Strategy generation and reality migration method and system for complex product multi-device collaborative intelligent assembly

Abstract

The invention discloses a strategy generation and reality migration method and system for complex product multi-device collaborative body intelligent assembly, and relates to the field of intelligent manufacturing and robot control. The method comprises the following steps: firstly, constructing a hierarchical assembly task network based on a deviation propagation model, and quantitatively analyzing multi-process error accumulation; based on an optimal transmission theory and a swarm intelligence algorithm, realizing optimal allocation and scheduling of multi-device tasks; and then a high-precision flexible assembly track fusing real-time deviation compensation and physical feedback gradient guidance is generated. And finally, realizing robust migration of the strategy to a physical environment through diversified simulation verification and domain adaptation technologies. According to the method, a whole-process closed loop from error perception, intelligent decision-making to precise execution is realized, the assembly precision, the automation level and the production line adaptability are remarkably improved, and the problems that a traditional method depends on manpower, the precision is insufficient, and virtual-real migration is difficult are solved.

Description

Technical Field

[0001] This invention belongs to the field of intelligent manufacturing and industrial robot control technology, and specifically relates to a method and system for generating and transferring multi-device collaborative assembly strategies to complex assembly scenarios. Background Technology

[0002] In high-end equipment manufacturing fields such as aerospace and automobile manufacturing, the assembly of large and complex components (such as aircraft sections and automobile hoods) is characterized by complex processes, high precision requirements, and close collaboration among multiple devices. Such assembly typically requires multiple devices with different functions (such as positioning robots, tightening robots, measuring robots, flexible tooling, etc.) to work collaboratively within a limited physical space to complete a series of precision operations such as gripping, alignment, fitting, and connection.

[0003] With the development of intelligent manufacturing technology, traditional multi-equipment collaborative assembly systems and methods are no longer able to meet the high-level requirements of modern production for flexibility, precision and intelligence, and mainly face the following three challenges.

[0004] First, the rigid task planning and scheduling makes it difficult to adapt to dynamic and high-precision requirements.

[0005] Existing multi-equipment collaborative assembly systems rely heavily on pre-defined process procedures or manual teaching programming for task planning. The timing logic between tasks and equipment assignment are often static or semi-static, unable to adaptively adjust to real-time operating conditions (such as part material deviations and equipment status fluctuations). While some research has employed digital twins, message buses, or evolutionary algorithm-based scheduling optimization to improve system integration and scheduling flexibility, these methods primarily focus on information flow integration, communication protocols, or macroscopic time / resource scheduling optimization. They lack quantitative modeling and closed-loop feedback for the core physical constraints of the assembly process—namely, geometric accuracy and the cumulative effect of multi-process errors. When faced with multi-variety, variable-batch production and unavoidable manufacturing and positioning errors, existing methods struggle to automatically generate optimal collaborative strategies that simultaneously meet the requirements of work cycle time and sub-millimeter or even higher assembly accuracy (such as surface differences and gaps), resulting in long reconfiguration cycles and limited levels of intelligence.

[0006] Second, the execution of control is rigid, lacking intelligent compliant strategies that integrate force perception.

[0007] In contact-based operations involving precision alignment, interference fits, or assembly of consumable parts (such as pin insertion and sealing strip press-fitting), traditional robot execution methods based on pure position control are prone to workpiece surface scratches, excessive assembly stress, or abnormal equipment shutdowns due to excessive system rigidity and inability to actively adapt to changes in contact force. Although some advanced systems have introduced force sensors, the control strategies are mostly simple threshold force control or admittance / impedance control, lacking adaptability and intelligence under complex contact conditions. While existing technologies can handle real-time compensation for computational tasks, their "delay compensation" and "redundant copy" mechanisms are designed for the reliability of communication and computing in the information world, and cannot be directly converted or applied to solve the problems of real-time trajectory correction and compliant contact control based on force-position hybrid information in physical assembly. Therefore, how to generate and execute high-precision motion trajectories that can actively avoid collisions, adapt to changes in contact force, and achieve "compliant fit" remains a critical technical bottleneck that needs to be overcome.

[0008] Third, a gap exists in the migration between virtual and real worlds, leading to instability in the physical deployment of simulation strategies.

[0009] Training and validating assembly strategies using digital twins or simulation environments is a crucial way to improve system intelligence and safety. However, because simulation models cannot perfectly reproduce all the complex factors in the real physical world (such as subtle differences in friction coefficients, sensor noise, actuator nonlinear characteristics, and minor environmental vibrations), intelligent strategies that perform well in simulation (such as strategies generated based on reinforcement learning) often experience significant performance degradation or even failure when directly deployed to the real production line—a "simulation-to-real" migration gap. While existing methods can perform multi-dimensional evaluation and integrity assessment of multi-device collaborative tasks, their core focus is on state assessment and diagnosis, rather than strategy generation and robust cross-domain migration. General simulation verification frameworks also fail to provide efficient migration techniques specifically designed to overcome the uncertainty of physical parameters for assembly accuracy control strategies. Therefore, there is an urgent need for a migration method and verification system that can ensure the robust one-time deployment of intelligent assembly strategies from the virtual environment to the physical production line without extensive on-site debugging.

[0010] Therefore, there is an urgent need in this field for a systematic approach that can connect the above-mentioned links and fundamentally improve the autonomy, accuracy and adaptability of multi-device collaborative assembly. Summary of the Invention

[0011] In view of this, the purpose of the present invention is to provide a method and system for strategy generation and real-world transfer in the collaborative intelligent assembly of complex products using multiple devices. By quantifying error propagation, generating high-precision smooth trajectories, and realizing robust transfer of strategies from the simulation environment to the physical production line, the present invention aims to solve the problems of low assembly accuracy, poor adaptability, and reliance on manual debugging.

