DexGarmentLab
收藏资源简介:
DexGarmentLab是一个专门为灵巧(尤其是双手)衣物操作设计的模拟环境。该环境包括15个任务场景,每个场景都有高质量的3D资产,以及2500多件来自ClothesNet数据集的衣物。DexGarmentLab通过利用衣物结构对应关系,仅用一个专家演示自动生成具有多样化轨迹的数据集,从而显著减少了手动干预的需求。此外,该环境还引入了层次化衣物操作策略(HALO),利用可供性(用于定位衣物操作区域)和扩散方法(基于衣物和场景生成轨迹),实现了比现有模仿学习算法更好的泛化性能。该数据集可用于研究灵巧双手操作衣物的算法,旨在解决现实世界中衣物操作的挑战。
DexGarmentLab is a simulation environment specifically designed for dexterous garment manipulation, especially dual-hand garment manipulation. This environment features 15 task scenarios, each equipped with high-quality 3D assets, and over 2500 garment items sourced from the ClothesNet dataset. DexGarmentLab leverages the structural correspondences of garments to automatically generate datasets with diverse trajectories using only a single expert demonstration, thereby significantly reducing the need for manual intervention. Additionally, this environment introduces the Hierarchical Garment Manipulation Strategy (HALO), which leverages affordances for locating garment manipulation regions and diffusion-based methods for generating trajectories conditioned on garments and the environment, achieving better generalization performance than existing imitation learning algorithms. This dataset can be used to research algorithms for dexterous dual-hand garment manipulation, aiming to address real-world garment manipulation challenges.




