HI-SLAM2: Geometry-Aware Gaussian SLAM for Fast Monocular Scene Reconstruction

Can a single RGB camera reconstruct a full 3D scene in real time? HI-SLAM2 says yes. This T-RO paper combines monocular depth priors with 3D Gaussian splatting to deliver geometry-aware SLAM that rivals RGB-D methods, all from monocular input. Being presented at #ICRA2026. This video presents HI-SLAM2, a novel monocular dense SLAM system that combines 3D Gaussian Splatting with real-time pose estimation for high-fidelity scene reconstruction. Our approach leverages the efficiency of explicit 3D Gaussian representations alongside robust online loop closure and global optimization strategies to achieve accurate geometry and photorealistic rendering, all while maintaining real-time performance. The video highlights the system’s capabilities in challenging real-world settings, including large-scale outdoor environments and complex indoor scenes. It demonstrates robust pose tracking and high-quality dense reconstruction under diverse conditions. This supplementary video also provides visual comparisons with state-of-the-art baselines, offering clear evidence of HI-SLAM2’s superior performance in both geometric accuracy and visual quality. Project: https://hi-slam2.github.io/ Authors: Wei Zhang, Qing Cheng, David Skuddis, Niclas Zeller, Daniel Cremers & Norbert Haala Paper URL: https://ieeexplore.ieee.org/document/... #ICRA2026 #SLAM #GaussianSplatting #3DReconstruction