[0012] To achieve the above objectives, the present invention provides the following technical solution: This invention first proposes a strategy generation and reality transfer method for complex product multi-device collaborative embodied intelligent assembly, including the following steps: S1: Construct a hierarchical assembly task network based on the deviation propagation model: Construct a tensor representation model of assembly relationships based on assembly process knowledge; establish a discrete state space model describing the cumulative propagation of deviations in multiple processes based on the representation model to identify key control paths and deviation sensitive points; construct a hierarchical task network model with task nodes and dependency edges based on the assembly process flow to generate an initial assembly task sequence. S2: Multi-device task planning and scheduling based on optimal transmission and swarm intelligence: Based on the results of step S1, a recursive decomposition algorithm is used to decompose high-level assembly tasks into low-level operation units. Based on the entropy regularized optimal transmission theory, a cost matrix between task requirements and equipment capabilities is constructed to solve for the optimal transmission plan and realize the mapping from tasks to equipment actions. Based on the task allocation results, a scheduling model with the goal of minimizing total completion time and load balancing is constructed. An improved swarm intelligence optimization algorithm is used to search for the optimal time scheduling sequence that satisfies the deviation constraint check in non-Euclidean space. S3: Generate a high-precision compliant assembly strategy: Based on the cumulative deviation vector calculated in real time in step S2, construct a dynamic motion primitive model with online deviation compensation to correct the robot's motion target; for the precision contact stage, use a conditional diffusion model to generate the trajectory, and inject a gradient guidance mechanism based on physical feedback during its reverse denoising process to ensure that the trajectory meets force control and obstacle avoidance constraints. S4: Simulation verification and real-world transfer of assembly strategy: Construct a verification sample set covering multiple working conditions using a deep generative model, evaluate and iteratively optimize the strategy generated in step S3 in a simulation environment; Employ domain adaptation technology to fine-tune the strategy network by minimizing the distribution difference between the simulation domain and the real domain in the feature space, thereby achieving robust transfer of the strategy to the physical environment.

[0013] Furthermore, in step S1, the assembly relationship tensor quantization representation model uses the assembly relationship tensor... The calculation method is as follows: in: , representing the feature embedding vectors of active and passive parts respectively, transforming geometric information into computable numerical features; Source: CAD Geometric Analysis The weight matrix of the assembly relationship has the following tensor product structure: in: Let be the nominal homogeneous transformation matrix, containing the nominal rotation matrix. With the nominal translation vector ; To constrain the masking operator; This is the activation function.

[0014] Furthermore, in step S1, the discrete state-space model is represented as: in: For the first Cumulative deviation vector after process; The adjoint transformation matrix; Translation vector The antisymmetric matrix is ​​used to... Item capture Abbe error; Input influence matrix; This is the input error vector for this process; This is the overall system disturbance term; The nominal translation vector defines the design position coordinates of the part in the coordinate system. For the first A rotation matrix for assembly relationships.

[0015] Furthermore, in step S2, the task requirements are quantified into a task difficulty vector. Its definition is: in: For the sake of precision and urgency; For design tolerances; Represents the geometric travel of the assembly operation; The complexity is expressed as degrees of freedom; For the first Cumulative deviation vector after process; To constrain the masking operator; The nominal translation vector; Equipment capabilities are quantified into equipment capability vectors. Its definition is: in: For accuracy capability; To ensure repeatability and accuracy; For wingspan ability; The radius of the workspace; For the ability of degrees of freedom; Number of motion axes; Elements of the cost matrix The formula for calculating the weighted Mahalanobis distance between the task difficulty vector and the device capability vector in the feature space is as follows: in: It is a diagonal weight matrix; Weights for bias sensitivity; Spatial dimension weights; Weights for degrees of freedom; For the first The accompanying transformation matrix of the process.

[0016] Furthermore, in step S2, the solution to the optimal transmission plan is achieved through the following optimization problem: in: These are elements of the cost matrix; Indicates the task Assigned to device The probability weights; This is the regularization coefficient, which controls the smoothness of the allocation scheme; This is an entropy penalty term.

[0017] Furthermore, in step S2, the update law of the improved swarm intelligence optimization algorithm is: in: Let be the position vector, representing the first position vector. The particle in the first The task execution sequence at the next iteration; The velocity vector represents the reordering operation steps; This is a difference operator used to compute sequence exchange steps; To apply operators to perform swap operations; , and These are the inertia weight and the learning factor, respectively. and These are the historical optimal sequence and the globally optimal sequence, respectively. and To introduce parameters for randomness, and .

[0018] Furthermore, in step S3, the motion equations of the dynamic motion element model with online deviation compensation are: in: Assemble coordinates in CAD format; For phase variables, initial value ; For nonlinear forced terms, determined by the Gaussian function and shape weights obtained through imitation learning constitute: ; and These represent the current position and velocity of the robot's end effector, respectively. and The gain coefficient defines the stiffness and damping characteristics of the system. This is the time scaling factor; The stable decay rate of the canonical system; For the first Cumulative deviation vector after process; This is the rate of change of position over time, i.e., velocity; This is the rate of change of velocity over time, i.e., acceleration; The rate of change of the phase variable with time is given by the canonical system equations. definition; This is the initial position of the robot's end effector.

[0019] Furthermore, in step S3, the gradient guidance mechanism corrects the trajectory during the inverse denoising process of the conditional diffusion model using the following formula: in: The current noisy trajectory; The output of the noise prediction network; To represent network parameters; and These are the single-step retention factor and the cumulative retention factor, respectively. The gradient of the cost function is constructed based on physical feedback; The real-time contact force captured by the sensor; To limit the contact force; is the robot's geometric Jacobian matrix, used to map contact forces in Cartesian space to corrected gradients in joint space; For guiding strength; This is the variance coefficient.

