RoboTransfer: Geometry-Consistent Video Diffusion for Robotic Visual Policy Transfer

Liu Liu, Xiaofeng Wang, Guosheng Zhao, Keyu Li, Wenkang Qin, Jiaxiong Qiu, Zheng Zhu, Guan Huang, Zhizhong Su

---

🔍 Abstract

RoboTransfer is a novel diffusion-based video generation framework tailored for robotic visual policy transfer. Unlike conventional approaches, RoboTransfer introduces geometry-aware synthesis by injecting depth and normal priors, ensuring multi-view consistency across dynamic robotic scenes. The method further supports explicit control over scene components, such as background editing, object identity swapping, and motion specification, offering a fine-grained video generation pipeline that benefits embodied learning.

---

🧠 Key Features

• 📐 Geometry-Consistent Diffusion: Injects global 3D cues (depth, normal) and cross-view interactions for multi-view realism.
• 🧩 Scene Component Control: Enables manipulation of object attributes (pose, identity) and background features.
• 🔁 Cross-View Conditioning: Learns representations from multiple camera views with spatial correspondence.
• 🤖 Robotic Policy Transfer: Facilitates domain adaptation by generating synthetic training data in target domains.

---

📖 BibTeX

@article{liu2025robotransfer,
  title={RoboTransfer: Geometry-Consistent Video Diffusion for Robotic Visual Policy Transfer},
  author={Liu, Liu and Wang, Xiaofeng and Zhao, Guosheng and Li, Keyu and Qin, Wenkang and Qiu, Jiaxiong and Zhu, Zheng and Huang, Guan and Su, Zhizhong},
  journal={arXiv preprint arXiv:2505.23171},
  year={2025}
}