A method and apparatus for an industrial agent to plan a path

Abstract

The application discloses a method and device for planning a path of an industrial agent, and relates to the field of industrial intelligent control and robot offline programming. The method is deployed in a local industrial agent system, and the method comprises: performing vectorization coding matching processing on business data and local knowledge base entries to form a unified task description body; obtaining a structured planning semantic layer result based on the unified task description body, and obtaining a semantic action mapping path from the structured planning semantic layer result; performing inverse kinematics solving processing and interpolation calculation processing on the semantic action mapping path to obtain a candidate path; performing simulation detection processing on the candidate path to obtain a simulation detection result, and generating a final planning path according to the simulation detection result and issuing the final planning path to a robot control module. In the application, an autonomous closed loop of planning, checking and correction is formed, the overall global re-planning calculation overhead can be reduced, the real-time performance and robustness are improved, and the production safety and process precision consistency are improved at the same time.

Description

Technical Field

[0001] This application relates to the fields of industrial intelligent control and robot offline programming technology, and in particular to a method and apparatus for planning paths for an industrial intelligent agent. Background Technology

[0002] In industrial scenarios such as discrete manufacturing, assembly, welding, and grinding, robots performing complex technological tasks typically rely on offline programming by humans and the experience of process experts, resulting in low efficiency, high maintenance costs, and poor adaptability. In recent years, task planning and code generation methods based on large language models have improved the efficiency of task understanding and planning. However, industrial sites often have network isolation requirements, data compliance constraints, and real-time requirements, rendering large cloud-based models unusable.

[0003] Existing offline industrial planning methods typically employ a global replanning strategy: when the simulation detects collisions, unreachability, or process deviations, the entire path is replanned. This approach is computationally expensive, slow to respond, and prone to introducing path jitter, making it difficult to guarantee production continuity and safety. Therefore, there is an urgent need for a method suitable for a fully offline environment, capable of completing the planning-verification-correction closed loop locally, and significantly reducing the computational cost of replanning. Summary of the Invention

[0004] The purpose of this application is to at least address one of the aforementioned technical deficiencies.

[0005] On one hand, embodiments of this application provide a method for planning paths in an industrial intelligent agent. This method is deployed in a local industrial intelligent agent system and includes: The system acquires the business data and local knowledge base entries corresponding to the industrial business to be executed, and performs vectorized encoding matching on the business data and local knowledge base entries to form a unified task description body. The business data includes production task text, workpiece model, design engineering model and process constraint information. The structured planning semantic layer result is obtained based on the unified task description body, and the structured planning semantic layer result is mapped to the preset action template and action constraints are injected to obtain the semantic action mapping path; The semantic action mapping path is solved by inverse kinematics to obtain the joint space target vector sequence, and the joint space target vector sequence is interpolated to obtain the candidate path; The candidate paths are simulated and tested to obtain simulation results. Based on the simulation results, the final planned path is generated and sent to the robot control module so that the robot control module can execute the industrial tasks to be performed.

[0006] Optionally, the business data and local knowledge base entries are vectorized and matched to form a unified task description body, including: The text of each production task and the text of each knowledge entry in the local knowledge base are vectorized and encoded by a local semantic encoder to obtain the semantic vector of each task and the vector of each knowledge entry. A unified task description body is obtained by filtering for relevance based on the semantic vectors of each task and the vectors of each knowledge item.

[0007] Optionally, a unified task description body is obtained by performing relevance filtering based on the semantic vectors of each task and the vectors of each knowledge item, including: Determine the cosine similarity between the semantic vector of each task and the vector of each knowledge entry, and obtain the cosine similarity corresponding to the semantic vector of each task; Obtain a preset relevance threshold, and determine the semantic vectors of tasks whose cosine similarity is greater than the preset relevance threshold as the target task semantic vectors; The target process parameter matrix, target knowledge entry vector, and target space constraint vector corresponding to the target task semantic vector are determined, and fusion encoding is performed based on the target task semantic vector, target process parameter matrix, target knowledge entry vector, and target space constraint vector to obtain a unified task description body.

[0008] Optionally, the semantic layer result of structured planning includes at least one of the following: process sequence, key pose set, and constraint set, wherein the key pose belongs to the pose space.

[0009] Optionally, inverse kinematics is performed on the semantic action mapping path to obtain a joint space target vector sequence, and interpolation is performed on the joint space target vector sequence to obtain candidate paths, including: Inverse kinematics is performed on the set of key poses included in the semantic action mapping path to obtain the joint space target vector sequence. Candidate paths are obtained by interpolating the target vector sequence in the joint space using either linear interpolation or sample interpolation.

[0010] Optionally, the simulation detection process includes collision detection, joint limit detection, and process accuracy detection. The simulation detection result is either verification passed or verification abnormal. Specifically, when at least one of the following is abnormal: the collision detection result corresponding to the collision detection, the joint detection result corresponding to the joint limit detection, and the precision detection result corresponding to the process precision detection, the simulation detection result is considered an anomaly.

