Identifying Qualitatively Different Experiences: Experiments with a Mobile Robot

We present an unsupervised learning method that allows a situated embodied agent to identify and represent qualitatively diierent experiences. The occurrence of events, such as the initiation of a particular action, triggers the collection of multivariate time series of sensor values. Those time series are clustered using Dynamic Time Warping as a measure of similarity , and prototypes are extracted from the clusters. Each prototype represents a distinct experience, such as one possible outcome of engaging in an activity. Prototypes can be used for oo-line planning and for on-line prediction by nding the best partial match among the prototypes to current sensor readings. Experiments with a Pioneer-1 mobile robot demonstrate the utility of the approach with respect to capturing the structure and dynamics of a complex, real-world environment.