The Stanford Cart crosses a room

The Stanford Cart began life in the early 1960s as a test of whether a lunar rover could be driven from Earth with a two-and-a-half-second radio delay. By 1979 Hans Moravec had made it drive itself. A television camera slid along a rail on top, taking nine pictures from different positions; the computer matched features between them to build a three-dimensional map, planned a path around the obstacles, moved a metre, and started again. Crossing a room full of chairs took about five hours, and the cart sometimes lost track of where it was when shadows moved.

It was the first vehicle to navigate by vision alone, and the ancestor of every self-driving car. The stereo matching Moravec used to find depth is the same problem the camera-only autonomy systems of the 2020s solve at thirty frames a second.

Moravec drew a broader conclusion from the experience, now called Moravec's paradox: the things that are hard for people, chess and calculus, are easy for computers, and the things a one-year-old does without effort, seeing and moving, are the hardest problems in the field. It took the deep-learning era to make the paradox look less absolute.