Published in RCIM iLSPR-Learning-Based Scene Point-Cloud Registration for Intelligent Manufacturing

We’re announcing our paper “iLSPR: A Learning-based Scene Point-cloud Registration Method for Robotic Spatial Awareness in Intelligent Manufacturing”(Yusen Wan, Xu Chen, University of Washington). The work focuses on industrial settings where robots need high-precision pose estimation and scene understanding, but data is often limited.

We propose iLSPR, a learning-based point-cloud registration framework that combines a robust point matching network (MF-RPMN), a geometry-primitive-based synthetic data generation method (GPDG) for pretraining, and a digital model library of target objects (CAD models). Given an RGB-D point cloud, iLSPR segments the target object and registers the corresponding model back into the scene to support reconstruction.

To evaluate the method, we build the ISOPR dataset in NVIDIA Isaac Sim, with 67 workpiece models and 2000 partial point-cloud samples. Experiments report strong registration accuracy and improvements over prior baselines, and we also validate the approach on real data using an Intel RealSense D405 camera.