Visual Servoing for Robotic Arm With Broad Optimization and Scene Dynamic Focusing Mechanism

Accurate and efficient visual servoing (VS) methods are crucial for enhancing the intelligence of industrial robots. However, how to drive the model for efficient learning and accurate prediction is still a challenging problem for deep learning-based VS. Moreover, imbalanced training data distributions can negatively impact algorithm performance. To address these problems, we proposed a two-stage VS network incorporating broad optimization and scene dynamic focusing mechanism. First, a broad optimizer is used to map the initial predictions of the network combined with random weights to the network’s nodes, and the nonlinear approximation ability of the broad learning is utilized to optimize the results of the pose regression. Second, a loss function based on the scene dynamic focus mechanism is designed to prioritize difficult examples during training, effectively improving the model’s accuracy and robustness. In addition, a rectangular distribution-based data generator is introduced to produce training datasets with more balanced distributions. Simulations and real-world experiments on the 6-DoF robotic arm demonstrate the effectiveness of the proposed method. The results show that the mean absolute error (MAE) and root-mean-square error (RMSE) errors of our method are reduced by 18.7% and 8.2%, respectively, compared with the popular backbone network. The code is available at https://github.com/zhaoweishuo/vs_bosdf.git