A welding control method and system for a dual-arm robot

Abstract

A welding control method and system of a dual-arm robot, which controls the dual-arm robot to complete a welding task on a non-standard workpiece, comprising the following steps: 1) performing structural analysis and disassembly on 3D data of the target workpiece to generate independent parts; 2) based on reinforcement learning, taking the dual-arm robot as an agent, training and optimizing an optimal assembly strategy of the parts, and outputting an assembly task plan and a welding task plan as a control strategy result; and 3) based on the control strategy result, deciding a motion trajectory of a mechanical arm of the dual-arm robot, and sequentially assembling and welding the multiple parts in the assembly order to obtain a finished product of the target workpiece. The present application realizes full-automatic welding control of the dual-arm robot on a complex non-standard workpiece, completes automatic analysis and disassembly of complex parts, and assembly and welding tasks, and no longer needs to rely on traditional methods such as manual welding and jig pre-clamping, and can further improve welding quality consistency.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of robots, and relates to a mechanical hand, in particular to a welding control method and system of a dual-arm robot. BACKGROUND

[0002] Currently, laser welding technology mainly includes handheld laser welding and robot welding. The handheld laser welding relies on the cooperative operation of the hands of an operator, and can flexibly cope with complex workpieces and small-batch production welding tasks. The robot welding is usually based on a single-arm collaborative robot or a single-arm industrial robot, and the system thereof is composed of a robot body, a welding workstation and a supporting fixture. In the robot welding process, pre-clamping or pre-spot welding positioning by a tooling fixture must be performed as a welding preparation, and then a welding task is realized by means of pre-programming. This welding method is mainly suitable for the welding manufacturing of batch and standardized products.

[0003] Robot welding is one of the important directions of automation development. However, when facing the welding task of complex non-standard workpieces, the traditional robot welding method is often difficult to perform. The main reason is that the welding forming process of complex workpieces is complex, involving the assembly of parts and multiple welding, and it is difficult to realize robot automatic welding by pre-programming. In addition, due to the characteristics of multiple varieties and small batches of non-standard parts, it is necessary to design different task plans for each different part when robot automatic welding is adopted. When welding deployment is performed in this way, there are problems such as long cycle and high deployment cost. Therefore, when facing the welding of complex non-standard workpieces, the general traditional robot welding method is no longer applicable.

[0004] In recent years, with the rapid development of perception technology and artificial intelligence, dual-arm robot technology has been more widely applied and developed. Dual-arm robots can not only complete basic tasks such as sorting and stacking, but also can complete more complex tasks such as folding clothes. However, the current application of dual-arm robots is mainly concentrated in the fields of sorting, grasping, stacking and assembly, and there are still technical bottlenecks in the task of welding, especially the assembly and welding of complex non-standard parts. This is because the welding task needs to disassemble the parts according to the structure of the complex workpiece, in addition to the welding task, the assembly task of the parts also needs to be solved, and the two tasks need to be combined according to the welding process to plan the detailed assembly and welding tasks. However, the existing dual-arm robots do not have this ability. SUMMARY

[0005] The technical problem to be solved by the present application is that in the existing robot welding task, the robot only focuses on the welding task, and for a workpiece product composed of multiple non-standard parts, the existing dual-arm robot control method cannot automatically identify the workpiece product and output the corresponding assembly and welding task plan.

[0006] The technical scheme of the present application is a welding control method of a dual-arm robot, which controls the dual-arm robot to complete a welding task on a non-standard workpiece, comprising the following steps:

[0007] 1) Call the three-dimensional data of the target workpiece product, based on the preset welding process parameters of the target workpiece, perform structural analysis and disassembly on the 3D model of the target workpiece product, generate the independent parts required to form the welding body, and prepare the real parts accordingly;

[0008] 2) Based on reinforcement learning, learn the optimal action strategy for assembling and welding parts with the dual-arm robot as the agent, one arm of the dual-arm robot performs the assembly task, and the other arm performs the welding task, combined with the structural analysis and disassembly results of the 3D model, plan the assembly task and the welding task as the control strategy result, the assembly task planning includes the assembly action trajectory of the assembly sequence and the pose of the corresponding parts, and the welding task planning includes the welding action trajectory of the pose of the corresponding parts;

[0009] 3) Based on the control strategy result, generate the motion trajectory of the dual-arm robot mechanical arm, wherein one arm of the dual-arm robot obtains the real parts according to the assembly task planning and adjusts the pose of the real parts, and controls the other arm to perform welding actions on the current real parts based on the welding task planning, through the iteration of assembly-welding, the multiple real parts are sequentially welded according to the assembly sequence and the pose, and the finished product of the target workpiece is obtained.

[0010] Further, in step 3), the robot operating system ROS platform is also used to simulate and verify the control strategy result, after the simulation verification is passed, the simulation obtained parameters and behavior data are deployed and debugged on the real dual-arm robot system, after the debugging is completed, the welding instruction is issued to the dual-arm robot system again, and the dual-arm robot is driven to execute the cooperative assembly and welding operation sequence in the real physical environment.

