Towards Physically Executable 3D Gaussian for Embodied Navigation

ABSTRACT

3D Gaussian Splatting (3DGS) is a photorealistic rendering method but lacks semantics and physical executability for Visual-Language Navigation (VLN). We propose SAGE-3D, a paradigm that upgrades 3DGS into an executable environment. It comprises Object-Centric Semantic Grounding, adding fine-grained annotations, and Physics-Aware Execution Jointing, which embeds collision objects into 3DGS. Experiments show that 3DGS scene data, while difficult to converge, exhibits strong generalizability, improving baseline performance by 31% on VLN-CE Unseen tasks. We provide a dense, differentiable scene representation that allows for significantly better trajectory planning in complex indoor environments.