Add dataset description, task category, and links

README.md

@@ -1,3 +1,17 @@
 ---
 license: mit
+task_categories:
+- robotics
 ---
+
+# RoboSPA
+
+RoboSPA (**Robo**t **S**patial-**P**rocedural **A**ssessment) is a large-scale robotic manipulation dataset and benchmark for diagnosing embodied reasoning in Vision-Language-Action (VLA) models.
+
+This dataset focuses on two core dimensions: **Fine-Grained Spatial Reasoning** and **Long-Horizon Procedural Planning**, covering 10 task categories and 56 base tasks, with each task instantiated across five difficulty levels (280 variants). It includes 527K trajectories across multiple embodiments and diverse scenes.
+
+For more details, please refer to:
+
+- Paper: [RoboSPA: Can VLA Models Go Beyond Simple Scenes and Short-Horizon Tasks?](https://huggingface.co/papers/2609.05324)
+- Project page: https://fanzhenxuan.github.io/RoboSPA
+- Code: https://github.com/fanzhenxuan/RoboSPA