Meet ‘Code-as-World’: An Agentic Loop That Rewrites Real Videos Into Executable MuJoCo Physics Programs
Researchers convert real-world video into executable physics simulations — a new way to train embodied AI reasoning.
Why it matters
Code-as-World demonstrates a novel capability milestone: extracting and verifying executable world models from unstructured video. This addresses a fundamental frontier challenge in embodied AI — bridging the sim-to-real gap by working backward from observation to simulation.
The key facts
10 to knowConverts real video into editable MuJoCo physics code (executable world representation)
Agentic loop that verifies and refines extracted code
Trains physical reasoning on verified synthetic worlds derived from real observation
Bridges video-to-simulation gap for embodied AI training
Research appears to focus on physical understanding and model capability
Recovers editable MuJoCo scene code from real video
Closes loop: video → code → verified worlds → physical reasoning training
Agentic loop structure suggests iterative refinement of world models
Published August 2026 (very recent)
Bridges computer vision and physics simulation for embodied AI
Go to the source
MarkTechPostmarktechpost.com
Publisher excerpt: Code-as-World recovers editable MuJoCo scene code from real video, then uses those verified worlds to train physical reasoning.