[0011] Optionally, the candidate paths are subjected to simulation detection processing to obtain simulation detection results, including: When the simulation detection process is collision detection, the minimum distance between the bounding body and the set of obstacles of each trajectory point in the candidate path in the corresponding posture is determined. When the minimum distance is less than the preset safe distance threshold, the collision detection result is judged to be abnormal. When the simulation detection process is a joint limit detection, if the angle of any joint in the candidate path exceeds the preset joint angle range, the joint detection result is determined to be abnormal. When the simulation detection process is used for process accuracy detection, the deviation between the actuator position and the reference position of the end trajectory point in the candidate path is calculated. When the deviation is greater than the process deviation threshold, the accuracy detection result is judged to be abnormal.

[0012] Optionally, based on the simulation results, a final planned path is generated and sent to the robot control module, including: When the simulation test result is a verification anomaly, an anomaly flag is generated and a local replanning process is triggered to reconstruct the path segment within the window and obtain the path segment outside the window; the path segment within the window and the path segment outside the window are spliced ​​together to obtain the final planned path and sent to the robot control module; When the simulation test result is that the verification is passed, the candidate path is used as the final planned path and sent to the robot control module.

[0013] Optionally, the window range in local replanning is determined by the initial outlier index and the window radius.

[0014] On the other hand, embodiments of this application provide an apparatus for planning paths in an industrial intelligent agent. This apparatus is deployed within a local industrial intelligent agent system and includes: The planning semantic output unit is used to obtain the business data and local knowledge base entries corresponding to the industrial business to be executed, and to perform vectorized encoding matching processing on the business data and local knowledge base entries to form a unified task description body. The business data includes production task text, workpiece model, design engineering model and process constraint information. The path construction unit is used to obtain the structured planning semantic layer result based on the unified task description body, map the structured planning semantic layer result to the preset action template and inject action constraints to obtain the semantic action mapping path; perform inverse kinematics solution on the semantic action mapping path to obtain the joint space target vector sequence, and perform interpolation calculation on the joint space target vector sequence to obtain the candidate path; The simulation detection unit is used to perform simulation detection processing on candidate paths and obtain simulation detection results. The local replanning unit is used to generate the final planned path based on the simulation test results and send it to the robot control module so that the robot control module can execute the industrial tasks to be performed.

[0015] In another aspect, embodiments of this application provide an electronic device, including a processor and a memory: The memory is configured to store machine-readable instructions that, when executed by the processor, cause the processor to perform any of the methods in an industrial agent's path planning approach.

[0016] The beneficial effects of the technical solutions provided in this application include at least the following: In this application, candidate paths can be determined by encoding and matching based on business data corresponding to the industrial business to be executed and entries in the local knowledge base. Then, the candidate paths are subjected to linkage constraints on collisions, joint limits, and process deviations in a simulation environment, ultimately yielding the path executed by the robot module. This process forms an autonomous closed loop of planning, verification, and correction, enabling task understanding, knowledge retrieval, plan generation, and simulation verification even in a network-isolated environment, meeting data compliance and on-site safety requirements. Furthermore, collisions, joint limits, and process deviations are uniformly converted into discrete judgment signals and anomaly identifiers are generated, improving auditability, reproducibility, and production process safety and consistency. Simultaneously, after verification failure, a local replanning window is constructed centered on the first anomaly point. Only the path fragments within the window are reconstructed and the paths outside the window are spliced ​​and reused, significantly reducing the computational cost of path replanning and improving the real-time performance, quality, and stability of the path. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. 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 method for planning a path for an industrial intelligent agent, provided in an embodiment of this application; Figure 2 This is a schematic diagram of the local industrial intelligent agent system structure provided in the embodiments of this application; Figure 3 A flowchart illustrating the thought process of an industrial intelligent agent path planning method provided in this application embodiment; Figure 4 A schematic diagram of the structure of an industrial intelligent agent path planning device provided in an embodiment of this application; Figure 5 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Detailed Implementation

[0019] The embodiments of this application are described in detail below. Examples of these embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain this application, and should not be construed as limiting the invention.

[0020] Those skilled in the art will understand that, unless specifically stated otherwise, the singular forms “a,” “an,” “the,” and “the” used herein may also include the plural forms. It should be further understood that the term “comprising” as used in this application means the presence of the stated features, integers, steps, operations, elements, and / or components, but does not exclude the presence or addition of one or more other features, integers, steps, operations, elements, components, and / or groups thereof. It should be understood that when we say an element is “connected” or “coupled” to another element, it can be directly connected or coupled to the other element, or there may be intermediate elements. Furthermore, “connected” or “coupled” as used herein can include wireless connections or wireless coupling. The term “and / or” as used herein includes all or any units and all combinations of one or more associated listed items.

[0021] To make the objectives, technical solutions, and advantages of this application clearer, the embodiments of this application will be described in further detail below with reference to the accompanying drawings.

[0022] The technical solution of this application and how the technical solution of this application solves the above-mentioned technical problems are described in detail below with specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments. The embodiments of this application will now be described with reference to the accompanying drawings.

[0023] Specifically, such as Figure 1 As shown, the method may include: Step S11: Obtain the business data and local knowledge base entries corresponding to the industrial business to be executed, and perform vectorized encoding matching processing on the business data and local knowledge base entries to form a unified task description body. The business data includes production task text, workpiece model, design engineering model and process constraint information.