[0011] Further, step 2) based on reinforcement learning, learning the optimal strategy for assembling and welding parts with the dual-arm robot as the agent includes:

[0012] Define a state description function, take the TCP pose, speed of the end of the mechanical arm and the welding assembly state as the state variables, construct the robot action state function of the assembly and welding process, and take it as the judgment basis for reinforcement learning training;

[0013] Define an assembly reward function for quantitatively evaluating the advantages and disadvantages of the part assembly state, the TCP pose of the robot end and the speed in the assembly and welding process;

[0014] Take the dual-arm robot as the agent, perform reinforcement learning on the optimal action strategy for assembling and welding the parts operated by the dual-arm robot, and output the assembly task planning and the welding task planning as the control strategy result.

[0015] Further, in the control strategy result, for the assembly task planning, the following steps are included:

[0016] Step A1, obtain the feature information of the virtual model parts disassembled from the 3D model in step 1) and the feature information of the real parts, match the feature information of the virtual model parts with the feature information of the real parts, and establish the correspondence between the virtual model parts and the real parts;

[0017] Step A2, determine the base part as the assembly basis, generate a part assembly strategy for the virtual model parts, and perform robot kinematics modeling, mathematically describe the motion relationship of each part during assembly driven by the mechanical arm, including joint motion and pose transformation of the mechanical arm, for controlling the motion trajectory of the robot;

[0018] Step A3, optimize the generated part assembly strategy and robot kinematics model using reinforcement learning, adjust the assembly strategy according to the assembly reward function through the interaction between the agent and the environment, and if the verification is passed, output the assembly task planning for the real part assembly; if the verification is not passed, return to the reinforcement learning stage for further adjustment and optimization of the strategy until the verification is passed;

[0019] Step A4, obtain the real base part, establish a workpiece coordinate system based on the base part, provide a reference coordinate for assembly operation, and combine the correspondence between the virtual model parts and the real parts to obtain the assembly motion trajectory of the dual-arm robot from the assembly task planning.

[0020] Further, in the control strategy result, for the welding task planning, the following steps are included:

[0021] Step B1, obtain the dual-arm robot and the information of the parts to be welded, including the size, shape, material of the parts, and the motion specification parameters of the welding robot;

[0022] Step B2, establish a welding simulation environment to simulate the relative position relationship and operation conditions of the welding robot and the parts, then identify the weld of the parts in the simulation environment to determine its position and shape characteristics;

[0023] Step B3, generate a welding path according to the identified weld, covering the full length of the weld and meeting the process specifications, then perform forward and inverse kinematics solving to calculate the pose of the welding torch and the motion instructions of the robot joints, and then judge whether the welding pose is reasonable through offline simulation, if not, return to the kinematics solving step for adjustment until the simulation result meets the requirements, and output the welding trajectory corresponding to the identified weld.

[0024] Meanwhile, the welding seam information of the current real part in the assembly process is obtained, matched with the identified welding seam in the simulation environment, the robot motion trajectory in the real environment is determined based on Sim2real, combined with the kinematics characteristics of the mechanical arm and the welding process parameters, the welding process parameter package is called, and the planned current welding trajectory is obtained;

[0025] Step B4, according to the current welding trajectory planned in step B3, the welding task of the current real part is completed, and the welding of the next part is carried out;

[0026] Step B5, repeating steps B3 to B4 until all parts are completely assembled and welded.

[0027] The application also provides a welding control system of a dual-arm robot, comprising a movable leveling platform, a dual-arm robot, a visual perception system, a welding system and an AI intelligent computing unit, the dual-arm robot comprising a left arm and a right arm, the welding system being arranged at the end of the right arm through an end gripping execution system, the AI intelligent computing unit storing at least one instruction or at least one program, and the at least one instruction or program being executed to realize the welding control method of the dual-arm robot.

[0028] Further, the AI intelligent computing unit comprises the following modules:

[0029] An assembly trajectory planning module is used to output the motion trajectory of one arm of the dual-arm robot according to the assembly task planning, and complete the assembly task of the parts, comprising:

[0030] An assembly mechanical arm trajectory generation unit is used to generate the assembly motion trajectory of the mechanical arm according to the current real part and the assembly task planning;

[0031] An assembly trajectory optimization unit is used to optimize the assembly motion trajectory of the mechanical arm by using smoothing processing, speed planning and collision detection algorithm;

[0032] A welding trajectory planning module is used to output the motion trajectory of the other arm of the dual-arm robot according to the welding task planning, and complete the welding task of the parts, comprising:

[0033] A welding mechanical arm trajectory generation unit is used to generate the welding motion trajectory of the mechanical arm according to the current real part and the welding task planning;

[0034] A welding trajectory optimization unit is used to adjust the smoothness, continuity, speed and acceleration parameters of the welding motion trajectory of the mechanical arm to meet the welding process requirements;

