jellyho/droid_subsets_scooping

这是DROID 1.0.1数据集的一个技能过滤子集，专注于舀取（scooping）技能。该子集基于LeRobot格式，通过droid_filter.py脚本从原始DROID数据集中筛选出与舀取相关的任务，使用关键词（如scoop up、use the spoon to等）进行过滤。数据集包含300个episodes、147394帧，每个episode平均长度为491.3帧，成功率为1.0。保留了原始DROID数据中的动作字段（如gripper_position、cartesian_position等），文件结构遵循LeRobot风格，包括元数据文件、任务数据和episodes索引。数据集主要用于轻量级的技能条件微调实验，不包含视频文件。

This is a skill-filtered subset of the local DROID 1.0.1 LeRobot-format dataset, focusing on the scooping skill. It was generated with the droid_filter.py script, filtering episodes based on skill keywords (e.g., scoop up, use the spoon to). The subset includes 300 episodes, 147394 frames, with a mean episode length of 491.3 frames and a success rate of 1.0. Action columns from the original DROID dataset (e.g., action.gripper_position, action.cartesian_position) are preserved. The repository maintains the LeRobot-style layout with metadata, task data, and episode indices. It is intended for lightweight skill-conditioned fine-tuning experiments and does not include video files.