ccwatson/metaworld_bin_transfer_baby_red_red_high_low_noise005_50eps
收藏资源简介:
这是一个小型LeRobot格式的MetaWorld数据集,用于双任务婴儿版本的bin-transfer基准测试。场景采用完整bin-transfer套件使用的三托盘设置:灰色源托盘包含一个红色块和一个蓝色干扰块,红色和蓝色目标托盘位于桌子上。该数据集仅要求将红色块移动到红色箱子,语言指令仅变化运输高度约束(高或低)。目的是测试策略在对象颜色和目标颜色固定的情况下,是否能区分高与低运输行为。数据集比完整的12任务bin-transfer数据集更小更简单,包含100个episodes、18032帧、2个任务/提示,所有episodes均通过最终验证。数据格式为LeRobot v2.1风格parquet episodes,兼容OpenPI MetaWorld LeRobot路径。包含两个摄像头视角(corner4.image和gripperPOV.image,均为224x224 RGB),状态和动作维度分别为[4]和[39]、[4]。帧率为24fps,数据集分割为train: 0:100。数据集用于快速的高vs低运输实验,适用于冒烟测试、提示风格测试、BAKU/OpenPI数据管道测试以及扩展到完整bin-transfer套件前的小型LoRA运行。
This is a small LeRobot-format MetaWorld dataset for a two-task baby version of the bin-transfer benchmark. The scene is the same three-tray setup used by the full bin-transfer suite: a gray source tray contains a red block and a blue distractor block, with red and blue destination trays on the table. This dataset only asks for the red block to be moved to the red bin, with language varying only the carriage-height constraint (high or low). The purpose is to isolate whether a policy can distinguish high versus low transport behavior when object color and destination color are held fixed. It is intentionally much smaller and simpler than the full 12-task bin-transfer dataset. It contains 100 episodes, 18032 frames, 2 tasks/prompts, with all episodes passing final validation. The format is LeRobot v2.1 style parquet episodes, compatible with the OpenPI MetaWorld LeRobot path. Includes two camera views (corner4.image and gripperPOV.image, both 224x224 RGB), state and action dimensions of [4] and [39], [4] respectively. Frame rate is 24fps, dataset split is train: 0:100. The dataset is intended for quick high-vs-low carriage experiments, useful for smoke tests, prompt-style tests, BAKU/OpenPI data plumbing, and small LoRA runs before scaling to the full 12-task bin-transfer suite.




