Data-Driven Method for Robotic Trajectory Error Prediction and Compensation Based on Digital Twin

Abstract

In addressing the limited absolute positioning accuracy of industrial robots, which stems from the discrepancy between the nominal kinematic model and the physical entity, this paper proposes a novel paradigm for online compensation based on data-driven error prediction. The present study utilized a KUKA KR4 R600 robot as the experimental platform to construct a high-fidelity digital twin system capable of real-time synchronization. Within this framework, a new machine learning model, termed the Global Configuration-Error Forest (GCE-Forest), was developed and validated. The fundamental principle of GCE-Forest, based on the Random Forest algorithm, is its offline learning of the complex, highly non-linear mapping from the robot’s six-dimensional joint space configuration to its three-dimensional end-effector Cartesian error space. This facilitates online, feedforward, and predictive compensation for the nominal trajectory during robot operation. Through rigorous comparative experiments, the superiority of the proposed GCE-Forest was established. The final outcomes of dynamic trajectory tracking validation demonstrate that the system, by accurately predicting a mean nominal error of 0.1977 mm, successfully reduced the average spatial positioning error of the end-effector to 0.0845 mm, achieving an accuracy improvement of 57.25%. This research provides comprehensive validation of the method’s robust performance, offering a low-cost, non-invasive, and highly effective solution for significantly enhancing robotic accuracy.