Double-arm robot cable wiring method and equipment based on digital twinning of electrical cabinet

Abstract

The invention discloses a double-arm robot cable wiring method and equipment based on digital twinning of an electrical cabinet, and relates to the technical field of industrial robot automation. The method comprises the following steps: constructing an electrical cabinet digital twinborn model; on the basis of the undirected connection diagram, global topology arrangement is carried out on the cable bundle by using the digital twin model of the electrical cabinet, and an offline planning path point sequence is generated; forming a wiring instruction; performing wiring and plugging operation on each cable through a double-arm robot; and determining whether the wiring and plugging operation of the current cable is completed or not, and dynamically adjusting the task sequence or the motion trail under the condition that the wiring and plugging operation is not completed, and returning to the step: executing the wiring and plugging operation for each cable through the double-arm robot. According to the invention, multi-constraint unified modeling and automatic layout planning are realized, and the wiring success rate is improved.

Description

Technical Field

[0001] This invention relates to the field of industrial robot automation technology, specifically to a cable routing method and equipment for a dual-arm robot based on an electrical cabinet digital twin. Background Technology

[0002] Electrical control cabinets, as core control units in fields such as power, automation, rail transit, and new energy, contain a large number of electrical components, including relays, circuit breakers, contactors, PLCs (programmable logic controllers), and terminal blocks. The cables connecting these components are numerous, diverse in specifications, and complex in path, creating a high-density wiring environment. Traditional electrical cabinet assembly heavily relies on skilled workers, facing numerous challenges such as low efficiency, poor quality consistency, high labor intensity, and high safety risks.

[0003] With the development of Industry 4.0 and intelligent manufacturing, the use of industrial robots for automated cabling has become an important trend. Dual-arm collaborative robots, due to their human-like operational characteristics, demonstrate unique advantages in electrical cabinet cabling tasks. However, current automated cabling systems still face several key technological bottlenecks in practical applications, severely restricting their large-scale commercial application.

[0004] Digital twin modeling lacks deep integration of electrical semantics and process rules: Existing digital twin technologies mostly focus on three-dimensional geometric visualization or only on mechanical stress analysis, without binding electrical attributes (voltage level, signal type) and wiring process specifications (safety spacing, strong and weak current isolation, minimum bending radius) with the three-dimensional model in a structured and computable manner. This results in the planning system "not understanding" electrical specifications, and the output scheme often violates industry standards, still requiring manual verification and adjustment.

[0005] Cable path planning is done on a single-line basis, lacking global collaborative optimization across multiple cables: Existing cabling planning methods mostly focus on finding the shortest path for individual cables, failing to consider the overall topology of the "cable bundle" for collaborative planning. This results in suboptimal total cabling length, cross-interference between high-voltage and low-voltage cables, low space utilization, and difficulties in later bundling and maintenance. Although ant colony optimization can be used for path optimization, its optimization objective is only to minimize stress, without integrating multi-dimensional constraints such as electrical isolation and process layout.

[0006] Low success rate of robot motion planning in high-density, confined spaces: The internal space of electrical cabinets is extremely limited, and the risk of collisions between robotic arms, cables, and cabinet components is extremely high. Traditional path planning algorithms based on sampling (such as RRT (Rapid Expanding Random Tree)) or discrete search are computationally inefficient in high-dimensional constrained spaces, easily getting trapped in local optima or failing to plan, and generating unsmooth trajectories that do not conform to the robot's dynamic characteristics.

[0007] The reliability of terminal plugging operations places stringent demands on sensing and control: existing plugging solutions mostly rely on high-precision vision servo systems, but in environments with uneven lighting, metal reflections, and severe obstructions, the stability of visual positioning is poor. Position control alone is prone to causing pin bending and connector damage due to minor deviations.

