Multi-Constraint Optimization-Based Perception and Trajectory Planning for Robotic Strand Cutting and Roughening in Box Girder Anchor Recesses

A robotic system that integrates prestressing-steel strand cutting with automated surface roughening was developed to address the high labor cost, harsh working conditions, and poor quality consistency encountered in anchor-hole operations for prefabricated box girders. To overcome perception bottlenecks in high-dust and strong-glare casting-yard environments, a coarse-to-fine, two-stage pose estimation method was proposed based on point-cloud semantic segmentation and geometric-feature constraints. A deep-learning network was employed to decouple and isolate regions of interest (ROIs), and a weighted ICP algorithm was applied to achieve submillimeter-level positioning. To mitigate mobile-chassis docking deviations and constrained workspace limitations, a multi-constraint trajectory-optimization model and an online re-planning strategy were implemented. Field experiments show that, under extreme conditions with point-cloud incompleteness up to 20%, the system yields a positioning error below 1.65 mm and a trajectory-tracking root-mean-square error (RMSE) consistently within 2.8 mm. While the automated process increases the operational speed by approximately 1.5 to 1.8 times compared to manual work, its primary value lies in eliminating occupational hazards and ensuring consistent processing quality, demonstrating significant engineering value.