This article presents the notion of qualitative descriptor, a theoretical tool which describes within the same formalism different approaches to transform quantitative data into qualitative data. This formalism is used with a grouping algorithm to extract qualitative phases from a data flow. Work on action perception, based on qualitative descriptors, is used to illustrate these ideas. The grouping algorithm generates a qualitative symbolic data flow from a video sequence. The ultimate aim is to provide an unsupervised learning algorithm working this qualitative flow to extract abstract description for common actions such as “take”, “push” and “pull”.