Seamless Integration and Coordination of Cognitive Skills in Humanoid Robots: A Deep Learning Approach

摘要

This study investigates how adequate coordination among the differentcognitive processes of a humanoid robot can be developed through end-to-endlearning of direct perception of visuomotor stream. We propose a deep dynamicneural network model built on a dynamic vision network, a motor generationnetwork, and a higher-level network. The proposed model was designed to processand to integrate direct perception of dynamic visuomotor patterns in ahierarchical model characterized by different spatial and temporal constraintsimposed on each level. We conducted synthetic robotic experiments in which arobot learned to read human's intention through observing the gestures and thento generate the corresponding goal-directed actions. Results verify that theproposed model is able to learn the tutored skills and to generalize them tonovel situations. The model showed synergic coordination of perception, actionand decision making, and it integrated and coordinated a set of cognitiveskills including visual perception, intention reading, attention switching,working memory, action preparation and execution in a seamless manner. Analysisreveals that coherent internal representations emerged at each level of thehierarchy. Higher-level representation reflecting actional intention developedby means of continuous integration of the lower-level visuo-proprioceptivestream.