This paper presents V-Dreamer, a fully automated framework for generating simulation-ready manipulation environments and executable robot trajectories from natural language instructions. The system leverages large language models and 3D generative models to construct physically grounded 3D scenes validated by geometric constraints for stable, collision-free layouts. Video generation models serve as rich motion priors for behavior synthesis, which are then mapped to executable robot trajectories through a Sim-to-Gen visual-kinematic alignment module using CoTracker3 and VGGT. This approach enables high visual diversity and physical fidelity for robotic manipulation training.