[0008] The rigid dual-arm collaboration mechanism leads to a significant bottleneck in the overall system efficiency: existing collaboration strategies are mostly fixed master-slave modes or static task partitioning, lacking a dynamic arbitration mechanism based on real-time working conditions (task progress, load, collision risk), resulting in the two arms waiting for each other and low resource utilization.

[0009] Therefore, there is an urgent need for a new generation of automated cabling solutions for dual-arm robots that can deeply integrate electrical specifications, process knowledge and physical constraints, and possess multi-cable harness collaborative planning, high-success-rate motion execution and intelligent dynamic collaboration capabilities. Summary of the Invention

[0010] This invention addresses the problems of high planning failure rate, low operation efficiency, unstable wiring success rate, and poor system flexibility in current automated cabling systems due to extreme space constraints, numerous constraints, complex coordination, and high precision requirements. It provides a dual-arm robot cable cabling method and equipment based on electrical cabinet digital twins, which realizes unified modeling of multiple constraints and automated layout planning, thereby improving the cabling success rate.

[0011] The present invention is achieved through the following technical solution.

[0012] In a first aspect, a cable routing method for a dual-arm robot based on an electrical cabinet digital twin is provided, the method comprising:

[0013] Based on the geometric model of the electrical cabinet, a digital twin model of the electrical cabinet is constructed;

[0014] Electrical connection relationships are extracted from the wiring file to generate an undirected connection graph. Based on the undirected connection graph, the digital twin model of the electrical cabinet is used to perform global topology layout for the cable bundle and generate an offline planned path point sequence.

[0015] Based on the offline planned path point sequence, the operation of each cable in the cable bundle is decomposed into tasks, the dual-arm collaborative timing is planned, and the workspace is pre-divided to form wiring instructions.

[0016] Using a dual-arm robot, wiring instructions and real-time sensing data are employed to perform wiring and splicing operations for each cable.

[0017] Determine whether the current cable has completed the wiring and plugging operation, and if the wiring and plugging operation has not been completed, dynamically adjust the task sequence or movement trajectory, and return to step: using a dual-arm robot, the wiring instructions and real-time sensing data are used to perform wiring and plugging operations for each cable.

[0018] In some embodiments, a digital twin model of the electrical cabinet is constructed based on the geometric model of the electrical cabinet, including:

[0019] An electrical cabinet geometric model is constructed based on the electrical cabinet's geometric structure, wherein the electrical cabinet's geometric structure includes: structural components and electrical elements;

[0020] Static attributes are added to each electrical component in the electrical cabinet geometric model, wherein the static attributes include semantic tags and associated electrical attributes, and the cabling process specification library is bound to the electrical cabinet geometric model in the form of computable logical rules;

[0021] Define the dynamic attributes, accessibility, and traversability of the electrical cabinet's geometric model to form a digital twin model of the electrical cabinet.

[0022] In some embodiments, based on the undirected connectivity graph, the global topology arrangement of the cable bundles is performed using the digital twin model of the electrical cabinet, including:

[0023] The mathematical constraints derived from the computable logical rules are integrated into the digital twin model of the electrical cabinet, and a comprehensive cost function for all cable path sets is established for the mathematically constrained digital twin model of the electrical cabinet.

[0024] ;

[0025] in, Indicates the total length of all cable paths; Indicates the total number of bends or the total bend radius penalty; This represents the penalty for undesirable proximity between high-voltage and low-voltage cables. Penalty indicating the required number of clamps; Indicates penalties for violating hard constraints; , , , , Indicates weight;

[0026] Using the hierarchical ant colony optimization algorithm and the digital twin model of the electrical cabinet, the comprehensive cost function corresponding to different cable bundle planning schemes is solved based on the undirected connection graph.

[0027] The cable bundle planning scheme that minimizes the comprehensive cost function is used as the final global topology layout.

