Grounding Abstract Action Words through the Hierarchical Organization of Motor Primitives

Cognitive developmental robotics is facing the challenge of building robots capable of working independently and/or with other agents in non-structured scenarios, which can autonomously react to dynamic changes that occur in the environment. Providing robots with the capability to comprehend and produce language in a "human-like" manner represents a powerful tool for flexible and intelligent interaction between robots and human beings. Robots endowed with linguistic capabilities, in fact, could better understand situations and exchange information; through language robots could cooperate and negotiate with human beings in order to accomplish shared plans. This work describes a neuro-robotics model for the acquisition of abstract action words in the iCub humanoid robot. We claim that the acquisition of concepts that refer to such abstract words (e.g. verbs like "use", "make") can be driven by action's organization. In the presented model the integration of low-level capabilities (e.g. perceptual and sensorimotor skills) enables the hierarchical organization of concepts that leads to the grounding of more general words.