Learning From Observation of Tabletop Knotting Using a Simple Task Model

摘要

This paper describes a method for teaching robots tabletop knotting tasks via a human demonstration. Simple teaching-and-playback approaches cannot be applied to rope manipulation tasks because the repeatability of a rope's shape is not high due to its deformation. In this study, we introduce the concept of unit manipulations that are basic operations in knotting procedures, to design a description method for tabletop knotting. The proposed method consists of two phases: a teaching phase and a reproduction phase. In the teaching phase, a human demonstrates a knotting procedure in front of a robot, and the robot interprets it as a sequence of unit manipulations. In the reproduction phase, the robot automatically reproduces the same knot from the sequence. The intermediate unit manipulation description enables this despite the rope's deformation and the kinematic differences between humans and robots. We evaluated the proposed method to confirm its effectiveness using a system with a dual-arm robot.