Stanford Cart: Autonomous Driving Pioneer
Stanford Cart achieves autonomous navigation under Hans Moravec, traversing obstacle courses using stereo vision.
Overview
The Stanford Cart was originally designed by James Adams in the 1960s and later developed by Hans Moravec in the 1970s. In 1979, Moravec achieved a milestone breakthrough: the Cart could capture multiple images through a camera mounted on a sliding track, compute stereo vision, identify obstacles, and plan a path across the room. Although it moved extremely slowly (taking hours to advance one meter), this was the first demonstration of computer vision-driven autonomous navigation. The Stanford Cart proved that vision-based autonomous driving was in principle feasible, laying the foundation for later autonomous driving research. Moravec later continued robotics research at CMU, proposing the famous "Moravec's paradox."
Study Notes
Background, mechanics, impact, and key takeaways in one read
Background & Context
From 1973 to 1979, Hans Moravec transformed the early Stanford Cart at the Stanford AI Lab (SAIL) into a vision-navigation experiment. The Cart had been a remote-controlled vehicle since the 1960s; Moravec's idea was to let a machine cross an ordinary, chair-filled room using only images from its onboard TV camera. Computer vision was then in its infancy and compute was minuscule, yet he insisted on a cautious loop: look carefully, plan a short segment, execute a little, stop, and think again. In 1979 the Cart crossed about 20 meters through a chair-strewn room with no human intervention—taking roughly five hours, but delivering the first real demonstration of visual autonomous navigation.
Key Points
- 1In 1979 Moravec's Cart crossed roughly 20 meters through a chair-filled room with no human aid.
- 2It relied entirely on onboard TV-camera images, using no external laser or radar sensors at all.
- 3The camera slid sideways to gain stereo vision, ranging dozens of object corners ahead in three dimensions.
- 4Every ~1 meter it stopped to rebuild a map and plan an obstacle-avoiding path—about 10–15 minutes per step.
- 5It refreshed the 3D map, route, and its own pose each meter; the full crossing took ~5 hours.
- 6It was an early milestone of visual autonomous navigation and of the sense-then-plan perception loop.
How It Works
The Cart's method was "dumb but solid": every ~1 meter it stopped, slid its camera sideways to take two images, and used stereo vision to compute the 3D coordinates of a few dozen object corners ahead, marking them on a sparse map as little stalks with height. The program then estimated its own pose, updated the map, and planned an obstacle-avoiding path, executing only a short segment before stopping to repeat. Picture a blindfolded person who, after each step, feels the surroundings, redraws the map in their head, and decides the next move. With only about a million operations per second—roughly one-thousandth of a 2002 PC—each meter's decision took 10–15 minutes, so the 20-meter crossing lasted about five hours, with small blunders every few tens of meters. The sparse map was barely adequate, yet it was the first time a vehicle "used its eyes" to walk itself through a cluttered space.
Impact & Legacy
The Stanford Cart proved that navigating a cluttered real environment using only vision was possible, laying conceptual groundwork for later autonomous driving—the DARPA Challenges, Stanley, and modern self-driving cars—and for robot vision. Its cautious perceive–map–plan–act loop is the prototype of today's SLAM and visual navigation. The limits are obvious: speed measured in hours, dependence on good lighting, extremely sparse maps, and no handling of dynamic scenes. But in 1979 it pulled "machine vision" from fantasy into the lab, its influence reaching the lidar and deep-learning perception eras.
Key Takeaways
- Autonomous navigation decomposes into a perceive–map–plan–act loop that improves step by step.
- Stereo vision uses two offset images to estimate depth—an early cheap 3D-perception scheme.
- Under compute limits, "slow but steady, think per step" beats blindly chasing speed.
Glossary
Estimating object distance by matching corresponding points in two viewpoint-offset images.
Simultaneous localization and mapping: building a map while estimating one's own pose; the Cart was its prototype.
The ability of a robot to perceive the environment, plan a path, and avoid obstacles autonomously using only onboard camera images.