[0035] An identification module is used to identify the parts and the welding seam, comprising:

[0036] Component identification unit: Used to identify and classify components by applying image processing and information tagging methods to the acquired component images, shape, and size feature information;

[0037] Weld seam recognition unit: Based on a visual perception system, it detects the position, shape and size of weld seams in real time, identifies weld seam information of real parts, and provides a basis for welding trajectory planning;

[0038] Control module: Used to output control commands to the dual-arm robot based on the control strategy results, including:

[0039] Robot posture control unit: Based on the motion trajectory planned for the assembly and welding tasks and the real-time feedback information from the robot, it controls the posture of the dual-arm robot and performs posture tracking by combining the dynamic model and external disturbance factors;

[0040] Welding process control unit: Sets process parameters such as welding current, voltage, speed, and wire feed speed, and monitors and adjusts them in real time to ensure the stability and consistency of welding process.

[0041] Existing robot welding control methods cannot determine the welding sequence for complex non-standard workpieces, making it difficult to automatically plan welding tasks. Even with pre-setting the welding sequence for non-standard parts, the flexibility and customization of these parts require constant adjustments to the tooling and fixtures to accommodate their assembly. This still makes it difficult to accurately coordinate and plan the motion tasks of a single-arm collaborative robot or a single-arm industrial robot. Furthermore, manual pre-programming is extremely labor-intensive and costly to debug. This invention proposes a welding control method and system for a dual-arm robot, which offers the following advantages:

[0042] 1. End-to-End Welding Task Planning and Execution: This invention automates the entire welding process through reinforcement learning, from task input to welding completion. It automatically plans the optimal path and adjusts welding parameters in real time to ensure welding quality and efficiency. This invention eliminates the need for manual analysis of non-standard workpieces, manual planning of welding schemes, and subsequent programming and deployment to the robot. First, it performs disassembly analysis of non-standard workpieces in a virtual environment and plans the robot's motion trajectory based on reinforcement learning. Then, it matches the strategy from the virtual environment to the real environment, achieving a fully automated welding control method for complex non-standard workpieces using a dual-arm robot: "automatic disassembly - planning assembly + planning welding - collaborative assembly and welding operation." It collaboratively plans the assembly and welding tasks for complex workpieces. For welding tasks of small batches of non-standard workpieces, only a 3D model of the non-standard workpiece needs to be constructed to automatically generate the cyclical collaborative operation task planning for one arm of the dual-arm robot (assembly) and the other arm (welding), completing the automatic analysis and assembly welding of complex components.

[0043] 2. Low-cost intelligent production of non-standard and small-batch products: With the help of Sim2Real technology, combined with random domain enhancement and multi-modal perception fusion, seamless migration from simulation training to real production is realized, production cost is reduced, and flexibility and efficiency in small-batch production are improved. Through the method of the application, 3D data generation is realized by 3D modeling method, which is used for simulation training, combined with task decision, fine-tuning model weld recognition and welding trajectory planning to realize training. Finally, the deployment and implementation of non-standard workpiece welding tasks are completed. Through the method of the application, on the one hand, a control strategy for the coordinated work of a dual-arm robot is automatically generated, and on the other hand, most of the debugging work of the control strategy is carried out in the simulation environment, and the debugging cost is low. The welding labor is released, and manual welding, jig pre-clamping and other means are no longer needed to realize welding, and the welding quality consistency is better, and is no longer limited by factors such as the operation level of traditional manual welders.

[0044] 3. Low-cost lightweight deployment: Optimize algorithm design and hardware configuration, use lightweight deep learning model Segformer and modular design, reduce computing resource demand, reduce hardware cost, improve system robustness and reliability, meet the production needs of modern manufacturing industry for non-standard and small-batch products. BRIEF DESCRIPTION OF DRAWINGS

[0045] Figure 1 It is a general flowchart of the welding control method of the application.

[0046] Figure 2 It is a preferred embodiment diagram of the application.

[0047] Figure 3 It is a flowchart of the welding control method of the application, based on reinforcement learning task planning.

[0048] Figure 4 It is a flowchart of the welding control method of the application, based on reinforcement learning task planning.

[0049] Figure 5 It is a flowchart of the welding control method of the application, based on reinforcement learning task planning.

[0050] Figure 6 It is a flowchart of the welding control method of the application, based on reinforcement learning task planning.

[0051] Figure 7 It is a structure diagram of the dual-arm robot welding control system of the application.

[0052] Figure 8 It is a structure diagram of the AI intelligent computing unit in the dual-arm robot welding control system of the application.

[0053] Figure 9 For a specific embodiment of the present application, the assembly and welding of the parts of the non-standard workpiece "blade" are shown in the assembly and welding process of the first blade.

