Abstract
Personal service robots are expected to help people in various scenes such as caring, daily support, and construction. Such assistive robots have to have an ability to collaborate with people. For a human and a robot to perform a collaborative work, it is necessary to teach the robot how to collaborate with the human. Such a teaching should be easy and accurate. A promising approach to teaching is Programming by Demonstration (PbD), in which humans demonstrate a task and a robot learns the task by observing the demonstration. In this paper, we develop a PbD system for collaborative assembly tasks. The target tasks we choose are assembly of a table or a chair. The system observes collaborative assembly by two person by using four RGB-D cameras. The system records RGB and depth images with the person skeleton data. All data are converted to a common world coordinate system. We use these data to detect movements of human hands and assembly parts, and then recognize actions performed.