Design Of Robot Assembly Line Automatic Loading/unloading System And Reinforcement Learning Motion Control

With the rapid progress of human society,more and more goods are needed to be loaded/unloaded and transported in factories,but human will feel tired and cannot always maintain a high degree of accuracy,in addition,certain working environment is full of danger and is not suitable for human operations.In order to improve the process efficiency of transporting and loading/unloading goods,in this article,a robot assembly line automatic loading/unloading system is built,a virtual simulation is also performed,in proportion to the physical system,to reduce the production risk before the industrial assembly line is deployed.In the actual crawling process of the robot,there will be many unstructured and diverse working environments,and it is necessary to study and strengthen the learning ability of the robot.First of all,this thesis introduces the hardware composition and software design of the robot assembly line automatic loading and unloading system.The hardware system consists of a linear conveyor and a storage turntable that can rotate 360°.The handling system uses KUKA KR6 R900 robot and KUKA LBR IIWA 7 R800 robot.The control system uses PLC(Programmable Logic Controller)control cabinet.The software design adopts the method of sub-modules,each module is separately programmed,the friendly HMI(Human Machine Interface)is designed,and the robot assembly line automatic loading/unloading system can be operated cyclically.Secondly,based on V-REP(Virtual Robot Experiment Platform),the physical model of KUKA KR6 R900 robot and storage turntable is established,and the robot assembly line automatic loading/unloading simulation system is reduced in proportion to the virtual experiment platform.The inverse kinematics analysis of the robot is carried out.The damped least squares method is used for the inverse solution calculation of the robot.The collision detection of the robot is also carried out.The four physics engines in V-REP are introduced.Through analysis and comparison,the Newton engine is finally selected from the stability of the entire system operation.The trajectory planning of the robot uses cubic spline interpolation in the joint space,which makes the displacement-time curve of the robot have good second-order derivability and smooth and accurate motion.Finally,the characteristics of various mainstream reinforcement learning algorithms are analyzed and compared.Based on the controllability and stability requirements of continuous motion,DDPG(Deep Deterministic Policy Gradient)algorithm is selected for reinforcement learning of robot fixed-point motion control.In the Python environment,a two-link robot is used as an example to train the designed deep reinforcement learning method.Using the method of comparative analysis of control variables discusses the difference between the convergence speed of the reward method for guiding the robot to approach the target point quickly and with or without constraints.The final result proves that the reinforcement learning training with constraints is better.