ExeDream: Grounding Video Generation Models in Physical World via Robot Execution

Video generation models have made impressive strides in synthesizing visually compelling content, yet their outputs remain confined to the virtual domain. A natural and important question follows: how well do these models reflect the physical world when their generated videos are taken out of the screen and into reality? We propose robotic manipulation as a concrete and measurable window onto this question. If a model has truly internalized the laws of the physical world, the motion it depicts should translate into executable robot behavior. We introduce ExeDream, an evaluation framework that operationalizes this criterion through a novel video-to-execution pipeline. Starting from a scene image and a task description, ExeDream uses each model to synthesize a manipulation video, converts the generated motion into robot trajectories, and executes them in a physics simulator. The resulting execution success provides a grounding signal that purely visual metrics cannot offer. Using this pipeline, we evaluate 8 models spanning frontier closed-source generators, open-source generators, and robot-specific policy models. Our benchmark covers over 100 manually curated manipulation tasks at three levels of physical complexity, with performance measured across visual quality, trajectory fidelity, and execution success. Our study yields an encouraging signal: several models achieve measurable execution success, suggesting that generative priors learned from internet-scale data already encode meaningful physical knowledge. Yet visual quality proves a poor predictor of executability, exposing a dimension of model capability that standard visual evaluations do not capture.