Related Experiment Video
Updated: Aug 14, 2026
06:54
Photorealistic Learned Landscapes for Augmented Reality
Published on: June 27, 2025
SonarReg-GS SLAM: Sparse Sonar-Guided Depth Regularization for Underwater Gaussian Splatting SLAM
Wen Yang1, Xiaolong Qian2, Xulin Liu1
1Ocean College, Zhejiang University, Zhoushan 316021, China.
Sensors (Basel, Switzerland)
|August 13, 2026
Summary
This study introduces SonarReg-GS SLAM, a novel framework for underwater robotic perception. It effectively integrates forward-looking sonar data with 3D Gaussian Splatting SLAM to enhance metric depth accuracy and mapping quality in challenging underwater environments.
Area of Science:
- Robotics
- Computer Vision
- Acoustic Sensing
Background:
- Underwater robotic perception relies on accurate 3D scene reconstruction.
- Monocular 3D Gaussian Splatting SLAM (3DGS SLAM) offers dense scene representation but struggles with metric depth estimation due to visual ambiguities.
- Forward-looking sonar (FLS) provides range data but is inherently sparse, noisy, and ambiguous.
Purpose of the Study:
- To develop an underwater visual-acoustic 3DGS SLAM framework that overcomes the limitations of monocular depth estimation.
- To leverage sparse sonar data as reliable metric depth anchors for improved SLAM accuracy and mapping.
- To enhance underwater robotic perception by integrating RGB imagery and FLS measurements.
Main Methods:
- Proposed SonarReg-GS SLAM framework integrating synchronized RGB and sonar inputs.
- Utilized object masks to constrain acoustic range association and employed object-aware sampling for selecting sparse metric anchors.
- Implemented sonar-guided depth regularization for Gaussian initialization, tracking, and an object-aware RGB mask loss for enhanced supervision.
Main Results:
- SonarReg-GS SLAM significantly improved tracking accuracy, reducing average ATE RMSE by up to 21.7% compared to Splat-SLAM.
- Enhanced mapping quality was demonstrated through increased average PSNR (up to 7.72 dB) and reduced average LPIPS (by up to 0.277).
- The framework successfully regularized monocular depth scale and provided metric depth priors, outperforming classical and Gaussian Splatting SLAM baselines.
Conclusions:
- SonarReg-GS SLAM provides a robust solution for underwater visual-acoustic SLAM, effectively addressing metric depth challenges.
- The integration of sparse sonar depth anchors with 3DGS SLAM significantly boosts perception system performance in underwater environments.
- This framework offers a promising advancement for underwater robotic navigation, mapping, and exploration tasks.
