Integration of behaviors and languages with a hierarchal structure self-organized in a neuro-dynamical model

This paper proposes an approach for robots to ac-quire language grounding in their robot's sensory-motor flow using neuro-dynamical models. We trained our neuro-dynamical model over a set of sentences represented as sequences of characters. For the integrated recognition, we introduced a cognitive hypothesis for integrated recognition where a human's brain separately processed the "structure" and "contents" of a sentence. Our model was trained with the spelling of words and their semantic role emerged in the first model. As a result of binding the model with sensory-motion patterns, we confirmed that it could associate proper word spellings with a sensory-motor flows and a semantic roles, even if an observed flow had not been learned.