Enhancing Embodied Reasoning and Grounding by Novel View Synthesis

Vision-Language Models (VLMs) provide strong image/video-level reasoning, but their use in embodied scene understanding remainslimited by the fixed viewpoints available in episodic RGB-D memories.Such observations can miss query-relevant evidence due to occlusion, ob-ject truncation, limited fields of view, or suboptimal view composition. Wepresent SplatReasoner, a framework that introduces novel view synthesisinto the VLM reasoning process by leveraging 3D Gaussian Splatting(3DGS). Given a user query about a 3D scene, SplatReasoner retrievesand synthesizes camera viewpoints that provide the most relevant visualevidence for answering the query and grounding the referred entitiesin 3D. Experiments show that query-conditioned novel view synthesisimproves embodied reasoning and grounding over fixed-view memory andlanguage-embedded 3DGS baselines.