High-precision assembly path planning method for artificial intelligence cooperating with industrial robots

Abstract

The present application relates to a kind of high-precision assembly path planning method of artificial intelligence cooperates industrial robot, the present application replaces the path planning of assembly system high-dimensional configuration space into the multi-group path planning of industrial robot low-dimensional configuration space without considering the collision of industrial robot, then iteratively search and detect the conflict between candidate path, using low-cost candidate path replacement and new candidate path generation, eliminate the conflict between candidate path, so that the collision-free candidate path between industrial robot, the collision-free candidate path combination obtains the assembly system path planning result of multiple groups of collision-free, select the cost optimal from assembly system path planning result as the final assembly system path planning.From the above design, the dimensionality reduction in path planning process is realized, the calculation efficiency of path planning is improved, and the real-time path planning of complex industrial robot assembly system is better satisfied.

Description

Technical Field

[0001] This invention relates to the field of robot path planning technology, and in particular to a high-precision assembly path planning method for artificial intelligence collaborative industrial robots. Background Technology

[0002] The statements in this section are merely background information relating to this disclosure and do not necessarily constitute prior art.

[0003] Each industrial robot has multiple joints, each representing a degree of freedom. The target task often requires an assembly system comprising industrial robots to collaboratively complete the task within the same workspace. When multiple industrial robots work together in the same workspace, the dimensionality of the entire industrial robot assembly system explodes exponentially. Path planning for each industrial robot in the assembly system requires finding a continuous path in the state space that avoids both self-collisions and mutual collisions while satisfying kinematic constraints. Furthermore, as the number of robotic arms in the assembly system increases, the proportion of safe paths in the total space shrinks drastically. The originally spacious feasible region is fragmented into isolated islands, instantly increasing computational complexity. Sampling-based path planners excel in low-dimensional spaces, but when faced with multi-arm assembly systems, the dimensionality of the joint space expands exponentially. This means that to find a collision-free path, the number of nodes the algorithm needs to explore will instantly exceed the computational power limit, leading to the curse of dimensionality. On the other hand, while optimization-based planners can generate smooth paths, they are prone to getting trapped in local minima in the complex non-convex space formed by multi-arm collaboration and are highly dependent on the quality of the initial solution. Summary of the Invention

[0004] To solve the above-mentioned technical problems, or at least partially solve them, the present invention provides a high-precision assembly path planning method for artificial intelligence collaborative industrial robots.

[0005] This invention provides a high-precision assembly path planning method for AI-assisted industrial robots. For an assembly system with multiple industrial robots, it coordinates the multiple industrial robots to perform assembly collaboratively, including: determining the target posture of the assembly actuator of each industrial robot in the world coordinate system according to the assembly target. The configuration space and feasible configuration space of each industrial robot are determined based on its base posture, all joints, joint range of motion, self-collision constraints, and environmental obstacles. Using a pre-trained first path generation model, and taking into account the historical motion sequence including the joint pose, base pose, link position, actuator pose, and target pose of the industrial robot, multiple action sequences with a set future time are independently generated for each industrial robot within the feasible configuration space, so as to achieve the target pose of the assembled actuator in the world coordinate system within a certain time range; the multiple action sequences within a certain time range correspond to multiple candidate paths of the industrial robot. Construct a search tree for all candidate paths of industrial robots, and iteratively detect conflicts between candidate paths of each node based on the search tree. When there is a conflict between candidate paths of nodes, try other candidate paths of the conflicting node, or when there is no feasible candidate path, use a pre-trained second path generation model with the configuration to be adjusted corresponding to a conflicting candidate path as input and the conflicting configuration corresponding to another repeated candidate path as condition to generate a new configuration to be adjusted. The final path is selected from the conflict-free candidate paths of the assembly system, based on the optimal cost in terms of path smoothness and target proximity.

