iLSPR: A Learning-based Scene Point-cloud Registration method for robotic spatial awareness in intelligent manufacturing

A critical capability for intelligent manufacturing is the ability for robotic systems to understand the spatial operation environment – the ability for robots to precisely recognize and estimate the spatial positions and orientations of objects in industrial scenes. Existing scene reconstruction methods are designed for general settings with low precision needs of objects and abundant data. However, manufacturing hinges on high object precision and operates with limited data. Addressing such challenges and limitations, we propose a novel L earning-based S cene P oint-cloud R egistration framework for automatic i ndustrial scene reconstruction (iLSPR). The proposed iLSPR framework leverages point cloud representation and integrates three key innovations: (i) a M ulti- F eature R obust P oint M atching N etwork (MF-RPMN) that learns from both raw data and deep features of the objects to accurately align point clouds, (ii) a G eometric- P rimitive-based D ata G eneration (GPDG) method for efficient synthetic data generation, and (iii) a digital model library of industrial target objects. During operation, vision sensors capture point clouds in the scenes, and the iLSPR method registers high-fidelity object models in the scenes using MF-RPMN, pre-trained with GPDG-generated data. We introduce an I ndustrial S cene O bject P oint-cloud R egistration (ISOPR) dataset in IsaacSim to benchmark performance. Experimental results demonstrate that iLSPR significantly outperforms existing methods in accuracy and robustness. We further validate the approach on a real-world robotic manufacturing system, demonstrating reliable digital reconstruction of industrial scenes. • We develop an iLSPR scene reconstruction method to address unique challenges in robotic precision manufacturing. • We propose a novel point-cloud matching model based on multi-feature fusion and robust point matching. • We create a geometric-primitive-based data generation method for model pre-training. • We release an open-source Industrial Scene Object Point-cloud Registration (ISOPR) Dataset and conduct a comprehensive benchmark of the proposed method against state-of-the-art techniques. • We implement the proposed method experimentally and build a prototype system.