[0024] Optionally, business data including production task text, workpiece model, design engineering model, and process constraint information can be acquired. Here, production task text refers to the text data corresponding to the industrial business to be executed; workpiece model refers to the workpiece CAD / point cloud model; design engineering model refers to the equipment / workstation model; and process constraint information refers to the constraint information related to the process during the execution of the industrial business. Furthermore, a workspace model can be formed based on the production task text, workpiece CAD / point cloud model, equipment / workstation model, and process constraint information. It is understood that this workspace model is not for one-time use but is used throughout the entire solution; simply put, it represents the part that needs to be reprogrammed based on the coordinates.

[0025] Optionally, local knowledge base entries can be obtained, and then the business data and local knowledge base entries can be vectorized and matched to form a unified task description body.

[0026] In optional embodiments of this application, the business data and local knowledge base entries are subjected to vectorized encoding matching processing to form a unified task description body, including: The text of each production task and the text of each knowledge entry in the local knowledge base are vectorized and encoded by a local semantic encoder to obtain the semantic vector of each task and the vector of each knowledge entry. A unified task description body is obtained by filtering for relevance based on the semantic vectors of each task and the vectors of each knowledge item.

[0027] Optionally, when performing vectorized encoding matching to form a unified task description, a local semantic encoder can be pre-configured locally. Then, based on the local semantic encoder, vectorized encoding processing can be performed on each production task text and each knowledge entry text in the local knowledge base, resulting in task semantic vectors corresponding to each production task text and knowledge entry vectors corresponding to each knowledge entry text. For example, a semantic vector can be defined in the following form: v_task = f_enc(T_task) ∈ R^d v_k^(i) = f_enc(T_k^(i)) ∈ R^d, i=1..N V_K = { v_k^(1), v_k^(2), …, v_k^(N)} Where f_enc is the local semantic encoder, T_task is the production task text, T_k^(i) is the knowledge entry text, d is the embedding dimension, v_task is the task semantic vector, v_k^(i) is the knowledge entry vector, and V_K is the computation object used to characterize which knowledge entry texts are most relevant to the current task.

[0028] Correspondingly, after determining the task semantic vector corresponding to each production task text and the knowledge entry vector corresponding to each knowledge entry text, relevance filtering can be performed to obtain a unified task description body.

[0029] In optional embodiments of this application, a unified task description body is obtained by performing relevance filtering based on the semantic vectors of each task and the vectors of each knowledge entry, including: Determine the cosine similarity between the semantic vector of each task and the vector of each knowledge entry, and obtain the cosine similarity corresponding to the semantic vector of each task; Obtain a preset relevance threshold, and determine the semantic vectors of tasks whose cosine similarity is greater than the preset relevance threshold as the target task semantic vectors; The target process parameter matrix, target knowledge entry vector, and target space constraint vector corresponding to the target task semantic vector are determined, and fusion encoding is performed based on the target task semantic vector, target process parameter matrix, target knowledge entry vector, and target space constraint vector to obtain a unified task description body.

[0030] Optionally, we can first determine the cosine similarity between each task semantic vector and each knowledge entry vector. This allows us to obtain the cosine similarity between each task semantic vector and each knowledge entry vector, i.e.: s_i = cos(v_task, v_k^(i)) = (v_task · v_k^(i)) / (||v_task||·||v_k^(i)||), i=1..N Where s_i is the cosine similarity corresponding to the semantic vector of each task, v_task is the semantic vector of the task, and v_k^(i) is the vector of each knowledge item.

[0031] Furthermore, a relevance threshold θ can be preset, and then it can be determined whether the cosine similarity corresponding to each task semantic vector is not less than the relevance threshold θ. The task semantic vectors that are not less than the relevance threshold are determined as the target task semantic vectors. For example, the target task semantic vector can be represented as I = {i | s_i ≥ θ}. The value of the relevance threshold θ can be configured according to the equipment type, process type, and on-site safety requirements, and this application does not limit it.

[0032] Accordingly, the target process parameter matrix, target knowledge entry vector, and target space constraint vector corresponding to the target task semantic vector are determined. Based on these, a unified task description is obtained through fusion encoding. This process can be represented as follows: x_s = g_s(v_task, {v_k^(i)}_{i∈I}) x_p = g_p({P^(i)}_{i∈I}) x_g = g_g({c^(i)}_{i∈I}, M_pc) x = Concat(x_s, x_p, x_g) Where v_task is the target task semantic vector, v_k^(i) is the vector of each target knowledge item, P^(i) is the target process parameter matrix, c^(i) is the target space constraint vector, M_pc is the workspace point cloud / geometric model, and x is the unified task description body.

[0033] Step S12: Obtain the structured planning semantic layer result based on the unified task description body, map the structured planning semantic layer result to the preset action template and inject action constraints to obtain the semantic action mapping path.

[0034] Optionally, in this application, a restricted-size large model with a parameter size not exceeding 30B can be deployed on a local inference device. In this case, a unified task description can be input into the restricted-size large model to obtain a structured planning semantic layer result. Then, the planning semantic layer result is mapped to an action template and constraints are injected to obtain a semantic action mapping path.

[0035] In an optional embodiment of this application, the structured planning semantic layer result includes at least one of a process sequence, a set of key poses, and a set of constraints, wherein the key poses belong to the pose space. The data structure of the structured planning semantic layer result is as follows: Y = F_LLM(x) { Π, Q, C} Π =<a_1, a_2, …, a_m> Q = { q_1, q_2, …, q_r}, q_j ∈ SE(3) C = { c_1, c_2, …, c_u} Where Y is the semantic layer result of structured planning, Π is the process sequence, Q is the set of key poses, and C is the set of constraints.

