GarmentFolding:Learning Efficient Garment Folding Based on Lift&Drag Dynamic Action

In robotic arm manipulation, folding clothes is highly challenging due to the garments' inherent infinite degrees of freedom. Our designed system can unfold garments from any configuration and automatically center them in the workspace during folding, thereby minimizing constraints on the robotic arm's motion range. High-quality garment unfolding significantly enhances folding efficiency. Therefore, this study introduces a novel action strategy, “Lift&Drag,” which involves high-speed dynamic actions to adjust garments from any state to a prede-fined configuration, ensuring superior unfolding. To bridge the simulation-reality gap and improve fabric keypoint detection ac-curacy, we collected extensive real-world fabric images, creating a dataset of approximately 2,000 images. This dataset notably enhances the performance of fabric folding keypoint detection. Experimental results show that the method reliably performs in real-world settings, achieving 91.3 % unfolding coverage and a 92.8 % folding success rate, particularly with heavily wrinkled t-shirts.