How do rescue drones find missing people in cluttered buildings?

Drones searching for targets in chaotic 3D spaces like collapsed buildings need both semantic smarts (knowing what matters) and smart exploration. This framework combines a viewpoint planner that predicts where targets are likely hiding—using reasoning about visible objects—with an active perception pipeline that spreads semantic priorities across unexplored regions. LLM similarity scores guide the search. Simulations show faster target discovery than baselines; real drone tests confirm it handles battery limits and sensor noise.