meituan/LIBERO-X
收藏Hugging Face2026-04-29 更新2026-03-29 收录
下载链接：
https://hf-mirror.com/datasets/meituan/LIBERO-X
下载链接
链接失效反馈官方服务：
资源简介：
LIBERO-X数据集通过更细粒度的任务级扩展，使模型能够接触到多样化的任务表述和工作空间配置，包括2,520个演示、600个任务和100个场景，确保在各种场景下的广泛泛化能力。特点包括：
- **多任务场景设计：** 每个场景平均包含6个不同的任务，相比原始LIBERO数据集的平均2.6个任务有显著增加，支持更复杂和真实的多目标学习。
- **属性条件操作：** 动作明确依赖于细粒度的对象属性（如大小、颜色、纹理），而不仅仅是广泛的类别。
- **空间关系推理：** 任务不仅要求目标定位，还需要理解和推理对象之间的空间关系，包括左/右、前/后和近/远。
- **人类演示收集：** 所有轨迹都是通过使用Meta Quest 3的VR遥操作由人类收集的。
LIBERO-X introduces finer-grained task-level extensions to expose models to diverse task formulations and workspace configurations, includeing 2,520 demonstrations, 600 tasks, and 100 scenes, ensuring broad generalization across diverse scenarios, featuring:
- **Multi-Task Scene Design:** Each scene averages 6 distinct tasks, a significant increase compared to the original LIBERO dataset’s average of 2.6 tasks per scene, enabling more complex and realistic multi-objective learning.
- **Attribute-Conditioned Manipulation:** Actions are explicitly conditioned on fine-grained object properties (e.g., size, color, texture) beyond broad categories.
- **Spatial Relationship Reasoning:** Tasks extend beyond target localization to require understanding and reasoning about spatial relationships among objects, including left/right, front/back, and near/far.
- **Human Demonstration Collection:** All trajectories were human-collected via VR teleoperation using a Meta Quest 3.
提供机构：
meituan