[0020] Furthermore, in step S4, minimizing the distribution difference between the simulation domain and the real domain in the feature space is achieved by minimizing the square of the maximum mean difference, which is calculated as follows: in: The trace of the matrix; , and The kernel matrix measures the similarity of samples within and across domains, and its bandwidth parameter is based on the random perturbation term in the simulation environment. The statistical variance is adaptively configured, and: This is the similarity matrix within the simulated samples; This is the similarity matrix within the real samples; This is the cross-domain similarity matrix between simulated samples and real samples; , and Let be the domain probability indicator matrix, and: All elements are ; All elements are ; All elements are ; This refers to the amount of simulation data. This represents the actual amount of data. Minimize by gradient descent Update the neural network parameters of S3. The policy network is forced to learn cross-domain invariant features, eliminating transfer errors caused by simplistic modeling of the simulation environment. in: For the first The set of neural network parameters at the next iteration; For the first The parameter set after the next iteration; This indicates an assignment operation; The learning rate; Indicates the opposite direction of the gradient; The original task loss function is defined; For the loss function with respect to parameters The vector of partial derivatives; For the domain-adaptive equilibrium coefficient; This is the front-end part of the neural network, which maps the original input into a high-dimensional feature vector.

[0021] This invention also proposes a strategy generation and reality migration system for complex product multi-device collaborative embodied intelligent assembly, including a processor and a memory. The memory stores a computer program. When the computer program is executed by the processor, it realizes the strategy generation and reality migration method for complex product multi-device collaborative embodied intelligent assembly as described above.

[0022] The beneficial effects of this invention are as follows: This invention provides a strategy generation and real-world transfer method for complex product multi-device collaborative embodied intelligent assembly. Through a closed-loop end-to-end process, it fundamentally optimizes the complex assembly process and achieves the following technical effects: (1) Breakthrough in precision and quality: By constructing a deviation propagation model and conducting quantitative analysis, the system can actively predict and compensate for the cumulative error of multiple processes. Combined with the compliant trajectory generation based on the diffusion model, the final assembly precision can be stably controlled at the sub-millimeter level, significantly improving product consistency and pass rate. (2) Leap in intelligence and automation: The solution realizes closed-loop intelligent decision-making throughout the entire process from task modeling, intelligent planning, trajectory generation to verification and migration, completely replacing the traditional model that relies on manual trial and error and expert experience, and forming a complete integrated intelligent system of "perception-decision-execution"; (3) Improved scheduling efficiency and flexibility: Based on optimal transmission theory and swarm intelligence, task planning and scheduling can dynamically generate optimal task allocation and execution sequence according to real-time accuracy requirements and equipment capabilities, greatly improving the efficiency of multi-device collaboration and resource utilization, and quickly adapting to multi-variety and variable batch production; (4) Reliability transfer and cost reduction and efficiency improvement: Through simulation verification for physical uncertainty and robust transfer based on domain adaptation, the intelligent strategy is successfully deployed to the real production line in one go, eliminating the dependence on repeated manual debugging on site, and ensuring high consistency and traceability of the process while reducing labor costs. Attached Figure Description

[0023] To make the objectives, technical solutions, and beneficial effects of this invention clearer, the following figures are provided for illustration: Figure 1 This is a flowchart of the strategy generation and reality transfer method for complex product multi-device collaborative embodied intelligent assembly of the present invention; Figure 2 A schematic diagram illustrating the principles of multi-device task planning and scheduling; Figure 3 A schematic diagram illustrating the principle of generating a high-precision compliant assembly strategy.

[0024] Figure 4 This is a schematic diagram illustrating the principle of simulation verification and real-world transfer of the assembly strategy.

[0025] Figure 5 This is a schematic diagram illustrating the principles of strategy generation and real-world transfer methods. Detailed Implementation

[0026] The present invention will be further described below with reference to the accompanying drawings and specific embodiments, so that those skilled in the art can better understand and implement the present invention. However, the embodiments described are not intended to limit the present invention.

[0027] This embodiment proposes a method and system for strategy generation and real-world transfer in the collaborative intelligent assembly of complex products using multiple devices. The aim is to address issues such as task planning's strong reliance on expert experience, rigid execution with poor flexibility, and insufficient strategy security and generalization during the assembly process. This is achieved through intelligent planning and scheduling of multi-device collaborative assembly tasks, high-precision compliant assembly strategy generation, and simulation verification and real-world transfer of assembly strategies. Specifically, the method and system for strategy generation and real-world transfer in the collaborative intelligent assembly of complex products using multiple devices in this embodiment mainly include: This embodiment takes the assembly of the left front door assembly of an automobile as an example to illustrate in detail the specific implementation process of the method and system of this embodiment.

[0028] Assembly items: left front door assembly, upper hinge, lower hinge.

[0029] Equipment A (handling robot): Heavy-duty robotic arm with a force-controlled gripper at the end, responsible for grasping and positioning the car door.

[0030] Device B (Tightening Robot): A lightweight collaborative arm equipped with a vision camera and a tightening gun, responsible for inserting and tightening hinge bolts.

[0031] Specifically, such as Figure 1 As shown in the figure, the strategy generation and reality transfer method for complex product multi-device collaborative embodied intelligent assembly in this embodiment includes the following steps.

[0032] S1: Construct a hierarchical assembly task network based on a deviation propagation model. A tensor representation model of assembly relationships is constructed based on assembly process knowledge; a discrete state-space model describing the cumulative propagation of deviations across multiple processes is established based on this representation model to identify key control paths and deviation-sensitive points; based on the assembly process flow, a hierarchical task network model with task nodes and dependency edges is constructed to generate an initial assembly task sequence.

[0033] This step aims to structure and quantify the physical characteristics and error propagation mechanisms of the assembled objects through mathematical models.