[0054] Figure 10 For the assembly and welding process of the second blade Figure 9 is shown. DETAILED DESCRIPTION

[0055] The present application provides a planning and control method and system for a dual-arm robot to handle non-standard welding tasks. In view of the challenges faced by complex non-standard part welding tasks, through the "Sim2Real" technology, the 3D model in the virtual simulation environment is seamlessly connected with the complex parts in the real physical environment, combined with the three-dimensional (3D) model and the pre-trained weld recognition model, the weld position and features in the complex parts are quickly and accurately identified, the complex parts are efficiently and accurately disassembled, and the part information for sub-task execution is generated. This process not only significantly improves the disassembly efficiency, but also provides accurate part information for subsequent welding task planning. In addition, the present application further introduces a "brain-like" task layer planning process based on reinforcement learning, which realizes the whole process planning from complex part disassembly to part assembly and welding, and provides an efficient, intelligent and flexible solution for automatic welding of complex non-standard parts.

[0056] At the task planning level, the present application adopts a "brain-like" task layer planning process. This process simulates the decision-making and planning mechanism of the human brain, starting from the disassembly of complex parts, gradually planning the assembly process of the parts, combining the actual motion constraints of the dual-arm robot, optimizing the motion trajectory of the robot, and finally formulating a detailed robot assembly and welding task execution scheme. This whole process planning method not only considers the complexity and diversity of the welding task, but also can dynamically adjust the task execution strategy according to the real working condition, to ensure the efficient completion of the welding task.

[0057] As shown in Figure 1 , the overall process of the dual-arm welding robot welding control method includes the following steps:

[0058] 1) Call the three-dimensional data of the target workpiece product, based on the preset welding process parameters of the target workpiece, the 3D model of the target workpiece product is analyzed and disassembled, the independent parts required to constitute the welding body are generated, and the real parts are prepared accordingly; wherein, the 3D model of the target workpiece is identified by parameters, and the key parameter information of each part is accurately obtained, such as size, shape, position, etc. Then, based on the welding process of the target workpiece, the 3D model is disassembled and decomposed into individual virtual model parts, ensuring the integrity and independence of each virtual model part. Real parts and virtual model parts correspond to each other, waiting for assembly and welding into the target workpiece.

[0059] 2) Based on reinforcement learning, the optimal action strategy of assembling and welding parts is learned by taking the dual-arm robot as the agent, one arm of the dual-arm robot performs the assembly task, and the other arm performs the welding task. Combined with the structure analysis and disassembly results of the 3D model, the assembly task and the welding task are planned as the control strategy results, the assembly task planning includes the assembly action trajectory of the corresponding part assembly sequence and pose, and the welding task planning includes the welding action trajectory of the corresponding part pose.

[0060] Specifically includes:

[0061] Define the state description function, take the TCP pose and speed of the robot arm and the assembly / welding state as the state variables, and construct the state function of the robot action in the assembly and welding process, which is used as the basis for judging the reinforcement learning training;

[0062] Define the assembly reward function, which is used to quantitatively evaluate the assembly state of the parts, the TCP pose and speed of the robot during assembly and welding. As an example, the reward function definition can include the smoothness, continuity of the trajectory, whether it meets the physical constraints of the robot arm, whether it matches the characteristics of the parts and the task requirements, and whether the motion distance is the shortest. If the trajectory is unreasonable, adjust it, if reasonable, save the action data.

[0063] 3) Generate the motion trajectory of the dual-arm robot based on the control strategy results, wherein one arm of the dual-arm robot controls the real parts according to the assembly task planning and adjusts the pose of the real parts, and the other arm controls the welding action on the current real part based on the welding task planning. Through the iteration of assembly-welding, multiple real parts are sequentially welded according to the assembly sequence and pose, and the finished product of the target workpiece is obtained.

[0064] For the same kind of non-standard workpiece, once reinforcement learning is used to obtain the control strategy results, the robot can be deployed to complete the batch production task.

[0065] Further, as Figure 2As shown, the present application also utilizes the Robot Operating System (ROS) platform to perform high-fidelity simulation verification on the planned task strategy. After simulation verification, based on the parameters and behavior data obtained through simulation, deployment and debugging are performed on the real robot system. After debugging is completed, welding instructions are issued to the robot system to drive the dual-arm to perform a collaborative assembly and welding operation sequence in a real physical environment. During the operation process, multi-modal sensor information such as visual sensing and force sensing is integrated to realize real-time closed-loop feedback control of the operation process. Finally, the entire process of automatic welding tasks from component assembly to weld formation is realized.

[0066] The following describes the specific implementation of the present application for planning assembly tasks and welding tasks based on reinforcement learning.