[0036] Optionally, the key pose belongs to the SE(3) pose space, where SE(3) is the pose (position + attitude) of a rigid body in three-dimensional space. The role of SE(3) is to limit the key pose to a rigid body pose expression in three-dimensional space, thereby providing a unified and computable space for inverse kinematics solution, trajectory interpolation and simulation detection.

[0037] Specifically, when mapping the semantic layer results of the planning to the action template and injecting constraints to obtain the semantic action mapping path, it can be based on the process sequence Π =<a1, a2, …, a_m> Type parsing is performed on each semantic action a_i, and it is compared with the preset action template set. Matching is performed to obtain the corresponding action template T_i. This action template defines parameter interfaces and constraint interfaces, and the parameter interfaces are extensible. Further, based on the key pose set Q, the process parameter matrix, and the unified task description body x, the action template T_i is parameterized to obtain the parameterized action representation A_i = T_i(θ_i), where the parameters in the parameterized action representation include pose parameters and execution parameters. Then, based on the constraint set C, constraints are injected into the parameterized action representation A_i to obtain the constraint-enhanced action representation A_i^*. This constraint injection includes at least one of parameter pruning, constraint fusion, or rule filtering. Finally, each A_i^* is combined according to the process sequence to generate the semantic action mapping path Π_A = ... for subsequent path solving and execution.<A1^, A2^, …, A_m^*> , .

[0038] Step S13: Perform inverse kinematics solution on the semantic action mapping path to obtain the joint space target vector sequence, and perform interpolation calculation on the joint space target vector sequence to obtain the candidate path.

[0039] Optionally, in the process of obtaining candidate paths, the semantic action mapping path can first be processed by inverse kinematics to obtain the joint space target vector sequence J_j, and then the joint space target vector sequence J_j can be processed by interpolation to obtain candidate paths.

[0040] In optional embodiments of this application, inverse kinematics processing is performed on the semantic action mapping path to obtain a joint space target vector sequence, and interpolation processing is performed on the joint space target vector sequence to obtain candidate paths, including: Inverse kinematics is performed on the set of key poses included in the semantic action mapping path to obtain the joint space target vector sequence. Candidate paths are obtained by interpolating the target vector sequence in the joint space using either linear interpolation or sample interpolation.

[0041] Optionally, since the semantic action mapping path is obtained by mapping the structured planning semantic layer result to a preset action template and injecting action constraints, and the structured planning semantic layer result includes a set of key poses, the semantic action mapping path also includes the semantic action mapping path. In this case, the key pose set included in the semantic action mapping path can be processed by inverse kinematics, which can be expressed based on the following formula: J_j = IK(q_j, φ), j=1..r Where φ represents the robot structural parameters configured for performing industrial tasks, J_j represents the joint space target vector sequence, and q_j represents the set of key poses.

[0042] Furthermore, the joint space target vector sequence is interpolated using either linear interpolation or sample interpolation to obtain candidate paths. This interpolation process is based on the following formula: T = { J(t_n)}_{n=1..L}, t_n = n·Δt J(t_n) = Interp(J_j, J_{j+1}, t_n) Where T is the complete sequence of robot joint trajectories, which is a time series representing a set of trajectories. Each element in this set is a "joint state vector". J(t_n) is the joint vector at time t_n. k is the number of joints (e.g., a 6-axis robot). Each θ is a joint angle. J_j and J_{j+1} are the joint solutions corresponding to two adjacent "key poses". J_j is the joint configuration of the j-th key point. t_n is the n-th sampling time point. Δt is the time step (e.g., 2ms / 10ms). n is the sequence number. L is the total number of sampling points, which is determined by the path length and Δt. nterp(·) is the interpolation function used to generate a smooth transition between J_j and J_{j+1}.

[0043] Optionally, when using linear interpolation, intermediate states can be generated by linearly combining adjacent joint vectors according to proportional coefficients. When using spline interpolation, a smooth trajectory is generated based on continuity constraints. The role of `Interp` is to generate a continuous trajectory between adjacent joint configurations. Essentially, it performs interpolation calculations on discrete path points in the target vector of the joint space, thereby obtaining an executable motion sequence that satisfies temporal continuity.

[0044] Step S14: Perform simulation detection processing on the candidate paths to obtain simulation detection results, and generate the final planned path based on the simulation detection results and send it to the robot control module so that the robot control module can execute the industrial business to be executed.

[0045] In optional embodiments of this application, the simulation detection process includes collision detection, joint limit detection, and process accuracy detection, and the simulation detection result is either verification passed or verification abnormal. Specifically, when at least one of the following is abnormal: the collision detection result corresponding to the collision detection, the joint detection result corresponding to the joint limit detection, and the precision detection result corresponding to the process precision detection, the simulation detection result is considered an anomaly.

[0046] Optionally, after obtaining the candidate path, collision detection, joint limit detection, and process accuracy detection can be performed point-by-point on the candidate path in the local simulation environment. This yields the collision detection results, joint limit detection results, and process accuracy detection results. Based on these results, it is determined whether the simulation result is a pass or an error. Specifically, if at least one of these results is abnormal, the simulation result is considered an error. The simulation result can be expressed as follows: δ(t_n) = δ_col(t_n) ∨ δ_lim(t_n) ∨ δ_acc(t_n) Where δ(t_n) is the simulation detection result, δ_col(t_n) is the collision detection result, δ_lim(t_n) is the joint detection result, and δ_acc(t_n) is the accuracy detection result.

