Tactile Object Classification Methods Based On Tactile Sequences

Tactile sensing is a kind of important feeling for robots, and the system of tactile sensing is a basic medium for environmental communication. The technology of tactile sensor in robotic decterous hand is improving nowadays, and providing more accurate descriptions for object surface. The robot dexterous operation ability will be improved by dealing with tactile senses, which is of great significance to build a more intelligent robot system. This work is based on State 863 project ’The Fine Operation of Robot Arm’, the tactile information is collected and algorithms are designed to realize the classification of tactile sequences.Our experimental platform is based on the Schunk robot arm with 7 degree of freedom fixed with the BH8-280 dexterous hand of 4 degree of freedom in the state key laboratory of department of computer science and technology of Tsinghua University. The database TSH-16 is formed with grasping 16 objects of different positions and poses, which is superior to other similar databases. The linear dynamic system with Martin distance is first used in the modeling of tactile information, and combined with k nearset neighbor to form the Martin-KNN algorithm which is used to classify the tactile sequences. The linear dynamic system is proved to perform better at the description of the temporal spatial characters of the tactile sequences when compared with the DTW-KNN algorithm. Then, after modeling the tactile information by the linear dynamic system with Martin distance, a bag-of-system model is used to describe the features of tactile sequences, and with the popular machine learning algorithm, support vectoer machine and extreme learning machine are used to train the classifiers, two algorithms called SVM and ELM are formed, and these algorithms are also first used in the classification of tactile sequences.The tactile database TSH-16 is used for testing those classification algorithms designed above, and the other four kinds of databases are also used in our paper, the complex matrixes at the best classification accuracy are plotted to show the recognition rate of each object. The results show that all the algorithms above can realize the classification of tactile sequences and among them, the machine learning algorithm based on extreme learning machine performs best. In the future researches, this kind of algorithm will be applied to real time operaction, and the object can be recognized during the grasping process.