Estimation of essential interactions to achieve a task by integrating demonstrations

To learn new everyday tasks, the system needs to detect which parts of the demonstration are essential to complete the task without task-dependent knowledge. In this paper, a novel technique is proposed, which can estimate almost any types of interactions, including stationary, periodic and dynamic interactions from multiple demonstrations. In this approach, a demonstrator needs to give an explicit signal once during each essential interaction. From visual information and these signals, the system automatically analyzes the essential parts of the task and their periods, and also detects which environmental objects are interacted with the manipulated object.