[0067] Reference Figure 3 An assembly task planning flowchart is provided. The specific process is as follows:

[0068] Step A1, obtain the feature information of the virtual model components disassembled from the 3D model in step 1) and the feature information of the real components, match the virtual model component feature information with the real component feature information, and establish a correspondence between the virtual model components and the real components;

[0069] Step A2, determine the base component as the basis for assembly, generate a component assembly strategy for the virtual model components, and perform robot kinematics modeling, mathematically describe the motion relationship of each component during assembly by the robot arm, including joint motion and pose transformation of the robot arm, for controlling the motion trajectory of the robot;

[0070] Step A3, use reinforcement learning to optimize the generated component assembly strategy and robot kinematics model, adjust the assembly strategy according to the assembly reward function through the interaction between the agent and the environment, and perform Sim2real verification on the reinforcement learning results, apply the learning results in the simulation environment to the real scene, if the verification is passed, output the assembly task planning for the assembly of real components; if the verification is not passed, return to the reinforcement learning stage for further adjustment and optimization of the strategy until the verification is passed;

[0071] Step A4, obtain the real base component, establish a workpiece coordinate system based on the base component to provide a reference coordinate for assembly operations, and obtain the assembly motion trajectory of the dual-arm robot from the assembly task planning based on the correspondence between the virtual model components and the real components.

[0072] During the assembly process, the dual-arm robot needs to recognize the real components and correspond them with the components in the simulation environment to determine how to act according to the assembly task planning. The correspondence between the virtual model components and the real components is the basis for component recognition, such asFigure 4 The implementation process is as follows.

[0073] In the part recognition, for the virtual model parts, the shape and size feature information is obtained according to the 3D model disassembly analysis result in step 1), including size, shape and surface texture, to accurately reflect the shape and position characteristics of the parts. At the same time, the geometric feature information of the real parts may be slightly different from the design model due to the deviation in the manufacturing process or the wear in the use, so it needs to be obtained and analyzed separately. By scanning or measuring, the feature information of the real parts is obtained, including shape, size and surface texture. As an embodiment, a robot equipped with a visual sensing device is used to obtain the image of the real part. The real image may be affected by factors such as light, angle, and occlusion. When obtaining the feature information of the real part through the image, a robust image processing algorithm is needed to ensure the accuracy of the feature information.

[0074] Then, the virtual model parts and the real parts are classified and labeled multiple times according to the obtained feature information. In this embodiment, this step is repeated three times to ensure the accuracy and comprehensiveness of classification and labeling. Each classification and labeling may be based on different feature dimensions, such as size, shape, function, etc., in order to distinguish and identify different types of parts more carefully. Through multiple classification and labeling, the accuracy of part recognition can be improved, and more detailed information can be provided for the subsequent matching process. After completing the multiple classification and labeling of the parts, by comparing and matching the corresponding feature information of the virtual model parts and the real parts, the accurate alignment and matching between the virtual model and the real object is realized, and the correspondence between the virtual model parts and the real parts is established, and their similarity and difference are determined. This step ensures the high consistency between virtual design and real production, and provides a reliable basis for subsequent manufacturing, assembly or quality control, etc.

[0075] Referring to Figure 5 The welding trajectory planning flowchart includes the following steps.

[0076] Step B1, obtain the information of the parts to be welded by the welding robot, including the size, shape, material of the parts, and the specification parameters of the welding robot, etc.

[0077] Step B2, establish a welding simulation environment to simulate the relative position relationship and operating conditions of the welding robot and the parts. Then, identify the weld of the parts in the simulation environment and determine its position and shape characteristics.

[0078] Step B3, generate a welding path according to the identified weld, covering the entire length of the weld and conforming to the process specifications. Then, forward and inverse kinematics solving is performed to calculate the pose of the welding gun and the motion instructions of the robot joints. Then, through offline simulation, it is judged whether the welding pose is reasonable. If not, return to the kinematics solving step for adjustment until the simulation result meets the requirements, and output the welding trajectory corresponding to the identified weld.

[0079] At the same time, the weld information of the current real parts during assembly is obtained to ensure its accuracy and integrity. Then match with the identified weld in the simulation environment, determine the robot motion trajectory in the real environment based on Sim2real, combine the kinematics characteristics of the robot arm and the welding process parameters, call the welding process parameter package, get the planned current welding trajectory, and guide the welding operation combined with the robot arm trajectory.

[0080] Step B4, according to the current welding trajectory planned in step B3, complete the welding task of the current real part, and proceed to the welding of the next part.

[0081] Step B5, repeat steps B3 to B4 until all parts are completely assembled and welded. The whole process ensures accurate planning of welding trajectory and efficient and stable welding operation.

