Abstract 

Real-time, multi-agent systems capable of jointly estimating motion and reconstructing dense 3D environments are key to the next generation of spatially intelligent robotics. SplatUnity, a fully decentralized SLAM framework, is introduced. It employs 3D Gaussian Splatting (3DGS) to enable high-fidelity mapping and robust pose estimation across multiple agents. Unlike prior learning-based SLAM systems that rely on centralized coordination or limited scalability, SplatUnity allows each agent to independently build local submaps, detect overlapping regions, and estimate relative poses using a bird’s-eye view (BEV) matching pipeline refined via colored ICP. Submaps are registered into a shared global frame and merged through a loop-closure-aware optimization strategy. Experimental results on both synthetic and real-world multi-agent datasets demonstrate that SplatUnity achieves superior performance in trajectory accuracy and rendering quality compared to state-of-the-art baselines, even under unknown initial pose conditions.