A Developmental Robotic Paradigm Using Working Memory Learning Mechanism

Traditional machine learning algorithms typically employ task specific methods and only the parameters predetermined by the human programmer are updated. These methods often fail to respond to the dynamically changing states of the uncontrolled environments. Additionally, such methods may not represent a developmental entity, such as a human mind. In contrast, an open-ended developmental robot system can learn simple behaviors and build up more complex behaviors by utilizing the previously learned behaviors. In this paper, we propose a basic framework for visual learning tasks that integrates a perceptual system into a biologically inspired working memory system. A main objective of this research is to provide a general framework for developmental learning and to investigate how well a neuro-computational PFC working memory model performs on a robotic platform in a real world environment with complex tasks. Experimental results given show impressive performance both in terms of accuracy and speed of learning.