Dual-Arm Robotic Fabric Manipulation With Quasi-Static and Dynamic Primitives for Rapid Garment Flattening

Abstract

Service robots across industrial, medical, and domestic environments face significant challenges in flattening garments, due to their nearly infinite degrees of freedom and complex dynamic properties. Traditional quasi-static methods for handling fabrics require extensive interactions and are inefficient. While high-velocity dynamic actions enable efficient unfolding, they are unable to perform fine-grained adjustments. To address this, we propose a self-supervised learning framework combining dynamic (Fling) and static (pick and place and pick and drag) manipulation primitives using a dual-arm robot. This method unfolds crumpled garments efficiently without expert demonstrations or hand-labeled data. We introduce a novel factorized reward function based on garment coverage, shape matching, and smoothness. In addition, we employ a spatial action maps network, leveraging a Vision Transformer, to accurately determine grasping points. Our approach significantly reduces interaction steps and improves fabric coverage. We conducted simulations and real-world experiments on 14 garments across six categories. Our approach quickly achieved over 80% coverage within just two actions on long sleeves, surpassing the state-of-the-art by about 9%, and reached 88% –92.9% coverage across all categories within five actions. Our zero-shot sim-to-real transfer has successfully been demonstrated on UR5 robots, proving the effectiveness of the model on various fabrics.