Research On The Design And Grasping Planning Of A New Three-finger Underactuated Hand

With the advancement of Made in China 2025,collaborative robots that can work collaboratively with people have developed rapidly.However,there is relatively little research on end fixtures for collaborative robots.On the one hand,the research hotspots for mechanical ontology tend to be more challenging humanoid dexterous hands,and there is a lack of research on end fixtures for industrial use;on the other hand,the research on gripping planning focuses more on the motion planning of the robotic arm,and there are few studies on the gripping strategy of the end fixture.This paper designs a new type of under-actuated smart gripper,and proposes a gripping planning based on its gripping characteristics.The main work and innovations are as follows:(1)Propose a brand new underactuated smart gripper structure.The under-actuated dexterous gripper drives 4 degrees of freedom with 10 degrees of freedom to achieve multiple gripping modes;the single-drive multi-degree-of-freedom underactuated finger grip is the key part of the dexterous gripper,and the dual-stage drive link uses elastic elements as knuckles additional weights are used to achieve the "minimum energy consumption" effect;the end knuckles remain perpendicular to the palm surface and are completely passively retracted,providing a "fingertip gripping" mode;for the state of finger fixtures in different contact modes,analyze the position and speed of the gripping knuckles to optimize the gripping space;analyze the gripping force,optimize the parameters of the elastic element,make the movement rules of different knuckles reasonably fit,flexible and stable.Use Adams to carry out the grab simulation experiment to verify the movement characteristics of the finger fixtures and the stability of the grab.After determining the mechanism,complete the overall structural design of the underdriven dexterous hand and complete the production of the prototype.(2)Research on grasping position planning based on force closure and maximum stability for planar polygons.Established a mathematical model for grasping analysis,studied the physical meaning of the stability index,and proposed an optimization algorithm for grasping position based on the stability index.The planning of the grabbing position of different plane polygons can get the grabbing position with better grabbing performance,which verifies the effectiveness of the algorithm.(3)For the grasping of curved contour objects,a grasping planning based on deep reinforcement learning is proposed.Comparing the effects of DQN and DDPG in the grabbing problem,for the problem that the image state dimension is large and overwhelms the gripping posture state,a non-fully connected network structure for state classification is proposed to verify its feasibility and the problems in this thesis advantage.(4)Research and experimental verification of grasp planning under V-REP simulation platform.Build Tensorflow deep reinforcement learning experiment platform and V-REP robot simulation platform and conduct joint simulation;for the input image layer of the captured image,three image processing methods were used to experiment,respectively,adding a convolutional neural network in front of the Actor-Critic network,extracting edge features with a single convolution kernel,and sharing the same convolutional neural network with the Actor-Critic network.Comparing the effects of different network structures on the success rate of crawling simulation experiments,it is verified that the way that the Actor-Critic network shares the same convolutional neural network has better training results.