Data-Driven Algorithms for Coil-Based Sensor Simulation and Experimental Contact Analysis

Gecko-inspired adhesives utilize microstructured fibrillar surfaces to achieve strong and reversible adhesion across various materials. Their ability to generate van der Waals forces without requiring external energy input makes them highly promising for robotic gripping applications. However, ensuring consistent adhesion, particularly on curved or irregular surfaces, necessitates precise contact formation. Even minor misalignments can lead to significant reductions in adhesion strength, impacting the reliability of gripping mechanisms. The effectiveness of these adhesives is largely influenced by their mechanical properties, which are often not explicitly available in standard material datasheets. As a result, accurate modeling of their mechanical response remains a key
challenge.
This thesis addresses the problem of material parameter estimation for finite element modeling (FEM) of a soft tactile sensor integrated with bioinspired adhesives. A key objective is to bridge the gap between simulated and experimental sensor responses, ensuring that the numerical models accurately replicate real-world behavior. Bayesian optimization is employed to determine the optimal Young’s modulus and Poisson’s ratio of the sensor’s foam layer by minimizing the error between simulated and experimental displacement data. This optimization ensures that the simulated deformations closely match physical measurements, enhancing the predictive reliability of the FEM model. Additionally, machine learning models are developed to predict the contact area and classify object shapes based on sensor displacement values, enhancing the interpretability of sensor responses in robotic applications.
To achieve this, a dataset of simulated sensor responses is generated through FEM, and Bayesian optimization is used to iteratively refine material properties to achieve realistic displacement predictions. The optimized model serves as a foundation for generating synthetic datasets, reducing the dependency on labor-intensive experimental data collection.The resulting synthetic dataset is then leveraged to train classification models, evaluating how sensor displacement data can infer contact conditions and object shape. Among the machine learning models tested, Random Forest and XGBoost classifiers exhibit the highest performance for shape prediction, with the best configuration achieving over 97% classification accuracy when utilizing a four-sensor setup. The findings demonstrate that integrating optimized FEM simulations with machine learning significantly improves predictive accuracy in sensor-based contact analysis.
The proposed approach not only enhances the calibration of soft tactile sensors but also facilitates data-driven analysis of bioinspired adhesives. These results underscore the potential of combining physics-based modeling with data-driven approaches to refine sensor performance and improve decision-making in robotic systems. These advancements contribute to the broader application of gecko-inspired adhesives in robotic gripping and automated material handling, where precise adhesion and object classification are critical for performance and reliability.