Add model card and metadata for StableVLA

This PR adds a model card for StableVLA, a robust Vision-Language-Action model.
It includes:
- Metadata with the `robotics` pipeline tag.
- Links to the [ICML 2026 paper](https://huggingface.co/papers/2605.18287), the project page, and the GitHub repository.
- A description of the model and its Information Bottleneck Adapter (IB-Adapter).
- A usage section demonstrating how to download the model weights using `huggingface_hub` as specified in the official documentation.
- Citation information.

README.md

@@ -0,0 +1,54 @@
+---
+license: apache-2.0
+pipeline_tag: robotics
+tags:
+- robotics
+- vla
+- vision-language-action
+- robust-ai
+---
+
+# StableVLA: Towards Robust Vision-Language-Action Models without Extra Data
+
+StableVLA is a robust Vision-Language-Action (VLA) model designed to maintain performance when encountering unseen real-world visual disturbances. It introduces the **Information Bottleneck Adapter (IB-Adapter)**, a lightweight module grounded in information theory that selectively filters noise from visual inputs.
+
+[**Paper**](https://huggingface.co/papers/2605.18287) | [**Project Page**](https://dagroup-pku.github.io/StableVLA/) | [**GitHub**](https://github.com/DAGroup-PKU/HumanNet/tree/main/src/model/StableVLA)
+
+## Model Description
+
+Current VLA models often suffer significant performance drops under visual disturbances not seen during training. StableVLA addresses this by integrating the IB-Adapter, which:
+- **Filters Noise:** Selectively removes potential visual noise with negligible parameter overhead (<10M).
+- **Improves Robustness:** Consistently improves over baselines by an average of 30% without extra training data or augmentations.
+- **Efficiency:** Even with a 0.5B backbone, it achieves robustness competitive with 7B-scale state-of-the-art VLAs.
+
+## Usage
+
+You can download the model weights using the `huggingface_hub` library. Depending on your needs, you can download the full repository or specific task checkpoints (e.g., spatial, object, goal, or long).
+
+```python
+from huggingface_hub import snapshot_download
+
+# Download everything
+snapshot_download(repo_id="beikui12345/stablevla", local_dir="./hf_weights")
+
+# Or download a specific task checkpoint (e.g., spatial)
+snapshot_download(repo_id="beikui12345/stablevla", local_dir="outputs/spatial",
+                  allow_patterns="spatial/*")
+```
+
+For detailed instructions on environment setup, training, and evaluation, please refer to the [official GitHub repository](https://github.com/DAGroup-PKU/HumanNet/tree/main/src/model/StableVLA).
+
+## Citation
+
+If you find StableVLA helpful in your research, please cite the following paper:
+
+```bibtex
+@inproceedings{fu2026stablevla,
+  title     = {StableVLA: Towards Robust Vision-Language-Action Models without Extra Data},
+  author    = {Fu, Yiyang and Zhang, Chubin and Gong, Shukai and Deng, Yufan and
+               Sun, Kaiwei and Min, Qiyang and Hou, Qibin and Tang, Yansong and
+               Wang, Jianan and Zhou, Daquan},
+  booktitle = {International Conference on Machine Learning (ICML)},
+  year      = {2026},
+}
+```