Learning from Uninterpreted Experience in the SSH
Benjamin J. Kuipers, Patrick Beeson, Joseph ModayU, Jefferson Provost · 2001
Suppose a robot is born into the world with sensors and effectors, but no knowledge of their relationship to the un-known external world. How can the robot learn an inter-pretation of its sensorimotor experience adequate for it to explore its environment and build a cognitive map? Pierce and Kuipers (Pierce & Kuipers 1997) solved a very restricted version of this problem, for a simulated mobile robot whose primary sensory input was a ring of sonar-like range sensors. We are working toward a more general solution, applied to physical robots with multiple sensory modalities. We are working within the framework of the Spatial Se-mantic Hierarchy (SSH) (Kuipers 2000), which is a lattice of related representations for large-scale space. The SSH is conducive to research on learning from uninterpreted expe-rience because it separates the interface representations--