SimWorlds Creates Dynamic 3D Scenes from Text with Multi-Agent AI.

Current methods for translating natural language into 3D scenes primarily focus on static outputs, leaving the creation of dynamic, physics-driven 4D scenes largely unexplored. Such dynamic scenes, featuring flowing liquids, particle emissions, and articulated movements, hold significant value for editable content and as physics-grounded training data for video generation and embodied AI. This paper introduces SimWorlds, a multi-agent system designed to address this gap. SimWorlds employs a planner-coder-reviewer workflow, leveraging Blender-specific procedural knowledge to coordinate spatial layout, multiple physics solvers, temporal sequencing, camera, and lighting. It includes a layered scene protocol with a deterministic verifier and runtime-state inspection tools to catch mechanism failures that rendered images might miss. Alongside SimWorlds, the researchers present 4DBuildBench, a new benchmark to assess both the visual fidelity and physical consistency of the generated dynamic 3D scenes. Experiments show SimWorlds significantly outperforms existing dynamic Blender generation baselines.