[0047] In an optional embodiment of this application, simulation detection processing is performed on the candidate path to obtain simulation detection results, including: When the simulation detection process is collision detection, the minimum distance between the bounding body and the set of obstacles of each trajectory point in the candidate path in the corresponding posture is determined. When the minimum distance is less than the preset safe distance threshold, the collision detection result is judged to be abnormal. When the simulation detection process is a joint limit detection, if the angle of any joint in the candidate path exceeds the preset joint angle range, the joint detection result is determined to be abnormal. When the simulation detection process is used for process accuracy detection, the deviation between the actuator position and the reference position of the end trajectory point in the candidate path is calculated. When the deviation is greater than the process deviation threshold, the accuracy detection result is judged to be abnormal.

[0048] Optionally, when performing collision detection on candidate paths, the minimum distance between the bounding volume and the set of obstacles at the corresponding pose of the robot's trajectory points on the candidate path can be determined. This minimum distance is then compared to a preset safe distance threshold. If the minimum distance is less than the preset safe distance threshold, it indicates that the robot will collide with other objects while working based on the candidate path, and the collision detection result can be determined to be abnormal. This collision detection process is expressed by the following formula: δ_col(t_n) = 1 when Dist(B(J(t_n)), O) < ε; otherwise δ_col(t_n) = 0 Where δ_col(t_n) = 1 indicates that the collision detection result is abnormal, Dist(B(J(t_n)), O) is the minimum distance between the bounding volume and the obstacle set in the corresponding pose, ε is the preset safe distance threshold, and δ_col(t_n) = 0 indicates that the collision detection result is normal.

[0049] Optionally, when performing joint limit detection on candidate paths, the joint angle of any joint in the selected path is determined, and then the joint angle of any joint is compared with a preset joint angle range. If the joint angle of any joint in the candidate path exceeds the preset joint angle range, the joint detection result is determined to be abnormal. This joint limit detection process is expressed by the following formula: δ_lim(t_n) = 1, when k: J_k(t_n) [J_k^min, J_k^max]; otherwise, δ_lim(t_n) = 0 Where δ_lim(t_n) = 1 indicates that the joint detection result is abnormal, and Dist(B(J(t_n)), O) is... k: J_k(t_n) is the joint angle, [J_k^min, J_k^max] is the preset joint angle range, and δ_lim(t_n)=0 indicates that the joint detection result is normal.

[0050] Optionally, when performing process accuracy detection on candidate paths, the deviation between the actuator position and the reference position at the end trajectory point in the candidate path is calculated. Then, this deviation is compared with a process deviation threshold. If the deviation between the actuator position and the reference position is greater than the process deviation threshold, the accuracy detection result is determined to be abnormal. This process accuracy detection process is expressed by the following formula: e(t_n) = || p(t_n) - p_ref(t_n) ||_2 δ_acc(t_n) = 1, when e(t_n) > η; otherwise δ_acc(t_n)=0 Where p(t_n) is the actuator position, p_ref(t_n) is the reference position, e(t_n) is the deviation between the actuator position and the reference position, η is the process deviation threshold, δ_acc(t_n) = 1 indicates that the accuracy detection result is abnormal, and δ_acc(t_n) = 0 indicates that the accuracy detection result is normal.

[0051] The parameters mentioned above, such as the safety distance threshold, process deviation threshold, interpolation step size, and preset joint angle range, can be configured according to the equipment type, process type, and on-site safety requirements, and this application does not impose any limitations on them. For example, the safety distance threshold can be 2mm to 20mm, the process deviation threshold can be 0.1mm to 2mm, and the interpolation step size can be 2ms to 20ms, etc.

[0052] In an optional embodiment of this application, generating a final planned path based on the simulation detection results and sending it to the robot control module includes: When the simulation test result is a verification anomaly, an anomaly flag is generated and a local replanning process is triggered to reconstruct the path segment within the window and obtain the path segment outside the window; the path segment within the window and the path segment outside the window are spliced ​​together to obtain the final planned path and sent to the robot control module; When the simulation test result is that the verification is passed, the candidate path is used as the final planned path and sent to the robot control module.

[0053] Optionally, when the simulation test result is a verification anomaly, an anomaly flag can be generated, which can be represented as follows: E =<type, n*, id> Where type∈{δ_col, δ_lim, δ_acc}, n* is the index of the first outlier, and id is the identifier of the corresponding constraint or path segment.

[0054] Furthermore, after generating an anomaly flag, local replanning can be triggered to reconstruct the path segments within the window and retain the path segments outside the window. Then, the reconstructed path segments within the window and the retained path segments outside the window are spliced ​​together to obtain the final planned path, which is then sent to the robot control module.

[0055] In an optional embodiment of this application, the window range in the local replanning process is determined by the initial outlier index and the window radius w, with the window range being [n*-w, n*+w]∩[1,L], and reconstruction is performed only on trajectory segments within the window range. The window radius w can be configured according to the equipment type, process type, and on-site safety requirements, and is not limited in this application.

