A memory-based system for autonomous indoor navigation is presented. T
he system was implemented as a follow-midline reflex on a robot that m
oves along the corridors of our institute. The robot estimates its pos
ition in the environment by comparing the visual input with images con
tained in its memory. Spatial positions are represented by classes. Me
mories are formed during a learning phase by encoding labeled images.
The output of the system is the a posteriori probability distribution
of the classes, given an input image. During performance, an image is
assigned to the class that maximizes the probability. This work shows
that extensive use of memory can reduce information processing to a si
mple and flexible procedure, without the need of complicated and speci
fic preprocessing. The system is shown to be reliable, with good gener
alization capability. With learning limited to a small part of a corri
dor, the robot navigates along the entire corridor. Furthermore, it is
able to move in other corridors of different shape, with different il
lumination conditions.