A Distributed Multi-UAV Exploration Framework via Communication-Sensing-Computation Co-Design

Abstract

Autonomous exploration using multiple unmanned aerial vehicles (UAVs) is of significant practical importance in GNSS-denied unknown environments, such as disaster response scenarios and underground space inspection. However, the practical deployment of multi-UAV systems remains challenging due to three major issues: 1) the vulnerability of centralized communication architectures to single-point failure, 2) the lack of reliable inter-UAV relative perception and localization in the absence of external positioning, and 3) the severe computational burden caused by simulating high-sampling-density LiDAR on simulation platforms. To address these challenges, this paper studies deployment-oriented adaptations for multi-UAV exploration under communication and computation constraints. At the communication layer, a lightweight decentralized middleware is adopted to decouple inter-UAV collaborative information from intra-UAV local data, thereby reducing the dependence on a single ROS Master. At the sensing layer, retroreflective markers are explored as a possible aid for inter-UAV relative observation in GNSS-denied deployment. At the computation and planning layer, a compute-adaptive trajectory planning strategy is introduced by reducing LiDAR sampling density, refining frontier removal logic, and incorporating dynamic constraints into path generation. Experimental results in complex simulated environments show that the resulting implementation significantly improves network robustness and increases the real-time factor (RTF) from approximately $0.3-0.4$ to $0.6-0.7$, corresponding to an approximately twofold improvement. In addition, the proposed method effectively eliminates the abnormal in-place spinning behavior caused by delayed frontier removal in simulation. These results demonstrate the efficiency and engineering practicality of the resulting implementation for collaborative multi-UAV exploration.