Behavior-Based Segmentation of Demonstrated Tasks

Robot learning from demonstration presents several challenges. Given a task demonstration, the robot must sense, understand, and learn appropriate task attributes. We propose a method for automatic segmentation of a complex demonstration into an ordered set of simpler behaviors. These behaviors present a significantly less complex domain for learning. Our method is based on empirically derived attributes of tasks and simple mathematical transformations that results in a fast and intuitive mechanism for automatic task segmentation. The method is validated on a simulated Pioneer 2 DX platform across four demonstrations that vary in environmental complexity, task complexity, and task strategy.