Autonomous docking experiments for underwater spherical robot based on visual guidance with improved detection transformer

This study focuses on the autonomous recovery docking of the underwater spherical robot (USR). First, a visual guidance docking system is developed according to the geometric configuration and motion characteristics of the self-developed USR. To achieve continuous guidance from autonomous homing to terminal capture, a visual-guidance scheme combining long-range target detection with close-range ArUco-marker-based pose estimation is proposed. Meanwhile, a docking station (DS) adapted to the spherical geometry is designed. Its unique enveloping guide structure and spring-compliant V-shaped capture mechanism provide passive misalignment tolerance, contact cushioning, and mechanical locking. Then, to address frequency-domain information degradation, spatial-detail loss, and complex background interference encountered during DS recognition in the USR recovery process, an end-to-end Spatial–Frequency and Sparse Attention Detection Transformer (SFSA-DET) is proposed to improve the DS detection accuracy while reducing the model parameter count. Furthermore, a Vision-Aided Geometric Capture Docking (VAGCD) strategy is proposed to coordinate the three progressive stages of visual homing, geometric alignment, and precision capture. Finally, the pool experiments show that the USR achieves a docking success rate of 64.7%, preliminarily validating the feasibility and effectiveness of the developed docking system.