[0056] Understandably, when the anomaly type is collision anomaly or joint over-limit anomaly, the anomaly identifier is sent to the planning generation module to trigger local path reconstruction. When the anomaly type is process deviation anomaly, the anomaly identifier is sent to the model inference module to update the action constraint parameters or process parameters before performing local replanning.

[0057] Correspondingly, when the simulation test result is that the verification is passed, it means that the generated candidate path meets the industrial requirements. At this time, the candidate path can be directly used as the final planned path and sent to the robot control module.

[0058] Specifically, when the final planned path is sent to the robot control module, the local encryption module can encrypt and encapsulate the final path before sending it to the robot control module, so as to achieve secure path delivery and execution in an offline environment.

[0059] In this application, candidate paths can be determined by encoding and matching the business data corresponding to the industrial business to be executed and the entries in the local knowledge base. Then, the candidate paths are subjected to linkage constraints on collision, joint limits, and process deviations in a simulation environment, and finally, the path executed by the robot module is obtained. In this process, even in a network-isolated environment, task understanding, knowledge retrieval, planning generation, and simulation verification can be achieved, meeting data compliance and on-site safety requirements. Furthermore, collision, joint limits, and process deviations are uniformly converted into discrete judgment signals and anomaly identifiers are generated, improving auditability, reproducibility, and production process safety and process consistency. At the same time, after verification failure, a local replanning window is constructed centered on the first anomaly point. Only the path fragments within the window are reconstructed and the paths outside the window are spliced ​​and reused, which significantly reduces the computational cost of path replanning and improves the real-time performance, quality, and stability of the path.

[0060] Optionally, the method provided in this application can be deployed and executed in a local industrial intelligent agent system, the structure of which and the data flow can be as follows: Figure 2 As shown, the specific components may include a service request module 101, a model reasoning module 102, a planning generation module 103, an execution verification module 104, a robot control module 105, a local knowledge base archiving module 106, a data encryption module 107, and a three-dimensional space measurement unit 108. Optionally, the steps of the method provided in this application can be as follows: Figure 3 As shown, it specifically includes: S1. Localized loading of industrial knowledge and spatial models; S2. Construct a semantic-geometric fusion task description body; S3, Constrained-size model local inference generates planning semantics; S4, Semantic-Action Mapping and Path Construction; S5, simulation verification and self-correction closed loop; S6, Encrypted Deployment and Production Execution.

[0061] This application provides an apparatus for planning paths for an industrial intelligent agent, such as... Figure 4As shown, the device may include: a planning semantic output unit 301, a path construction unit 302, a simulation detection unit 303, a local replanning unit 304, and an anomaly feedback unit 304, wherein, The planning semantic output unit is used to obtain the business data and local knowledge base entries corresponding to the industrial business to be executed, and to perform vectorized encoding matching processing on the business data and local knowledge base entries to form a unified task description body. The business data includes production task text, workpiece model, design engineering model and process constraint information. The path construction unit is used to obtain the structured planning semantic layer result based on the unified task description body, map the structured planning semantic layer result to the preset action template and inject action constraints to obtain the semantic action mapping path; perform inverse kinematics solution on the semantic action mapping path to obtain the joint space target vector sequence, and perform interpolation calculation on the joint space target vector sequence to obtain the candidate path; The simulation detection unit is used to perform simulation detection processing on candidate paths and obtain simulation detection results. The local replanning unit is used to generate the final planned path based on the simulation test results and send it to the robot control module so that the robot control module can execute the industrial tasks to be performed.

[0062] Optionally, when the semantic output unit performs vectorized encoding matching on business data and local knowledge base entries to form a unified task description, it is specifically used for: The text of each production task and the text of each knowledge entry in the local knowledge base are vectorized and encoded by a local semantic encoder to obtain the semantic vector of each task and the vector of each knowledge entry. A unified task description body is obtained by filtering for relevance based on the semantic vectors of each task and the vectors of each knowledge item.

[0063] Optionally, when the semantic output unit is used to obtain a unified task description body by performing relevance filtering based on the semantic vectors of each task and the vectors of each knowledge item, it is specifically used for: Determine the cosine similarity between the semantic vector of each task and the vector of each knowledge entry, and obtain the cosine similarity corresponding to the semantic vector of each task; Obtain a preset relevance threshold, and determine the semantic vectors of tasks whose cosine similarity is greater than the preset relevance threshold as the target task semantic vectors; The target process parameter matrix, target knowledge entry vector, and target space constraint vector corresponding to the target task semantic vector are determined, and fusion encoding is performed based on the target task semantic vector, target process parameter matrix, target knowledge entry vector, and target space constraint vector to obtain a unified task description body.

[0064] Optionally, the semantic layer result of structured planning includes at least one of the following: process sequence, key pose set, and constraint set, wherein the key pose belongs to the pose space.

[0065] Optionally, when the path construction unit performs inverse kinematics solving on the semantic action mapping path to obtain a joint space target vector sequence, and then performs interpolation calculations on the joint space target vector sequence to obtain candidate paths, it is specifically used for: Inverse kinematics is performed on the set of key poses included in the semantic action mapping path to obtain the joint space target vector sequence. Candidate paths are obtained by interpolating the target vector sequence in the joint space using either linear interpolation or sample interpolation.

