A proposal of extracting of motion primitives by analyzing tracked data of hand motion from human demonstration

This paper presents a proposal to achieve motion primitives in the execution of manipulation actions from human demonstration. Human hand motion contains the most important information in the execution of manipulation actions. We design a method by using Kinect sensor to capture hand motion of each demonstrated action with a three-color-marker glove. The hand motion tracking data is calculated with three types: hand 3D position, hand orientation and hand states. Then, these tracked data is segmented to extract the motion primitives which then are used for building robot program that executes the action. We categorize three types of motion primitives including translation, rotation and state changing. In this study, we combine segmentation techniques based on mean square velocity and the change of hand state to extract the primitives of translation and state changing in the execution of action 'pick a cup'. In the experiment, we implement our design for tracking hand motion and analyze the tracked data to confirm our proposed segmentation techniques.