Decomposition of Movement Data of Quadruped Robot by Using Autoencoder

In this presentation, the gait patterns of a quadruped robot are analyzed by decomposing and integrating motions by means of feature quantity analysis based on machine learning from motion data. Currently, robot control is a complicated problem, and the development of a system that can realize various movements in a simple manner is required. Although many researchers have been attempting to generate new motions by using quadruped locomotion robots, the controller of which consists of central pattern generators, discussions about the generation of gait patterns remain open. Here, we decompose a gait pattern in latent space by means of auto-encoders based on behavioral approach in which behavior data are obtained online from all sensors and actuators.