A Fundamental Experiment on Contact Position Estimation on Vision based Dome-type Soft Tactile Sensor using Ready-made Medium

Tactile sensors are critical components in robotics fields. Recently, soft tactile sensor utilizing vision is actively developed for safe human machine interaction. Some researches use novel custom-made medium in order to achieve tactile sensing. Deep learning can recognize pattern from any vision data when it has sufficient dataset, i.e., the system does not require specific pattern embedded hardware for the pattern recognition. To achieve soft tactile sensor's economical application for robot fingers, this paper presents a fundamental experiment on contract position estimation on vision based dome-type soft tactile sensor utilizing ready-made silicon as a medium and convolutional neural network. In order to estimate and classify the contact position, convolutional neural network (CNN) was applied. The modified VGGNet architecture was coded using Tensorflow and Keras. 1000 images were taken to train the modified VGG network; 200 images were taken for each neutral, left, right, lower, upper direction. For each direction, fingertip, pencil, ruler, and table corner were utilized to capture various situations. After checking the results of the test set, the trained model was applied to the embedded board and checked the contact position estimation in real-time. The experiment showed high accuracy on classifying the con-tact position of the vision based dome-type soft tactile sensor in real time. This contact position estimation system will be critical for the finger-typed robots since the system is reasonably small and it will reduce significant amount of manufacturing cost for the safe human machine interaction system. For the future work, we will acquire more image data and apply more advanced network architecture to improve accuracy.