Dynamic Airspace: A Benchmark for 4D Semantic Occupancy Generation for UAV

World models are becoming important for UAV embodied intelligence, as they enable agents to anticipate future environmental states before taking actions. However, many existing UAV world models still formulate world prediction primarily as 2D video generation, which overlooks the inherently 3D nature of UAV motion. Unlike ground agents, UAVs move freely in 3D space, and their future prediction should therefore capture 3D dynamic airspace, particularly changes in altitude along the vertical axis. Thus, in this paper, we present UAVX, a platform collecting large-scale data of dynamic 3D airspace for UAV world modeling. Specifically, we formulate future airspace prediction as 4D semantic occupancy generation, where the 3D occupancy and semantic state of the environment are predicted over time. We simulate UAV motion with 6-DoF control in large-scale 3D environments. As UAVX moves along 6-DoF trajectories, our system continuously records UAV observations together with the corresponding 4D semantic occupancy evolution, forming a large-scale benchmark for learning and evaluating future airspace prediction. Experimentally, we benchmark six representative methods on UAVX. The results show that existing methods generally struggle with 4D semantic occupancy generation under complex 6-DoF UAV motion, with temporal consistency remaining a major challenge. Even the best-performing method, GenieDrive, achieves only 65.31% on the temporal consistency metric. Meanwhile, our benchmark shows that methods that explicitly model the temporal evolution of UAV motion tend to achieve better performance, suggesting that motion-aware temporal modeling is a promising direction for future UAV world models.