Intelligent Control and Collaboration Techniques for Industrial Robots

Abstract

Industrial robot collaboration requires precise decisions about whether an operation has completed normally before another robot continues the sequential process. This paper investigates a novel execution supervisor that combines force and torque classification with an explicit permission condition for a shared transfer zone. Thirty statistical signal features and five task indicators support a robust ensemble classifier, while a training-only threshold procedure dynamically adjusts the balance between abnormal release and unnecessary holding. The evaluation utilizes 463 publicly available robot execution records containing 251 distinct sensor trajectories. Identical trajectories are systematically grouped across task files throughout repeated outer evaluation and inner threshold selection. The statistical random forest model achieves a mean balanced accuracy of 0.9669 and a macro F1 score of 0.9687 across five grouped evaluation repetitions. Furthermore, cost-aware threshold selection effectively reduces the abnormal-release rate from 1.50% to 1.14%, while increasing the normal-hold rate from 5.12% to 9.61%. The resulting mean weighted penalty changes only slightly, from 0.0683 to 0.0678. Tests on excluded task types yield balanced accuracy between 0.5250 and 0.9776, demonstrating limited transferability of the default threshold. A Boolean interlock is checked independently of the learned decision. These results strongly support a bounded design for post-observation execution supervision and sequential robot coordination, though physical contact safety and production throughput remain outside the current experimental scope.