Acquiring manipulation skills through observation

Abstract

Currently, most robot programming is done either by manual programming or using a teach pendant as part of the "teach-by-showing" method. Both of these methods have been found to have several drawbacks. To solve the problems, the assembly-plan-from-observation (APO) method was proposed, which has the capability of observing a human performing an assembly task, understanding the task, and subsequently generating a robot program to achieve the same task. This system, however, cannot observe a trajectory of a human performance. Necessary trajectories are generated from CAD models. Later, in order to overcome this problem, a direct observation method based on a trajectory of a human performance was proposed to project human trajectory to robot trajectory. Though its implementation is relatively easy, the system is susceptive against observation noise. This paper proposes a method to make the robust observation against noise using symbolic representations such as face contact transitions. The system divides the trajectory into small segments based on the face contact analysis, allocates an operation element, referred to as sub-skill, to those segments. By using this system, we can decompose large motion templates, employed in the previous system, into sets of smaller sub-skills.