Deep learning-based predicting and compensating method for the pose deviations of parallel robots

Abstract

The static pose accuracy is regarded as a crucial performance indicator for parallel robots, which is inevitably affected by the geometric and nongeometric errors. However, an accurate error model for parallel robots is highly complex, especially for mathematical modeling and identifying non-geometric errors. This limitation hampers the application of parallel robots in high-precision processing and manufacturing fields. To enhance the performance of parallel robots and address the complexities in establishing their error models, a deep learning-based predicting and compensating method is investigated, which can accurately predict their pose deviations. This method ingeniously integrates the spatial feature extraction capabilities of convolutional neural networks, the bidirectional dependency capturing prowess of bilateral long short-term memory networks, and the feature significance discernment offered by squeeze-and-excitation networks. Subsequently, the pose deviations are definitely predicted by this method and are pre-compensated into nominal pose so as to improve the performance. A pose-deviation pairs dataset is established to accomplish the comparison of the investigated method with the two commonly used artificial neural network and long short-term memory on the 6-UPS parallel robot. Experiment on the 6-UPS parallel robot against traditional kinematic calibration methods demonstrate the investigated method achieves satisfactory pose accuracies, with the ability to reduce pose deviations by 88.60 %. Meanwhile, the effectiveness and general applicability of this method have also been validated by successful implementation on a 6-UCU parallel robot.