Segmentation, Memorization, Recognition and Abstraction of Humanoid Motions Based on Correlations and Associative Memory

In order to self-organize symbols from observed motion patterns, it is necessary to temporally segment the continuous motion pattern flows into meaningful chunks. For reusability of the acquired information, repeatedly observed patterns are important, which means that segmentation, memorization, recognition and abstraction depend on each other. From this point of view, we propose methods for motion patterns of humanoid robots observed as a continuous flow using pattern correlations and associative memory. Initially, patterns are segmented by pattern correlations and then stored into the associative memory. Afterwards, only new kinds of motions are fed through this process. Associative memory is capable of segmentation, recognition and abstraction, and has ease in incremental update of the stroge for new patterns