A machine learning–based tool for enhancing position accuracy in industrial robots with a reduced dataset

Industry X.0 robotic manufacturing demands higher accuracy and flexibility, enabling continuously adaptive processes designed and optimized through simulations and Digital Twins. To achieve this level of flexibility and productivity in high value-added processes, where limited robot position accuracy becomes a critical constraint, advanced engineering methods and digital tools are required. These solutions must predictively compensate for inevitable robot positional accuracy errors, eliminating the need for manual pose refinement and enabling the generation of “first-time-right” robot code. This work aims to address these challenges by introducing an engineering tool capable of predictively correcting robot positioning inaccuracies across the workspace, enabling accurate point-to-point motion generation. It is intended for tasks with limited process interaction forces, where positioning errors are dominated by geometric, compliance, and joint-related effects. The tool leverages a multi-parameter Machine Learning (ML) error predictor trained on a reduced experimental dataset, minimizing data acquisition time and production downtime. Realized as a Python-based framework, it can be seamlessly integrated into commercial offline programming environments to automatically generate validated robot programs. The paper details the framework structure, focusing on the definition of the ML-based position error predictor, and its implementation on a robotic cell equipped with a high-payload KUKA robot and a FARO laser tracker. A preliminary experimental analysis identified payload, approach direction, and point location as the key operational parameters, accessible at the code level, that influence positioning accuracy. These insights guided feature selection and the design of reduced training datasets. In particular, a uniform spatial grid of only 64 points, corresponding to about one hour of measurement time, was sufficient to achieve near-optimal model accuracy. Several ML algorithms were compared, with the Tabular Prior-data Fitted Network achieving superior generalization on small datasets. Experimental validation on the KUKA robot showed up to a 98.4 % reduction in positioning error and consistent performance across all tested points, confirming the tool robustness and suitability for deployment across different industrial environments. All datasets, source code, and implementation scripts are openly released to enable reproducibility and facilitate industrial deployment.