Attention-Guided Reinforcement Learning for Visual Servoing Control of Multirotor UAVs

To address the challenges of dynamic perception, real-time decision-making, and control stability in UAV visual tracking tasks, this study proposes an Attention-guided Visual Servoing Reinforcement Learning (AVSRL) framework. Unlike conventional Image-Based Visual Servoing (IBVS) methods that rely on analytical Jacobian control, AVSRL employs a virtual camera mechanism to normalize image observations and extract geometry-aware visual features as state inputs to a deep reinforcement learning (DRL) agent. The proposed framework integrates model identification, attention-enhanced actor-critic learning, and multi-source visual-environment encoding to enable robust and adaptive UAV control in dynamic and complex environments. A multi-head attention module selectively processes spatial-temporal observations from both static and dynamic objects, while a centralized critic facilitates cooperative optimization across agents. Extensive simulations and real-world experiments were conducted, including average reward analysis, hyperparameter sensitivity evaluation, trajectory tracking of complex geometric paths, and pedestrian following in outdoor scenarios. Results show that AVSRL outperforms baseline DRL and IBVS controllers in terms of tracking accuracy, control smoothness, and adaptability to visual and environmental variability. The findings validate the AVSRL framework as a promising solution for real-time UAV navigation and tracking tasks under uncertain and complex visual conditions.