We have seen much recent progress in task-specific clothes manipulation, but generalizable clothes manipulation is still a challenge. Clothes manipulation requires sequential actions, making it challenging to generalize to unseen tasks. Besides, a general clothes state representation method is crucial. In this paper, we adopt language instructions to specify and decompose clothes manipulation tasks, and propose a large language model based hierarchical learning method to enhance generalization. For state representation, we use semantic keypoints to capture the geometry of clothes and outline their manipulation methods. Simulation experiments show that the proposed method outperforms the baseline method in terms of success rate and generalization for clothes manipulation tasks.

翻译：近年来，任务特定的衣物操作已取得显著进展，但可泛化的通用衣物操作仍面临挑战。衣物操作需要序列化动作，这使其难以泛化至未见任务。此外，通用的衣物状态表征方法至关重要。本文采用语言指令来指定和分解衣物操作任务，并提出一种基于大语言模型的分层学习方法以增强泛化能力。对于状态表征，我们使用语义关键点来捕捉衣物的几何结构并勾勒其操作方法。仿真实验表明，所提方法在衣物操作任务的成功率与泛化性能方面均优于基线方法。