The Stanford Cart began as a way to study driving a lunar vehicle from Earth. Around 1960, researchers gave a wheeled test vehicle a control link that could imitate the 2.6-second Earth–Moon radio delay. Later versions followed a white line; by 1979, the cart was choosing a route around obstacles.[1]
That progression changed the job of its camera. An operator first needed a view from which to steer despite the delay. A line-following machine needed to keep a visible track in the right place. Autonomous navigation required a richer answer: where are the obstructions, and where can the vehicle go?[1]
One camera, several viewpoints
In Hans Moravec’s 1980 thesis, the camera takes nine pictures along a 52 cm track, at 6.5 cm intervals. Software matches distinctive patches across the images. Nearby features shift farther than distant ones, supplying depth.[2]
Nine pictures provide 36 distinct pairs: 9 × 8 ÷ 2. The program combined their distance estimates; consistent matches reinforced one another. Shared images make these correlated measurements.[2]
After driving roughly a metre, the cart photographed the scene again. Matching previously located features revealed its actual movement, allowing it to revise the path.[2]
Each stop took 10–15 minutes to process. Outdoors, movement-estimation errors placed obstacles at conflicting map positions; the cart struck cardboard scenery. A navigation map depended on locating its observer.[2]
A map can preserve uncertainty
Moravec’s separate 1996 work described a fuller representation: a three-dimensional grid accumulating evidence about small volumes of space. A hand-collected set of 20 stereo image pairs produced an office map with 256 × 256 × 64 cells. Chairs and a cabinet became visible in reconstructed slices.[3]
Three different meanings inside the map
| State | Meaning |
|---|---|
| Occupied | Evidence suggests material is present.Source 3 |
| Empty | Observations support a clear volume.Source 3 |
| Unknown | The map has insufficient evidence.Source 3 |
Keeping the third state matters: an unseen part of a room cannot become free space merely because no obstacle has been recorded there. The three-dimensional representation also separates heights, so an overhanging tabletop need not be flattened into the same description as the floor beneath it.[3]
The office demonstration used collected images; it was a later mapping experiment. Together, the projects show two connected achievements: recovering structure from changing views, then retaining what those observations justify in a spatial model.[3]
Sources & further reading
- Stanford’s robotics legacy Stanford University · 16 Jan 2019 · Accessed 19 Oct 2026
- Obstacle Avoidance and Navigation in the Real World by a Seeing Robot Rover — Chapters 3, 7 and 9 Hans Moravec / Stanford University; author’s archive at Carnegie Mellon University · Accessed 19 Oct 2026
- Robot Spatial Perception by Stereoscopic Vision and 3D Evidence Grids Hans Moravec / Carnegie Mellon University · Accessed 19 Oct 2026