Resilient AAV Trajectory Planning via Few-Shot Meta-Offline Reinforcement Learning

Reinforcement learning (RL) has emerged as a powerful tool for intelligent decision-making in wireless networks. However, most existing RL frameworks rely on continuous online interaction with the environment, which is often impractical in real-world scenarios due to safety risks, high costs, and lack of scalability across different environments. To overcome these limitations, we propose a novel few-shot meta-offline RL framework for resilient UAV trajectory and scheduling optimization in wireless IoT networks. Our approach integrates conservative Q-learning (CQL) with model-agnostic meta-learning (MAML), enabling UAVs to learn effective policies from static offline datasets and rapidly adapt to unseen environments with only a few training shots. The proposed framework is resilience-aware, allowing robust operation under unpredictable conditions such as weather-induced link failures. Numerical results show that the proposed algorithm consistently outperforms baseline schemes, including deep Q-networks (DQNs), CQL, and DQN-MAML. Specifically, it converges significantly faster, achieves lower joint age-of-information (AoI) and transmission power, and maintains reliable performance in outage-prone environments. Moreover, the meta-learning component enhances scalability by achieving more than 50% gain in return for policies trained over new objectives with limited data.