[0034] (1) Quantitative representation of the assembly process knowledge base: By integrating historical assembly data, process manuals, and expert experience, a knowledge graph is constructed that includes the geometry, material properties, assembly tolerances, and connection relationships of nodes such as the left front door assembly, upper hinge, lower hinge, body A-pillar mounting surface, and mounting bolts. The correlation strength between entities under specific assembly relationships is quantified through a bilinear transformation mechanism. The core is to construct a relation weight matrix coupled with a nominal homogeneous transformation matrix and a degree-of-freedom masking operator. This masking operator explicitly defines the rigid constraints of assembly constraints in six spatial degrees of freedom, thereby transforming semantic alignment into a strict geometric constraint expression.

[0035] To transform the semantic knowledge in the assembly scenario into a computable mathematical model, the entities (parts, fixtures) and their constraints are first mapped to a high-dimensional feature space. An assembly relation tensor is defined. This is used to quantify the strength of physical associations between entities in a knowledge graph under specific assembly relationships. This formula measures the entity. and In assembly relationship The physical correlation strength under the given conditions.

[0036] in: : These represent feature embedding vectors for active parts (such as locating pins) and passive parts (such as pin holes), respectively, which transform geometric information into computable numerical features.

[0037] Source: CAD Geometric Analysis The weight matrix of the assembly relationships. To mathematically represent the physical constraints, it is derived as a tensor product structure: .

[0038] : The nominal homogeneous transformation matrix contains the rotation matrix. With translation vector .

[0039] : The nominal rotation matrix is ​​determined by Euler angles ( Expanding on this, defining what revolves around... The nominal angle when the shaft is installed.

[0040] : The nominal translation vector defines the design position coordinates of the part in the coordinate system.

[0041] : For the constraint mask operator, the definition is in Physical constraint strength for 6 degrees of freedom in the displacement and rotation directions. .

[0042] Activation function, mapping association to Probability space.

[0043] (2) Establishing a deviation propagation model: For the assembly object, analyze the manufacturing and positioning errors of local features. Establish a discrete state-space model describing the flow of deviations in multiple processes to quantify how local errors accumulate and affect the overall assembly accuracy. Use the adjoint transformation matrix to accurately describe error propagation. This matrix contains antisymmetric matrix forms of rotation transformations and translation vectors to capture the Abbe error effect. Based on this model, identify the key control paths and deviation-sensitive connection points that have the greatest impact on the final assembly accuracy. For example, the "upper hinge mounting hole" is identified as a deviation-sensitive connection point, and its accuracy directly determines the amount of sag at the rear end of the door.

[0044] Based on the above geometric benchmarks, a discrete state-space model describing the cumulative transmission of deviations across multiple processes is established: : No. The cumulative deviation vector after the process (including 3 translation errors and 3 rotation errors).

[0045] in: The accompanying transformation matrix is ​​responsible for propagating the deviation from the previous process to the current process; its core is... Item, using antisymmetric matrix The magnification error caused by the capture lever arm.

[0046] here It is a translation vector An antisymmetric matrix. The term precisely describes the Abbe error, which is the linear displacement amplification induced at the end of the long lever arm of the current process by a small rotational deviation in the previous process.

[0047] : Input influence matrix, which defines the contribution rate of input error in this process to the total deviation.

[0048] : Input errors determined in this process (such as fixture repeatability accuracy).

[0049] The system's overall disturbance term represents uncontrollable random noise (such as environmental vibration).

[0050] (3) Generating a hierarchical task network: The assembly process is described based on a directed acyclic graph, with tasks as nodes and dependencies as directed edges. By calculating the deviation sensitivity of each task node, critical quality paths are identified, and an initial hierarchical task sequence is generated. For example: Task X1 (Equipment A picks up the door) - Task X2 (Equipment B performs visual positioning) - Task X3 (Equipment A performs alignment) - Task X4 (Equipment B performs tightening). The system calculation found that the path from X2 to X3 has the greatest impact on the final gap and is marked as the critical quality path.

[0051] S2: Multi-device task planning and scheduling based on optimal transmission and swarm intelligence. Based on the results of step S1, a recursive decomposition algorithm is used to decompose high-level assembly tasks into low-level operation units. Based on the entropy regularized optimal transmission theory, a cost matrix between task requirements and equipment capabilities is constructed, and the optimal transmission plan is solved to realize the mapping from tasks to equipment actions. Based on the task allocation results, a scheduling model is constructed with the goal of minimizing total completion time and load balancing. An improved swarm intelligence optimization algorithm is used to search for the optimal time scheduling sequence that satisfies the deviation constraint check in non-Euclidean space.

[0052] Specifically, such as Figure 2 As shown, this embodiment identifies key control paths and deviation-sensitive points by constructing an assembly process knowledge base and a deviation propagation model that transforms local deviations into global assembly accuracy. Combined with the process flow, a hierarchical task network model with process logic and task dependencies is constructed based on a directed acyclic graph to generate an initial assembly task sequence. Furthermore, a recursive decomposition algorithm integrating the hierarchical task network and optimal transport theory is used to achieve a fine mapping from processes to operations. Finally, combining subtask priorities, cycle time, and equipment capacity constraints, a multi-equipment scheduling model based on a swarm intelligence optimization algorithm is constructed to achieve reasonable task allocation, execution sequence optimization, and collaborative cycle time improvement.

[0053] Specifically, this step mainly includes the following:

[0054] (1) Task Recursive Decomposition and Optimal Matching: A recursive decomposition algorithm is used to refine high-level tasks, and entropy regularization optimal transmission theory is introduced to solve the task allocation problem. A cost matrix is ​​constructed to quantify the matching distance between task requirements and equipment capabilities. By solving the problem of minimizing transmission costs with entropy regularization terms, the abstract task distribution is smoothly mapped to the underlying equipment action distribution, thereby transforming the process planning into specific operational units that can be executed by the robot. For example, the high-level task of "installing car doors" is decomposed into "grasping", "positioning", "aligning", and "tightening". The algorithm maps the tasks of "grasping" and "aligning" to equipment A with a very high probability, and maps the tasks of "positioning" and "tightening" to equipment B.

