Research On Robot Autonomous Assembly Technology Based On Multimodal Perception And Learning

With the continuous development of intelligent manufacturing and Industry 4.0,robots are gradually being applied to complex assembly tasks.However,traditional robotics technology cannot meet actual application requirements.In recent years,under the premise of continuous breakthroughs in artificial intelligence technology,the industry and academia are exploring how to realize autonomous assembly operations of robots through artificial intelligence technology.The existing robot autonomous assembly technology relies heavily on fixed assembly actions or preset processes.There are bottlenecks such as low assembly success rate,insufficient task generalization and poor environmental adaptability.In view of this,based on the theory of multi-modal perception and deep reinforcement learning,this paper carries out research on autonomous assembly technology of robots based on multi-modal perception and learning.Modal perception technology performs feature extraction and fusion of multi-mode heterogeneous sensor information such as vision,force/torque,and robot body kinematics in the robot assembly process,and uses the deep deterministic strategy gradient algorithm(DDPG)to perform the robot assembly strategy model Learn to realize autonomous assembly of robots.The main research contents of this paper include:(1)An analysis of the research status of autonomous robot assembly technology and multimodal perception technology;the basic knowledge of deep reinforcement learning theory and its application in autonomous robot assembly tasks.(2)The robot autonomous assembly platform is built,kinematics modeling,and the visual positioning system is designed according to the requirements of the assembly task.The internal and external parameters of the camera are calibrated,and a coordinate system with the robot base as the base standard is established.Through Kinect V1 camera,Kinova jaco 6robot,computer host and other related components,an experimental platform for robot shaft hole assembly and electric control cabinet component assembly was built.(3)Based on multi-modal perception technology and DDPG algorithm to realize robot autonomous assembly operations,focusing on the research of multi-modal perception technology network and deep deterministic strategy gradient algorithm network and process.(4)Construct a robot shaft hole assembly system and an electrical cabinet component assembly system,including assembly process design,platform construction and calibration,data acquisition and processing,and result analysis.In view of the influence of different learning rate parameter settings on the robot assembly reward value,the relative distance of the shaft hole and the success rate.The results show that when the learning rate is set to 0.001,the simulation robot shaft hole assembly has the highest success rate of 90%;the robot shaft hole assembly has the highest success rate of 85%;the robot electrical cabinet component assembly success rate is 80%.