[0006] Furthermore, for any industrial robot k, it has If there are 1 joint, then the dimension is 1. vector space Generalized coordinates of configuration space The configuration of industrial robot k is represented; further considering the constraints of joint mobility range, self-collision constraints and environmental obstacle constraints of industrial robot, the feasible configuration space is determined. Based on the configuration space, the movable range of the joint is used to construct the outer boundary of the feasible configuration region to obtain the bounded configuration space. , ; i represents the configuration space of industrial robot k. arbitrary generalized coordinates Dimension index; In bounded configuration space Within the region, self-collision constraints are used to obtain the self-collision obstacle region. These constraints restrict the industrial robot from self-collision, and the self-collision is represented as follows: ; in, This represents the self-collision obstacle region for industrial robot k. These are the robot linkage numbers, ranging from 0 to... ,in As a base, To assemble the actuator, Limiting non-connected links They represent the configurations respectively. Below, the geometry of the u-th and v-th links in the work group space; Obstacles in the workspace of an industrial robot are mapped to its configuration space using the Minkowski algorithm, resulting in the configuration space obstacle region. ; Based on the bounded configuration space, the self-collision barrier region, and the configuration space barrier region, the feasible configuration space is obtained as follows: .

[0007] Furthermore, the process of the homogeneous transformation matrix of the link relative to the base coordinate system includes: For each link n of the industrial robot k, the coordinate system is defined according to the DH convention: Z-axis: along the axis of the nth joint, X-axis: along the common perpendicular of the axes of link n and the (n+1)th joint, and Y-axis is determined by the right-hand rule, making the coordinate system a right-hand system. Determine the DH parameters and use them to write the DH transformation matrix for a single link: For each pair of adjacent joints, define the following four DH parameters: joint angles. Linkage offset Link length and connecting rod torsion angle The homogeneous transformation from the link n-1 coordinate system to the link n coordinate system is as follows: ; Multiplying the homogeneous transformation matrices of the preceding links' coordinate systems yields the homogeneous transformation matrix of the link relative to the base coordinate system: .

[0008] Furthermore, the first path generation model employs a conditional diffusion model. During training, paths are randomly selected from the path dataset and subjected to S-step noise diffusion. A denoising iteration step s is randomly chosen, and the noise corresponding to each iteration step is obtained based on the diffusion process. ; Training requires the first-path generation model to predict the added noise; the process of the first-path generation model generating action sequences is modeled as a stepwise denoising process. This process denoises pure noise samples that follow a Gaussian distribution through S iterations, generating a series of intermediate actions with progressively lower noise levels, ultimately resulting in noise-free actions. Each step of the denoising process follows the following equation: ; in It is the first path generation model to be learned and optimized, with parameters as follows: , This indicates that the first path generation model, guided by the action sequence conditions at time step t in the s-th iteration step, predicts noisy actions based on the actions predicted in the s-th iteration step. The predicted noisy action at the (s-1)th iteration step Noise between them; Gaussian noise is added in each iteration; action is the change in feasible configuration; parameters , and The noise scheduling coefficient is related to the number of iteration steps s; The training loss function defines the random noise sampled. The mean square error between the predicted noise and the actual noise is: ; Minimizing the loss function is equivalent to minimizing the variational lower bound of the KL divergence between the path distribution and the path distribution sampled from the diffusion model during the diffusion process, thus enabling the first-condition model to predict feasible paths.

[0009] Furthermore, the paths in the path dataset are generated by the bidirectional fast expanding random tree algorithm. The bidirectional fast expanding random tree algorithm simultaneously grows two trees outward from any initial configuration corresponding to the initial pose and any target configuration corresponding to the target pose of the assembly actuator in the world coordinate system. The two trees are grown by randomly sampling joint parameters. When the distance between the two trees in the configuration space is less than a set threshold and there is no collision, the two trees are connected to form a complete path from the initial configuration to the target configuration.

[0010] Furthermore, the action sequence conditions are mapped into condition coefficients and condition biases by the linear layer. The condition coefficients and condition biases are applied to the actions encoded by each action encoding layer of the first path generation model. The action encoding layer can be a convolutional layer or a fully connected layer.

[0011] Furthermore, the second path generation model has the same structure as the first path generation model, but introduces an additional conflict cost loss into the loss function. The conflict cost loss is the L2 norm between the new configuration to be adjusted and the corresponding conflict configuration. Correspondingly, in the denoising process, an L2 norm gradient correction between the new configuration to be adjusted and the corresponding conflict configuration is introduced.