[0028] In some embodiments, using a hierarchical ant colony optimization algorithm and the digital twin model of the electrical cabinet, based on the undirected connectivity graph, the comprehensive cost function corresponding to different cable bundle planning schemes is solved, including:

[0029] The channels available for wiring are identified from the digital twin model of the electrical cabinet, and the channels are spatially discretized to form a coarse topology map, and a fine-grained local fine mesh is established within each channel.

[0030] Using ant colony optimization, cable bundle-level routing planning is performed on the coarse topology map to obtain coarse bundle backbone paths;

[0031] In the local fine grid of the coarse bundle backbone path, the position of each cable in its respective cable bundle is planned for each cable;

[0032] Curve fitting is performed on the spatially discretized path points of each cable under the minimum bending radius constraint to obtain a smooth continuous path, and the location of the cable's fixed points is determined based on the maximum spacing constraint of the fixed points and the available support structure.

[0033] In some embodiments, ant colony optimization is used to perform cable bundle-level routing planning on the coarse topology map to obtain coarse bundle backbone paths, including:

[0034] Design and initialize pheromones;

[0035] Let the probability be: An ant starts from the set of starting points of the cable bundle and chooses the next node from the current node.

[0036] ;

[0037] in, For path superior The pheromone concentration at any given time For heuristic information, As a forward-looking compatibility factor, , , Indicates the importance coefficient;

[0038] Check if the selected node violates hard constraints;

[0039] Select the chosen node without violating the hard constraints;

[0040] Update the pheromones;

[0041] After an ant has traversed a path, it releases pheromones locally.

[0042] After all ants have completed one iteration, select the best solution of this iteration and the best solution in the global history to update the path.

[0043] In some embodiments, a dual-arm robot performs wiring and splicing operations for each cable using the wiring instructions and real-time sensing data, including:

[0044] Using the global trajectory guide and the wiring instructions, a path reference trajectory for each cable is generated.

[0045] A local real-time motion planner based on deep reinforcement learning generates locally optimized action instructions for wiring and splicing based on the path reference trajectory, robot body state, target and progress state, real-time perception data, cable state, and historical information.

[0046] In some embodiments, a local real-time motion planner based on deep reinforcement learning generates locally optimized action instructions for wiring and splicing based on the path reference trajectory, robot body state, target and progress state, real-time perception data, cable state, and historical information, including:

[0047] The insertion process of a single terminal is divided into a finite state machine, and the states of the finite state machine include:

[0048] Coarse approach: Based on offline wiring planning and visual coarse positioning, move to a safe distance above the target point;

[0049] Visual alignment: Identify the target terminal hole and calculate the offset;

[0050] Force-controlled compliant insertion: A variable impedance control model is used to control the insertion force. The variable impedance control model is as follows:

[0051] ;

[0052] ;

[0053] in, Indicates virtual quality; Indicates virtual damping; Represents the adaptive stiffness matrix; This represents the actual position vector of the robot's end effector. Represents the desired position vector of the robot's end effector; This represents the actual velocity vector of the robot's end effector. This represents the desired velocity vector at the robot's end effector. This represents the actual acceleration vector at the robot's end effector. This represents the desired acceleration vector at the robot's end effector. This represents the external contact force sensed by the robot's end effector; Indicates the stiffness during the search phase; Indicates the stiffness during the insertion phase; Indicates the insertion depth; Indicates the alignment status; Indicates depth Weighting coefficients; Indicates alignment status Weighting coefficients;

[0054] Position determination and holding: monitoring axial force and acoustic emission signals;

[0055] Abnormal handling: When the axial force exceeds the safety threshold or continuous jamming occurs, the yielding strategy is triggered.

[0056] In some embodiments, the method further includes:

[0057] Real-time monitoring of impedance parameters during wiring and connection processes;

[0058] The adaptive stiffness matrix is ​​updated based on the impedance parameters monitored during the wiring and connection process.

