Visuomotor learning for achieving cognitive robot behaviors

Abstract

Endowing a robot with human-like cognitive capabilities is one of the long-term goals of artificial intelligence and robotics. In this dissertation, we introduce our approach to building a cognitive robot from a visuomotor associative learning perspective. We assumed that cognitive robot behaviors would emerge by learning from sensorimotor experience acquired from the interaction with its environment. Particularly, we focused on the four key principles: learning from experience, sensorimotor integration, hierarchical computation, and prediction error minimization. Based on these principles, we proposed deep neural network models that can learn from large-scale visuomotor patterns in an end-to-end manner. Then, we conducted a series of synthetic robotic experiments from reaching-and-grasping to imitation tasks to understand the underlying roles of visuomotor association in various tasks. The experimental results verified that the proposed model learned the tutored skills and generalized them to novel situations. The model developed and coordinated a set of cognitive skills including visual perception, working memory, action preparation and execution in a seamless manner. Furthermore, the proposed model was able to develop a predictive model of the world through predictive learning of visuo-proprioceptive patterns. Consequently, the model could perform mental simulation of action with given intention. In addition, the model was also able to infer intention behind observations by minimizing prediction error. To conclude, this dissertation illustrates a cognitive neurorobotics approach to building cognitive agents based on the key principles of the brain.