A Hand Image Instruction Learning System Using Transient-SOM

摘要

This paper presents a practical hand image instruction learning system for partner robots using an advanced selforganizing map“Transient-SOM”. The construction of the proposal system is designed with 4 layers: a feature map to classify input images, an action map to acquire correct actions using a reinforcement learning algorithm, a memory layer to preserve “Best Matching Unit (BMU)” after one instruction training is finished, and a feeling map to express robot's internal state during learning. Experiments using an entertainment robot, AIBO (ERS-7, Sony Co., 2003), showed high effectiveness of the proposed system.