Dual-Stage correction for high-precision pose estimation of electric vehicle charging ports

Purpose This study aims to improve the accuracy of robotic arm alignment with EV charging ports under outdoor conditions, addressing lighting changes, pose variation and detection errors. Design/methodology/approach A Dual-Stage Correction method is proposed. It combines a learnable correction module after YOLOv8 detection and a differentiable homography-based pose estimation enhanced by stereo depth and 3D template priors. Findings Experiments show a success rate over 80%, with an average translation error of 1.38 mm and rotation error of 1.33°, outperforming baseline methods. Originality/value This method integrates geometry and learning in a fully differentiable pipeline, improving robustness and precision for real-world robotic charging.