Research On Object Recognition Based On Robot Tactile Perception

Perception of objects is the premise for the robot to perform fine operation tasks.In various perception modes,touch can not only help the robot interact with objects,but also can reflect the physical properties of objects.With the rapid development of sensor technology,tactile sensor has gradually become one of the important components of robot.The tactile sensor with high sensitivity and high resolution makes the robot obtain more information and can perform a wider range of tasks.Therefore,the analysis and research of acquiring,processing and using tactile information are of great significance to enhance the perception of robot and improve the level of robot intelligent interaction.Focusing on the subject of object recognition based on robot tactile perception,this thesis established a robot grasping recognition experimental system including Kinova manipulator,Kinova dexterous hand and Numa Tac tactile sensor.The object recognition problems of soft/hard,category,multi-attribute and robot grasping based on haptics were studied.For the objects recognition problem of soft/hard and category based on haptics,a tactile dataset was established by collecting and analyzing the contact signals between Numa Tac tactile sensor and 30 objects,and a multi-scale convolution neural network model was proposed.The model was trained and tested on the dataset.The results showed that the recognition accuracy of the model for soft/hard of objects was 97.0%,and the recognition accuracy of objects categories was 97.3%.The recognition accuracy of comparison model Mobile Net＿v2 for soft/hard of objects was 92.6%,and the recognition accuracy of objects category was 92.8%.The recognition accuracy of comparison model Res Net18 for soft/hard of objects was 95.0%,and the recognition accuracy of objects category was 94.1%.The above comparison proved that effect of multi-scale convolution neural network model was better.To address the problem of multi-attribute recognition of objects based on haptics,human perception experiment was first designed.The purpose of the experiment was to analyze and determine the multi-attribute labels of objects to form a tactile dataset containing 15 objects.In order to recognize the temporal information in the haptic samples,a parallel long short-term memory neural network model and a parallel onedimensional convolutional neural network model were proposed.The models were trained and tested.The experimental results were as follows: The Auc of the parallel long short-term memory neural network model reached 0.884 for multi-attribute recognition and the highest Auc of single attribute reached 0.960.The Auc of the parallel onedimensional convolutional neural network model reached 0.988 for multi-attribute recognition and the highest Auc of single attribute reached 1.000.The above comparison proved that effect of parallel one-dimensional convolutional neural network model was better.To address the the problem of robot grasping recognition based on touch.In order to make the robot have the ability to recognize objects in the real environment,the Kinova robot motion node for planning fixed-point grasping action,the Numa Tac tactile sensor sensing node for transmission of tactile data,the data processing node and the neural network node for identifying tactile samples were developed through the robot operating system.All nodes communicate with each other to manipulate the robot to complete the task of fixed-point grasping and recognition of objects.The experimental results showed that the robot’s soft/hard recognition accuracy of objects reached 84.0%,the recognition accuracy of object categories reached 92.0%,and the Auc of object multi-attribute recognition reached 0.800.It showed that the robot based on touch can effectively complete the task of object recognition.