Vision-Based Anchor Detection and Localization for Autonomous Coreless Filament Winding

• Shokhin Sharipov, Mechanical Engineering, University of Delaware

• Kelvin Fu, Mechanical Engineering, University of Delaware

Coreless filament winding enables the fabrication of lightweight, freeform composite structures by routing continuous fiber tow around discrete anchor points. Reliable robotic winding requires accurate knowledge of the as-placed scaffold geometry, because small deviations between designed and actual anchor locations can lead to path errors, anchor misses, or unstable fiber placement. This project focuses on a vision-based anchor detection and localization module for integration into an autonomous coreless filament-winding workflow. RGB-D images of the winding scaffold were collected and labeled to train a YOLO-based object detection model for identifying individual anchors. The trained detector achieved an approximate average precision of 0.85, with example anchor confidence scores ranging from 0.95 to 0.99. Detected anchor centers were then combined with depth information to estimate anchor locations in the robot base coordinate frame. Across 25 tested anchor positions, the localization module achieved a mean residual 2D robot-frame error of 1.59 ± 1.09 mm, with an RMSE of 1.92 mm. These results indicate that the detection and localization module can provide sufficiently accurate anchor coordinates for reconstructing the as-placed scaffold geometry and supporting downstream robotic path generation. By reducing dependence on ideal CAD placement, this perception module provides an important foundation for adaptive, vision-guided autonomous filament winding.