The increasing demand for elderly care presents substantial challenges for global healthcare systems, necessitating scalable and adaptive technological solutions. Humanoid robots hold significant potential in assisting elderly users with mobility support, medication management, and social engagement. However, their effectiveness is often constrained by current robotic learning methods, which predominantly rely on scripted or teleoperated behaviors, limiting their adaptability in dynamic caregiving contexts. This conceptual paper introduces a structured Learning from Demonstration (LfD) framework. It leverages Virtual Reality (VR) to capture realistic, embodied demonstrations tailored explicitly for elderly care tasks. We critically discuss key challenges inherent to translating VR-trained behaviors to real-world robot deployment—such as embodiment discrepancies, sensor limitations, and environmental variability—and propose iterative methods to mitigate these issues. This framework explicitly aims to stimulate scholarly discussion, paving the way for subsequent empirical validation and prototyping. Our work thus establishes foundational guidelines and specific research questions for future studies on VR-driven robotic adaptation in elderly care scenarios.