CG-World: New Dataset for World Models and Embodied AI.

World models, which aim to learn the joint dynamics of states, actions, events, and observations, are often limited by existing datasets that capture only partial aspects of this complex structure. To address this, researchers have introduced CG-World, a comprehensive, large-scale dataset and protocol derived from industrial computer graphics production pipelines. CG-World explicitly records a rich array of intermediate states, including multimodal semantics, spatial configurations, skeletal and controller states, motion curves, camera and lighting parameters, physics caches, and contact events. This detailed information is organized into unified spatiotemporal samples, providing a robust foundation for training advanced world models. Crucially, CG-World also defines a branch lineage that covers factual trajectories, various types of interventions (observation, action, mechanism), and strict counterfactual branches. This design supports intervention learning and counterfactual reasoning by explicitly recording intervention targets, invariants, and alternative outcomes. Initial evaluations show that CG-World offers reusable structured supervision for geometry-conditioned video generation, action prediction, and closed-loop vision-language-action policy transfer, paving the way for advancements in Physical AI and embodied intelligence.