[0055] 1) Parameter definition : To meet the task requirements ( ) and equipment capabilities ( The set of all possible assignments constrained by S1. The high-level task distribution generated by S1. (Including accuracy requirement features) Deconstructing into equipment motion distribution (Including device capability vectors) ): The task difficulty vector, derived from step S1, is represented as: in: : Indicates the urgency of accuracy. If step S1 predicts the cumulative deviation Approaching design tolerances The denominator approaches zero. The value surged, forcibly assigned to high-precision equipment.

[0056] Design tolerances (e.g., 0.05 mm) are derived from the knowledge graph.

[0057] This indicates the geometric travel distance of the assembly motion. For example, if a car door needs to be installed across a distance of 1.5m, this value would be 1500, requiring a robot with a long reach.

[0058] : Degrees of freedom complexity. If the rank of the matrix is ​​1, it is a simple planar fit; if it is 6, it is a fully constrained precise insertion.

[0059] The device capability vector, derived from the device library, is represented as: in: Precision capability. (Source: Equipment manual) This refers to the repeatability accuracy (e.g., 0.02mm).

[0060] Arm span capability. Equipment kinematic parameters. Workspace radius.

[0061] Degrees of freedom capability. Equipment configuration. Number of motion axes, such as 6 axes or 7 axes.

[0062] 2) Calculate the weighted Mahalanobis distance in the feature space. Measuring the mismatch between task difficulty and equipment precision: in: : A diagonal matrix that adjusts the weights of the three dimensions.

[0063] Bias sensitivity weight. , Normalization coefficient.

[0064] and These represent the spatial dimension weight and the degree of freedom weight, respectively.

[0065] 3) Utilize optimal transmission to find the best matching solution between the task and the robot: : The optimal transmission plan matrix, whose elements Indicates the task Assigned to device The probability weights.

[0066] in: : Cost matrix elements, usually Mahalanobis distance.

[0067] Regularization coefficient: controls the smoothness of the allocation scheme.

[0068] Entropy penalty term, used to accelerate algorithm convergence and optimize computational efficiency.

[0069] (2) Multi-device collaborative scheduling optimization: After the task allocation is determined, the system needs to generate the optimal time scheduling sequence. A scheduling model with the goal of minimizing the total completion time and load balancing is constructed. An improved swarm intelligence algorithm is used to solve the problem, and a discrete update law for non-Euclidean space is introduced. The algorithm uses difference operators and application operators to search for the optimal sequence in the permutation group space. The search process embeds the deviation constraint verification mechanism of step S1: the sequence is considered a valid solution only when the cumulative deviation corresponding to the scheduling sequence does not exceed the preset threshold, thereby ensuring that the scheduling scheme meets both efficiency and accuracy requirements. For example: the initial scheduling suggestion is that after device A is in place, device B should tighten immediately. The deviation model predicts that if device A remains stationary without secondary force control compensation, the final door gap deviation will exceed 2mm due to the downward gravity of the robotic arm. The scheduling algorithm will require device A to perform an "upward compensation" action first, and only after the accuracy prediction is met will device B be allowed to intervene.

[0070] In determining the allocation Next, a time-dimensional scheduling sequence needs to be generated. An improved particle swarm optimization algorithm is adopted, the core of which lies in defining a discrete update law for non-Euclidean space: in: : Position vector. An ordered list of integers representing the position vector. The particle in the first The task execution sequence during the next iteration.

[0071] : Velocity vector. It is not a number, but a set of basic swap sequences that represent the steps involved in reordering.

[0072] Difference operators. Calculate the steps required to transpose sequences.

[0073] : Apply operator. Indicates that a set of swap operations are performed sequentially on the sequence.

[0074] Inertia weight and learning factor.

[0075] : Key parameters of randomness, and .

[0076] Historical optimal sequence, global optimal sequence.

[0077] S3: Generate a high-precision compliant assembly strategy. Based on the cumulative deviation vector calculated in real time in step S2, a dynamic motion primitive model with online deviation compensation is constructed to correct the robot's motion target. For the precision contact stage, a conditional diffusion model is used to generate the trajectory, and a gradient guidance mechanism based on physical feedback is injected during its inverse denoising process to ensure that the trajectory meets force control and obstacle avoidance constraints.

[0078] like Figure 3 As shown, this embodiment extracts basic motion patterns such as precise alignment and force-controlled pushing through expert instruction, and constructs a reusable assembly basic motion skill model by combining imitation learning and deep reinforcement learning methods. Based on this, and using an assembly probability diffusion model, safety constraints such as force control thresholds, attitude boundaries, and collision avoidance are introduced into the trajectory generation process, and these are formalized as differentiable penalty terms or logical barriers, thereby generating a high-precision compliant motion trajectory that satisfies complex constraints.

[0079] Specifically, this step mainly includes the following:

[0080] (1) Dynamic Motion Element Extraction and Deviation Compensation: Combining imitation learning to extract expert skills, a second-order dynamic system with deviation compensation is constructed. This system uses the cumulative deviation vector calculated in real time in step S1 to directly correct the nominal target point in the dynamic equation, enabling the robot to move to the actual aligned coordinates after deviation compensation, thus achieving automatic correction. At the same time, the system constructs a nonlinear forced term through the Gaussian function to reproduce the force control characteristics of the expert at the moment of precision operation. For example, the vision system detects that the A-pillar of the current vehicle body is concave inward by 1mm compared to the standard CAD model. The system uses the cumulative deviation vector calculated in S1 to directly modify the target point of the dynamic equation. When device A approaches the vehicle body, it automatically corrects the target coordinates inward by 1mm and reproduces the speed curve of the worker handling the vehicle gently through the Gaussian function.

