This is the official code for the paper "Point Bridge: 3D Representations for Cross Domain Policy Learning".

Point Bridge is a framework for learning robot manipulation policies using 3D point representations that enable transfer across simulation, real-world, and different camera viewpoints.

Website: https://pointbridge3d.github.io/

Paper: https://arxiv.org/abs/2601.16212

For business inquiries, please submit this form: NVIDIA Research Licensing

🌟 Key Features

• View-Invariant Representations: Uses 3D point clouds instead of pixels, enabling transfer across camera configurations
• Cross-Domain Transfer: Train in simulation, deploy in real-world with minimal visual gap
• Vision Foundation Model Integration: Leverages Molmo, SAM 2, and Foundation Stereo for robust point extraction
• Modular Architecture: Easy to adapt to new environments, robots, and tasks

📚 Documentation

Getting Started

• Getting Started Guide - Quick start tutorial and first experiment

Running Experiments

• Simulation Experiments - Training and evaluation in MimicLabs
• Real-World Deployment - Running on Franka FR3 robot

Architecture & Customization

• Architecture & Codebase Overview - System architecture, components, and code organization
• Adaptation Guide - Comprehensive guide for adapting to new environments and robots

🤝 Contributing

We welcome contributions! See Contributing Guide for details.

License

The code is released under the NVIDIA Source Code License.

📄 Citation

If you find this work useful, please cite the paper using the following bibtex:

@article{haldar2026pointbridge,
  title={Point Bridge: 3D Representations for Cross Domain Policy Learning},
  author={Haldar, Siddhant and Johannsmeier, Lars and Pinto, Lerrel and Gupta, Abhishek and Fox, Dieter and Narang, Yashraj and Mandlekar, Ajay},
  journal={arXiv preprint arXiv:2601.16212},
  year={2026}
}