A Skin-Inspired PDMS Optical Tactile Sensor Driven by a Convolutional Neural Network

• Published in: IEEE sensors journal Vol. 24; no. 6; pp. 8651 - 8660
• Main Authors: Yue, Shichao, Xu, Minzhi, Che, Zifan
• Format: Journal Article
• Language: English
• Published: New York IEEE 15.03.2024; The Institute of Electrical and Electronics Engineers, Inc. (IEEE)
• Subjects: Adaptive optics; Algorithms; Artificial neural networks; Automation; Cameras; Data processing; Electromagnetic interference; Force; Light emitting diodes; Light sources; Machine learning; Manufacturing engineering; Neural network; Neural networks; optical fiber; Optical fibers; Polydimethylsiloxane; Robot sensing systems; Robotics; Sensors; Spatial resolution; tactile sensor; Tactile sensors; Tactile sensors (robotics)
• ISSN: 1530-437X; 1558-1748; 1558-1748
• DOI: 10.1109/JSEN.2024.3355555

Tactile sensors play a crucial role in enhancing the integration of automation, robotics, and biomedical equipment, particularly in perceptual functions. Optical fiber-based tactile sensors have gained significance due to their robustness and immunity to electromagnetic interference. However, existing optical fiber-based tactile sensors face limitations related to bio-imitation, scalability, and precise data processing algorithms. This study introduces a novel skin-inspired polydimethylsiloxane (PDMS)-manufactured tactile sensor utilizing a structured light source with low-cost light-emitting diodes and a multimode optical fiber, coupled with tactile information processing through a trained convolutional neural network (CNN). Specklegram images captured from the optical fiber are analyzed for force amplitude and tactile location. The CNN is trained, validated, and tested, achieving accuracies of 99.6%, 99.5%, and 99%, respectively. The tactile sensor demonstrates a spatial resolution of 2 mm and a force-sensing range up to 3 N. The confusion matrix, based on classification results, reveals only three misclassifications out of 315 tests, indicating a mean absolute error (MAE) of 0.95%. The spatial resolution and force-sensing capabilities, coupled with the machine learning approach of the proposed tactile sensor, showcase promising potential for future applications in tactile embodiment.