MVISTA-4D: View-Consistent 4D World Model with Test-Time Action Inference for Robotic Manipulation
This work introduces MVISTA-4D, a novel embodied 4D world model for robotic manipulation that generates geometrically consistent RGBD scenes from single views and infers actions. It achieves this by explicitly fusing cross-view and cross-modality features to ensure appearance and geometry align across different camera perspectives. A test-time optimization strategy refines predicted trajectories into executable actions, enabling robust 4D scene generation and downstream manipulation.