Uncertainty-Driven 3-D Gaussian Splatting for Robust Real-Time RGB-D SLAM

Abstract

3D Gaussian Splatting has emerged as a promising map representation that allows autonomous robots to localize themselves while reconstructing a photorealistic environment. However, its robustness in real-world deployments remains limited: the lack of effective noise mitigation and efficient loop closure leads to the accumulation of errors from motion blur and sensor noise. This paper presents $\mu $ SLAM, a robust, real-time RGB-D Simultaneous Localization and Mapping (SLAM) system addressing these limitations through a synergistic, uncertainty-driven framework. It builds on a hybrid map integrating a sparse feature-based backbone with an uncertainty-aware Gaussian field. Local noise is mitigated through information-theoretic keyframe selection, coupled with a confidence-modulated adaptive optimizer for map refinement. Global consistency is ensured by a novel coarse-to-fine loop closure that leverages the embedded sparse map for robust initial alignment and the dense field’s rendering capabilities for high-precision refinement. Extensive experiments on public benchmarks and real-world robotic sequences demonstrate that $\mu $ SLAM achieves state-of-the-art tracking accuracy and reconstruction quality, with superior robustness to noise and drift. Crucially, it is the only evaluated system delivering this level of performance while consistently operating in real-time (over 30 frames per second), making it a comprehensive solution for practical robotic applications. Note to Practitioners—This paper is motivated by the dual demand in indoor service robotics and AR/VR industries for creating high-fidelity, photorealistic 3D maps while maintaining precise, globally consistent self-localization in real-time. A major practical bottleneck is the reliance on professional stabilization equipment; using low-cost cameras on mobile robots or handheld devices often results in trajectory drift and map corruption caused by rapid motion blur and camera shake. This work addresses these issues by introducing $\mu $ SLAM, a system that ensures robust pose estimation and flexible data collection in noisy environments. Our solution improves system performance in three key aspects: 1) Efficiency: By quantifying map “confidence,” the system intelligently selects only the most informative frames for processing; 2) Quality: It employs targeted optimization strategies to reconstruct only the regions that are not fully modeled; 3) Reliability: It integrates a sparse geometric backbone with loop closure mechanisms to guarantee accurate, drift-free trajectory estimation for both autonomous robots and human operators, even during aggressive motion. The primary contribution is a quantified approach to handling 3DGS noise, achieving an optimal balance between real-time performance and robustness. A current limitation is the hardware requirement (high-end GPUs like RTX 3090). Future work will explore better theoretical confidence algorithms to enable deployment on resource-constrained platforms and extend the system to Active SLAM. Potential applications include providing reliable pose and confidence data for robotic grasping and navigation, as well as constructing realistic scenes for training Vision-Language-Action (VLA) models.