The long tail of driving
Uber will put self-driving cars on Pittsburgh streets next month, six weeks after the first Autopilot fatality became public. The hard part of driving is the rare case.
Uber announced on Thursday that customers in Pittsburgh will be able to hail self-driving Volvos starting later this month, with an engineer in the driver’s seat ready to take over. The same week, Ford said it plans fully driverless cars with no steering wheel for ride sharing by 2021.
At the end of June, Tesla disclosed that a Model S driver had been killed in May while Autopilot was engaged. A tractor trailer crossed the highway in front of the car. Tesla’s explanation was that neither the camera system nor the driver noticed the white side of the trailer against a brightly lit sky, and the car drove under it.
These two stories together describe the problem well. Getting a car to drive itself 99% of the time is a problem the industry mostly knows how to solve. Lane keeping, following distance, highway merges, even a lot of city driving. Most of the miles are boring and look like each other. The remaining 1% is where everything happens: a white trailer against a white sky, a cyclist doing something odd, a police officer waving you through a red light, a plastic bag versus a rock on the road, a child’s ball rolling out from between parked cars.
Each of those cases is rare, but there are so many different ones that together they come up all the time. You can’t write rules for all of them, and a learned system only handles the ones that looked something like its training data. Machine learning is good at the fat middle of a distribution and weakest at the tail, and in driving the tail is the part that kills people.
What humans bring to the tail isn’t better eyesight. It’s a general model of how the world works. You’ve never seen that exact trailer at that angle, but you know trucks are solid, that they don’t disappear against the sky, and that something big crossing a highway is worth slowing down for even if you can’t tell what it is. The car has none of that. It has a classifier for things it was trained on and a set of rules about what to do with them.
I think fully driverless ride-sharing will happen, but on a longer timeline than 2021 for anything beyond limited areas with good weather and lots of mapping. The long tail gets shorter slowly. The companies that get there will be the ones collecting the most miles of weird cases, and they’ll need systems that understand scenes well beyond labeling what’s in them.