[0012] Furthermore, to ensure that the final path is selected from the conflict-free candidate paths of the assembly system based on the optimal cost (indicating path smoothness and target proximity), when constructing candidate paths for the assembly system using a search tree, the cost corresponding to each candidate path of a node is calculated. When selecting a candidate path for a node, the path with the lowest replacement cost is selected as the candidate path for the assembly system to construct the final path. When there is a conflict between candidate paths of a node, when trying other candidate paths of the conflicting node, the path with the second-lowest cost is tried instead.

[0013] Secondly, the present invention provides a high-precision assembly path planning device for an artificial intelligence collaborative industrial robot, comprising: at least one processing unit, the processing unit being connected to a storage unit via a bus unit, the storage unit storing a computer program that can run on a processor, and the processing unit implementing the high-precision assembly path planning method for the artificial intelligence collaborative industrial robot by running the computer program stored in the storage unit.

[0014] Thirdly, the present invention provides a computer-readable storage medium storing a computer program, which, when executed, implements the high-precision assembly path planning method for the artificial intelligence collaborative industrial robot.

[0015] The technical solutions provided in the embodiments of the present invention have the following advantages compared with the prior art: This invention proposes a high-precision assembly path planning method for AI-assisted industrial robots. It replaces the path planning in the high-dimensional configuration space of the assembly system with multiple sets of path planning in the low-dimensional configuration space of the industrial robots, without considering collisions between them. Then, iteratively searches and detects conflicts between candidate paths, eliminating conflicts through low-cost candidate path replacement and new candidate path generation, thereby obtaining collision-free candidate paths. For all industrial robots in the assembly system, their collision-free candidate paths are combined to obtain multiple sets of assembly system path planning results. The path with the lowest cost is selected as the final assembly system path plan. This design achieves dimensionality reduction in the path planning process, improves computational efficiency, and better meets the real-time path planning requirements of complex industrial robot assembly systems. Attached Figure Description

[0016] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with the invention and, together with the description, serve to explain the principles of the invention.

[0017] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 A flowchart illustrating a high-precision assembly path planning method for an AI-assisted industrial robot provided in an embodiment of the present invention; Figure 2 A schematic diagram of an assembly system for an industrial robot provided in an embodiment of the present invention; Figure 3A flowchart for determining the feasible configuration space provided in an embodiment of the present invention; Figure 4 A flowchart for determining the homogeneous transformation matrix of the target link relative to the base coordinate system, provided in an embodiment of the present invention; Figure 5 This is a schematic diagram of a high-precision assembly path planning device for an artificial intelligence collaborative industrial robot, provided as an embodiment of the present invention. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0021] Example 1 This invention provides a high-precision assembly path planning method for AI-assisted industrial robots, such as... Figure 2 As shown, for an assembly system with multiple industrial robots, the multiple industrial robots are coordinated to perform assembly collaboratively; the assembly system includes K industrial robots formed by robotic arms.

[0022] like Figure 1 As shown, a high-precision assembly path planning method for AI-assisted industrial robots includes: S100 determines the target orientation of the assembly actuator of each industrial robot in the assembly system in the world coordinate system based on the assembly target.

[0023] S200 determines the configuration space and feasible configuration space of each industrial robot based on its base posture, all joints, joint range of motion, self-collision constraints, and environmental obstacles.

[0024] Configuration space is a mathematical space describing all possible configurations of an industrial robot. Mapping all configurations of the industrial robot to points in configuration space, the robot's configuration is represented by the positions of these points; it is determined by the positions (angles or displacements) of all joints of the industrial robot. For any industrial robot k, it has… There are 1 joint; therefore, its configuration space has a dimension of . vector space That is, the generalized coordinates of configuration space. The dimension is In practice, if the joint is a rotary joint, the parameter of the corresponding dimension of the generalized coordinate system is the angle; if the joint is a translational joint, the parameter of the corresponding dimension of the closed coordinate system is the displacement.

[0025] Further consideration is given to constraints on the joint range of motion, self-collision constraints, and environmental obstacle constraints of industrial robots to determine the feasible configuration space, such as... Figure 3 As shown, it includes: Based on the configuration space, the joint's range of motion constructs the outer boundary of the feasible configuration region, thus obtaining the bounded configuration space. , ; i represents the configuration space of industrial robot k. arbitrary generalized coordinates The dimensional index. That is, the product of the movable range of each joint of the industrial robot k defines the outer boundary of the configuration space.

