Real Embodied Dataset (RED)

Real Embodied Dataset (RED)是由上海交通大学计算机科学与工程系创建的大型多视角数据集，旨在解决机器人视觉系统在杂乱场景中抓取物体时的挑战。该数据集包含31000个对齐的高质量RGB-D图像，采集自173个杂乱场景，涵盖17种家用物品，并配有相应的相机姿态和手工注释的模态分割掩码。RED不仅支持视觉推理，还支持决策制定，适用于机器人抓取任务的训练和评估。通过密集采样上部半球的视角，RED能够为机器人提供真实的视觉反馈，从而提高模型在未知杂乱场景中的适应性和抓取成功率。

Real Embodied Dataset (RED) was developed by the Department of Computer Science and Engineering, Shanghai Jiao Tong University, as a large-scale multi-view dataset designed to address the challenges encountered by robotic vision systems when grasping objects in cluttered scenes. This dataset contains 31,000 aligned high-quality RGB-D images collected from 173 cluttered scenes, covering 17 household items, and is accompanied by corresponding camera poses and manually annotated modal segmentation masks. RED supports both visual reasoning and decision-making, and is suitable for the training and evaluation of robotic grasping tasks. By densely sampling viewpoints from the upper hemisphere, RED can provide realistic visual feedback for robots, thereby improving the adaptability and grasping success rate of models in unknown cluttered scenes.