Video2World: turning videos into executable simulators with GPT-6 Astra
A research team introduces Video2World, a benchmark where coding agents reconstruct scene, objects, physics and interaction logic end-to-end from an RGB interaction video and replay the task in a simulator.
Embodied AI has long been bottlenecked by data. What if any real-world video could become a physically executable, interactive simulator for generating scalable embodied data? With coding agents like GPT-6 Astra advancing rapidly, we asked whether they could themselves become a new form of world model. We introduce Video2World, a benchmark for autonomous video-to-simulation. Given an RGB interaction video, an agent reconstructs the scene, objects, physics, and interaction logic end-to-end, then executes the demonstrated task in the simulator.