[0059] In some embodiments, the method further includes:

[0060] The following information is monitored in real time: current task status of each robotic arm, estimated remaining time, joint torque / power load, end effector status, distance between the end effects of the two arms, and short-term trajectory predicted by the local planner;

[0061] Construct a conflict prediction and performance evaluation model, and based on the monitoring information, use the conflict prediction and performance evaluation model to predict the probability of a collision occurring within a predetermined time in the future, the total task completion time, and the load balancing degree of the two arms, continuing the current task allocation scheme.

[0062] If a key indicator in the monitoring information exceeds a threshold or reaches a preset decision point, the arbitrator is triggered to arbitrate.

[0063] The arbitrator selects or combines response strategies from a predefined strategy library to make a decision, wherein the response strategies include: task swapping, timing fine-tuning, spatial resegmentation, and inserting avoidance actions.

[0064] Based on the decision made by the arbitrator, new task instructions or constraints are sent to the local real-time motion planners of each arm of the dual-arm robot, and the local real-time motion planners replan the subsequent trajectory based on the new instructions or constraints.

[0065] Secondly, a dual-arm robot cable routing device based on an electrical cabinet digital twin is provided, the device comprising:

[0066] The digital twin model building module is used to: build a digital twin model of the electrical cabinet based on the geometric model of the electrical cabinet;

[0067] The global topology layout generation module is used to: extract electrical connection relationships from the cabling file, generate an undirected connection graph, and based on the undirected connection graph, use the digital twin model of the electrical cabinet to perform global topology layout for the cable bundle and generate an offline planned path point sequence.

[0068] The cabling instruction generation module is used to: based on the offline planned path point sequence, perform task decomposition, dual-arm collaborative timing planning, and workspace pre-segmentation for the operation of each cable in the cable bundle to form cabling instructions;

[0069] The wiring and splicing operation module is used to perform wiring and splicing operations for each cable by means of a dual-arm robot, using the wiring instructions and real-time sensing data;

[0070] The action adjustment module is used to: determine whether the current cable has completed the wiring and plugging operation, and if the wiring and plugging operation has not been completed, dynamically adjust the task sequence or movement trajectory, and return to the wiring and plugging operation module for execution.

[0071] Compared with the prior art, the present invention has the following advantages and beneficial effects.

[0072] 1. Significantly improves the first-time connection success rate: The learning impedance controller can automatically adapt to different terminal types and cable characteristics, eliminating the need for manual parameter tuning and improving the first-time connection success rate.

[0073] 2. Possesses continuous evolution capability: As operational data accumulates, the controller is continuously optimized, and the plug-in efficiency (average time) can be gradually reduced.

[0074] 3. Strong environmental adaptability: Through online learning and meta-learning, the system can quickly adapt to unforeseen disturbances (such as terminal manufacturing tolerances and robot wear).

[0075] 4. Reduced reliance on precise models: No need for precise terminal-socket friction models; contact dynamics are learned autonomously through data-driven processes.

[0076] 5. Form a process knowledge base: The impedance strategies learned can be archived as "process knowledge packages" for rapid deployment and inheritance of similar terminals. Attached Figure Description

[0077] To more clearly illustrate the technical solutions of the exemplary embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly described below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation of the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0078] Figure 1 This is a flowchart illustrating a cable routing method for a dual-arm robot based on an electrical cabinet digital twin, according to an embodiment of the present invention.

[0079] Figure 2 This is a flowchart illustrating the overall cabling process of a dual-arm robot based on an electrical cabinet digital twin according to an embodiment of the present invention.

[0080] Figure 3 This is a schematic diagram of a multi-constraint cable bundle topology planning according to an embodiment of the present invention.

[0081] Figure 4 This is a diagram illustrating the working mechanism of the online execution layer according to an embodiment of the present invention.

[0082] Figure 5 This is a structural block diagram of a dual-arm robot cable routing device based on an electrical cabinet digital twin according to an embodiment of the present invention. Detailed Implementation

[0083] To make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the embodiments and accompanying drawings. The illustrative embodiments and descriptions of the present invention are only used to explain the present invention and are not intended to limit the present invention.

