Interpretable structure-aware perception for local docking of mobile Rebar-Tying robots in precast manufacturing

In automated precast manufacturing, precise local docking of mobile rebar-tying robots relative to repetitive, texture-less reinforcement cages remains a critical challenge due to severe perceptual aliasing. Conventional methods are often unreliable in such environments, since they rely on unique-feature assumptions. To address this, a structure-aware perception framework is proposed that leverages geometric regularity as a robust estimation constraint. First, a hierarchical semantic abstraction module extracts structural primitives via deep learning-based crosspoint detection. Next, a novel Structural Congruence RANSAC (SC-RANSAC) algorithm integrates orthogonality and global completeness metrics to resolve local fitting ambiguities inherent in repetitive geometries. Finally, a task-driven projection step decouples the spatial pose into heading and distance variables for planar servoing. Experiments demonstrated the framework’s robustness against occlusion and sensing noise. On a 300-frame Random Viewpoint Benchmark, the proposed framework achieved mean pose estimation errors of 9.25 mm in distance and 0.71° in heading. Comparative evaluation against four widely adopted robust estimators (MLESAC, MAGSAC++, LO-RANSAC, and NAPSAC) showed that these approaches did not identify the correct target plane in 48% to 64% of the 300 frames owing to the lack of semantic discrimination; even on the frames where they succeeded, the proposed framework achieved lower orientation and translation errors. A case test in a precast factory showed that the proposed perception framework yielded pose estimates within the tolerance of the iterative servo loop and supported subsequent tying execution.