OmniDexDataGen
收藏资源简介:
OmniDexDataGen是由汉堡大学、Agile Robots SE和慕尼黑工业大学联合开发的语义丰富的灵巧抓取数据集生成框架。该数据集通过集成抓取分类法引导的配置采样、功能可供性接触点采样、分类法感知的微分力闭合抓取采样以及基于物理的优化和验证,系统覆盖了多样化的抓取类型。数据集内容包含多维度语义信息,如抓取分类法、接触结构和功能可供性,旨在为下游推理和生成任务提供强先验。其应用领域聚焦于机器人灵巧抓取生成,解决现有方法在语义多样性和可控性方面的不足,推动任务感知的机器人操作研究。
OmniDexDataGen is a semantically-rich dexterous grasping dataset generation framework jointly developed by the University of Hamburg, Agile Robots SE, and the Technical University of Munich. This framework systematically covers a diverse range of grasp types by integrating grasp taxonomy-guided configuration sampling, affordance-based contact point sampling, taxonomy-aware differential force-closure grasp sampling, as well as physics-based optimization and validation. The dataset contains multi-dimensional semantic information, including grasp taxonomies, contact structures, and affordances, aiming to provide strong priors for downstream inference and generation tasks. Its application scenarios focus on robotic dexterous grasp generation, addressing the shortcomings of existing methods in terms of semantic diversity and controllability, and advancing research on task-perceptive robotic manipulation.




