Continuous learning of contact episodes from proprioceptive sensors in industrial assembly scenarios using Adaptive Resonance Theory
Vinzenz Bargsten, Dimitrios Rakovitis, Vamsi Krishna Origanti, Adrian Danzglock, Frank Kirchner
In Proceedings of the 7th International Conference on Intelligent and Fuzzy Systems - INFUS2025, (INFUS-2025), 29.7.-31.7.2025, Istanbul, Springer, pages 1-8, Jul/2025.

Zusammenfassung (Abstract) :

Robotic automation plays a crucial role in modern industrial production, yet many assembly tasks still require manual intervention. Unlike standard pick-and-place operations, which can be executed with position-controlled manipulator arms, assembly tasks inherently involve physical contact between components, requiring precise force and torque management. Accurately assessing whether tightly fitting parts are correctly aligned or whether flexible components (e.g., plastic covers, cables) are properly assembled is challenging due to inevitable uncertainties in material properties and positioning. Traditional solutions rely on hand-crafted thresholds set by experts, which are costly and impractical for frequently changing product variants. To address this challenge, we present a machine learning approach based on Adaptive Resonance Theory (ART) for real-time, continuous learning and adaptation of contact episodes using proprioceptive sensor data. Our method processes joint torque and end-effector force-torque measurements, encoding these time-series signals into a frequency domain representation using Short-Time Fourier Transform (STFT). The ART-based module dynamically classifies contact patterns, identifying the most suitable learned category while detecting novel situations that deviate from prior experience, enabling adaptive control strategies. The proposed approach provides a scalable and cost-effective solution by reducing reliance on predefined heuristics and enabling online adaptation to new product configurations. The system is experimentally validated in two industrial assembly scenarios, where it demonstrates robust classification accuracy, real-time responsiveness, and adaptability compared to static threshold-based methods. Results highlight its potential for seamless integration into industrial control workflows, allowing robots to autonomously adjust assembly strategies or escalate novel contact situations for further inspection.

Files:

20250507_manuscript.pdf


© DFKI GmbH
zuletzt geändert am 27.02.2023