jae0311/serving_ab

该数据集是LeRobot v3.0版本中用于OpenManipulator-X（OMX）机器人咖啡馆服务任务的数据集，名为serving_ab，合并了类别A（长片段）和类别B（原子演示），用于完整的GR00T N1.7微调计划。数据集包含425个片段（A: 265个，B: 160个），总计219,877帧，时长约2小时2分钟，帧率为30 fps。摄像头包括顶部和手腕视角（640×480分辨率，30 fps）。状态和动作维度均为6（单臂5维，夹爪1维），动作范围为16。任务包括三个：拾取盘子并放置到目标区域（任务索引0）、拾取杯子并放置到目标区域（任务索引1）、返回起始位置（任务索引2）。类别A涵盖多步骤拾取和放置场景，从起始位置开始并结束，包括单杯、单盘、双盘、双杯、一杯一盘、两杯一盘等场景；类别B专注于技能演示，如单个盘子和杯子的拾取放置、连续两个盘子、连续2-3个杯子、重试恢复等，以提高基础抓取可靠性和多对象处理能力。数据集的标签边界方案基于夹爪状态事件自动检测，用于划分多任务片段。数据集结构包括数据文件、元信息和视频文件。推荐训练计划包括两个阶段：首先使用serving_b数据集进行原子拾取放置可靠性训练，然后使用serving_ab数据集进行长片段性能训练。This dataset is the LeRobot v3.0 dataset for the OpenManipulator-X (OMX) café serving task, named serving_ab, which merges Category A (Long Episode) and Category B (Atomic Demos) for the full GR00T N1.7 fine-tuning schedule. It contains 425 episodes (A: 265, B: 160), with a total of 219,877 frames, approximately 2 hours and 2 minutes in duration at 30 fps. Cameras include top and wrist views (640×480 resolution, 30 fps). State and action dimensions are 6 each (single_arm × 5, gripper × 1), with an action horizon of 16. Tasks include three: pick up plate and place at target zone (task index 0), pick up cup and place at target zone (task index 1), and return to home (task index 2). Category A covers multi-step pick-and-place scenarios starting and ending at the home pose, including one cup, one plate, two plates, two cups, one cup + one plate, two cups + one plate, and empty scenarios. Category B focuses on skill demonstrations such as plate single pick-and-place, cup single pick-and-place, sequential two plates, sequential 2-3 cups, and retry recovery, to strengthen foundational grip reliability and multi-object handling. The label boundary scheme segments multi-task episodes based on gripper state events. The dataset structure includes data files, metadata, and video files. A recommended training schedule involves two stages: first, using serving_b for atomic pick-and-place reliability, then using serving_ab for long-episode performance from the Stage 1 checkpoint.