Cross-Modal Fusion-Based Bolt Pose Estimation for Tightening Robot

Accurate bolt pose estimation is essential for improving the efficiency and precision of intelligent bolt tightening in automotive frame assembly. To address this requirement, a cross-modal fusion-based bolt pose estimation method is presented. First, to enhance bolt localization in lowtexture and reflective environments, a dynamic hybrid IoU loss function is introduced, and a DH-YOLO bolt detection network is developed. By dynamically combining multiple loss functions, adaptive bounding-box regression is achieved. Second, a dualconstraint DBSCAN-based point cloud segmentation algorithm is proposed, integrating curvature and normal vector direction consistency with an adaptive neighborhood radius strategy to accurately segment the bolt head, sidewalls, and frame plane. In addition, to reduce circular fitting errors caused by local point cloud gaps and uneven edge distributions, a sector-division circular fitting algorithm is employed to obtain a stable 3D center of the bolt head. Experimental results demonstrate that the proposed method achieves an mAP50 of 97.7% for bolt detection and a point cloud segmentation mIoU of 93.8%. The robotic tightening system attains an average success rate of 98.9% under diverse operating conditions, demonstrating its effectiveness for flexible automotive frame assembly.