Code-as-World Translates Real Videos into Physics Programs
A new agentic framework designated as Code-as-World translates real-world video footage into executable MuJoCo physics programs, according to a recent report by MarkTechPost. The approach addresses the challenge of grounding video models in physical reality by structuring environment representations as programmatic code rather than purely pixel-based predictions.
Traditional video generation models often struggle with long-term physical consistency because they lack underlying mechanical constraints. The Code-as-World method uses an iterative agentic loop that analyzes video inputs, infers physical properties, and writes functional simulation scripts. By structuring the output as standard physics engine programs, builders can verify, edit, and interact with the reconstructed environments programmatically.
According to MarkTechPost, the technique bridges the gap between passive video observation and active simulation infrastructure. This architecture allows AI agents to translate visual data into structured environments that can run directly inside physics engines. For developers working on robotics, simulation pipelines, and world models, this programmatic translation offers a path toward generating testable environments straight from raw visual data without manual 3D asset creation.
Based on reporting by www.marktechpost.com.
