Analysis method without task model for interpreting human demonstration

Abstract

Robots have been controlled by various methods. Teaching robot task from human demonstrations has been proposed to reduce robot programming cost. Some prototype systems showed that they are valid when task models are given in advance to analyze human demonstrations. Interpreting demonstration data without task dependent model is required to expand their applicability. This paper discusses the task independent methods of analyzing data. First, demonstration data are matched by minimizing distance between patterns using translational and rotational displacement. Next, data is wavelet analyzed to detect points where motions changed. The result of segmentation is discussed for key-latch task.