Until our ever-smarter cars can finally take the wheel, the main problem will be to get them to cooperate with us—to sense our intentions, exploit our abilities, and anticipate our errors.
Swedish researchers say they've made a step in that direction by explaining the longstanding problem of why people drive so jerkily. Until now, the consensus was that it had to do with our difficulties in tracking the road; the researchers say that it is better understood as a reaching problem.
It turns out we human beings spend about as much time reaching for something close as for something farther away, going slow or fast as the case may be. It’s just the way we are. The result here is oversteering, usually a small annoyance but one that can mean danger, for instance when driving on icy roads.
Ola Benderius, a doctoral student in machine and vehicle systems at Chalmers University, and his colleague, Gustav Markkula, an engineer with Volvo, modeled 1,000 hours of real-life steering data and found that 95 percent of the corrections made while following the road could be explained by the reaching theory.
Bendarius used driving simulators to compare 12-year-old children with their parents and found that even kids without any driving experience at all were very good at keeping in their lanes. “From these two examples, one observed from a child and one from an adult, it is quite clear that drivers are neurologically hard-wired in their response to unexpected steering wheel disturbances,” he says in his doctoral thesis.
It may seem strange that human behavior should have evolved to meet such mediocre specifications, but this may be just another example of the perfect being the enemy of the good-enough. In the terminology of the late cognitive scientist Herbert Simon, we do not seek to optimize, merely to “satisifice.” After all, as Bendarius observes, it’s easy to add corrections to steering when the driver notices that the car is reaching some threshold—say, that of straying from its lane—“something which is arguably more realistic than continuous, error-minimizing control.”
The point of all this work is to improve the models that active-safety programs use to figure out what the driver is going to do and whether to offer advice or even to take charge. Examples include forward collision avoidance, which sounds an alarm to wake a driver who’s coming dangerously close to the car in front. If that doesn’t work, it may then shake the wheel, then even slam on the brakes to minimize the effects of impact.
Today these programs try to predict what people will do on the basis of logic. Tomorrow’s will take account of our illogical side, as well.