Effect of Pattern on Resolution of Tactile Image Sensor Based on Deep Learning

Tactile image sensors are efficient and useful in the robotics and health-care fields due to their advantage of being impervious to electromagnetic fields. The functionality and applications of the sensors have been extensively researched. Furthermore, the pattern with the arrangement of sizes, colors, shapes, and densities is like the “heart” of tactile image sensors, which immediately determines the results of detecting the position and the force. However, the effect of patterns on the resolution of tactile sensors has not been fully investigated yet. Hence, we experimentally studied the effect of different patterns on the resolution of tactile image sensors due to difficulties in theoretical analysis. Optical fibers and cameras were used to integrate flexible sensors with different features, which are divided into 61 dots. The sensors are put through their trials in various positions and forces, and the data is processed using deep learning methods. The sensors with four different patterns were all successful in distinguishing pressure, shear, and torque. More crucially, the root mean square error (RMSE) values were used to compare the impact of the various patterns on the sensors. All sensors with four types of patterns exhibited outstanding resolution for positions, normal force, shear forces, and torsion, which should be due to the optimization of the neural network and sensor structure. The seven position and force parameters of these four patterned sensors have the same order of magnitude of RMSEs, which suggests that the pattern complexity has negligible influence on the tactile image sensor resolution for these factors in our sensor structure. Moreover, we have demonstrated the outstanding resolution of the sensor with a pure color pattern. This work not only guides the pattern designs of forthcoming tactile image sensors but also enables the mass manufacture of flexible sensors for intelligent robots.