VITRA-TeleData
收藏资源简介:
VITRA遥操作数据集包含使用配备灵巧手和头戴式RGB摄像头的7自由度机械臂收集的真实世界机器人遥操作演示。每个演示提供同步的数值状态/动作数据和视频记录。该数据集用于微调项目VITRA(可扩展的视觉-语言-动作模型预训练,用于机器人操作与真实人类活动视频)。数据集中的每个演示由两个同步文件组成:一个HDF5文件包含机器人状态、动作、运动学和元数据,另一个MP4文件包含头戴式摄像头记录的RGB视频流。数据集支持单臂和双臂演示,当前发布的版本仅包含右臂数据。数据集的文件结构详细描述了元数据、运动学、观察、状态、动作和掩码的组织方式。此外,数据集还定义了多个坐标系和转换关系,包括arm_base、ee_urdf、hand_mount和head_camera坐标系,并提供了详细的姿态表示和转换矩阵说明。
The VITRA Teleoperation Dataset contains real-world robotic teleoperation demonstrations collected using 7-degree-of-freedom (7-DOF) robotic arms equipped with dexterous hands and head-mounted RGB cameras. Each demonstration provides synchronized numerical state/action data and video recordings. This dataset is used for fine-tuning the project VITRA: Scalable Vision-Language-Action Model Pretraining for Robotic Manipulation and Real Human Activity Videos. Each demonstration in the dataset consists of two synchronized files: one HDF5 file containing robotic states, actions, kinematics, and metadata, and another MP4 file containing the RGB video stream recorded by the head-mounted camera. The dataset supports single-arm and dual-arm demonstrations, while the currently released version only contains data for the right arm. The file structure of the dataset details the organization of metadata, kinematics, observations, states, actions, and masks. Additionally, the dataset defines multiple coordinate systems and transformation relationships, including the arm_base, ee_urdf, hand_mount, and head_camera coordinate systems, and provides detailed descriptions of pose representations and transformation matrices.




