Research On Deep Learning-based Grasp Control Methods For Five-fingered Humanoid Robot Hands

Grasping an unknown object is one of the basic abilities of humanoid robot hands to perform various complex tasks.However,increasingly complex task environments place greater requirements on robotic-grasping research,which aims to control robots so as they can grasp objects smoothly and accurately.In this thesis,two five-fingered humanoid robot hands are designed as the end-effectors of the manipulator,and the deep learning-based objectdetection method is proposed to apply in the grasping detection.In addition,two intelligent devices,data gloves and a brain-computer interface can effectively improve the flexibility of robotic grasp.Firstly,according to the human hand structure this thesis first designs a five-fingered humanoid robot hand model with 21-degrees of freedom(DOF).The forward and inverse kinematics models are established by D-H(Denavit-Hartenberg Matrixmatrixs),it realizes the models motion control.The classification and regression tree(CART)mapping methods map the glove data to the joint-rotation angles to achieve the real-time motion control.In addition,the decision tree model is also trained for the gesture control of the humanoid robot hands.Secondly,two object-detection networks,named Faster R-CNN(Faster Regions with Convolutional Neural Network)and SSD(Single Shot Multibox Detector)are trained for grasping detection.And then the contour recognition and four-layer convolutional neural network are applied to determine the optimal five-dimensional representation of robotic grasp.What's more,to meet the requirements of real-time grasping detection,this thesis trained a lightweight semantic segmentation network ENet(Efficient neural network).The edge and gravity generation methods are proposed to generate candidate regions.By comparing the three kinds of neural networks,it is shown in the simulation that the deep learning-based objectdetection method can effectively improve the accuracy of the robotic grasp.In addition,to satisfy various industrial control requirements,this thesis proposes a brain-computer interface control method.The ID3 random forest and neural network command-recognition model are trained for analyzing EEG(electroencephalogram).Finally,this thesis designs a five-fingered humanoid robot hand grasping simulation system.The system integrates the five-fingered humanoid robot hand model,data glove,brain-computer interface and deep learning-based grasping detection methods.All the proposed research methods are validated in the simulation system.In the simulation,cooperating with a 6-DOF manipulators,the five-fingered humanoid robot hands can correctly complete the grasping operation.