<![CDATA[Decentralized navigation and inter-agent communication are key requirements for the safety of UAV swarms. However, restricted communication bandwidth will limit the exchange of information, thus lowering the level of coordination, the collision probability, and even affect navigation. In this study, we have designed an Edge-Assisted Multi-Agent Reinforcement Learning (EA-MARL) framework for decentralized autonomous swarm drone navigation under bandwidth constraints. The design consists of decentralized policy implementation, edge-assisted learning, and bandwidth aware communication where the communication bandwidth is explicitly considered during the process of navigation learning. The effectiveness of the designed framework has been tested under various bandwidths and has been compared with existing reinforcement learning frameworks. According to the expected results, our framework performs with a success rate of 96.1%, a collision rate of 2.8%, an average path length of 107.6 units and an average reward of 93.5. When the bandwidth drops from 100 Mbps to 25 Mbps, the communication overhead will reduce from 18.5 MB to 10.7 MB per episode with a drop of 42.2%, while the navigation success still maintains 89.6%. Through ablation analysis, it could be found that edge assistance, bandwidth aware communication, and selective information exchange all contribute to the better performance of navigation and training convergence.]]>
Copyrights © 2027