[0066] Optionally, the simulation detection process includes collision detection, joint limit detection, and process accuracy detection. The simulation detection result is either verification passed or verification abnormal. Specifically, when at least one of the following is abnormal: the collision detection result corresponding to the collision detection, the joint detection result corresponding to the joint limit detection, and the precision detection result corresponding to the process precision detection, the simulation detection result is considered an anomaly.

[0067] Optionally, when the simulation detection unit performs simulation detection processing on the candidate path and obtains the simulation detection results, it is specifically used for: When the simulation detection process is collision detection, the minimum distance between the bounding body and the set of obstacles of each trajectory point in the candidate path in the corresponding posture is determined. When the minimum distance is less than the preset safe distance threshold, the collision detection result is judged to be abnormal. When the simulation detection process is a joint limit detection, if the angle of any joint in the candidate path exceeds the preset joint angle range, the joint detection result is determined to be abnormal. When the simulation detection process is used for process accuracy detection, the deviation between the actuator position and the reference position of the end trajectory point in the candidate path is calculated. When the deviation is greater than the process deviation threshold, the accuracy detection result is judged to be abnormal.

[0068] Optionally, the anomaly feedback unit generates an anomaly flag when the simulation detection result is a verification anomaly; the local replanning unit, when generating the final planned path based on the simulation detection result and sending it to the robot control module, is specifically used for: The local replanning process is triggered to reconstruct the path segments within the window and obtain the path segments outside the window; the path segments within the window and the path segments outside the window are concatenated to obtain the final planned path and sent to the robot control module; When the simulation test result is that the verification is passed, the candidate path is used as the final planned path and sent to the robot control module.

[0069] Optionally, the window range in local replanning is determined by the initial outlier index and the window radius.

[0070] The device for planning the path of an industrial intelligent agent in this embodiment can execute the method for planning the path of an industrial intelligent agent shown in the embodiment of this application. The implementation principle is similar and will not be described again here.

[0071] This application provides an electronic device, which includes: a processor; and a memory configured to store machine-readable instructions that, when executed by the processor, cause the processor to perform a method for planning a path for an industrial intelligent agent.

[0072] This application provides an electronic device, such as... Figure 5 As shown, Figure 5 The illustrated electronic device includes a processor 2001 and a memory 2003. The processor 2001 and the memory 2003 are connected, for example, via a bus 2002. Optionally, the electronic device 2000 may further include a transceiver 2004. It should be noted that in practical applications, the transceiver 2004 is not limited to one type, and the structure of this electronic device 2000 does not constitute a limitation on the embodiments of this application.

[0073] Processor 2001 may be a CPU, a general-purpose processor, a DSP, an ASIC, an FPGA, or other programmable logic device, transistor logic device, hardware component, or any combination thereof. It may implement or execute the various exemplary logic blocks, modules, and circuits described in conjunction with the disclosure of this application. Processor 2001 may also be a combination that implements computational functions, such as including one or more microprocessor combinations, a combination of a DSP and a microprocessor, etc.

[0074] Bus 2002 may include a pathway for transmitting information between the aforementioned components. Bus 2002 may be a PCI bus or an EISA bus, etc. Bus 2002 can be divided into address bus, data bus, control bus, etc. For ease of representation, Figure 5 The bus is represented by a single thick line, but this does not mean that there is only one bus or one type of bus.

[0075] The memory 2003 may be ROM or other type of static storage device capable of storing static information and instructions, RAM or other type of dynamic storage device capable of storing information and instructions, or EEPROM, CD-ROM or other optical disc storage, optical disc storage (including compressed optical discs, laser discs, optical discs, digital universal optical discs, Blu-ray discs, etc.), magnetic disk storage media or other magnetic storage devices, or any other medium capable of carrying or storing desired program code in the form of instructions or data structures and accessible by a computer, but not limited thereto.

[0076] The memory 2003 stores the application code that executes the scheme of this application, and its execution is controlled by the processor 2001. The processor 2001 executes the application code stored in the memory 2003 to implement... Figure 4 The illustrated embodiment provides the operation of an industrial intelligent agent path planning device.

[0077] It should be understood that although the steps in the flowcharts of the accompanying figures are shown sequentially as indicated by the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the accompanying figures may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily completed at the same time, but can be executed at different times, and their execution order is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the sub-steps or stages of other steps.

[0078] The above description is only a partial embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for path planning by an industrial intelligent agent, characterized in that, The method is deployed in a local industrial intelligent agent system, and the method includes: Obtain the business data and local knowledge base entries corresponding to the industrial business to be executed, and perform vectorized encoding matching processing on the business data and the local knowledge base entries to form a unified task description body. The business data includes production task text, workpiece model, design engineering model and process constraint information. Based on the unified task description body, the structured planning semantic layer result is obtained, and the structured planning semantic layer result is mapped to the preset action template and action constraints are injected to obtain the semantic action mapping path; The semantic action mapping path is subjected to inverse kinematics solution to obtain a joint space target vector sequence, and the joint space target vector sequence is subjected to interpolation calculation to obtain candidate paths; The candidate paths are subjected to simulation detection processing to obtain simulation detection results, and the final planned path is generated based on the simulation detection results and sent to the robot control module so that the robot control module can execute the industrial business to be executed; The step of performing vectorized encoding matching processing on the business data and the local knowledge base entries to form a unified task description body includes: The text of each production task and the text of each knowledge entry in the local knowledge base are vectorized and encoded by a local semantic encoder to obtain the semantic vector of each task and the vector of each knowledge entry. Based on the semantic vectors of each task and the vectors of each knowledge item, a relevance screening is performed to obtain a unified task description body; The process of performing relevance filtering based on the semantic vectors of each task and the vectors of each knowledge item to obtain a unified task description body includes: Determine the cosine similarity between each task semantic vector and each knowledge entry vector to obtain the cosine similarity corresponding to each task semantic vector; A preset relevance threshold is obtained, and the task semantic vectors with a cosine similarity greater than the preset relevance threshold are determined as the target task semantic vectors. The target process parameter matrix, target knowledge entry vector, and target space constraint vector corresponding to the target task semantic vector are determined, and the unified task description body is obtained by fusing and encoding based on the target task semantic vector, the target process parameter matrix, the target knowledge entry vector, and the target space constraint vector.

