Efficient template-based robotic sorting with one-shot multi-object pose estimation algorithm

In industrial sorting, the efficiency and accuracy of robotic systems are crucial for optimizing manufacturing and logistics processes. Traditional template-based methods such as point pair feature (PPF) and iterative closest point (ICP) are widely used to obtain robust pose estimation solutions. However, in practice, their efficiency is highly dependent on the templates of workpieces. In this paper, we developed an efficient robotic sorting system with a template-based pose estimation algorithm for static sorting tasks. Considering the redundancy of the nominal template, we optimized the template by sparsifying its point clouds and cutting the hidden part. The proposed algorithm can achieve a one-shot multi-object pose estimation functionality, which includes point cloud segmentation, cluster, and registration. The random sample consensus (RANSAC) is used to segment the point clouds of workpieces from the background. The principal component analysis (PCA) is used to distinguish object categories. The PPF and ICP are used to achieve the coarse and fine registration, respectively. We conducted 60 groups of experiments to compare the optimized template with the nominal template. The experimental results showed that the average time taken by the optimized template to register the wheel and cube workpieces was 5.78 s and 7.73 s, respectively. However, the average time to register with the nominal template was 105.18 s and 191.09 s, respectively. The proposed method can generally achieve a 20-time efficiency improvement and a good performance in actual sorting experiments.