Interactive Categorization of Containers and Non-Containers by Unifying Categorizations Derived from Multiple Exploratory Behaviors

The ability to form object categories is an important milestone in human infant development (Cohen 2003). We propose a framework that allows a robot to form a unified object categorization from several interactions with objects. This framework is consistent with the principle that robot learning should be ultimately grounded in the robot’s perceptual and behavioral repertoire (Stoytchev 2009). This paper builds upon our previous work (Griffith et al. 2009) by adding more exploratory behaviors (now 6 instead of 1) and by employing consensus clustering for finding a single, unified object categorization. The framework was tested on a container/non-container categorization task with 20 objects. Initial attempts at robotic object categorization have produced limited results as they assume that robots will explore objects using only a single behavior (Griffith et al. 2009). Research with animals, however, has shown that some birds use almost their entire behavioral repertoire to explore a novel object (Lorenz 1996). This suggests that robots should do the same when categorizing objects. Indeed, an object categorization derived from multiple exploratory behaviors may contain more information compared to one derived from a single behavior. Further work is necessary, however, to determine how a robot can combine its observations from multiple behaviors to come up with one unified categorization for a set of objects, instead of having a separate categorization for each behavior. This paper tests the hypothesis that a robot can use consensus clustering to form a single categorization for a set of objects after it interacts with them using multiple exploratory behaviors. Our robot performed a sequence of 6 exploratory behaviors during multiple interaction trials with 20 objects (10 containers and 10 non-containers). The robot extracted features from its interaction history with each object and then employed unsupervised clustering to form 6 different categorizations. Consensus clustering was used to combine the 6 different categorizations into a unified object categorization. This resulted in a meaningful separation of containers from non-containers, even in the presence of noisy clusterings.