[0026] In bounded configuration space Within a given industrial robot k, during operation, it is required that its components do not collide with each other; that is, there should be no collisions between links or between links and the base. The self-collision constraint is expressed as follows: ; in, The self-collision obstacle region of industrial robot k is a bounded configuration space. A subset of , containing all configurations that would cause collisions between links within the industrial robot itself, and between links and the base. This indicates the existence of a pair of linked indexes. , These are the robot linkage numbers, from 0 to... ,in As a base, To assemble the actuator, This applies to non-connected links, because adjacent links are connected by joints, and their contact is normal and not considered a collision. They represent the configurations respectively. Let the geometry of the u-th and v-th links in the workspace be given. The pose of each link in the workspace is calculated using the mapping relationship from configuration space to workspace. To establish the mapping relationship between configuration space and workspace, based on the geometric relationships of the industrial robot base, joints, and links, and the given industrial robot configuration, the homogeneous transformation matrix of the link relative to the base coordinate system is calculated: ; in: For rotation matrix (attitude). For position vectors, This is a transformation from the link's n-coordinate system to the base coordinate system. For example... Figure 4 As shown, the process includes: For industrial robot k, establish a link coordinate system: For each link n, define the coordinate system according to the DH convention: Z-axis: along the axis direction of the nth joint, X-axis: along the common perpendicular direction of the axes of link n and the (n+1)th joint, Y-axis is determined by the right-hand rule, making the coordinate system a right-hand system; Determine the DH parameters and use them to write the DH transformation matrix for a single link: For each pair of adjacent joints, define the following four parameters: joint angles. Linkage offset Link length and connecting rod torsion angle The homogeneous transformation from the link n-1 coordinate system to the link n coordinate system is as follows: ; Multiplying the homogeneous transformation matrices of the target link's preceding link coordinate system yields the homogeneous transformation matrix of the target link relative to the base coordinate system: .

[0027] The base pose serves as the reference origin for the transformation from the workspace to the configuration space; it does not affect the structure and dimensions of the configuration space, but only the mapping relationship between the configuration space and the workspace. Changing the base pose will translate / rotate the workspace, but the configuration space of the industrial robot itself remains unchanged.

[0028] Obstacles in the workspace of an industrial robot are mapped to its configuration space using the Minkowski algorithm, resulting in the configuration space obstacle region. For industrial robots, obstacles in the workspace are considered only as static obstacles.

[0029] Obstacle O in the workspace is mapped to the configuration space via Minkowski and: , in, Let q be the set of geometric bodies of the industrial robot. This is a Minkowski operation. This operation expands industrial robots with complex shapes into points, and obstacles are enlarged accordingly, simplifying collision detection.

[0030] The feasible configuration space is then: .

[0031] Through the above process, the complex workspace industrial robot planning problem is transformed into a path planning problem of moving within a configuration space while avoiding obstacle zones. A generative diffusion model can be used to generate the path between the industrial robot's initial and target postures.

[0032] S300 utilizes a pre-trained first path generation model, taking into account historical motion sequences including joint poses, base poses, link positions, actuator poses, and target poses of the industrial robot, to independently generate future motion sequences for each industrial robot within the feasible configuration space, so as to achieve the target pose of the assembled actuator in the world coordinate system within a certain time range; the motion sequences within a certain time range form the candidate paths of the industrial robot.

[0033] The first path generation model adopts a conditional diffusion model. The dataset used to train the first path generation model comes from paths generated by the bidirectional fast expanding random tree algorithm. The bidirectional fast expanding random tree simultaneously grows two trees outward from any initial configuration corresponding to the initial pose and any target configuration corresponding to the target pose of the assembly actuator in the world coordinate system. The two trees grow by randomly sampling joint parameters. When the distance between the two trees in the configuration space is less than a set threshold and there is no collision, the two trees are connected to form a complete path from the initial configuration to the target configuration. The obtained path is then subjected to noise diffusion. The distribution sampled during the diffusion process is used to obtain the distribution of each path.

