CG-World Releases 850k-Segment CG Dataset to Reshape World Models with Counterfactuals

TheTuringPost · x · 2026-08-01

Existing video and robotics datasets often capture only partial physical structures, making it difficult to train truly usable world models. To address this, CG-World introduces a new data protocol based on industrial computer graphics (CG), treating pixels as observations generated by an underlying state.

The v1 dataset contains approximately 850,000 temporally aligned segments of 1-5 seconds. Beyond RGB alignment, it comprehensively records object geometry, actions, contact events, and physics states.

Its core innovation lies in introducing counterfactual branches: starting from the same initial state, it records alternative outcomes by altering a single variable such as action, mass, or friction. This design shifts the focus of world models from merely "predicting the next frame" to "understanding how interventions alter underlying states." While it includes 5,000 action and 5,000 mechanism branches, the full raw data currently requires controlled access due to industrial constraints.

Related event: CG-World Releases 850k Dataset to Reshape World Models(2 posts)→

Original post →

More from Embodied

Embodied channel →