2. The method according to claim 1, characterized in that, The structured planning semantic layer result includes at least one of a process sequence, a set of key poses, and a set of constraints, wherein the key poses belong to the pose space.

3. The method according to claim 1, characterized in that, The process of performing inverse kinematics solving on the semantic action mapping path to obtain a joint space target vector sequence, and then performing interpolation calculations on the joint space target vector sequence to obtain candidate paths, includes: Inverse kinematics processing is performed on the set of key poses included in the semantic action mapping path to obtain the joint space target vector sequence; The candidate path is obtained by interpolating the joint space target vector sequence using either linear interpolation or sample interpolation.

4. The method according to claim 1, characterized in that, The simulation detection process includes collision detection, joint limit detection, and process accuracy detection. The simulation detection result is either verification passed or verification abnormal. Specifically, when at least one of the collision detection result corresponding to the collision detection, the joint detection result corresponding to the joint limit detection, and the precision detection result corresponding to the process precision detection is abnormal, the simulation detection result is a verification anomaly.

5. The method according to claim 4, characterized in that, The simulation detection process for the candidate path, to obtain the simulation detection results, includes: When the simulation detection process is collision detection, the minimum distance between the bounding body and the set of obstacles of each trajectory point in the candidate path in the corresponding posture is determined. When the minimum distance is less than the preset safe distance threshold, the collision detection result is determined to be abnormal. When the simulation detection process is a joint limit detection, if any joint angle in the candidate path exceeds the preset joint angle range, the joint detection result is determined to be abnormal. When the simulation detection process is a process accuracy detection, the deviation between the actuator position and the reference position of the end trajectory point in the candidate path is calculated. When the deviation is greater than the process deviation threshold, the accuracy detection result is determined to be abnormal.

6. The method according to claim 4, characterized in that, The step of generating the final planned path based on the simulation detection results and sending it to the robot control module includes: When the simulation detection result is a verification anomaly, an anomaly identifier is generated and a local replanning process is triggered to reconstruct the path segment within the window and obtain the path segment outside the window; the path segment within the window and the path segment outside the window are spliced ​​together to obtain the final planned path and sent to the robot control module; When the simulation test result is verified, the candidate path is used as the final planned path and sent to the robot control module.

7. The method according to claim 6, characterized in that, The window range in the local replanning process is determined by the initial outlier index and the window radius.

8. A device for planning paths for an industrial intelligent agent, characterized in that, The device is deployed in a local industrial intelligent agent system, and the device includes: The planning semantic output unit is used to obtain the business data and local knowledge base entries corresponding to the industrial business to be executed, and to perform vectorized encoding matching processing on the business data and the local knowledge base entries to form a unified task description body. The business data includes production task text, workpiece model, design engineering model and process constraint information. The path construction unit is used to obtain the structured planning semantic layer result based on the unified task description body, map the structured planning semantic layer result to a preset action template and inject action constraints to obtain a semantic action mapping path; perform inverse kinematics solving on the semantic action mapping path to obtain a joint space target vector sequence, and perform interpolation calculation on the joint space target vector sequence to obtain a candidate path; The simulation detection unit is used to perform simulation detection processing on the candidate path to obtain simulation detection results; A local replanning unit is used to generate a final planned path based on the simulation detection results and send it to the robot control module so that the robot control module can execute the industrial business to be executed. The step of performing vectorized encoding matching processing on the business data and the local knowledge base entries to form a unified task description body includes: The text of each production task and the text of each knowledge entry in the local knowledge base are vectorized and encoded by a local semantic encoder to obtain the semantic vector of each task and the vector of each knowledge entry. Based on the semantic vectors of each task and the vectors of each knowledge item, a relevance screening is performed to obtain a unified task description body; The process of performing relevance filtering based on the semantic vectors of each task and the vectors of each knowledge item to obtain a unified task description body includes: Determine the cosine similarity between each task semantic vector and each knowledge entry vector to obtain the cosine similarity corresponding to each task semantic vector; A preset relevance threshold is obtained, and the task semantic vectors with a cosine similarity greater than the preset relevance threshold are determined as the target task semantic vectors. The target process parameter matrix, target knowledge entry vector, and target space constraint vector corresponding to the target task semantic vector are determined, and the unified task description body is obtained by fusing and encoding based on the target task semantic vector, the target process parameter matrix, the target knowledge entry vector, and the target space constraint vector.

Images