[0081] In the dynamic motion primitive framework, the real-time deviation predicted in step S1 is introduced. Compensation, correcting the robot's motion target: in: CAD nominal assembly coordinates (design values).

[0082] Phase variable, initial value .

[0083] : Nonlinear forced term, simulating the force compensation characteristics in expert teaching.

[0084] Gaussian function.

[0085] Shape weights are obtained through imitation learning by regressing from expert-taught trajectories. They preserve the expert's operating style.

[0086] : The current position and velocity of the robot's end effector.

[0087] Gain coefficient: Defines the stiffness and damping characteristics of the system.

[0088] Time scaling factor.

[0089] : The steady decay rate of a regular system.

[0090] The rate of change of position over time, i.e., velocity.

[0091] The rate of change of velocity over time, i.e., acceleration.

[0092] The rate of change of the phase variable with time is given by the canonical system equations. definition.

[0093] : The initial position of the robot's end effector.

[0094] This is the core control equation of the execution layer. The nominal target pose is set to [value]. When the S1 model detects a pose offset from the previous workstation, the target point is automatically reset to [value]. This enables closed-loop online correction.

[0095] (2) Trajectory generation based on diffusion model: In the micro-contact stage, a conditional diffusion model is used to generate fine trajectories for "grabbing", "positioning", "aligning", and "tightening". During the reverse denoising (generation) process, a gradient guidance mechanism based on physical perception is injected. A physical cost function containing collision risk and excessive contact force is constructed, and its gradient with respect to the trajectory is calculated. Using the robot's geometric Jacobian matrix, the Cartesian contact force sensed by the end effector force sensor is mapped to a corrected gradient in the joint space. For example, when device B attempts to forcefully push the bolt in, the calculated gradient indicates "excessive contact force", and the trajectory generated by the diffusion model will immediately adjust along the opposite direction of the gradient (slightly retreat and adjust the angle), achieving "obstacle avoidance before contact and smooth operation after contact", ensuring that the bolt is screwed in smoothly.

[0096] During the precision contact phase, a conditional diffusion model is used to generate trajectories that satisfy safety constraints. At each step of the inverse denoising process, gradient guidance based on physical feedback is introduced. in: : Current noisy trajectory.

[0097] Noise prediction network. It is the core output of the neural network. The network parameters are obtained by learning from a large amount of expert teaching data.

[0098] Single-step retention coefficient, cumulative retention coefficient (noise scheduling parameter).

[0099] The gradient of the cost function. If the sensor senses too much downward pressure, the gradient will indicate that the joint should lift upward.

[0100] : The real-time contact force captured by the sensor.

[0101] : The set limit on contact force.

[0102] The robot geometric Jacobian matrix establishes a mapping between joint space and Cartesian force space.

[0103] : Guiding strength.

[0104] : Coefficient of variance.

[0105] S4: Simulation verification and real-world transfer of assembly strategy: Construct a verification sample set covering multiple working conditions using a deep generative model, evaluate and iteratively optimize the strategy generated in step S3 in a simulation environment; Employ domain adaptation technology to fine-tune the strategy network by minimizing the distribution difference between the simulation domain and the real domain in the feature space, thereby achieving robust transfer of the strategy to the physical environment.

[0106] This embodiment establishes a systematic strategy verification method, utilizing deep generative models such as generative adversarial networks and variational autoencoders to generate diverse verification sample sets covering ideal, mass production, and extreme operating conditions. An evaluation system centered on assembly accuracy, cycle time, and first-pass yield is constructed to verify and evaluate the strategy under multiple operating conditions. To overcome the virtual-to-real migration gap, a domain randomization method is used to randomly configure key environmental parameters, improving the strategy's adaptability to operating condition fluctuations. Furthermore, by introducing a small amount of real data for domain adaptation fine-tuning, robust migration of the strategy to the real physical environment is ultimately achieved.

[0107] Specifically, this step mainly includes the following:

[0108] (1) Construct a diverse set of verification samples: Using generative adversarial networks or variational autoencoders, generate a set of verification test cases covering different feature distributions based on limited historical data, including ideal test cases (nominal values), batch production test cases (deviations that conform to statistical distributions), and extreme test cases (extreme deviations, noise interference). For example, generate 10,000 virtual assembly scenarios, including "perfect car body", "car body with extremely large welding deformation", "situation where the sensor is noisy", "situation where the robotic arm has slight vibrations", etc.

[0109] (2) Strategy Evaluation and Iterative Optimization: A virtual test environment is established in the physical simulation engine. The generated strategy is input and evaluated using assembly accuracy, work cycle time, and first-pass yield as core indicators. The model parameters are then optimized based on the evaluation results. The strategy is run and the indicators are evaluated in the physical simulation engine (such as Isaac Sim or Mujoco). If the assembly success rate of a certain type of deformable body is low, the system will automatically adjust the diffusion model parameters in S3.

[0110] (3) Robust Transfer Based on Domain Adaptation: Domain adaptation technology is applied to achieve the transfer of simulation strategies to the physical environment. Specifically, the maximum mean difference is used as the loss function for optimization, and the distribution distance between simulation data and a small amount of real data in the feature space is calculated using a multi-scale Gaussian kernel function. By minimizing this distance, the policy neural network is forced to ignore non-essential differences caused by simulation modeling errors and focus on extracting essential contact features that are invariant to physical noise, thereby achieving seamless deployment and robust operation of the strategy on the real production line.

[0111] like Figure 4-5 As shown, in this embodiment, transfer is achieved by minimizing the difference in feature distribution between the simulation domain and the real domain. To ensure efficient computation during neural network training, this embodiment calculates the square of the maximum mean difference (MMD distance) as the trace of the kernel matrix: in: Simulation data volume Actual data volume : The trace of a matrix.