[0034] During training, unmodified paths are first randomly selected from the dataset. For each selected path, a denoising iteration step s is randomly chosen, and the noise corresponding to the iteration step is obtained according to the diffusion process. The training requires the first-path generation model to predict the added noise. Specifically, the process of the first-path generation model generating action sequences is modeled as a progressive denoising process. This process takes pure noise samples that follow a Gaussian distribution, denoises them through S iterations, and generates a series of intermediate results with progressively lower noise levels. Ultimately, the ideal noise-free output is obtained. Each step of the denoising process follows the equation below: ; in It is the first path generation model to be learned and optimized, with parameters as follows: , This represents the action sequence condition of the first path generation model at time step t in the s-th iteration. Guided by the noisy action predicted in the s-th iteration step The predicted noisy action at the (s-1)th iteration step Noise between them; To diffuse the Gaussian noise introduced in each iteration, in this application, the action is the change in the feasible configuration space. Parameters , and The noise scheduling coefficient is related to the number of iteration steps s.

[0035] The training loss function defines the random noise sampled. The mean square error between the predicted noise and the actual noise is: ; Minimizing the loss function is equivalent to minimizing the variational lower bound of the KL divergence between the path distribution and the path distribution sampled from the diffusion model during the diffusion process, thus enabling the first-condition model to predict feasible paths.

[0036] At time step t, the first path generation model receives the historical motion sequence conditions within the most recent conditional time window. As input, it predicts a sequence of actions within a predetermined time frame; the first Te steps of the future action sequence are executed directly until the assembly actuator reaches the target posture in the world coordinate system. The action sequence within a certain time frame forms the candidate path for the industrial robot.

[0037] Candidate paths are generated in a low-dimensional industrial robot configuration space using the first path generation model.

[0038] The action sequence conditions are mapped into condition coefficients and condition biases by a linear layer. The condition coefficients and condition biases are applied to the actions encoded by each action encoding layer of the first path generation model. The action encoding layer can be a convolutional layer or a fully connected layer.

[0039] Taking the Unet architecture convolutional model as an example, the first path generation model uses each convolutional layer of the Unet encoder as an action encoding layer. The action sequence conditions are mapped into conditional coefficients and conditional biases by the linear layer. The conditional coefficients and conditional biases are applied to the actions encoded by each action encoding layer of the first path generation model.

[0040] S400 constructs a search tree of all candidate paths for industrial robots and iteratively detects conflicts between all candidate paths of all pairs of nodes based on the search tree. When there is a conflict between candidate paths of nodes, other candidate paths of the conflicting nodes are tried, or when there is no feasible candidate path, a new configuration to be adjusted is generated by using a pre-trained second path generation model with the configuration to be adjusted corresponding to a conflicting candidate path as input and the configuration to be adjusted corresponding to another repeated candidate path as a condition.

[0041] The second path generation model has the same structure as the first path generation model, but introduces an additional conflict cost loss into the loss function. The conflict cost loss is the L2 norm between the new configuration to be adjusted and the corresponding conflict configuration. Correspondingly, in the denoising process, an L2 norm gradient correction between the new configuration to be adjusted and the corresponding conflict configuration is introduced.

[0042] This process generates conflict-free candidate paths for each industrial robot.

[0043] S500 combines the conflict-free candidate paths of each industrial robot to obtain the assembly system candidate path, and selects the one with the best cost based on path smoothness and target proximity as the final path planning of the assembly system.

[0044] To select the most cost-optimal path from conflict-free candidate paths in the assembly system, based on path smoothness and target proximity, the cost of each candidate path for a node is calculated when constructing candidate paths using a search tree. The path with the lowest cost is selected as the final path. Furthermore, when candidate paths conflict, other candidate paths for the conflicting node are considered, and the path with the second-lowest cost is replaced. This eliminates the need for conflict detection and handling of numerous high-cost candidate paths. In other words, conflict handling is performed directly in S400 only on candidate paths with lower costs, directly yielding the final path plan for the assembly system.

[0045] Example 2 like Figure 5 As shown, this embodiment of the invention provides a high-precision assembly path planning device for an AI-assisted collaborative industrial robot, comprising: at least one processing unit, the processing unit being connected to a storage unit via a bus unit, the storage unit serving as a computer-readable storage medium for storing software programs, computer-executable programs, and modules, such as the software program, computer-executable program, and modules corresponding to the high-precision assembly path planning method for an AI-assisted collaborative industrial robot in this embodiment of the invention. The processing unit implements the aforementioned high-precision assembly path planning method for an AI-assisted collaborative industrial robot by running the software program, computer-executable program, and modules stored in the storage unit.

