Case-Based Similarity for Social Robots

. Social robots must operate in a world of cases that are richer and more varied than those used in conventional case-based systems, since they must be able to represent their internal state, plus their environment, plus the states of humans around them. The attributes describing this environment are qualitative, with multivariate dependencies among them, and at any particular moment those attributes that are relevant will typically form a small subset of all possible attributes. We design and implement a case-based system to deal with this complexity. It is characterized by (a) decomposition of attributes into subtypes that have qualitatively different similarity structures (in particular, affordances, contexts, properties), and (b) an inference technique for cases that is automatically derived from natural-language descriptions so that a robot can infer it from what it sees (e.g. signage, menus) and (over)hears, rather than being limited to a pre-built ontology. We demonstrate the effectiveness of this system by generating subtype clusters with high semantic relevance.