Geometry-driven perception and action for topology-agnostic robotic rebar tying

Robust detection of rebar joints remains difficult due to geometric diversity, irregular intersection configurations, and noise introduced by real-world sensing. Prior approaches predominantly depend on appearance-based cues and topology-dependent supervision, which constrains their ability to generalize. This paper reframes rebar joint perception as a geometry-dominated structural inference task and introduces a two-stage learning framework that promotes the acquisition of geometric invariants without explicit domain adaptation. In the first stage, a synthetic dataset containing only geometric information is generated to train a detector, which is subsequently fixed and used as an automatic annotator to produce reliable pseudo-labels for simple cross-type intersections in real-world images. In the second stage, a new model is trained from scratch using these pseudo-labeled real samples, thereby incorporating realistic geometric variations without manual labeling effort. Building on this formulation, a unified perception-to-action pipeline is developed that operates independently of intersection topology for robotic rebar tying. • Geometry-dominated representation enables structural reasoning across diverse rebar joints. • Geometric invariants reduce reliance on appearance modeling and significantly shrink dataset size. • Two-stage sim-to-real learning leverages geometry-only training and offline pseudo-labeling. • Topology-agnostic perception-to-action mapping enables collision-free robotic rebar tying.