What the Ego Needs: Receiver-Conditioned Primitive Collaboration for Multi-UAV Perception

Abstract

Collaborative perception in low-altitude wireless networks (LAWNs) enables unmanned aerial vehicle (UAV) swarms to enhance scene understanding through cross-view observations. The key challenge lies in identifying receiver-relevant semantics that can compensate for the ego UAV's incomplete and viewpoint-constrained evidence. In this paper, we propose a demand-driven framework that recasts multi-UAV collaboration as receiver-conditioned evidence organization. The ego UAV first generates an on-demand semantic request to expose its insufficiently supported task evidence. Conditioned on this request, neighboring UAVs restructure their local observations into role-aware collaborative primitives, including target-centric primitives for object-level evidence and support-centric primitives for contextual compensation. This design converts fragmented token-level interaction into compact and functionally structured semantic collaboration. Experiments on UAV3D and CoPerception-UAV show that the proposed method achieves a favorable trade-off between 3D detection performance and communication cost, consistently improving detection accuracy and geometric reliability.