[0112] Kernel matrix: Measures the similarity of samples within and across domains. The bandwidth parameter is based on the random disturbance term. The statistical variance is adaptively configured.

[0113] : The similarity matrix within the simulated samples.

[0114] : Similarity matrix within the real samples.

[0115] Cross-domain similarity matrix between simulated samples and real samples.

[0116] : Domain probability indicator matrix, used to weight and align samples from different domains, normalization factor, and eliminate the effects of sample imbalance.

[0117] All elements are .

[0118] All elements are .

[0119] All elements are .

[0120] Minimize by gradient descent Update the neural network parameters in step S3. This forces the policy network in step S3 to learn cross-domain invariant features, eliminating the transfer error caused by the simplified modeling of the simulation environment.

[0121] in: : No. The set of neural network parameters (including all weights and biases) at the next iteration.

[0122] : No. The parameter set after the next iteration.

[0123] Assignment operation.

[0124] The learning rate determines the magnitude of parameter updates.

[0125] The opposite direction of the gradient.

[0126] : The original task loss function is defined.

[0127] Loss function with respect to parameters The vector of partial derivatives.

[0128] : Domain adaptive equilibrium coefficient, a non-negative scalar hyperparameter.

[0129] The front part of a neural network (feature extraction layer) maps the original input into a high-dimensional feature vector.

[0130] This embodiment also proposes a strategy generation and reality migration system for complex product multi-device collaborative embodied intelligent assembly, including a processor and a memory. The memory stores a computer program. When the computer program is executed by the processor, it implements the strategy generation and reality migration method for complex product multi-device collaborative embodied intelligent assembly as described above.

[0131] The strategy generation, real-world migration, and system for complex product multi-device collaborative intelligent assembly in this embodiment offer the following advantages over existing technologies: Full-process automation and intelligence: It realizes closed-loop intelligent decision-making and execution of the entire process from assembly task modeling, multi-equipment collaborative planning, high-precision trajectory generation to simulation verification and migration, completely replacing manual trial and error and experience adjustment, and realizing an integrated intelligent closed loop of "perception-decision-execution" in complex assembly scenarios.

[0132] Ultra-high precision and superior quality: Based on the deviation propagation model and high-precision compliance strategy, the system can quantitatively analyze the source of error, predict the propagation path, and generate a fine trajectory that meets multiple safety constraints through the diffusion model. Finally, the assembly accuracy (such as gap and surface difference) can be stably controlled at the sub-millimeter level, which is significantly better than the traditional manual or semi-automatic adjustment method, greatly improving product consistency and pass rate.

[0133] High flexibility and rapid changeover adaptability: Through a hierarchical task network and optimal transmission task planning, the system can automatically decompose and schedule assembly tasks based on different product models, batch tolerances, and process requirements. When the production line switches products, only the digital twin model and process parameter library need to be updated to quickly generate a new assembly strategy, significantly shortening the production line debugging and reconfiguration time and improving the ability to produce multiple products on mixed lines.

[0134] Eliminating reliance on manual adjustments and reducing costs while increasing efficiency: Through simulation-based strategy verification and robust real-world migration, the system achieves first-pass assembly qualification, completely eliminating the repetitive adjustments required by skilled workers in traditional assembly. While reducing labor costs and intensity, it ensures high consistency and traceability of the process, providing reliable technical support for fully automated, high-quality assembly.

[0135] The embodiments described above are merely preferred embodiments for fully illustrating the present invention, and the scope of protection of the present invention is not limited thereto. Equivalent substitutions or modifications made by those skilled in the art based on the present invention are all within the scope of protection of the present invention. The scope of protection of the present invention is defined by the claims.

Claims

1. A method for generating and transferring strategies for collaborative intelligent assembly of complex products using multiple devices, characterized in that: Includes the following steps: S1: Construct a hierarchical assembly task network based on the deviation propagation model: Construct a tensor representation model of assembly relationships based on assembly process knowledge; establish a discrete state space model describing the cumulative propagation of deviations in multiple processes based on the representation model to identify key control paths and deviation sensitive points; construct a hierarchical task network model with task nodes and dependency edges based on the assembly process flow to generate an initial assembly task sequence. S2: Multi-device task planning and scheduling based on optimal transmission and swarm intelligence: Based on the results of step S1, a recursive decomposition algorithm is used to decompose high-level assembly tasks into low-level operation units. Based on the entropy regularized optimal transmission theory, a cost matrix between task requirements and equipment capabilities is constructed to solve for the optimal transmission plan and realize the mapping from tasks to equipment actions. Based on the task allocation results, a scheduling model with the goal of minimizing total completion time and load balancing is constructed. An improved swarm intelligence optimization algorithm is used to search for the optimal time scheduling sequence that satisfies the deviation constraint check in non-Euclidean space. S3: Generate a high-precision compliant assembly strategy: Based on the cumulative deviation vector calculated in real time in step S2, construct a dynamic motion primitive model with online deviation compensation to correct the robot's motion target; for the precision contact stage, use a conditional diffusion model to generate the trajectory, and inject a gradient guidance mechanism based on physical feedback during its reverse denoising process to ensure that the trajectory meets force control and obstacle avoidance constraints. S4: Simulation verification and real-world transfer of assembly strategy: Construct a verification sample set covering multiple working conditions using a deep generative model, evaluate and iteratively optimize the strategy generated in step S3 in a simulation environment; Employ domain adaptation technology to fine-tune the strategy network by minimizing the distribution difference between the simulation domain and the real domain in the feature space, thereby achieving robust transfer of the strategy to the physical environment.

2. The method for strategy generation and real-world transfer of complex product multi-device collaborative embodied intelligent assembly according to claim 1, characterized in that: In step S1, the assembly relationship tensorization representation model is trained by assembling relationship tensors The implementation is calculated as follows: in: , representing the feature embedding vectors of active and passive parts respectively, transforming geometric information into computable numerical features; Source: CAD Geometric Analysis The weight matrix of the assembly relationship has the following tensor product structure: wherein: is a nominal homogeneous transformation matrix, comprising a nominal rotation matrix and a nominal translation vector ; is a constraint mask operator; is an activation function.

