Anomaly Detection for Insertion Tasks in Robotic Assembly Using Gaussian Process Models

Component insertion is a common task in robotic assembly, and is widely used for manufacturing a variety of electronic devices. This task is generally characterized by low tolerances, thus requiring high precision during assembly. An early detection of a fault in the mating during the insertion process enables quality control of the end products, as well as safeguards the robotic equipment. We propose to use Gaussian Process Regression-based methods to learn the force profile during successful insertions, as well as quantify permissible deviations from this profile. The GPR model is then used to detect anomalies in case the observed force profile deviates significantly from the expected range. Apart from the standard GPR formulation, we consider two other variants – the Heteroscedastic GPR and the local GPR for better modeling accuracy and computational time efficiency, respectively. We report an accuracy of 100% in differentiating between normal and faulty insertions. The modeling and detection results indicate that our approach is accurate and robust to severe uncertainties due to process (e.g., force drift) and measurement noise.