GeneralVLA-2: Geometry-Aware Reconstruction and Governed Memory for Robot Planning

GeoFuse-MV3D replaces monocular SAM3D object reconstruction in GeneralVLA with a multi-view geometry-prior approach, reducing pose hallucination and improving 3D shape stability when calibrated cameras are available. A redesigned KnowledgeBank adds conflict resolution, confidence scoring, and geometric relevance filtering to memory retrieval, replacing the original append-only semantic matching. The system targets more reliable end-effector trajectory planning from RGB-D observations and language instructions.