In late July 1995, Carnegie Mellon researchers Dean Pomerleau and Todd Jochem drove a 1990 Pontiac Trans Sport minivan, Navlab 5, from Pittsburgh toward San Diego in a demonstration they called “No Hands Across America.” The car steered itself for the great majority of the trip using only a windshield-mounted camera and a portable computer. As their own project page describes the setup, the gear was modest enough to be “powered from the cigarette lighter, just like a radar detector.”
The system worked by vision: “The computer takes a picture of the scene ahead. Using this picture, it finds the position and orientation of the road” and issued steering commands to stay in lane. The humans kept their feet on the throttle and brake, but the computer did the steering. A preliminary Pittsburgh-to-Washington run had already shown the car could drive itself 96 percent of the way (290 of 302 miles), and the cross-country attempt extended that to roughly 2,850 miles with the vehicle steering for more than 98 percent of the distance.
What makes this milestone striking is its date. A decade before the DARPA Grand Challenges, a camera-and-laptop rig was already handling highway lane-keeping across most of a continent. The hard parts left undone, throttle, braking, lane changes, traffic, and unmapped roads, are exactly the parts that took the next thirty years to chip away at.
For a general reader, No Hands Across America is a reminder that the easy 90 percent of autonomy arrived surprisingly early and cheaply, and that the remaining few percent, the messy edge cases, are what made self-driving a multi-decade, multi-billion-dollar effort rather than a 1990s product.