KANLoc: Learning to Anchor Visual Odometry with KAN-Based Pose Regression for Planetary Landing

Official implementation of "Learning to Anchor Visual Odometry: KAN-Based Pose Regression for Planetary Landing"

Accepted to IEEE Robotics and Automation Letters (RA-L), December 2025

Xubo Luo¹, Zhaojin Li², Xue Wan², Wei Zhang², Leizheng Shu²

¹University of Chinese Academy of Sciences   ²Technology and Engineering Center for Space Utilization, Chinese Academy of Sciences

---

🎯 Overview

KANLoc is a monocular localization framework designed for autonomous lunar landing that tightly couples visual odometry (VO) with a lightweight Kolmogorov-Arnold Network (KAN)-based absolute pose regressor. By fusing high-frequency relative tracking with sparse but robust global anchors, KANLoc achieves:

• 32% reduction in translation error
• 45% reduction in rotation error
• Real-time performance at ≥15 FPS
• Remarkable parameter efficiency (35.6M parameters)

---

📐 Architecture

---

📝 Citation

If you find this work useful, please cite:

@article{luo2025kanloc,
  title={Learning to Anchor Visual Odometry: KAN-Based Pose Regression for Planetary Landing},
  author={Luo, Xubo and Li, Zhaojin and Wan, Xue and Zhang, Wei and Shu, Leizheng},
  journal={IEEE Robotics and Automation Letters},
  year={2025},
  publisher={IEEE}
}

---

🙏 Acknowledgments

We thank the following open-source projects:

• ORB-SLAM3 for visual odometry
• DINO-v2 for visual features
• g2o for bundle adjustment
• AirSim for synthetic data generation

---

📄 License

This project is licensed under the MIT License - see the LICENSE file for details.

---

Made with ❤️ for the lunar exploration community