Rynn4DDataset 1.0
收藏资源简介:
Rynn4DDataset 1.0是由阿里巴巴达摩院等机构联合构建的大规模4D具身视频数据集,旨在为机器人操作提供几何与运动显式表达的训练基础。该数据集包含超过2.544亿视频帧,整合了以人为中心的第一视角活动视频与多样化机器人操作数据,并通过先进模型生成了高质量的深度与光流伪标注。其构建过程融合了多模态标注流水线,包括视频描述生成、光流估计与深度预测,确保了数据在时空维度上的一致性。该数据集主要应用于4D具身世界模型的训练,致力于解决开放世界机器人操作中场景动态预测与精确动作生成的难题,为具身智能的发展提供了关键的数据支撑。
Rynn4DDataset 1.0 is a large-scale 4D embodied video dataset jointly constructed by Alibaba DAMO Academy and other institutions. It aims to provide a training foundation with explicit geometric and motion representations for robotic manipulation. This dataset contains over 254.4 million video frames, integrates first-person activity videos centered on humans and diverse robotic manipulation data, and generates high-quality pseudo-annotations for depth and optical flow via advanced models. Its construction process integrates a multi-modal annotation pipeline, including video caption generation, optical flow estimation and depth prediction, to ensure the consistency of the data in both spatial and temporal dimensions. This dataset is mainly applied to the training of 4D embodied world models, and is committed to solving the challenges of scene dynamic prediction and precise action generation in open-world robotic manipulation, providing critical data support for the development of embodied intelligence.

- 1RynnWorld-4D: 4D Embodied World Models for Robotic Manipulation阿里巴巴集团·达摩院; 香港具身智能实验室; 香港中文大学; 湖畔实验室 · 2026年



