Add dataset card content, task category, and paper link

README.md

@@ -1,4 +1,23 @@
 ---
 license: mit
 pretty_name: DGS-Net
----
+task_categories:
+- robotics
+---
+
+# DGS-Net: Learning Dexterous Grasps from Single-View Point Clouds via a Multi-Object Scene Dataset
+
+This dataset accompanies the paper [DGS-Net: Learning Dexterous Grasps from Single-View Point Clouds via a Multi-Object Scene Dataset](https://huggingface.co/papers/2603.15410).
+
+It contains 307 objects, 240 multi-object scenes, and over 350,000 validated grasps for dexterous grasping in multi-object scenes. The dataset includes scene-level RGB-D images, segmentation masks, camera poses, object visibility information, and grasp annotations with pre-grasp and final joint configurations.
+
+- **Project page:** [https://4taotao8.github.io/DGS-Net/](https://4taotao8.github.io/DGS-Net/)
+- **Code:** [https://github.com/4taotao8/DGS-Net](https://github.com/4taotao8/DGS-Net)
+
+## Dataset Structure
+
+The dataset is organized into `scenes/` and `grasp_labels/` folders. The `scenes/` directory contains multi-object scene data, while `grasp_labels/` contains object-level grasp records. See the [GitHub repository](https://github.com/4taotao8/DGS-Net) for detailed download and usage instructions.
+
+## License
+
+The dataset is released under the MIT License. Please respect the rights of third-party assets as noted in the repository.