Research On Robotic Grasping And Intelligent Assembly Based On Deep Reinforcement Learning

The rapid development of artificial intelligence has promoted industrial transformation in all walks of life,which has changed the way of life and work for people.To meet the development needs of industrial transformation,accelerating the application of artificial intelligence has become the top priority of development planning for industry.Due to the increase in labor costs,robots are widely used in industry,service industry,medical industry,aerospace and other industries to replace humans to complete many tasks.Traditional robots work in a structured environment,and traditional robot control systems not only need to establish an accurate control model,but also need to perform tedious program compilation and a large number of parameter deployment tasks.However,the fluctuations of control accuracy and sensor errors are inevitable in practical work,which will lead to the failure of the established model.In addition,building an accurate model may be an impossible task for some grasping and assembly problems in unstructured environments.Therefore,improving robot control strategies with artificial intelligence is a research with great research value and application significance,which make robots no longer rely on the model establishment and parameter deployment,and can learn work skills autonomously through training to complete complex work tasks in the unstructured environment.To solve the difficulty of process description for grasping and assembly work in the unstructured environment,this paper analyzes the interaction characteristics between robot and environment in the working process according to the work requirements of assembly task in the unstructured environment.The intelligent grasping and assembly of robot systems based on deep reinforcement learning were constructed.The robot learning grasping and assembly skills and perception fusion were studied.The main research contents and innovations of this paper include the following aspects:(1)The robot system of intelligent grasping and assembly.For the characteristics of assembly task for peg-in-hole in the unstructured environment,such as uncertainty of position,diversity of alignment attitude and uncertainty of contact state,the working principle of compliant action in the assembly task and the control principle of the impedance control system are analyzed,and the assembly sub-tasks of each stage are divided.The working characteristics of visual perception and tactile perception are studied,as well as the difficulty of applying them in each sub-task of peg-in-hole assembly.The various contact states between the assembly hole and the assembly peg are deeply analyzed,as well as corresponding robot control strategies.In addition,the constraints between the assembly parts and the environment in the process of grasping and assembly are considered,as well as the constraints of safe operation in the working process.The simulation and physical experiment platform of intelligent grasping and assembly for the robot is constructed,which provided an experimental platform for strategy training and verification of subsequent research.(2)Intelligent grasping based on deep reinforcement learning.When robots select a grasping position,the traditional grasping strategy only considers whether the robot can successfully grasp the target object,but does not consider the impact of the grasping position on the alignment adjustment of the peg-in-hole assembly after successful grasping.Aiming at the peg-in-hole assembly task in unstructured environments,an intelligent grasping strategy based on a deep reinforcement learning is proposed,and a fully convolutional neural network model is constructed.On the basis of ensuring that the robot can successfully grasp the target part,a reward function is designed to introduce assembly constraints into the grasping decision,so that the robot can consider the impact of the grasping position on the assembly efficiency when deciding the grasping position.The strategy training and testing are performed on the built simulation platform and physical experiment platform.The test results show that the robot’s grasping position selection is compressed into a smaller restricted area after the introduction of assembly constraints,and the robot can obtain higher assembly efficiency by selecting the grasping position within this area.(3)Assembly research on combined perception of single-view vision and tactility.It increases the difficulty of perceiving environment information for the robot due to the diversity of alignment attitude and contact state and the uncertainty of disturbance torque fluctuation,which makes it difficult for the robot to accurately establish the task model.In addition,robots can only use one perception method in a certain assembly stage according to its working characteristics because of the respective limitations of visual perception and tactile perception,which further increases the difficulty of describing the assembly process.To solve this difficult problem,this paper proposes a combined perception assembly strategy based on deep reinforcement learning,which can establish a mapping relationship between visual perception and tactile perception in the alignment stage of assembly,and can solve the difficulty of describing the assembly process to improve the robot work efficiency of peg-in-hole assembly.(4)Assembly research on combined perception of multi-view vision and tactility.In the process of peg-in-hole assembly task,the visual perception may lead to the lack of visual feature information due to the visual occlusion of assembly parts,which has affected the adjustment judgment of the decision-making system for the alignment between the peg and the hole.To solve the problem of lack of visual feature information,an assembly strategy of combined perception with multi-view based on deep reinforcement learning is proposed.According to the requirements of the alignment adjustment for peg-in-hole assembly,the proposed strategy added the assistant view of visual perception to compensate for the lack of visual feature information,thereby improving the assembly efficiency of the robot.The strategy training and test experiment are executed on the simulation system and the physical experiment platform.The experimental results show that the assembly strategy of combined perception with multi-view can make up for the lack of visual feature information,and can further improve the work efficiency and stability of the peg-in-hole assembly.