Active Boundary Component Models for robotic dressing assistance

Abstract

The dynamics of deformable objects, especially that of highly flexible articles of clothing, is difficult to model. This is due to their vast number of degrees of freedom in addition to the noisy and incomplete measurements robots have to cope with. Therefore, we suggest focusing on the structures and object parts which are relevant to the task at hand. The openings (e.g., at the waist, leg or sleeve ends) characterize garments surprisingly well, not only from a topological perspective, but also in terms of their inherent function, namely dressing. We model openings as closed, oriented chains of movable points which we refer to as Active Boundary Component Models (ABCMs). Compared with the hardly predictable motions of an overall piece of clothing, relatively strict assumptions regarding the dynamics of these contour models can be made. We express these assumptions through position-based constraints which drastically restrict the degrees of freedom. In the present paper, we show how ABCMs can be initialized exploiting geometric prior knowledge of garments, and how they can be tracked visually using 3D point cloud data. Additionally, we consider the task of sliding a rod through a pant leg as a first step toward robotic dressing assistance for physically handicapped persons.