10Kh-RealOmin-OpenData
收藏资源简介:
该数据集是目前业内规模最大的开源具身智能数据集,累计超过 10,000 小时真实交互数据与 100 万段以上任务片段,覆盖 3,000 余个真实家庭环境和近 10,000 个细粒度操作目标,显著提升模型在开放世界场景下的泛化能力。数据来自自然人类操作过程,围绕 10 类家庭场景与 30 项核心技能系统构建,避免技能碎片化与重复,并完整记录双臂协同与长时序复杂任务的全过程行为。数据包含大视场鱼眼图像、精确轨迹、关节与末端执行器状态、IMU 与触觉等多模态信息,并通过统一标定与轨迹重建保证时空对齐与测量精度。当前已发布 Stage 1 版本,覆盖 4 个场景下的 12 项技能,总时长约 950 小时、约 3.45TB,采用 mcap 格式存储并配套提供格式说明、可视化工具与加载接口,支持具身智能与机器人操作学习等研究与应用。
This dataset is the largest open-source embodied intelligence dataset currently available in the field. It contains over 10,000 hours of real-world interactive data and more than 1 million task segments, covering more than 3,000 real home environments and nearly 10,000 fine-grained manipulation targets, which significantly enhances the generalization capability of models in open-world scenarios. The data is collected from natural human manipulation processes, constructed based on 10 types of home scenarios and 30 core skill systems to avoid skill fragmentation and redundancy, and fully records the entire behavioral process of dual-arm collaboration and long-sequence complex tasks. The dataset includes multimodal information such as wide-field fisheye images, precise trajectories, joint and end-effector states, IMU and tactile data, and ensures spatiotemporal alignment and measurement accuracy through unified calibration and trajectory reconstruction. Currently, the Stage 1 version has been released, covering 12 skills across 4 scenarios, with a total duration of approximately 950 hours and a size of about 3.45 TB. It is stored in mcap format, and is accompanied by format specifications, visualization tools and loading interfaces, supporting research and applications such as embodied intelligence and robotic manipulation learning.




