Visual Memory-Based Learning for Mobile Robot Navigation
                              Daniel Nikovski


Vision can be very useful for mobile robot navigation, because of the 
richness of visual representations. Learning can be very useful too,
because it eliminates the need to program recognition tasks explicitly.
Both vision and learning, however, are usually computationally
intensive and hard to use in real-time control and navigation.

I will describe a simple way to get fast learning and fast visual 
processing at the same time. The idea of the method is to do
memory-based learning by means of Fourier spectra. For typical 
image sizes, this results in hundred-fold speed-ups and permits
real-time visual servo-control on Xavier and Amelia. I will 
discuss several implementations of the control loop for the 
task of finding a door and exiting/entering a room.