Feature Extraction and Generation of Robot Writing Motion Using Encoder-Decoder Based Deep Neural Network

Recent improvement of control technologies allows robots to precisely follow given commands. Robots will be able to expand their executable tasks by saving the motion data that humans demonstrated beforehand. However, it is difficult to deal with the saved motion data when the number of the data increases. In this paper, we focus on writing task and propose an encoder-decoder based neural network for generating time series of motion as a framework for giving motion command directly to the robot. Motions of the robots can be represented as time series data of position data, which contains features of the motions. We can get low dimensional expression of the motion data that is called latent variables by training the network using the saved motion data. We can deal with and generate the saved motion data by using the latent variables and the decoder network. In the experiments, we collected data of writing a Kanji, trained the network using the saved data. We experimentally validated the generated data from the trained network by giving it to the robot.