Multi-Touch Characterization of the SoftMag Tactile Sensor

1st Zhengyin Du† Soft BioRobotics Perception Lab Istituto Italiano di Tecnologia Genova, Italy [email protected] 2nd Chengjin Du†* Soft BioRobotics Perception Lab Istituto Italiano di Tecnologia Genova, Italy https://orcid.org/0000-0002-9879-9409 3rd Lucia Beccai Soft BioRobotics Perception Lab Istituto Italiano di Tecnologia Genova, Italy [email protected] Abstract—This study presents a learning-based framework for multi-touch characterization of the SoftMag sensor. A combined classification and regression approach is employed to simultaneously identify contact regions and estimate corresponding interaction forces. The proposed method enhances tactile resolution and supports robust perception for integration into compliant robotic systems. Keywords—Tactile sensor; Soft magnetic sensor; Sensing characterization I. INTRODUCTION Soft tactile sensors are essential for enabling compliant robotic systems to interact safely and effectively with delicate and deformable objects, particularly in agricultural applications such as post-harvest fruit handling. Conventional tactile sensing technologies, including optical [1], capacitive [2], and piezoresistive [3] methods, offer good sensitivity and spatial resolution but often face challenges when integrated into soft robotic platforms. These challenges include structural rigidity, mechanical hysteresis, and sensitivity to environmental disturbances [4][5]. Magnetic-based tactile sensors have emerged as a promising alternative. By embedding permanent magnets within deformable elastomeric substrates and combining them with magnetic field sensors such as Hall-effect chips, these systems provide robust, multi-axis force measurements while maintaining mechanical compliance and structural simplicity [6][7][8]. Their resilience to moisture, surface irregularities, and large deformations makes them particularly suitable for unstructured and organic environments. This includes automated fruit handling and packaging systems, where traditional sensors often fail to operate reliably [9][10]. In this paper, we explore multi-touch sensing capabilities of the SoftMag tactile sensor. A method is proposed that combines classification and regression to distinguish concurrent contact locations and estimate associated normal forces. Through structured experimental evaluation, the approach demonstrates effective recognition and interpretation of complex contact conditions, supporting its application in tasks that require rich tactile feedback. II. BACKGROUND AND TEST SETUP This study investigates the multi-touch response behavior of the SoftMag tactile sensor, as illustrated in Fig. 1, building upon the sensor architecture introduced in [11]. The same experimental setup and data processing pipeline in [11] were adopted to ensure consistency. Fig. 1 Five independently fabricated SoftMag prototypes A. Multi-Touch Contact Configurations Five custom-designed probes were developed to realize 15 multi-touch configurations, representing all non-empty combinations of simultaneous contact across the SoftMag sensor’s four defined regions. The total number of configurations was calculated as: 푁�=�퐶� �, � ��� (1) Each configuration, representing a unique contact scenario from single to four-point simultaneous indentations, underwent 20 cyclic tests using a hemispherical indenter applied at 0.2 mm/s via a three-axis platform. B. Data Acquisition and Preprocessing Force and magnetic flux data were sampled at 100 Hz and processed using a standardized MATLAB pipeline. This included Chebyshev Type I low-pass filtering (3 Hz passband), cycle segmentation, and normalization based on the Z-axis magnetic flux. The resulting labeled dataset supported subsequent analyses, including superposition evaluation and machine learning-based prediction of contact location and force magnitude. III. MULTI-TOUCH TESTING AND RESPONSE ANALYSIS A. Superposition Analysis Fig. 2 Data synthesized by superposing corresponding single-touch trials Fig. 3 Actual flux and force data from multi-touch experiments To evaluate whether the superposition principle holds for force and flux measurements, 11 synthetic datasets were generated by summing the corresponding single-touch datasets involved in each multi-touch case. These superimposed datasets, as shown in Fig. 2, were then This work was supported by the RAISE Project “Robotics and AI for Socio-economic Empowerment” (ECS_00000035 J33C22001220001). † These authors contributed equally to this work. ∗ Corresponding author Magnetic Flux Density (G) Reference Force (N) Magnetic Flux Density (G) Reference Force (N) 2025 I-RIM Conference October 17-19, Rome, Italy ISBN: 9788894580570 10.5281/zenodo.17629734 109 compared to the actual multi-touch data, as shown in Fig. 3. Qualitative observations suggested that the superposition principle captures general trends in both force and flux responses. However, quantitative evaluation revealed limitations. Root-mean-square error (RMSE) and standard deviation (STD), as shown in Fig. 4 and Fig. 5, between the true and superposed datasets were calculated and normalized by the respective maximum values of each signal. While force responses showed moderate deviations ranging from 14–25%, the discrepancies in flux signals were more pronounced, indicating reduced reliability of superposition as the number of simultaneous contact points increases. Fig. 4 RMSE and STD between actual and superimposed data: (a) multitouch force; and (b) magnetic flux components Fig. 5 Comparison of RMSE and STD