[0082] Further, in the above process, a lightweight model is used to obtain weld information from real parts, as shown in the flowchart Figure 6 The specific implementation process is as follows:

[0083] B3.1, carry out high-quality image acquisition work, use professional acquisition equipment to obtain high-resolution images of the weld, and ensure that the images can clearly reflect the details of the weld, providing high-quality data source for subsequent processing. In this embodiment, the image containing the weld to be welded of the current part is acquired by a perception camera;

[0084] B3.2, pre-process the image, including converting the image to a grayscale image to simplify the data structure and highlight the shape features, and using a denoising algorithm to remove random noise in the image to improve the clarity and usability of the image;

[0085] B3.3, weld feature extraction, using image processing techniques to extract key features of the weld from the pre-processed image, including edge contour and texture pattern, etc., and then using a lightweight Segformer segmentation model for image segmentation processing, which is based on deep learning technology and can efficiently and accurately segment the weld area from the complex background, and accurately define the boundaries and scope of the weld, providing a clear visual basis for subsequent detailed analysis. Then identify and analyze the segmented weld, including quantitative analysis of the weld's geometric shape, surface quality, and defect characteristics, to evaluate whether the weld meets the process standards and quality requirements, generate an analysis report, and ensure the accuracy and integrity of the weld for determining whether to weld;

[0086] B3.4, output the results of weld identification and analysis, including the location, size, and quality level of the weld, which are used to determine the real environment robot motion trajectory based on Sim2real in combination with the weld identified in the simulation environment.

[0087] As Figure 7 shown, the present application also provides a welding control system for a dual-arm robot, which includes a movable leveling platform 1, an end gripping execution system 2, a left arm industrial robot 3, a visual perception system 4, a right arm industrial robot 5, a welding system 6, and an AI intelligent computing unit 7. During operation, the movable leveling platform 1 is used to maintain the position of the dual-arm robot, allowing it to operate stably in the field. The visual perception system 4 is responsible for obtaining the state information of the task product, and the AI intelligent computing unit 7 makes decision analysis according to the above control method, finally driving the left arm industrial robot 3 and the right arm industrial robot 5 of the dual-arm industrial robot to operate cooperatively. The end gripping execution system 2 obtains and grips the parts, and the welding system 6 performs the welding task.

[0088] Referring to Figure 8 , the AI intelligent computing unit in the dual-arm robot welding control system of the present application includes the following modules:

[0089] Assembly trajectory planning module: used to output the motion trajectory of one arm of the dual-arm robot according to the assembly task planning, to complete the assembly task of the parts, including:

[0090] Assembly robot trajectory generation unit: generates the assembly motion trajectory of the robot according to the current real part and assembly task planning; in this embodiment, the left arm is used as the assembly robot.

[0091] Assembly trajectory optimization unit: uses smoothing processing, speed planning, and collision detection algorithms to optimize the assembly motion trajectory of the robot.

[0092] Welding trajectory planning module: for planning the other arm motion trajectory of the dual-arm robot according to the welding task, to complete the welding task of the parts, including:

[0093] Welding robot trajectory generation unit: generate the welding motion trajectory of the robot according to the current real part and the welding task planning; in this embodiment, the right arm is taken as the welding robot.

[0094] Welding trajectory optimization unit: adjust the smoothness, continuity, speed and acceleration parameters of the welding motion trajectory of the robot to meet the welding process requirements.

[0095] Recognition module: for recognizing the parts and welds, including:

[0096] Part recognition unit: for recognizing and classifying the parts by using image processing and information marking methods on the acquired part image, shape and size feature information.

[0097] Weld recognition unit: based on the visual perception system, real-time detect the weld position, shape and size, recognize the real part weld information, and provide basis for the welding trajectory planning.

[0098] Control module: for outputting the control instructions of the dual-arm robot according to the control strategy results, including:

[0099] Robot pose control unit: control the pose of the dual-arm robot according to the motion trajectory of the assembly task and the welding task planning and the real-time feedback information of the robot, and combine the dynamics model and external disturbance factors to perform pose tracking.

[0100] Welding process control unit: set the welding current, voltage, speed, wire feed speed process parameters, and real-time monitor and adjust to ensure the stability and quality consistency of the welding process.

[0101] Figure 9 And Figure 10 The figure is an example of controlling the dual-arm robot to assemble and weld a complex non-standard workpiece. The non-standard workpiece is a fan blade workpiece. First, based on the 3D model feature information of the fan blade workpiece, the weld position is determined based on the welding process of the target workpiece according to the forming characteristics of the parts, and the weld is taken as the division line to split the independent virtual model parts, and then the part features are extracted, and the real parts are prepared at the same time. Then, based on reinforcement learning, the control strategy is generated based on the dual-arm robot parameters and the virtual model parts in the simulation environment, the assembly task and the welding task are planned, and the dual-arm motion trajectory is generated. After judging that the welding trajectory is reasonable in combination with obstacle avoidance and welding process parameters, the robot control program is generated, and the operation of the dual-arm robot is verified on the robot operating system ROS platform, and then the test is performed on the dual-arm robot system, and then the batch real parts are assembled and welded.

[0102] As Figure 9 shown, in the assembly task planning, the center piece of the fan-shaped workpiece is taken as the base part, and the dual-arm robot identifies and controls the left arm of the robot to obtain the first blade according to the assembly task planning, drives the first blade to be connected with the center piece, completes the assembly, and clamps the irregular first blade to stabilize the pose, waiting for the right arm to perform welding. In this process, the dual-arm robot identifies the parts according to the shape, size and other characteristics through the visual perception system. At the same time, the welding task planning controls the robot to identify the weld of the real part, corresponds to the weld obtained from the simulation environment, calls the welding process parameter package, controls the right arm to drive the welding system to move according to the welding trajectory, and welds the first blade to the center piece, completing the assembly and welding task of the first blade. Figure 10 For the assembly and welding of the subsequent blades, the control process is the same as that of the first blade.

