Robotic Machining Errors Prediction via In-Process Force and Vibration Sensors and Mixture of Experts

Abstract

High-precision robotic machining remains challenging, as achieving accurate dimensional and geometric features requires feedback control based on reliable error prediction. Effective solutions must estimate geometric errors accurately while minimizing inspection costs and ideally enabling real-time corrective control. An established approach relies on costly laser tracker position sensing with data-driven modeling, limiting large-scale industrial deployment. This article presents a machine learning framework for error prediction using in-process low-cost force and vibration sensor data, and robot controller signals. A mixture-of-experts (MoE) architecture is adopted, where the robot controller dynamically weights predictions from multiple experts, enabling complementary information fusion. Both handcrafted features and deep-learned features are developed and evaluated. Experiments on square and circular trajectories demonstrate that the proposed method is comparable to laser-tracker-learning-based models and surpasses direct laser tracker measurements, reducing the mean absolute error sevenfold from 71.4 to 9.2 microns relative to a coordinate measuring machine (CMM). The proposed framework enables sub-5 ms latency and an energy-delay product of 760 microjoule-second on a CPU board. These results indicate that low-cost sensing combined with machine learning can match expensive metrology and position sensing systems, providing a practical and scalable approach for high-precision manufacturing.