SimWorlds: A Multi-Agent System for Dynamic 3D Scene Creation

SimWorlds presents a multi-agent framework for generating editable dynamic 4D scenes from text. Its planner-coder-reviewer workflow uses Blender-specific procedural knowledge and a fixed ordered construction process to coordinate spatial layout, multiple physics solvers, temporal sequencing, cameras, and lighting. A layered scene protocol and deterministic verifier enforce structural constraints, while runtime-state inspection detects failures that rendered frames may hide. The authors also introduce 4DBuildBench, a benchmark covering both visual fidelity and physical consistency. The abstract reports that SimWorlds outperforms prior dynamic Blender-generation baselines, but provides no quantitative results in the supplied material.