[0103] After the welding is completed, the dual-arm robot compares the completed task through the visual perception system, confirms that the assembly and welding workpiece is correct compared with the 3D model of the target fan-shaped workpiece, and completes all welding process tasks.

Claims

1. A welding control method for a dual-arm robot, characterized in that: Controlling a dual-arm robot to complete welding tasks on non-standard workpieces includes the following steps: 1) Call the 3D data of the target workpiece, and based on the preset welding process parameters of the target workpiece, perform structural analysis and disassembly on the 3D model of the target workpiece to generate the independent parts required to form the welded body, and prepare the corresponding real parts. 2) Based on reinforcement learning, the optimal action strategy for assembling and welding parts is learned by using a dual-arm robot as the intelligent agent. One arm of the dual-arm robot performs the assembly task and the other arm performs the welding task. Combining the structural analysis and disassembly results of the 3D model, the assembly task and welding task are planned as the control strategy results. The assembly task planning includes the assembly sequence and pose of the corresponding parts and the assembly action trajectory. The welding task planning includes the welding action trajectory of the pose of the corresponding parts. 3) Generate the motion trajectory of the dual-arm robot based on the control strategy results. One arm of the dual-arm robot acquires real parts and adjusts the pose of the real parts according to the assembly task planning. The other arm is controlled to perform welding action on the current real parts based on the welding task planning. Through the iterative cycle of assembly-welding, multiple real parts are welded in sequence according to the assembly order and pose to obtain the finished product of the target workpiece. The control strategy results, specifically the grouping task planning, include the following steps: Step A1: Obtain the feature information of the virtual model parts and the feature information of the real parts from the 3D model disassembly in Step 1), match the feature information of the virtual model parts with the feature information of the real parts, and establish the correspondence between the virtual model parts and the real parts. Step A2: Determine the base components that form the basis for assembly, generate component assembly strategies for the virtual model components, and perform robot kinematic modeling. Mathematically describe the motion relationships of the robotic arm driving the various components during the assembly process, including the joint motion and pose transformation of the robotic arm, to control the robot's motion trajectory. Step A3: Optimize the generated component assembly strategy and robot kinematic model using reinforcement learning. Through the interaction between the agent and the environment, adjust the assembly strategy according to the assembly reward function. The reinforcement learning results are verified using Sim2real, applying the learning results in the simulated environment to the real scene. If the verification is successful, the assembly task plan is output for assembling real components. If the verification fails, return to the reinforcement learning stage to further adjust and optimize the strategy until the verification is successful. Step A4: Obtain the actual base components, establish a workpiece coordinate system based on the base components to provide reference coordinates for the assembly operation, and combine the correspondence between the virtual model components and the actual components to obtain the assembly motion trajectory of the dual-arm robot from the assembly task planning. For welding task planning, the following steps are included: Step B1: Obtain information about the dual-arm robot and the parts to be welded, including the size, shape, material of the parts, and motion specifications of the welding robot. Step B2: Establish a welding simulation environment to simulate the relative positional relationship and operating conditions of the welding robot and the parts. Then, identify the weld seams of the parts in the simulation environment and determine their position and shape characteristics. Step B3: Generate a welding path based on the identified weld, covering the entire length of the weld and conforming to the process specifications. Then, perform forward and inverse kinematics solutions to calculate the pose of the welding torch and the motion commands of the robot joints. After that, use offline simulation to determine whether the welding posture is reasonable. If it is not reasonable, return to the kinematics solution step for adjustment until the simulation results meet the requirements, and output the welding trajectory corresponding to the identified weld. Meanwhile, the weld information of the real parts during the assembly process is obtained and matched with the welds identified in the simulation environment. Based on Sim2real, the robot's motion trajectory in the real environment is determined. Combining the kinematic characteristics of the robotic arm and the welding process parameters, the welding process parameter package is called to obtain the planned current welding trajectory. Step B4: Based on the current welding trajectory planned in step B3, complete the welding task for the current actual component and proceed to the welding of the next component; Step B5: Repeat steps B3 to B4 until all parts are assembled and welded.

2. The welding control method for a dual-arm robot according to claim 1, characterized in that: In step 3), the control strategy results are simulated and verified using the robot operating system ROS platform. After the simulation verification is successful, the parameters and behavioral data obtained from the simulation are deployed and debugged on the real dual-arm robot system. After the debugging is completed, welding commands are issued to the dual-arm robot system to drive the two arms to perform a collaborative assembly and welding operation sequence in the real physical environment.

