Researchers from the University of Michigan, including Onur Bagoren, Seth Isaacson, Sacchin Sundar, Yung-Ching Sun, Anja Sheppard, Haoyu Ma, Abrar Shariff, Ram Vasudevan, and Katherine A. Skinner, have developed a novel framework for real-time underwater Simultaneous Localization and Mapping (SLAM) that leverages stereo cameras and other sensors to enhance underwater robot perception. Their work addresses the significant challenges posed by underwater environments, where image degradation and lack of texture make depth estimation and 3D reconstruction particularly difficult.
The team’s research focuses on overcoming the limitations of stereo depth estimation in underwater settings. Traditional stereo estimation networks trained on above-water data fail to perform accurately underwater due to light attenuation, visual artifacts, and dynamic lighting conditions. To tackle this, the researchers propose a framework that enables sim-to-real training of underwater stereo disparity estimation networks. This approach uses simulated data for initial training and then fine-tunes the networks using self-supervised learning on real-world data.
One of the key innovations in their work is the development of SurfSLAM, a novel framework for real-time underwater SLAM. SurfSLAM integrates stereo camera data with Inertial Measurement Unit (IMU), barometric, and Doppler Velocity Log (DVL) measurements to achieve accurate trajectory estimation and 3D reconstruction. This integration is crucial for navigating and mapping complex underwater environments, such as shipwreck sites.
To validate their approach, the researchers collected a comprehensive dataset featuring over 24,000 stereo image pairs, along with high-quality photogrammetry models and reference trajectories. This dataset was used to evaluate the performance of their framework in real-world scenarios. Through extensive experiments, the team demonstrated significant improvements in stereo depth estimation and SLAM accuracy, highlighting the effectiveness of their sim-to-real training approach.
The practical applications of this research are substantial for the marine sector. Accurate underwater SLAM is essential for autonomous underwater vehicles (AUVs) and remotely operated vehicles (ROVs) tasked with exploration, inspection, and mapping missions. By enhancing the perception capabilities of these robots, the framework developed by the University of Michigan researchers can improve the efficiency and safety of underwater operations. This technology could be particularly valuable for applications such as offshore oil and gas inspections, underwater archaeological surveys, and marine habitat monitoring.
Furthermore, the framework’s ability to generate detailed 3D reconstructions of underwater environments can support scientific research and conservation efforts. Accurate mapping of shipwrecks, coral reefs, and other underwater structures provides valuable data for historians, archaeologists, and marine biologists. The insights gained from these reconstructions can inform conservation strategies and help protect fragile marine ecosystems.
In summary, the researchers from the University of Michigan have made significant strides in advancing underwater robot perception through their novel SLAM framework. By addressing the unique challenges of underwater environments, their work paves the way for more robust and accurate autonomous underwater operations. The practical applications of this technology are vast, offering numerous benefits for the marine industry, scientific research, and environmental conservation. Read the original research paper here.

