The Geometric Nature of Deformable Object Manipulation

Ziel

Manipulating deformable objects, such as cables, garments, bags, sponges, or food items, is an essential and recurring part of human daily routines. We interact with such objects with seamless ease, usually unaware of the complex perception, planning, and control machinery naturally provided by our brain and body to do so. In contrast, the ability to robustly manipulate deformable objects remains a critical frontier in robotics, often being considered as one of the most important challenges in robotic manipulation.

The tenet of GEOMANIP is that the key to a revolution in robotic manipulation of deformable objects will come from rethinking their representations to uncover and exploit their fundamental geometric nature. GEOMANIP proposes a paradigm shift that integrates the differential geometric structure of deformable objects into novel continuum physics-inspired neural representations. Specifically, we will build upon the differential geometry underlying the governing principles of continuum mechanics to design a new class of physics-inspired neural networks that generalize across a broad spectrum of deformable object types, structures, and properties. These representations will serve as the bedrock of a novel physically-grounded manipulation pipeline including perception, policy learning, and control, providing robots with advanced abilities to autonomously manipulate a wide variety of deformable objects. By uncovering the geometric nature of deformable object manipulation, GEOMANIP will reconcile the accuracy and generalizability of continuum mechanics with the expressivity and versatility of data-driven deep learning architectures, thus overcoming one significant and long-standing barrier in robotics.