A drone pilot has to hold two questions in mind at once. Where am I, how fast am I moving, which way am I facing? And separately: what is happening down there, what changed since my last pass? Most tests for vision-language models on aerial footage only ask the second kind of question. They check whether the model can describe a scene or recognise an event, which quietly assumes the hard part is the world and not the aircraft.

This work argues that the two belong together, and calls the pair dual cognition: reasoning about the drone's own state and about the environment, across several viewpoints and across time. The authors build a benchmark, UAV-DualCog, that deliberately tests both at once rather than letting a model score well by being good at scene description alone.

The reason this matters is that a drone that understands the ground perfectly but has lost track of its own heading is not much use, and neither is the reverse. Measuring the two separately can hide that gap. Measuring them jointly is how you find out whether a model could actually fly something.