Interest in physical AI and robotics in health care is increasing, but the #nursing literature shows that the evidence base remains early, #nurse-centered applications are underdeveloped, and real-world experiential evidence is limited. #nurses are more likely to accept robots that reduce physically demanding and repetitive work while preserving the interpersonal and judgment-intensive core of #nursing practice. This conceptual paper proposes a #nursing-centered framework in which virtual reality functions not merely as a simulator but as a scaffolded training infrastructure for supportive physical AI systems, enabling #nurse augmentation, site-specific adaptation through digital twins, and staged simulation-to-real transfer. The framework was developed through a conceptually integrative and implementation-aware synthesis drawing on #nursing robotics, AI in #nursing, immersive simulation, digital twins, human-in-the-loop learning, and physical AI development literature. It is organized around 5 linked elements and supported by a 4-layer technical architecture that outlines functional requirements and implementation pathways. Four key propositions ground the framework: (1) #nursing robot training should focus on competency formation, not on decontextualized data accumulation; (2) training should proceed through progressive fidelity and staged autonomy; (3) digital twins should function as operational bridges for local ward adaptation; and (4) simulation-to-real transfer should be governed by explicit #nursing-relevant validation criteria and retained human accountability. The proposed #nurse-in-the-loop, site-specific virtual reality framework offers a #nursing-centered complement to general-purpose physical AI pipelines by making workflow fit, role boundaries, local adaptation, and governed transfer explicit design requirements.