[ICLR 2026 Oral] MomaGraph: State-Aware Unified Scene Graphs with Vision-Language Models for Embodied Task Planning - View it on GitHub