MirroS releases Code-as-World to convert videos into executable MuJoCo physics programs

MirroS has released Code-as-World, a paradigm and set of models that represent physical scenes as executable code rather than pixels or latents. The system uses an agentic loop to recover these representations from real footage, enabling the generation of training data with exact physical labels.

• MirroS released two checkpoints, Code-as-World-VL-4B and Code-as-World-VL-9B, under Apache 2.0, fine-tuned from Qwen3.5 models.
• The agentic loop performs up to five rounds of propose, instantiate, execute, render, and verify steps to recover executable world representations.
• Verified worlds serve as training supervision, with the 9B model scoring 55.4 MRA on QuantiPhy-validation, surpassing Gemini-3.1 Flash at 54.8.
• The approach treats pixels as evidence of a scene rather than its ontology, compiling scenes into scene.json files for MuJoCo execution.

This method provides exact physical labels that real video lacks, allowing models to learn precise mass, contact, and gravity dynamics through supervised fine-tuning and reinforcement learning.