China's Orca world model matches specialized robotics systems without ever seeing a single action label

Researchers at the Beijing Academy of Artificial Intelligence have developed Orca, a world model capable of predicting abstract physical states from video data alone. By training on 125,000 hours of footage without labeled action sequences, the system achieves performance levels comparable to specialized robotics models. This development is significant because it demonstrates that robots can learn to understand environmental dynamics without the labor-intensive process of manual data annotation, potentially accelerating the development of autonomous systems.