TASTE-Rob
收藏资源简介:
TASTE-Rob是由香港科技大学等机构创建的大型数据集,包含100,856个针对任务导向的手-物交互视频。这些视频均以静态摄像机视角录制,每个视频对应一个语言任务指令,确保动作与指令的精确对应。数据集覆盖了多种环境和任务类型,包括厨房、餐桌、办公室等场景,以及拾取、放置、推动等多种手-物交互任务。TASTE-Rob的设计旨在为视频生成模型提供高质量的训练数据,进而改善机器人模仿学习中的视频演示质量,提升机器人操作的泛化能力。
TASTE-Rob is a large-scale dataset developed by The Hong Kong University of Science and Technology and other institutions. It contains 100,856 task-oriented hand-object interaction videos, all recorded from a static camera perspective. Each video is paired with a corresponding linguistic task instruction, ensuring precise alignment between the executed actions and the instructions. The dataset covers diverse environments and task types, including scenarios such as kitchens, dining tables, offices, and various hand-object interaction tasks like picking up, placing, and pushing. TASTE-Rob is designed to provide high-quality training data for video generation models, thereby improving the quality of video demonstrations in robot imitation learning and enhancing the generalization ability of robotic manipulation.