3. A welding control method for a dual-arm robot according to claim 1 or 2, characterized in that: Step 2) Based on reinforcement learning, the optimal strategy for assembling and welding parts is learned using a dual-arm robot as the intelligent agent, including: Define a state description function, using the TCP pose, velocity, and welding assembly state of the robotic arm end effector as state variables, to construct the robot action state function for the assembly and welding process, which serves as the basis for judging reinforcement learning training. Define an assembly reward function to quantitatively evaluate the quality of the assembly status of parts, the TCP pose of the robot end effector, and the speed during the assembly and welding process; Using a dual-arm robot as the intelligent agent, the optimal motion strategy for assembling and welding the parts of the dual-arm robot is reinforced through learning, and the assembly task plan and welding task plan are output as the control strategy results.

4. The welding control method for a dual-arm robot according to claim 1, characterized in that: The specific steps for establishing the correspondence between virtual model components and real components are as follows: Extract the feature information of virtual model parts decomposed from the 3D model in step 1), including size, shape and surface texture, to represent the shape and position features of the parts; 3D geometric feature information of real parts, including size, shape and surface texture, can be obtained by scanning or measuring. Based on the acquired feature information, the virtual model parts and real parts are classified and labeled multiple times. Each classification and labeling is based on different feature dimensions, including size, shape and part function. Through multiple classifications and labels, the recognition accuracy of individual part features is improved. Finally, based on the categorized and summarized components, the virtual model components are compared with the real components in turn to achieve alignment and matching between the virtual model components and the real components, which is used for the robot to identify the real components.

5. The welding control method for a dual-arm robot according to claim 1, Its characteristic is that step B3, which involves obtaining weld information for the actual components during the assembly process, specifically includes: B3.

1. Acquire images containing the weld seams of the actual components to be welded using a sensing camera; B3.2 Preprocess the image, including converting the image to grayscale and using a denoising algorithm to remove random noise from the image; B3.3 Weld Feature Extraction: Using image processing technology, key features of the weld, including edge contours and texture patterns, are extracted from the preprocessed image. The Segformer segmentation model is then used to segment the image, and the segmented welds are identified and analyzed. This includes quantitative analysis of the weld geometry, surface quality, and defect features to assess whether the weld meets the process standards and quality requirements, and to determine whether welding should proceed. B3.4 Output the results of weld identification and analysis, including the location, size and quality grade of the weld, to be used in conjunction with the welds identified in the simulation environment to determine the robot's motion trajectory in the real environment based on Sim2real.

6. A welding control method for a dual-arm robot according to claim 1 or 2, characterized in that the three-dimensional geometric data of the target workpiece is retrieved from the process database of the target workpiece, or established online through three-dimensional reconstruction technology.

7. A welding control system for a dual-arm robot, characterized in that: The invention includes a movable leveling platform, a dual-arm robot, a vision perception system, a welding system, and an AI intelligent computing unit. The dual-arm robot includes a left arm and a right arm. The welding system is located at the end of the right arm via an end-effector. The AI ​​intelligent computing unit stores at least one instruction or at least one program. When the at least one instruction or the program is executed, it implements the welding control method of the dual-arm robot according to any one of claims 1-6.

8. A welding control system for a dual-arm robot according to claim 7, characterized in that: The AI ​​intelligent computing unit includes the following modules: Assembly trajectory planning module: Used to plan and output the motion trajectory of one arm of the dual-arm robot according to the assembly task, to complete the assembly task of parts, including: Assembly robot arm trajectory generation unit: Generates the robot arm assembly motion trajectory based on the current real parts and assembly task plan; Assembly trajectory optimization unit: Utilizes smoothing, speed planning, and collision detection algorithms to optimize the robotic arm's assembly motion trajectory; Welding trajectory planning module: Used to plan and output the motion trajectory of the other arm of the dual-arm robot according to the welding task, to complete the welding task of the parts, including: Welding robot arm trajectory generation unit: Generates the welding motion trajectory of the robot arm based on the current real parts and welding task planning; Welding trajectory optimization unit: Adjusts the smoothness, continuity, speed, and acceleration parameters of the robotic arm's welding motion trajectory to meet welding process requirements; Identification module: used to identify parts and welds, including: Component identification unit: Used to identify and classify components by applying image processing and information tagging methods to the acquired component images, shape, and size feature information; Weld seam recognition unit: Based on a visual perception system, it detects the position, shape and size of weld seams in real time, identifies weld seam information of real parts, and provides a basis for welding trajectory planning; Control module: Used to output control commands to the dual-arm robot based on the control strategy results, including: Robot posture control unit: Based on the motion trajectory planned for the assembly and welding tasks and the real-time feedback information from the robot, it controls the posture of the dual-arm robot and performs posture tracking by combining the dynamic model and external disturbance factors; Welding process control unit: Sets process parameters such as welding current, welding voltage, welding speed, and wire feed speed, and monitors and adjusts them in real time to ensure the stability and consistency of welding process.

Images