Techniques to Control Robot Action Consisting of Multiple Segmented Motions using Recurrent Neural Network with Butterfly Structure

In the field of robot control, there have been several studies on humanoid robots operating in remote areas. We propose a methodology to control a robot using input from an operator with fewer degrees of freedom than the robot. This method is based on the concept that time-continuous actions can be segmented because human intentions are discrete in the time domain. Additionally, machine learning is used to determine components with a high correlation to input data that are often complex or large in quantity. In this study, we implemented a new structure on a conventional neural network to manipulate a robot using a fast Fourier transform. The neural network was expected to acquire robustness for amplitude and phase variations. Thus, our model can reflect a fluctuating operator input to control a robot. We applied the proposed neural network to manipulate a robot and verified the validity and performance compared with traditional models.