A Hybrid Deep Reinforcement Learning Strategy With Visual Servoing and Expert Guidance for Uncrewed Aerial Vehicle Autonomous Docking of Cylindrical Splice Sleeves

To address the absence of reference paradigms, limited generalization, and low learning efficiency in unmanned aerial vehicle (UAV) autonomous docking with transmission line splice sleeves under dynamic environments, this paper proposes an Expert-Guided Hybrid Deep Reinforcement Learning (EG-HDRL) strategy integrating intelligent visual servoing. The approach combines a model-based Visual Servo Proportional–Integral–Depth Control (VSPIDC) analytical controller—explicitly defined by mathematical models and requiring no training—with the Soft Actor-Critic (SAC) algorithm for high-dimensional policy optimization, enhancing VSPIDC’s generalization capability and improving DRL training efficiency. In the framework, VSPIDC serves as a fixed expert policy to generate reference trajectories, guiding SAC through a five-stage training pipeline: expert demonstration, behavior cloning, DAgger augmentation, offline fine-tuning, and dynamic hybrid training, enabling a smooth transition from analytical control to data-driven learning. A PX4 + GAZEBO + ROS simulation platform with wind disturbance modeling is developed. Experimental results show that EG-HDRL achieves a 91% docking success rate under baseline wind conditions, outperforming SAC (87%) and VSPIDC (71%), and reduces training time by 10.8% compared with SAC. The visual perception module’s mAP@50~90 significantly impacts convergence stability and state estimation accuracy, while generalization tests confirm robustness across varying wind speeds and directions. This study provides a viable reference for high-precision UAV docking in dynamic environments with strong potential for engineering applications.