Research On Flexible Robotic Assembly Method For Complex Components

Robotic flexible assembly technology plays an important role for improving the production quality and efficiency of industrial products,which provides technical support for the automation and intelligent development of assembly industry.During the assembly process of complex components with irregular shape and diverse materials,it is difficult to model robot manipulation due to the complex matching relation,the compact space and rich contact state.Therefore,how to deal with the contact diversity and state uncertainty is one of the most important problems in the robot flexible manipulation of of complex components.For the robot flexible assembly requirements of complex components,the contact interaction mechanism has been deeply analyzed and the robot flexible assembly system was constructed.In addition,some methods have been studied in this paper,including skill learning of robot assembly posture adjustment,assembly skill optimization and online assembly strategy.Some problems,including the diversity of contact state,posture uncertainty and information incompleness,have been effectively solved.The methods were verified on the robot flexible assembly platform which was constructed.The experimental results showed that this technique could perform quite well.The main work of this paper is as follows:(1)The robot flexible assembly system of complex components.The assembly types of complex parts are diverse,mainly by clamping,inserting,mounting and embedding.Robotic assembly operation has the characteristics of the inaccurate position of components,the change of system stiffness and the complexity contact state.The interaction mechanism was explored under multiple constraints such as geometry,force and environment.The assembly contact characteristics were analyzed.The constraint conditions such as the object characteristic and operation safety are emphatically studied in the robot assembly process.The flexible assembly simulation and physical verification platform based on KUKA LBR iiwa was built,which provided platform support for algorithm verification in subsequent chapters.(2)A robot skill learning method of assembly posture adjustment based on deep Q-network.During the robot assembly process,it is difficult to establish an accurate physical model due to the diversity randomness of contact states.Markov model was constructed from assembly environment,contact state and assembly action space in assembly contact state.In addition,due to the position deviation of gripping device and target parts,contact deformation,the robot does not have the ability to adjust its behavior action based on current state.A skill learning method of robot assembly posture adjustment based on deep Q network was proposed.The policy network optimization model was established,and the reward functions of assembly displacement were designed.The algorithm model was verified in simulation and physical platform.The results show that the robot assembly posture adjustment skill was acquired,and the environmental adaptability of the robot flexible assembly was improved.(3)Flexible assembly skill optimization of robot based on prior knowledge.In essence,the robot skill learning based on data-driven is to build a nonlinear mapping from state to action,which needs mass data.There is a problem of low learning efficiency in robot learning because of large-scale sampling of physical robot systems.A flexible assembly skill optimization method incorporating prior knowledge was proposed.The optimization was carried out from data storage and training of experience pool and update of prior knowledge base.The ineffective exploration was reduced in the process of robot assembly skill learning,and the data utilization effiency was improved.The fasten assembly experiment was carried out in the physical assembly platform.The results show that the robot skill optimization algorithm with assembly prior knowledge is effective.The leraning efficiency was 30%higher than traditional DQN,and 16.7%higher than Replay-DQN algorithm.(4)A robot assembly strategy based on multimode information description.Aiming at the problem of incomplete information of the assembly process,a robot assembly strategy with multimode information description was established based on vision and force information.The method of robot assembly skill acquisition with continuous action space was studied.The physical experimental verification was performed in the flexible clamping assembly of small circuit breakers.In the non-contact stage,the robot manipulator moves at a higher speed in free space to locate the initial pose of workpieces with vision information.The information changed to be force to adjust the robot assembly posture in contact stage.And assembly quality was detected by visual image.The results showed the assembly strategy with multimode information description effectively improved the robotic flexible assembly capability.