Heterogeneous Embodied Data Sources

该数据集包含了来自28种不同实体的一百万个演示数据，整合了多个开源机器人集合，并标准化为包含第三人称视觉观察和语言指令，同时保持了动作的异质性。该数据集允许在异构数据源上进行训练，以探索通用动作空间，从而促进跨实体的泛化能力。规模上，它涵盖了来自28种实体的百万级演示。任务上，该数据集用于训练和评估Uniact框架，以提升各种实体在执行性能上的表现。

This dataset contains one million demonstration samples from 28 distinct entities. It integrates multiple open-source robotic datasets and is standardized to include third-person visual observations and language instructions while preserving action heterogeneity. This dataset enables training on heterogeneous data sources to explore universal action spaces, thereby facilitating cross-entity generalization capabilities. In terms of scale, it covers one million demonstration samples across 28 distinct entities. For task applications, this dataset is used to train and evaluate the Uniact framework to improve the execution performance of various entities.