Large vision-language-action models are changing how robots connect language, perception and motion. For unmanned aerial vehicles, however, an open task often starts with a command that names a situation rather than a next safe motion. A request such as “inspect the damaged roof” or “search for a missing person” does not specify which fact must be checked first, from which viewpoint it can be checked, or whether the aircraft can obtain that evidence while keeping enough control margin.
In a Perspective article in National Science Review, researchers from Zhejiang University and Differential Robotics (Hangzhou) Technology Co., Ltd. argue that aerial embodied intelligence should be organized as command-measurement-action. The core idea is that a UAV should first turn a human command into executable measurement queries, use flight to obtain or reject task evidence, and then choose short-horizon actions based on evidence that has actually been checked.
The authors define a measurement query as more than a sensor reading. It is a task-level question that links a semantic proposition with a required viewpoint, confidence threshold and remaining safety margin. In roof inspection, for example, the useful question may be whether a crack on a particular tile can be confirmed from an oblique view with sufficient image scale and enough room to retreat. In search and rescue, the query may ask whether an occluded sector has truly been covered or whether a heat source can be accepted as human evidence.
This framing connects semantic goals to technical constraints that are especially important in flight. Moving closer may improve visual detail but reduce visual-inertial robustness. Changing altitude may reveal a target but increase wind or communication uncertainty. Waiting at a viewpoint may consume battery needed for a safe return. A useful aerial system therefore has to judge measurement feasibility together with motion feasibility.
The same view affects data generation, computation and evaluation. Expert trajectories alone are not enough if they do not record what uncertainty was reduced after each move. Synthetic and reconstruction-based data can be useful, but only when they include labels for visibility, dynamic feasibility, sensing noise and control conditions. Onboard, edge and cloud computation should also be selected according to the current task state, because extra reasoning is useful only when its delay, communication cost and energy use do not make the evidence stale or the action unsafe.
The article concludes that larger models alone will not solve the evidence-validity problem on aerial platforms. Future aerial robots should use flight to turn incomplete commands into reliable evidence, and their actions should depend on checked facts rather than assumptions.
National Science Review
Systematic review