lance-format/lerobot-pusht-lance

LeRobot PushT（Lance格式）是一个机器人学数据集，源自Diffusion Policy论文中的标准PushT基准。该数据集以Lance格式打包，采用三表布局（frames、episodes、videos），便于帧级训练、轨迹级工作和原始视频访问。关键特性包括内联MP4视频段、懒加载的二进制大对象处理、稳定的索引列以支持连接和时间迭代，以及模式演化友好性，允许后续添加新列而不重写数据。数据集包含机器人状态向量、动作向量、时间戳和视频元数据，适用于机器人策略学习、世界模型训练等任务。

LeRobot PushT (Lance format) is a robotics dataset derived from the standard PushT benchmark in the Diffusion Policy paper. This dataset is packaged in Lance format and adopts a three-table schema (frames, episodes, videos), which enables frame-level training, trajectory-level data processing, and direct access to raw videos. Key features include inline MP4 video segments, lazy-loaded binary large object (BLOB) handling, stable indexed columns that support table joins and temporal iteration, and schema evolution compatibility, which allows subsequent addition of new columns without overwriting existing data. The dataset contains robot state vectors, action vectors, timestamps, and video metadata, and is suitable for tasks such as robotic policy learning and world model training.