An operation plane using a neural network for intuitive generation of robot motion

Abstract

An operation plane is developed that greatly simplifies the task of editing motion of humanoid robots. The plane uses onomatopoeias, which are words that mimic the appearance or sound of things to produce richly realistic expressions. To create the plane, first, features of known motions for which there are onomatopoeias are extracted by P-type Fourier descriptors. Second, the similarity relationship between the features is learned by using a five-layer auto-associative neural network. Finally, the network's third layer, which has two units, is used as an operation plane. Using this plane, even people who are unfamiliar with robotics can edit motion of humanoid robots intuitively.