Physically Grounded Multimodal World Models for Embodied Robot Interaction via 3D Representation Alignment. JAAIR 2026, 1 (1).