czxlovesu03/ArenaVlaSafety

ArenaVlaSafety数据集是Arena/VLABench语义安全演示的一个精选子集，包含多个任务，每个任务最多有50个SAFE1和SAFE2的演示。数据集提供了多视角RGB-D图像、点云、本体感觉和低级控制动作，存储为HDF5格式。每个演示还附带一个MP4视频文件用于快速检查。数据集涵盖了多种语义安全任务，如将餐具放入容器、安全使用微波炉、药品和食物的安全放置、垃圾与贵重物品的分类以及液体倒入安全容器等。每个HDF5文件包含一个或多个演示，每个演示包括自然语言指令、元信息、观察数据（如RGB和深度图像、点云、机器人状态等）和低级控制动作轨迹。

ArenaVlaSafety is a curated subset of Arena/VLABench semantic safety demonstrations. Each task contains up to 50 SAFE1 and SAFE2 demonstrations where available, paired with multi-view RGB-D, point clouds, proprioception, and low-level control actions stored in HDF5 format. Every episode also includes an MP4 video file for quick inspection. The dataset covers various semantic safety tasks such as placing utensils into containers, safely using a microwave, safely placing medicines and food, sorting trash and valuables, and pouring liquids into safe receptacles. Each HDF5 file stores one or more episodes, each containing natural language instructions, meta-information, observation data (e.g., RGB and depth images, point clouds, robot states), and low-level control action trajectories.