3. The method for strategy generation and real-world transfer of complex product multi-device collaborative embodied intelligent assembly according to claim 1, characterized in that: In step S1, the discrete state-space model is represented as: in: For the first Cumulative deviation vector after process; The adjoint transformation matrix; Translation vector The antisymmetric matrix is ​​used to... Item capture Abbe error; Input influence matrix; This is the input error vector for this process; This is the overall system disturbance term; The nominal translation vector defines the design position coordinates of the part in the coordinate system. For the first A rotation matrix for assembly relationships.

4. The method for strategy generation and real-world transfer of complex product multi-device collaborative embodied intelligent assembly according to claim 1, characterized in that: In step S2, the task requirements are quantified into a task difficulty vector. Its definition is: in: For the sake of precision and urgency; For design tolerances; Represents the geometric travel of the assembly operation; The complexity is expressed as degrees of freedom; For the first Cumulative deviation vector after process; To constrain the masking operator; The nominal translation vector; Equipment capabilities are quantified into equipment capability vectors. Its definition is: in: For accuracy capability; To ensure repeatability of positioning accuracy; For wingspan ability; The radius of the workspace; For the ability of degrees of freedom; Number of motion axes; Elements of the cost matrix The formula for calculating the weighted Mahalanobis distance between the task difficulty vector and the device capability vector in the feature space is as follows: in: It is a diagonal weight matrix; Weights for bias sensitivity; Spatial dimension weights; Weights for degrees of freedom; For the first The accompanying transformation matrix of the process.

5. The strategy generation and reality transfer method for complex product multi-device collaborative embodied intelligent assembly according to claim 1 or 4, characterized in that: In step S2, the solution to the optimal transmission plan is achieved through the following optimization problem: in: For cost matrix elements; Indicates the task Assigned to device The probability weights; This is the regularization coefficient, which controls the smoothness of the allocation scheme; This is an entropy penalty term.

6. The strategy generation and reality transfer method for complex product multi-device collaborative embodied intelligent assembly according to claim 1 or 4, characterized in that: In step S2, the update law of the improved swarm intelligence optimization algorithm is: in: Let be the position vector, representing the first position vector. The particle in the first The task execution sequence at the next iteration; The velocity vector represents the reordering operation steps; This is a difference operator used to compute sequence exchange steps; To apply operators to perform swap operations; , and These are the inertia weight and the learning factor, respectively. and These are the historical optimal sequence and the globally optimal sequence, respectively. and To introduce parameters for randomness, and .

7. The strategy generation and reality transfer method for complex product multi-device collaborative embodied intelligent assembly according to claim 1 or 4, characterized in that: In step S3, the motion equations of the dynamic motion element model with online deviation compensation are: in: Assemble coordinates in CAD format; For phase variables, initial value ; For nonlinear forced terms, determined by the Gaussian function and shape weights obtained through imitation learning constitute: ; and These represent the current position and velocity of the robot's end effector, respectively. and The gain coefficient defines the stiffness and damping characteristics of the system. This is the time scaling factor; The stable decay rate of the canonical system; For the first Cumulative deviation vector after process; This is the rate of change of position over time, i.e., velocity; This is the rate of change of velocity over time, i.e., acceleration; The rate of change of the phase variable over time; This is the initial position of the robot's end effector.

8. The method for strategy generation and real-world transfer of complex product multi-device collaborative embodied intelligent assembly according to claim 1, characterized in that: In step S3, the gradient guidance mechanism corrects the trajectory during the inverse denoising process of the conditional diffusion model using the following formula: in: The current noisy trajectory; The output of the noise prediction network; To represent network parameters; and These are the single-step retention factor and the cumulative retention factor, respectively. The gradient of the cost function is constructed based on physical feedback; The real-time contact force captured by the sensor; To limit the contact force; is the robot's geometric Jacobian matrix, used to map contact forces in Cartesian space to corrected gradients in joint space; For guiding strength; This is the variance coefficient.

9. The method for strategy generation and real-world transfer of complex product multi-device collaborative embodied intelligent assembly according to claim 1, characterized in that: In step S4, minimizing the distribution difference between the simulation domain and the real domain in the feature space is achieved by minimizing the square of the maximum mean difference, which is calculated as follows: in: The trace of the matrix; , and The kernel matrix measures the similarity of samples within and across domains, and its bandwidth parameter is based on the random perturbation term in the simulation environment. The statistical variance is adaptively configured, and: This is the similarity matrix within the simulated samples; This is the similarity matrix within the real samples; This is the cross-domain similarity matrix between simulated samples and real samples; , and Let be the domain probability indicator matrix, and: All elements are ; All elements are ; All elements are ; This refers to the amount of simulation data. This represents the actual amount of data. Minimize by gradient descent Update the neural network parameters of S3. The policy network is forced to learn cross-domain invariant features, eliminating transfer errors caused by simplistic modeling of the simulation environment. in: For the first The set of neural network parameters at the next iteration; For the first The parameter set after the next iteration; This indicates an assignment operation; The learning rate; Indicates the opposite direction of the gradient; The original task loss function is defined; For the loss function with respect to parameters The vector of partial derivatives; For the domain-adaptive equilibrium coefficient; This is the front-end part of the neural network, which maps the original input into a high-dimensional feature vector.

10. A strategy generation and reality transfer system for multi-device collaborative embodied intelligent assembly of complex products, characterized in that: It includes a processor and a memory, the memory storing a computer program that, when executed by the processor, implements the strategy generation and reality migration method for complex product multi-device collaborative embodied intelligent assembly as described in any one of claims 1-9.

Images