Motion Generalization from a Single Demonstration Using Dynamic Primitives

In recent years, several studies have suggested that improved performance of modern robots can arise from encoding motor commands in terms of dynamic primitives. In this context, dynamic movement primitives (DMPs) have been proposed as a powerful tool for motion planning based on demonstrated examples. In this work, we focus on generalizing discrete and periodic movements from a single demonstration. Here, we argue that geometric invariance in itself may be useful to provide an initial representation of movements in an incremental process of learning from experience. The purpose of the current study is to portray the generalization performance of this approach, both using simulated and human motion capture data. The generalization performance is evaluated and the feasibility of the approach is discussed.