HM3D

该数据集包含了一系列照片级真实的环境，其中含有丰富的、可通行的、模拟的室内3D扫描数据，这些数据是从现实世界中的建筑物收集而来的。该数据集为可导航区域贡献了超过15万平方米的空间，这个面积显著大于MP3D的训练场景。该数据集的规模包括800个训练场景，旨在支持视觉与语言导航任务。

This dataset includes a series of photorealistic environments, featuring abundant, traversable simulated indoor 3D scan data collected from real-world buildings. It provides over 150,000 square meters of navigable space, which is significantly larger than the training scenes of MP3D. Comprising 800 training scenes, this dataset is designed to support vision-and-language navigation tasks.