[0046] Of course, the computer program stored in the memory of the high-precision assembly path planning device for an artificial intelligence collaborative industrial robot provided in the embodiments of the present invention is not limited to the method operation described above, but can also execute related operations in the high-precision assembly path planning method for an artificial intelligence collaborative industrial robot provided in any embodiment of the present invention.

[0047] Example 3 This invention provides a computer-readable storage medium storing a computer program, characterized in that, when the computer program is executed, it implements the high-precision assembly path planning method for artificial intelligence collaborative industrial robots.

[0048] In the embodiments provided by this invention, it should be understood that the disclosed structures and methods can be implemented in other ways. For example, the structural embodiments described above are merely illustrative. For instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another assembly system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be an indirect coupling or communication connection through some interfaces, structures, or units, and may be electrical, mechanical, or other forms.

[0049] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0050] Furthermore, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0051] The above description is merely a specific embodiment of the present invention, enabling those skilled in the art to understand or implement the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the present invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features claimed herein.

Claims

1. A high-precision assembly path planning method for AI-assisted industrial robots, for assembly systems with multiple industrial robots, coordinating the collaborative assembly of multiple industrial robots, characterized in that... include: Determine the target orientation of the assembly actuator of each industrial robot in the assembly system in the world coordinate system based on the assembly target. The configuration space and feasible configuration space of each industrial robot are determined based on its base posture, all joints, joint range of motion, self-collision constraints, and environmental obstacles. Using a pre-trained first path generation model, and taking into account the historical motion sequence including the joint pose, base pose, link position, actuator pose, and target pose of the industrial robot, multiple action sequences with a set future time are independently generated for each industrial robot within the feasible configuration space, so as to achieve the target pose of the assembled actuator in the world coordinate system within a certain time range; the multiple action sequences within a certain time range correspond to multiple candidate paths of the industrial robot. The first path generation model employs a conditional diffusion model. During training, paths are randomly selected from the path dataset and subjected to S-step noise diffusion. A denoising iteration step s is randomly chosen, and the noise corresponding to each iteration step is obtained based on the diffusion process. ; Training requires the first-path generation model to predict the added noise; the process of the first-path generation model generating action sequences is modeled as a stepwise denoising process. This process denoises pure noise samples that follow a Gaussian distribution through S iterations, generating a series of intermediate actions with progressively lower noise levels, ultimately resulting in noise-free actions. Each step of the denoising process follows the following equation: ； in It is the first path generation model to be learned and optimized, with parameters as follows: , This represents the action sequence condition of the first path generation model at time step t in the s-th iteration. Guided by the noisy action predicted in the s-th iteration step The predicted noisy action at the (s-1)th iteration step Noise between them; The Gaussian noise added in each iteration during diffusion; the action is the change in the feasible configuration; the parameters , and The noise scheduling coefficient is related to the number of iteration steps s; The training loss function defines the random noise sampled. The mean square error between the predicted noise and the actual noise is: ； Minimizing the loss function is equivalent to minimizing the variational lower bound of the KL divergence between the path distribution and the path distribution sampled from the diffusion model during the diffusion process, thus enabling the first-conditional model to predict feasible paths. Construct a search tree for all candidate paths of industrial robots, and iteratively detect conflicts between candidate paths of each node based on the search tree. When there is a conflict between candidate paths of nodes, try other candidate paths of the conflicting node, or when there is no feasible candidate path, use a pre-trained second path generation model with the configuration to be adjusted corresponding to a conflicting candidate path as input and the conflicting configuration corresponding to another repeated candidate path as an incremental condition to generate a new configuration to be adjusted. The candidate paths of each industrial robot without conflict are combined to obtain the candidate paths of the assembly system. The candidate paths of the assembly system with the best cost, based on the smoothness of the path and the proximity to the target, are selected as the final path planning of the assembly system.

