In automotive factories, articulated robots and conveyance systems require simulation within a high-fidelity virtual environment to ensure collision-free operation. Point clouds captured by terrestrial laser scanners (TLSs) are ideal for creating this “as-is” environment, but they contain both moving and stationary objects. For effective motion planning, the point clouds of moving equipment must be accurately replaced by their kinematic computer-aided design (CAD) models. A major challenge in this registration is the discrepancy between idealized CAD models and real assets. Robots are often outfitted with nonmodeled components like wire harnesses and covers, which act as outliers and degrade standard registration. We propose a robust fitting methodology using edge-weighted registration to address this issue. We hypothesize that geometric edges are structurally consistent between the CAD model and the point cloud, whereas irregular nonmodeled components are less likely to be detected. We introduce a fast edge detection algorithm that leverages the structured nature of TLS point clouds. By assigning higher weights to these stable edges, our method achieves robust alignment even with significant outliers. Our approach uniformly describes link mechanisms using the Unified Robot Description Format (URDF). It accommodates diverse kinematics: for single-chain mechanisms with revolute joints, posture is estimated by fitting links sequentially; for branched-chain mechanisms with prismatic joints, pose is determined by satisfying translational constraints. We evaluated the proposed method using virtual point clouds generated from a simulated scanner. The results show that the edge-weighted registration improves the robustness of pose estimation.
Kumazawa et al. (Sat,) studied this question.