A research team has built a multi-agent system that converts plain text into editable 3D scenes where things actually move.
Called SimWorlds, the framework targets what the researchers call the "4D" problem: generating scenes where liquids flow, particles emit, rigid bodies collide, and mechanisms articulate — all from a text prompt. It runs on Blender and uses a planner-coder-reviewer workflow to step through a fixed construction sequence. A deterministic verifier checks the scene's layered protocol, and a runtime inspection tool catches physics failures that would be invisible in a rendered still image. The team also released 4DBuildBench, a benchmark designed to score both visual fidelity and physical consistency — a gap that existing static-scene benchmarks don't fill.
Most text-to-3D work stops at static output: a model generates a mesh, and that's it. Coordinating spatial layout, multiple physics solvers, camera, lighting, and time in one coherent scene is a harder problem, and it matters because physically grounded synthetic data is increasingly valuable for training video generation models and embodied AI systems. SimWorlds outperformed prior dynamic Blender generation baselines in experiments, according to the paper.
The benchmark may prove as useful as the system itself — right now there's no standard way to measure whether a generated scene is physically plausible, and whoever sets that standard tends to shape what the field optimizes for.