between actual and synthetic data: (a) multi-touch force; and (b)-(d) magnetic flux components B. ML-Based Multi-touch Inference Modeling To evaluate the SoftMag sensor’s ability to recognize and quantify simultaneous contacts, a classification-regression framework was implemented. All multi-touch datasets were standardized using the mean and standard deviation of saturation data from the active contact zones and labeled according to their corresponding touch configurations. A classification analysis was conducted using a K-nearest neighbor (KNN) classifier with one neighbor and the Minkowski distance metric, trained on the standardized magnetic flux data. To enhance generalization and avoid overfitting, 10-fold cross-validation was applied. The classifier demonstrated excellent performance, achieving an overall accuracy of 99.7%, with most diagonal entries in the confusion matrix exceeding 98%, indicating highly consistent class-wise predictions. The positive predictive value (PPV) across all classes remained above 98.9%, while the false discovery rate (FDR) was consistently below 1%, confirming robust precision and low misclassification. Building on the classification output, a shallow feedforward neural network (FFNN) was developed using Bayesian Regularization to provide continuous predictions of contact position and force. The input matrix (4 ×1,404,930) included the three normalized flux components and the class label predicted by the KNN model. The dataset was divided into training (75%), validation (15%), and testing (15%) subsets. A grid search was performed to optimize the network architecture by varying the hidden layer size from 20 to 68 (in increments of 4) and the learning rate from 0.00001 to 0.001 (in increments of 0.0001), with the number of training epochs fixed at 300. The bestperforming configuration, 54 hidden units and a learning rate of 0.00071, resulted in a mean squared error (MSE) of 0.1030 and a correlation coefficient (R) of 0.9977. These findings demonstrate the sensor’s strong capability to detect and interpret multiple simultaneous contacts, supporting its deployment in real-world tactile sensing applications.. IV. CONCLUSION This study explored a learning-based approach for multitouch characterization using a soft magnetic tactile sensor. An experimental framework was developed to evaluate contact localization and force estimation across multiple simultaneous touch configurations. While signal superposition offers a rough approximation, data-driven models achieved higher accuracy. The results confirm the feasibility of reliable multipoint sensing with minimal hardware, suitable for soft robotics. However, limitations remain in sensitivity to misalignment, scalability to larger areas, and performance under dynamic or noisy conditions. Future work will address these to improve real-world deployment. ACKNOWLEDGMENT The research was partially funded by the European Union - NextGenerationEU and by the Ministry of University and Research (MUR), National Recovery and Resilience Plan (NRRP), Mission 4, Component 2, Investment 1.5, project “RAISE - Robotics and AI for Socio-economic Empowerment” (ECS00000035). REFERENCES [1] Guo J, Yang C, Dai Q, Kong L. Soft and stretchable polymeric optical waveguide-based sensors for wearable and biomedical applications. Sensors. 2019 Aug 30;19(17):3771. [2] Ha KH, Huh H, Li Z, Lu N. Soft capacitive pressure sensors: trends, challenges, and perspectives. ACS Nano. 2022 Mar 9;16(3):3442-3448. [3] Kim S, Shin H, Song K, Cha Y. Flexible piezoelectric sensor array for touch sensing of robot hand. In: 2019 16th International Conference on Ubiquitous Robots (UR) 2019 Jun 24 (pp. 21-25). IEEE. [4] Stassi S, Cauda V, Canavese G, Pirri CF. Flexible tactile sensing based on piezoresistive composites: A review. Sensors. 2014 Mar;14(3):5296-5332. [5] Ma Z, Zhang Y, Zhang K, Deng H, Fu Q. Recent progress in flexible capacitive sensors: Structures and properties. Nano Materials Science. 2023 Sep 1;5(3):265-277. [6] Man J, Chen G, Chen J. Recent progress of biomimetic tactile sensing technology based on magnetic sensors. Biosensors. 2022 Nov 21;12(11):1054. [7] Chen Y, Huang Q, Khawaja AH. An interference-rejection strategy for measurement of small current under strong interference with magnetic sensor array. IEEE Sensors Journal. 2018 Oct 17;19(2):692-700. [8] Elzwawy A, Rasly M, Morsy M, Piskin H, Volmer M. Magnetic Sensors: Principles, Methodologies, and Applications. In: Handbook of Nanosensors: Materials and Technological Applications 2024 May 25 (pp. 891-928). Cham: Springer Nature Switzerland. [9] Gongal A, Amatya S, Karkee M, Zhang Q, Lewis K. Sensors and systems for fruit detection and localization: A review. Computers and Electronics in Agriculture. 2015 Aug 1;116:8-19. [10] Ribeiro P, Cardoso S, Bernardino A, Jamone L. Fruit quality control by surface analysis using a bio-inspired soft tactile sensor. In: 2020 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) 2020 Oct 24 (pp. 8875-8881). IEEE. [11] Du, Chengjin, et al. “Design and Optimization of a Soft Magnetic Tactile Sensor.” 2023 IEEE SENSORS. IEEE, 2023. Comparable Observation (Fz) Comparable Observation (Bx) Comparable Observation (By) Comparable Observation (Bz) Flux Superposition Observation Flux Superposition Observation Force (N) Flux (Guess) # Dataset of multi-touch cases # Dataset of multi-touch cases # Dataset of multi-touch cases # Dataset of multi-touch cases # Dataset of multi-touch cases # Dataset of multi-touch cases Percentage (%) Percentage (%) Percentage (%) Percentage (%) 110