[0084] This invention adopts the core design concept of "digital twin-driven, virtual-physical closed-loop linkage," consisting of three interconnected and data-sharing layers: a digital twin and offline planning layer, an online sensing and execution control layer, and a data management and learning optimization layer. These three layers communicate through a high-speed industrial network and a unified data bus, forming a complete closed loop from virtual space planning to physical space execution, and then feedback data to optimize the virtual model.

[0085] On the one hand, the present invention provides a cable routing method for a dual-arm robot based on an electrical cabinet digital twin. Figure 1 This is a flowchart illustrating a cable routing method for a dual-arm robot based on an electrical cabinet digital twin, according to an embodiment of the present invention. (Reference) Figure 1 The cable routing method for a dual-arm robot based on the digital twin of the electrical cabinet includes: S10 to S50.

[0086] Figure 2This is a flowchart illustrating the overall cabling process of a dual-arm robot based on an electrical cabinet digital twin according to an embodiment of the present invention. Figure 3 This is a schematic diagram of a multi-constraint cable bundle topology planning according to an embodiment of the present invention. Figure 4 This is a diagram illustrating the working mechanism of the online execution layer according to an embodiment of the present invention. The following references... Figures 1 to 4 The present invention provides a detailed description of the cable routing method for a dual-arm robot based on electrical cabinet digital twin.

[0087] In S10, a digital twin model of the electrical cabinet is constructed based on the geometric model of the electrical cabinet.

[0088] In S20, electrical connection relationships are extracted from the cabling file to generate an undirected connection graph. Based on the undirected connection graph, a global topology layout is performed on the cable bundle using the digital twin model of the electrical cabinet to generate an offline planned path point sequence.

[0089] In S30, based on the offline planning path point sequence, the operation of each cable in the cable bundle is decomposed into tasks, the dual-arm collaborative timing planning is performed, and the workspace is pre-divided to form wiring instructions.

[0090] In the S40, a dual-arm robot performs wiring and splicing operations for each cable using wiring instructions and real-time sensing data.

[0091] In S50, determine whether the wiring and plugging operations for each cable have been completed. If the wiring and plugging operations have not been completed, dynamically adjust the task sequence or movement trajectory and return to step S40.

[0092] S10 to S30 can be implemented through digital twins and offline planning layers, while S40 to S50 can be implemented through online sensing and execution control layers. The following provides a detailed explanation of the digital twin and offline planning layers, and the online sensing and execution control layers.

[0093] 1.1 Digital Twin and Offline Planning Layer

[0094] The digital twin and offline planning layer are the "brain" and "prior knowledge base" of the system. They mainly run before the start of the task and their output provides global guidance and initial strategies for online execution.

[0095] 1.1.1 Construction of Digital Twin Model

[0096] In S10, a digital twin model of the electrical cabinet is constructed.

[0097] In some embodiments, constructing a digital twin model of an electrical cabinet based on its geometric model includes: constructing a geometric model of the electrical cabinet based on its geometric structure, wherein the geometric structure of the electrical cabinet includes structural components and electrical elements; adding static attributes to each electrical element, wherein the static attributes include semantic tags and associated electrical attributes, and binding a cabling process specification library to the model electrical cabinet geometric model in the form of computable logical rules; and defining the dynamic attributes, accessibility, and navigability of the electrical cabinet's geometric structure to form the digital twin model of the electrical cabinet.

[0098] Specifically, the geometric model import and processing supports importing precise CAD (Computer-Aided Design) models (STEP, IGES format) of electrical cabinets or models reconstructed from 3D laser scanning point clouds. This model includes the precise 3D geometry, installation position, and orientation of all mounting plates, mounting rails, door panels, ventilation holes, and each electrical component (circuit breaker, contactor, PLC (Programmable Logic Controller), terminal block, etc.). Furthermore, semantic tags (such as component type, model number, and tag number) are automatically added to each component.

