GEM-250K
收藏资源简介:
GEM-250K是一个用于具身智能的大规模视觉语言数据集,隶属于GEM(Generative-supervised Embodied vision-language Model)项目。该数据集旨在通过集成生成式监督(特别是深度图生成任务)来增强视觉语言模型在具身环境中的性能,弥合高级语义理解与低级空间物理知识之间的差距。数据集包含约250,800个样本,涵盖三种核心任务类型:时空规划、具身接地和空间推理。数据以两种配置组织:默认配置包括图像、问题、答案和问题类型字段,适用于图像-文本到文本任务;空间推理配置则包含视频、问题、答案和问题类型字段。所有数据均为英文,规模在100万到1,000万样本之间。该数据集支持视觉语言模型和视觉-语言-动作框架的训练与评估,特别适用于机器人技术中的接地、推理和规划等具身智能任务。
GEM-250K is a large-scale vision-language dataset for embodied intelligence, belonging to the GEM (Generative-supervised Embodied vision-language Model) project. This dataset aims to enhance the performance of vision-language models in embodied environments by integrating generative supervision, particularly depth map generation tasks, and bridge the gap between high-level semantic understanding and low-level spatial physical knowledge. The dataset contains approximately 250,800 samples, covering three core task types: spatio-temporal planning, embodied grounding, and spatial reasoning. The data is organized in two configurations: the default configuration includes fields such as image, question, answer, and question type, which is suitable for image-text to text tasks; the spatial reasoning configuration contains video, question, answer, and question type fields. All data is in English, with a scale ranging from 1 million to 10 million samples. This dataset supports the training and evaluation of vision-language models and vision-language-action frameworks, and is particularly suitable for embodied intelligence tasks such as grounding, reasoning, and planning in robotics.
数据集概述
数据集名称:GEM-250K
发布机构:清华大学与腾讯混元
许可证:MIT
语言:英文
目的与应用
GEM-250K 是 GEM(Generative-supervised Embodied vision-language Model) 项目的一部分,旨在通过引入深度图生成任务作为视觉-语言模型预训练的监督信号,弥补标准文本引导预训练范式与具身智能所需低层空间、物理知识之间的鸿沟。该数据集支持具身智能中的语义理解与物理操作能力提升。
数据规模与组成
- 总下载大小:约 2.50 GB
- 总数据集大小:约 2.54 GB
- 样本数量:约 250,800 个(标注为 1M<n<10M 类别)
数据集包含两个配置(config):
-
default(默认配置)
- 特征:question_type(字符串)、image(图像)、question(字符串)、answer(字符串)
- 子集:
spatiotemporal_planning:49,800 个样本,约 666.69 MBembodied_grounding:100,000 个样本,约 1.87 GB
- 任务类型:图像-文本到文本
-
spatial_reasoning(空间推理配置)
- 特征:video(字符串)、question(字符串)、answer(字符串)、question_type(字符串)
- 子集:
spatial_reasoning:101,000 个样本,约 36.65 MB - 任务类型:图像-文本到文本(注意:特征中视频为字符串路径)
数据内容与任务
- 具身接地(embodied_grounding):包含图像及对应问题-答案对,用于地面实体与任务理解。
- 时空规划(spatiotemporal_planning):包含图像及规划类问答,涉及时间与空间推理。
- 空间推理(spatial_reasoning):以视频路径形式(字符串)提供视频文件,附带空间推理问答。
数据格式
-
文件格式:Parquet
-
目录结构:
GEM-250K/ ├── assets/ │ ├── GEM-demo.mp4 │ ├── overview.png ├── data/ │ ├── embodied_grounding-00000-of-00002.parquet │ ├── embodied_grounding-00001-of-00002.parquet │ ├── spatial_reasoning-00000-of-00001.parquet │ └── spatiotemporal_planning-00000-of-00001.parquet └── README.md
标签与分类
- 任务类别:image-text-to-text
- 标签:hunyuan、vision-language-model、vision-language-action、embodied、Spatial-Intelligence
引用
如需引用该数据集,请参考以下 BibTeX 格式:
@article{zhao2026gem, title={GEM: Generative Supervision Helps Embodied Intelligence}, author={Zhao, Ruowen and Li, Bangguo and Liu, Zuyan and Liang, Yinan and Ye, Junliang and Liu, Fangfu and Wu, Diankun and Wang, Zhengyi and Yu, Xumin and Rao, Yongming and others}, journal={arXiv preprint arXiv:2605.28548}, year={2026} }




