XR-1-Stage2

[Project Page] [Paper] [GitHub]

This repository contains the Stage 2 weights for the XR-1 (X Robotic Model 1) project. This is the large-scale Vision-Language-Action (VLA) policy trained on top of the Unified Vision-Motion Codes (UVMC).

🤖 Model Description

XR-1-Stage2 is a generative Transformer-based policy. While Stage 1 focused on tokenizing sensory and action data, Stage 2 performs the high-level reasoning required to map natural language instructions and visual observations into executable robotic actions.

Key Features

• VLA Architecture: Integrates vision, language, and action into a unified Transformer framework.
• Token-Based Policy: Operates on the discrete latent space defined by the XR-1-Stage1-UVMC tokenizer.
• End-to-End Control: Predicts the next set of action tokens based on the current visual state and text prompt.

🧱 Model Hierarchy

1. Stage 1 (UVMC): Encodes images and actions into discrete tokens.
2. Stage 2 (VLA): [This Model] Predicts action tokens given instruction and image tokens.

🛠 Usage

To deploy this policy, you need both the Stage 1 tokenizer and these Stage 2 weights. Please refer to the XR-1 GitHub Repository for inference scripts and environment setup.

📝 Citation

If you use the XR-1 model in your research, please cite our work:

@article{fan2025xr,
  title={XR-1: Towards Versatile Vision-Language-Action Models via Learning Unified Vision-Motion Representations},
  author={Fan, Shichao and Wu, Kun and Che, Zhengping and Wang, Xinhua and Wu, Di and Liao, Fei and Liu, Ning and Zhang, Yixue and Zhao, Zhen and Xu, Zhiyuan and others},
  journal={arXiv preprint arXiv:2411.02776},
  year={2025}
}

📜 License

This project is licensed under the MIT License.

---

Contact: For questions, please open an issue on our GitHub or contact us at [email protected].