MirroS introduces code-as-world: transforming real videos into executable physics simulations

MirroS has unveiled Code-as-World, a groundbreaking approach that converts real-world videos into executable MuJoCo physics programs. The system represents physical scenes through executable code, enabling agents to verify and edit simulations. Code-as-World-VL-9B, trained on 73,335 image-space QA pairs, achieved a 55.4 MRA score on QuantiPhy-validation, surpassing Gemini-3.1 Flash and other baselines. The model uses an agentic loop to recover programs from videos in up to five rounds, leveraging tools like SAM 3 and VGGT-Omega. Training involved eight NVIDIA H100 GPUs and fine-tuning from Qwen3.5 models. The system also excels in object and trajectory alignment, with improvements noted after world-space reinforcement learning. MirroS has released the GitHub repo and two checkpoints under Apache 2.0, inviting collaboration for product releases and webinars. The approach addresses limitations in video models and 3D reconstruction by focusing on physical mechanisms rather than visual data. Code-as-World represents scenes as executable world representations (EWRs), compiled into scene.json files for MuJoCo simulation. The method combines abductive search with feedback loops to refine simulations against real-world data, achieving results comparable to top models like ChatGPT-5.1 and Qwen3-VL-32B-Instruct.