AtlasVLA: Persistent World-Ego State Modeling for Vision-Language-Action Models | Guiyu Zhao et al. | ResearchPod