Dynamic Generation and Switching of Object Handling Behaviors by a Humanoid Robot Using a Recurrent Neural Network Model

摘要

The present study describes experiments on a ball handling behavior learning that is realized by a small humanoid robot with a dynamic neural network model, the recurrent neural network with parametric bias (RNNPB). The present experiments show that after the robot learned different types of behaviors through direct human teaching, the robot was able to switch between two types of behaviors based on the ball motion dynamics. We analyzed the parametric bias (PB) space to show that each of the multiple dynamic structures acquired in the RNNPB corresponds with taught multiple behavior patterns and that the behaviors can be switched by adjusting the PB values.