Research On Robot Self-learning And Deep Reinforcement Learning Algorithm For Assembly

The robot replaces the manual and intelligently completes the assembly task which is the goal pursued by mankind at the beginning of the robot.At present,robots are mostly used in the rough processing fields of handling,welding,grinding,etc.in industrial production,while robots have not been widely used in the field of assembly requiring high precision.In the current era,assembly is increasingly requiring robots to be intelligent,especially in the 3C industry,where small batches,customization,and short cycles become the biggest challenges for robot assembly.With the development of artificial intelligence technology,AI(Artificial Intelligence)algorithm can give robots a variety of human capabilities to achieve the ideal robot.Among them,deep reinforcement learning is a representative of general-purpose artificial intelligence algorithms,which is currently a research hotspot in this field.The deep reinforcement learning algorithm,the unique training mechanism of selflearning and active exploration,fundamentally solves the dilemma of data acquisition difficulties.On the other hand,the end-to-end learning method has great development room in the field of robot crawling and assembly,which is difficult to establish a math model.As an important component of robot assembly,we use a variety of visual information processing algorithms to compare robot visual capture algorithms.In order to improve the ability of robots to handle messy environments,we use multiple operations to coordinate task such as cleanup,capture,etc.The robot simulation environment is built and the effectiveness of the proposed algorithm is verified on the simulation and real physical robots.In complex assembly tasks,it is time-consuming and laborious to train the robot to master the corresponding assembly skills from zero-based by the deep reinforcement learning algorithm.To solve the problem,demonstration learning that provides a priori knowledge of the robot becomes the best choice.At present,although there are still many challenges in the field of demonstration learning,it is still a hot research topic in the field of robot operation,especially in the case of increasing demand for industrial applications.In the intelligent assembly of robots based on demonstration learning,we solve the problems of demonstrating the difficulty of fine motion recognition of human hand and the uncertainty of robotic manipulation environment.We propose a demonstration understanding algorithm based on multi-source information fusion to process bioelectrical and inertial information collected by the presenter during the demonstration.Inspired by the different tasks of the different components of the arms when the human complete the assembly task,we divide the presentation into two parts:large movements and fine movements.The inertial information and the forearm surface myoelectric signal are thus used to capture the presenter’s operation,then the pre-trained operational gesture classification network is used to identify the fine motion of the presenter’s hand.Finally,according to the deep reinforcement learning control algorithm designed in the thesis,the robot’s operational adaptability in an unknown environment is improved.The experimental results show that the method can complete relatively complex robot assembly tasks with only one single demonstration,and the intelligent assembly of robot is realized with high success rate under slight error.