Nonrigid Point Cloud Registration Based on Local Features for Geometrical Calibration of Industrial Robots

Abstract

High-fidelity digital models are crucial for the accuracy of process simulation results of industrial robots. However, because of the influence of spatial pose errors, digital models cannot truly reflect the information of the physical world. To reduce the impact of pose errors on process simulation, this article proposes the nonrigid transformer (NRT), a high degree-of-freedom digital resource (i.e., industrial robots) pose calibration method based on nonrigid point cloud registration in multirobot assembly scenarios. The proposed method focuses on the characteristics of industrial robots and constructs local rigid feature extraction and fusion feature extraction modules. Furthermore, a multiple feature fusion module based on a gated recurrent unit (GRU) is proposed to explore the synergistic relationship and extract effective features to improve registration accuracy significantly. Experiments were conducted on the Fusion and Welding Station datasets, respectively. The results show that: 1) the proposed method is superior to existing nonrigid registration methods; 2) the Fusion dataset results show a remarkable 20% improvement in Chamfer distance (CD) and a significant 10.9% improvement in RMSE compared with the state-of-the-art (SOTA) method; and 3) the results of the Welding Station dataset demonstrate an outstanding 29.3% improvement in CD and a remarkable 31.6% improvement in RMSE compared with the SOTA method.