Extraction and inversion of abstract sensory flow representations

Abstract

We present a technique for analysing extracted abstract sensory ow representations. Mobile robot systems generally have a very limited memory storage capacity, and can thus only store information about a very short period of time. However, instead of storing a very detailed picture of the past, a more abstract representation can be extracted and stored, providing a much longer, but less detailed, picture of the past to the control system. In our technique, the components of such an abstract representation are automatically extracted; they correspond to distinct and stable inputs which arrive at the sensory systems. To a distal observer, the extracted components actually correspond to concepts such as corridors, walls and corners. However, the analysis of the extracted components has previously required that the experimenter constantly watches as the robot does something and notes which concepts get activated. This is cumbersome, and such an inspection may even be i...