Programming everyday task using primitive skills and generative model of movement demonstrated by human

Abstract

This manuscript describes a method to program everyday manipulation tasks for the robots from human demonstration by using a movement generating model and primitive skills. The hand movement in the demonstrated task is recorded by using a hand motion tracker with a Kinect camera and a color-marker glove. The recorded movement is segmented to sub-actions and then mapped to primitive skills which can be built independently with the task. To adapt with the change of the goal position, Dynamic Movement Primitives model was applied to generate movement trajectory which follows the demonstrated trajectory. In experiment, we considered the task “dispensing water” from a water thermos pot performed by a robot arm. We implemented DMPs model for the sub-movement such as ‘approaching object’ to confirm the adaptation to new goal of movement trajectory. We proposed a list of necessary common skills to use for accomplishing the demonstrated task and can be reused in other everyday manipulation tasks.