WONDER framework boosts UAV swarm coverage with radio world models

When disasters knock out terrestrial towers, UAV swarms can provide emergency wireless coverage. But each drone only sees local geometry and limited radio information, creating a gap between individual movement choices and swarm-wide coverage. The WONDER framework, proposed by Jiahao Huang, Rongpeng Li, and colleagues, attacks this problem with a Joint-Embedding Predictive Architecture (JEPA)-based radio world model that learns to predict the incremental radio effect of each candidate trajectory from deployment-available data.

Instead of forcing UAVs to share huge raw sensor streams, WONDER runs multi-round negotiation: each UAV ranks candidate trajectories, commits to one, then re-evaluates remaining options under the updated context. This is coupled with a Proximal Policy Optimization (PPO)-style actor that alternates between world-model training and policy improvement. To validate it, the team built RadioDynamics, a full simulation environment integrating UAV mobility, radio propagation, inter-UAV communication, and ray-traced digital-twin geometry across 62 metropolitan scenes. On 11 testing scenes, WONDER reached a balanced score of 0.870—beating STACCA by 0.162—while preserving 100% inter-UAV connectivity. The work suggests world models can make multi-agent coordination both smarter and more robust in real-world deployments.