3-D Gaussian SLAM Based on Joint Source-Channel Coding and Multi-Scale Constraints

Abstract

Leveraging its exceptional rendering efficiency and photorealistic reconstruction capabilities, 3D Gaussian Splatting (3DGS) has emerged as a promising approach for real-time scene reconstruction and Simultaneous Localization and Mapping (SLAM) applications. Its compact scene representation and computational efficiency offer compelling solutions to communication bottlenecks inherent in embodied intelligence systems, where traditional perception methods suffer from prohibitive overhead due to sluggish environmental modeling and massive data transmission requirements. Despite these advantages, conventional 3DGS approaches exhibit critical limitations in input noise resilience and multi-scale modeling adaptability, constraining their deployment reliability in real-world embodied intelligence scenarios. This paper presents a novel 3D Gaussian scene representation framework that jointly integrates Joint Source-Channel Coding (JSCC) with multi-scale optimization to enhance modeling precision and robustness in complex operational environments. Our approach introduces two fundamental innovations: We adopt a unified JSCC framework that jointly optimizes RGB-D data compression and transmission through coupled source-channel coding. This integrated design preserves salient scene characteristics while eliminating redundant information, simultaneously enhancing communication efficiency and channel noise robustness to alleviate transmission burdens in embodied intelligence architectures. We combine a Gaussian pyramid-based multi-scale decomposition scheme that processes both rendered outputs and JSCC-enhanced RGB-D data. Under multi-scale loss constraints, this mechanism enables coordinated optimization of global structures and fine-grained details, maintaining computational efficiency while substantially improving representation fidelity. Extensive experimental validation across public datasets and real-world deployments demonstrates that our method effectively mitigates noise-induced artifacts and scale-related deviations. The proposed framework achieves superior performance compared to state-of-the-art 3DGS approaches in Absolute Trajectory Error (ATE) metrics while preserving real-time rendering capabilities. Furthermore, it significantly enhances multi-scale robustness, establishing a viable technological pathway for field deployment of embodied intelligence systems.