Add model card, robotics pipeline tag, and links to paper/code

README.md

@@ -1,3 +1,84 @@
----
-license: cc-by-nc-4.0
----
+---
+license: cc-by-nc-4.0
+pipeline_tag: robotics
+---
+
+# StereoVLA: Enhancing Vision-Language-Action Models with Stereo Vision
+
+StereoVLA is the first Vision-Language-Action (VLA) model to leverage rich geometric cues from stereo vision. It employs a Geometric-and-Semantic (GeoSem) vision encoder that extracts geometric cues from subtle stereo-view disparities for precise spatial perception, while simultaneously capturing semantic features from pixel observations to support language-conditioned manipulation.
+
+- **Paper:** [StereoVLA: Enhancing Vision-Language-Action Models with Stereo Vision](https://huggingface.co/papers/2512.21970)
+- **Project Page:** [StereoVLA Webpage](https://shengliangd.github.io/StereoVLA-Webpage/)
+- **Repository:** [GitHub - shengliangd/StereoVLA](https://github.com/shengliangd/StereoVLA)
+
+---
+
+## Model Server Setup
+
+### Environment Setup
+
+1. Create a directory for the model code and weights, then `cd` into it.
+
+2. Create a new conda environment:
+   ```bash
+   conda create -p ./env python=3.12
+   conda activate ./env
+   ```
+
+3. Download the model weights locally:
+   ```bash
+   huggingface-cli download --local-dir ./storage shengliangd/StereoVLA
+   ```
+
+4. Clone the repository and install dependencies:
+   ```bash
+   git clone https://github.com/shengliangd/StereoVLA.git
+   cd StereoVLA
+   pip install -r requirements.txt
+   ```
+
+5. Clone the [modified FoundationStereo repo](https://github.com/shengliangd/StereoVLA-FoundationStereo) and install it:
+   ```bash
+   git clone https://github.com/shengliangd/StereoVLA-FoundationStereo
+   cd StereoVLA-FoundationStereo
+   pip install -e .
+   ```
+
+### Running the Server
+
+Run the model server from the cloned repository directory, pointing to the Hugging Face download directory:
+
+```bash
+STORAGE_PATH=`readlink -f ../storage` python -m vla_network.scripts.serve --path ../storage/stereovla/checkpoint/model.safetensors --port 6666
+```
+
+For faster inference, add `--compile` to the command. This speeds up inference by around 50%, at the cost of slower initial model loading.
+
+---
+
+## Supported Instructions
+
+The model accepts the following language instructions:
+
+- `pick up {object}`
+- `stack {color} bowl onto {color} bowl`
+- `stack {color} cube onto {color} cube`
+- `move {object} to {container}`
+- `move {object} to {color} {container}`
+- `pick up {color} {object}`
+
+---
+
+## Citation
+
+```bibtex
+@misc{deng2025stereovlaenhancingvisionlanguageactionmodels,
+      title={StereoVLA: Enhancing Vision-Language-Action Models with Stereo Vision}, 
+      author={Shengliang Deng and Mi Yan and Yixin Zheng and Jiayi Su and Wenhao Zhang and Xiaoguang Zhao and Heming Cui and Zhizheng Zhang and He Wang},
+      year={2025},
+      eprint={2512.21970},
+      archivePrefix={arXiv},
+      primaryClass={cs.RO},
+      url={https://arxiv.org/abs/2512.21970}, 
+}
+```