human social behavior also extends beyond other humans. We recognize threatening or friendly or needy behavior from dogs. Hunters and birders gear their efforts to the behavior of their prey or subjects. The kiskadees in my neighborhood know if they call an alert from my back tree, I’ll come to run off the stray cat.
Elon Musk is putting out the call for “hardcore software engineers.” What may be more difficult is finding the AI experts and social scientists who can develop the necessary social model for an automated vehicle. I suspect both Tesla and Google soon will run smack into that problem. It’s not an easy one. It requires machine learning, because a model of human drivers, bicyclists, and pedestrians, is something that we cannot program from scratch. It may have to be adaptable, because human behavior varies from town to country, and from region to region.
And it will be necessary for a broad variety of other kinds of helper robots, beyond automated vehicles.