libero-mem/LIBERO-Mem

LIBERO-Mem是一个针对机器人操作的长时间跨度的以对象为中心的数据集。它包含了任务级别的元数据和同步的RGB-D观测、分割图、本体感知以及控制动作的演示数据。数据集适用于模仿学习和强化学习，特别是针对厨房操作任务。元数据文件为metainfo.json，包含了任务的初始状态、描述、关键物体名词以及边界框信息。演示数据存储在HDF5格式的文件中，包含了执行任务时的动作数据、完成标志以及不同视角的观测数据。

LIBERO-Mem is a long-horizon object-centric dataset for robot manipulation. It includes task-level metadata and synchronized RGB-D observations, segmentation maps, proprioception, and control action demonstrations. The dataset is suitable for imitation learning and reinforcement learning, particularly for kitchen manipulation tasks. The metadata is stored in a metainfo.json file, which contains the initial state, description, key object nouns, and bounding box information of the tasks. Demonstration data is stored in HDF5 files, which include action data, completion flags, and observations from different perspectives during task execution.