Recognizing Assembly Tasks Through Human Demonstration

As one of the methods for reducing the work of programming, the Learning-from-Observation (LFO) paradigm has been heavily promoted. This paradigm requires the programmer only to perform a task in front of a robot and does not require expertise. In this paper, the LFO paradigm is applied to assembly tasks by two rigid polyhedral objects. A method is proposed for recognizing these tasks as a sequence of movement primitives from noise-contaminated data obtained by a conventional 6 degree-of-freedom (DOF) object-tracking system. The system is implemented on a robot with a real-time stereo vision system and dual arms with dexterous hands, and its effectiveness is demonstrated. KEY WORDS—assembly planning, learning from observation, movement primitives