China's Orca world model learns robotics signals without action labels

The Beijing Academy of Artificial Intelligence has released Orca, a world model designed to predict abstract world states rather than tokens or pixels, The Decoder reported. The model was trained on 125,000 hours of video without action labels.

That approach is important because robotics data is expensive to label and often tied to specific hardware. A model that learns useful structure from unlabeled video could reduce the cost of building more adaptable embodied AI systems.

The result also shows how world-model research is moving closer to practical robotics, where understanding physical change can matter more than text generation.