深層学習を用いた実ロボットの反射動作学習

日立-早大の共同研究開発事例

抄録

A motion generation method using deep learning has been proposed that is robust against various environmental changes. In this paper, we describe a reflection motion method that responds immediately to environmental changes. The reflection motion can be developed using a simple feed-forward neural network. The tasks of verifying the reflex movement were ”grip force control” that grips various objects without crushing them, and ”obstacle avoidance” that avoids unknown objects during autonomous movement. Results of an experiment using a real robot confirmed a reflection motion could be generated.