Omnipush

Omnipush数据集由麻省理工学院创建，旨在提供一个高多样性的平面推动行为数据集。该数据集包含250个对象，每个对象进行250次推动，总计62500条记录，所有推动均通过RGB-D视频和高精度跟踪系统记录。数据集设计用于系统探索影响推动的关键因素，如物体形状和质量分布，并允许研究模型学习的泛化能力。Omnipush数据集包括动态模型元学习基准，需要算法进行准确预测并估计自身不确定性。此外，数据集还提供了RGB视频预测基准，并提出了其他相关任务，适用于该数据集。

The Omnipush dataset, developed by the Massachusetts Institute of Technology (MIT), aims to provide a high-diversity planar pushing behavior dataset. It comprises 250 distinct objects, with 250 pushing trials conducted for each object, resulting in a total of 62,500 records. All pushing operations are recorded using RGB-D videos and a high-precision tracking system. This dataset is designed to systematically investigate key factors influencing planar pushing, such as object shape and mass distribution, and to support research on the generalization capabilities of learned models. The Omnipush dataset includes a dynamic model meta-learning benchmark, which requires algorithms to generate accurate predictions and estimate their own uncertainty. Additionally, the dataset provides an RGB video prediction benchmark and proposes other relevant tasks applicable to this dataset.