Submitted:
11 November 2026
Posted:
12 November 2026
You are already at the latest version
Abstract
This study presents a CAD-free bolt pose-estimation method for robotic tightening that uses a single structured-light point cloud and requires no bolt-specific 3D CAD during inference or deployment. After point-cloud filtering and DBSCAN extraction, relaxed geometric rules generate candidate bolt top faces. Each candidate is represented by 19 scale-normalized features plus within-set mean deviations and percentile ranks, forming a 57-dimensional contextual representation. A compact 1,922-parameter dual-head multilayer perceptron is trained with ListNet ranking and soft-target binary cross-entropy. On 310 independent test frames from held-out scenes and workpiece types, M3 achieved 96.77% success, compared with 94.52% for a PointNet scorer and 74.2% for handcrafted scoring. All ten M3 failures arose from missing acceptable proposals; no ranking failure occurred when a valid candidate was retained. The complete perception pipeline required 83.138 ms per frame. In 3,366 factory tightening operations, the deployed M3 system achieved 99.67% first-pass success, with one failure attributed to perception. The results show that contextual candidate ranking can provide interpretable and low-cost pose estimation while removing bolt-specific CAD from runtime inference.
Keywords:
CAD-free pose estimation
; bolt pose estimation
; structured-light point cloud
; learning to rank
; robotic tightening
; industrial robotics
Copyright: This open access article is published under a Creative Commons CC BY 4.0 license, which permit the free download, distribution, and reuse, provided that the author and preprint are cited in any reuse.