Current video-generating AI systems struggle with physical rules, for example, they might show a ball falling sideways or objects appearing in the wrong place. This happens because these systems only learn from watching videos, never understanding the 3D world or the laws of motion.
The researchers built an AI agent that, given a text prompt, writes and runs a short computer program. That program describes a 3D scene with objects, their positions, and how they move. The agent then executes the code to produce a rough draft of the video, which provides clear spatial and motion cues to a standard video generator.
Testing showed the approach boosted physical accuracy by around 7 points on one benchmark and 24 on another. This means videos are far more likely to show objects moving naturally and staying where they belong, making AI-generated videos more reliable for tasks like simulation or content creation.