Manar Samri, Jonathan Thiemecke, Eva Prinz, René Hensel, Eduard Arzt

Abstract

The exceptional properties of fibrillar adhesives will accelerate their integration in various applications, such as pick-and-place robotics for delicate object handling. Despite their advantageous switchable and strong adhesion, their performance is affected by several factors, notably the alignment imperfections with the target object and unbalanced stress distributions at the gripping interface. In addition, defects on the fibrillar level that are present at the interface lead to a variation of the adhesion strength across a fibrillar array.

To ensure an efficient and reliable handling, we propose an in-line monitoring system that combines in-situ observation of the contact between PDMS fibrillar adhesives and a glass substrate, and machine learning algorithms for adhesion prediction. Visual features were extracted from the recorded images at maximum compressive preload and used to train different supervised learning models.

Based on a threshold force corresponding to the object's weight, different classifiers were able to predict attachment and detachment of the object with high accuracy of about 90%. The prediction of the exact value of the adhesion force using regression models showed that the Boosted Tree demonstrates a higher predictive power and outperformed an analytical model reported in literature.