Pose Error Prediction, Compensation Method, and Applicable Condition Determination of Parallel Motion Platform Based on Transfer Learning.

Abstract

Collecting a large amount of measured configuration data for robots entails high costs and time, which restricts the widespread use of neural networks for robot error prediction and compensation. In this study, a "transfer network" is established by taking the motion transmission characteristics inherent in the ideal kinematic model as prior knowledge and transferring it to a network trained based on the actual poses. Compared with the traditional back propagation (BP) neural network trained by actual poses alone, the transfer network shows significant performance advantages, effectively solving the problems of low prediction accuracy and weak generalization ability in the case of the small-sample measured data. Considering this, a method for determining the applicability of transfer learning is proposed. This method is achieved by evaluating the similarity between the learning tasks and then revealing the coupling effect of task similarity and the sample number of actual poses on the performance of the transfer network. Experiments are conducted on a six degrees of freedom (6-DOF) parallel robot. The results verify the superiority of transfer learning applied in robot precision compensation and the effectiveness of the proposed determination method.