Tr0612/ShortMetaWorld

Short-MetaWorld-VLA (v2+v3)数据集是一个用于视觉语言机器人技术和多任务强化学习的短MetaWorld集合。它包含24个任务文件（12个v2版本和12个v3版本），每个任务有100个轨迹，每个轨迹有20或50个步骤，总计84,000个步骤样本。每个步骤的数据包括RGB图像、39维浮点向量状态、4维浮点向量动作、任务语言指令和任务标识符。数据集结构包括图像目录和r3m处理后的数据文件。任务包括篮球、按钮按下、门打开、抽屉打开/关闭、peg插入侧面、拾取放置、推动、到达、清扫、窗口打开/关闭等。该数据集适用于多任务强化学习、模仿学习、视觉语言机器人技术、元学习和机器人策略训练等研究应用。

The Short-MetaWorld-VLA (v2+v3) dataset is a short MetaWorld collection designed for vision-language robotics and multi-task reinforcement learning. It includes 24 task files (12 v2 + 12 v3), with 100 trajectories per task and 20 or 50 steps per trajectory, totaling 84,000 step samples. Each step contains an RGB image, a 39D float vector state, a 4D float vector action, a task language instruction, and a task identifier. The dataset structure comprises image directories and r3m-processed data files. Tasks include basketball, button-press-topdown, door-open, drawer-open/close, peg-insert-side, pick-place, push, reach, sweep, window-open/close, among others. This dataset is suitable for research applications such as multi-task reinforcement learning, imitation learning, vision-language robotics, meta-learning, and robot policy training.