Intuitive Knowledge Representations for Interactive Robot Programming

Abstract

Programming by Demonstration (PbD) is an intuitive method to transfer knowledge from a non-expert human teacher to a robot. To allow the non-expert user to intuitively understand what the robot has learned from the demonstration, we propose a framework that detects online which skills the human is demonstrating and builds from that a graph that describes how the task is performed. The skill recognition is achieved by a segmentation algorithm that combines symbolic skill segmentation, which makes use of pre- and postconditions to identify skills, with data-driven segmentation, which uses Support Vector Machines to learn to classify the skills from data. The framework is thus able to detect force-based skills in addition to manipulation skills, to allow the flexible use of robots in assembly production lines. The intuitiveness of the framework is evaluated in a user study that compares the task graph representation of our framework to the time-line based representation of an existing PbD framework that does not make use of skill recognition.