AERIS: Offline Policy Improvement for Multi-UAV Integrated Sensing and Communication

A support-aware offline multi-agent RL framework that learns safe ISAC control from fixed flight logs, improving return by 29.3% while cutting collision-risk events by 54.2%

Ziyuan Wang; Yifan Sui; Wei Wei; Wenjie Xin; Zekai Zhang; Xiangwang Hou; Xiao-Ping; Zhang· 2026· DOI 10.48550/arXiv.2608.25477

The core problem

Unmanned aerial vehicle (UAV)-enabled integrated sensing and communication (ISAC) is positioned as a promising paradigm for 6G networks, where aerial platforms must simultaneously deliver communication service and perform sensing tasks. The central challenge addressed by this work is dynamic multi-UAV ISAC control: a controller must jointly balance communication quality, sensing reliability, and flight safety under stochastic mobility. The authors observe that existing optimization methods typically require repeated global non-convex solving, which is computationally heavy and poorly suited to real-time, stochastic multi-agent settings. On the other hand, online reinforcement learning (RL) offers adaptivity but depends on risky trial-and-error flights that may cause sensing loss or collision-risk events during training. To avoid these safety and cost issues, the paper proposes AERIS, an offline policy improvement framework for multi-UAV ISAC. AERIS learns from fixed flight logs under centralized training and decentralized execution (CTDE), so each UAV acts from local histories while training uses logged global information to assess team-level effects. The paper further introduces S

Innovation

The authors evaluate STAR-CRDT against baselines on the multi-UAV ISAC control task. The headline result is a 29.3% improvement in the main ISAC objective return over the strongest baseline. Beyond the aggregate return, the method improves several component metrics: communication sum rate increases by 3.4%, sensing pass rate by 4.8%, and sensing margin by 69.1%. Safety is also improved, with collision-risk events reduced by 54.2%. These numbers indicate that the offline policy improvement does not merely trade safety for performance; rather, it improves both task performance and safety relative to the strongest baseline. Generalization is tested on unseen real-road maps built from OpenStreetMap data. On these maps, STAR-CRDT still obtains the best return, suggesting that the learned policies transfer beyond the training distribution of flight logs. The paper reports these results as evidence that support-aware rectification and trusted distillation can yield reliable improvements without online trial-and-error flights.

Unmanned aerial vehicle (UAV)-enabled integrated sensing and communication (ISAC) is positioned as a promising paradigm for 6G networks, where aerial platforms must simultaneously deliver communication service and perform sensing tasks. The central challenge addressed by this work is dynamic multi-UAV ISAC control: a controller must jointly balance communication quality, sensing reliability, and flight safety under stochastic mobility. The authors observe that existing optimization methods typically require repeated global non-convex solving, which is computationally heavy and poorly suited to real-time, stochastic multi-agent settings. On the other hand, online reinforcement learning (RL) offers adaptivity but depends on risky trial-and-error flights that may cause sensing loss or collision-risk events during training. To avoid these safety and cost issues, the paper proposes AERIS, an offline policy improvement framework for multi-UAV ISAC. AERIS learns from fixed flight logs under centralized training and decentralized execution (CTDE), so each UAV acts from local histories while training uses logged global information to assess team-level effects. The paper further introduces STAR-CRDT, an offline multi-agent RL algorithm that performs support-aware local action rectification and distills only trusted improvements into the decentralized actor. The authors prove an offline-support policy improvement guarantee and validate the approach experimentally, reporting a 29.3% improvement in the main ISAC objective return over the strongest baseline, along with gains in communication sum rate, sensing pass rate, and sensing margin, and a reduction in collision-risk events.

AERIS is designed as an offline policy improvement framework that operates on a fixed dataset of flight logs rather than interacting with the environment during learning. The architecture follows the centralized training and decentralized execution paradigm: during training, a centralized critic can access logged global information to evaluate team-level effects, while each UAV's decentralized actor conditions only on its local history at execution time. This separation is intended to preserve scalability and practical deployability while still capturing cooperative effects.

Why it matters

The results support the central claim that offline policy improvement can be a practical route to safe multi-UAV ISAC control. By learning from fixed flight logs, AERIS avoids the risky exploration inherent in online RL, which is particularly important when sensing loss or collisions are unacceptable. The CTDE design allows training to exploit global logged information while keeping execution decentralized, matching the operational reality of UAV teams with limited inter-agent communication. The reported gains—29.3% in main return, 3.4% in communication sum rate, 4.8% in sensing pass rate, 69.1% in sensing margin, and a 54.2% reduction in collision-risk events—suggest that the support-aware rectification and trusted improvement distillation mechanisms are effective at extracting useful policy improvements from offline data. The offline-support policy improvement guarantee provides a theoretical rationale for why such improvements can be trusted, though the practical tightness of the bound and its dependence on dataset coverage remain important considerations. Generalization to unseen OpenStreetMap-based real-road maps is a notable strength, as it indicates that the approach is not merely memorizing the training logs. Future work may explore scaling to larger teams, richer sensing modalities, and more diverse environments, as well as quantifying the sensitivity of the method to the quality and coverage of the offline dataset. Overall, AERIS and STAR-CRDT offer a promising offline, safety-conscious framework for multi-UAV ISAC control in 6G systems.

Who should read this

CS practitioners and researchers

Opening member content…