2. The high-precision assembly path planning method for AI-assisted industrial robots according to claim 1, characterized in that, For any industrial robot k, it has If there are 1 joint, then the dimension is 1. vector space Generalized coordinates of configuration space The configuration of industrial robot k is represented; further considering the constraints of joint mobility range, self-collision constraints and environmental obstacle constraints of industrial robot, the feasible configuration space is determined. Based on the configuration space, the movable range of the joint is used to construct the outer boundary of the feasible configuration region to obtain the bounded configuration space. , ; i represents the configuration space of industrial robot k. arbitrary generalized coordinates Dimension index; In bounded configuration space Within the region, self-collision constraints are used to obtain the self-collision obstacle region. These constraints restrict the industrial robot from self-collision, and the self-collision is represented as follows: ； in, This represents the self-collision obstacle region for industrial robot k. These are the robot linkage numbers, ranging from 0 to... ,in As a base, To assemble the actuator, Limiting non-connected links They represent the configurations respectively. Below, the geometry of the u-th and v-th links in the work group space; Obstacles in the workspace of an industrial robot are mapped to its configuration space using the Minkowski algorithm, resulting in the configuration space obstacle region. ; Based on the bounded configuration space, the self-collision barrier region, and the configuration space barrier region, the feasible configuration space is obtained as follows: 。 3. The high-precision assembly path planning method for AI-assisted industrial robots according to claim 2, characterized in that, The process of performing a homogeneous transformation of the link relative to the base coordinate system includes: For each link n of the industrial robot k, the coordinate system is defined according to the DH convention: Z-axis: along the axis of the nth joint, X-axis: along the common perpendicular of the axes of link n and the (n+1)th joint, and Y-axis is determined by the right-hand rule, making the coordinate system a right-hand system. Determine the DH parameters and use them to write the DH transformation matrix for a single link: For each pair of adjacent joints, define the following four DH parameters: joint angles. Linkage offset Link length and connecting rod torsion angle The homogeneous transformation from the link n-1 coordinate system to the link n coordinate system is as follows: ； Multiplying the homogeneous transformation matrices of the preceding links' coordinate systems yields the homogeneous transformation matrix of the link relative to the base coordinate system: 。 4. The high-precision assembly path planning method for AI-assisted industrial robots according to claim 1, characterized in that, The paths in the path dataset are generated by the bidirectional fast expanding random tree algorithm. The bidirectional fast expanding random tree algorithm grows two trees outward from any initial configuration corresponding to the initial pose and any target configuration corresponding to the target pose of the assembly actuator in the world coordinate system. The two trees are grown by randomly sampling joint parameters. When the distance between the two trees in the configuration space is less than a set threshold and there is no collision, the two trees are connected to form a complete path from the initial configuration to the target configuration.

5. The high-precision assembly path planning method for AI-assisted industrial robots according to claim 1, characterized in that, The action sequence conditions are mapped into condition coefficients and condition biases by a linear layer. The condition coefficients and condition biases are applied to the actions encoded by each action encoding layer of the first path generation model. The action encoding layer can be a convolutional layer or a fully connected layer.

6. The high-precision assembly path planning method for AI-assisted industrial robots according to claim 1, characterized in that, The second path generation model has the same structure as the first path generation model, but introduces an additional conflict cost loss into the loss function. The conflict cost loss is the L2 norm between the new configuration to be adjusted and the corresponding conflict configuration. Correspondingly, in the denoising process, an L2 norm gradient correction between the new configuration to be adjusted and the corresponding conflict configuration is introduced.

7. The high-precision assembly path planning method for AI-assisted industrial robots according to claim 1, characterized in that, To select the most cost-optimal path from the conflict-free candidate paths of the assembly system, with path smoothness and target proximity as indicators, when constructing candidate paths for the assembly system using a search tree, the cost of each candidate path for a node is calculated. When selecting candidate paths for a node, the path with the lowest replacement cost is selected as the candidate path for the assembly system to construct the final path. When there is a conflict between candidate paths for a node, try replacing other candidate paths for the conflicting node with the less optimal path in terms of cost.

8. A high-precision assembly path planning device for an artificial intelligence collaborative industrial robot, comprising: At least one processing unit, the processing unit being connected to a storage unit via a bus unit, the storage unit storing a computer program that can run on a processor, characterized in that the processing unit implements the high-precision assembly path planning method for an artificial intelligence collaborative industrial robot as described in any one of claims 1-7 by running the computer program stored in the storage unit.

9. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed, it implements the high-precision assembly path planning method for artificial intelligence collaborative industrial robots as described in any one of claims 1-7.

Images