[0099] Electrical Attribute and Rule Attachment: Establish a scalable database of electrical component attributes. Associate electrical attributes with each component or port, such as operating voltage level (e.g., 24VDC, 220VAC, 10kV), current level, and signal type (digital input DI, digital output DO, analog input AI / AO, communication bus). Simultaneously, bind a cabling specification library to the model; specifications exist in the form of computable logical rules, for example:

[0100] Safety distance rules: ;

[0101] Cable grouping rules:

[0102] ;

[0103] Fixed spacing rules:

[0104] ;

[0105] in, : Indicates the type identifier of two different cables or electrical components.

[0106] The minimum safe straight-line distance, usually measured in millimeters (mm). This is a hard constraint in cabling planning.

[0107] This indicates the operating voltage ratings of object A and object B (e.g., 24VDC, 220VAC, 10kV). The greater the voltage difference, the greater the required insulation distance is generally required.

[0108] : Indicates environmental factors, which may include cabinet temperature, humidity, or electromagnetic interference levels. These factors can affect the safety factor of the insulation distance.

[0109] : Indicates the voltage level of the cable, used to distinguish between high voltage and low voltage.

[0110] This indicates the type of signal transmitted by the cable. Cables with the same signal type or similar interference resistance will be grouped together to form a "cable bundle" for path planning.

[0111] : Indicates the cable type, which determines the applicable physical constraint model for that cable.

[0112] : Indicates the maximum allowable spacing between two fixed points (clamps / cable ties) as calculated.

[0113] : Indicates the cable diameter; thicker cables typically allow for greater support spacing.

[0114] : Indicates the stiffness coefficient (or flexibility) of the cable. The higher the stiffness, the stronger the self-supporting ability and the larger the allowable fixing spacing; flexible cables require denser fixing points to prevent sagging.

[0115] Dynamic Attributes and Accessibility Analysis: In addition to the static model, dynamic attributes are defined, such as the opening and closing state of the cabinet door (affecting the reachable space of the robotic arm) and the sliding rail position of the movable module. Based on the geometric model and the robot model, accessibility and mobility analysis are performed to pre-calculate the set of collision-free postures of the robotic arm end effector to each connection point, as well as the "channel" areas where cables may pass through, providing heuristic information for subsequent planning.

[0116] 1.1.2 Wiring Logic Analysis and Connection Diagram Generation

[0117] In S20, electrical connection relationships are extracted from the wiring file to generate an undirected connection diagram.

[0118] Specifically, the input interface supports direct import of wiring tables (CSV and XML formats) or schematic files generated by electrical design software (such as EPLAN and AutoCAD Electrical), as well as manual input or graphical configuration.

[0119] Logical parsing: Parses the input file, automatically extracts all electrical connections, and generates an undirected connection graph.

[0120] vertex : Represents each connection point, including attributes: the component it belongs to, port number, spatial coordinates (mapped from the twin model), electrical attributes, and the cable specifications (wire diameter, color, terminal type).

[0121] side : Represents a physical cable to be laid, including attributes: cable ID, start point V_s, end point V_t, cable type (single-core, multi-core, shielded), cross-sectional area, color, and minimum bending radius R_min.

[0122] Topology simplification and preprocessing: Preprocessing complex connection graphs, such as identifying common paths (multiple cables with similar start and end points can be bundled together), provides input for subsequent bundle-level planning.

[0123] 1.1.3 Multi-constraint cable harness topology collaborative planning

[0124] In S20, global topology layout is performed for cable bundles.

[0125] Specifically, firstly, problem modeling: the problem of "finding a path for each cable" is escalated to an optimization problem of "finding the global topology for the cable bundle". The decision variable is defined as the set of all cable paths. The objective is to minimize the overall cost function. The overall cost function is a weighted sum of multiple sub-costs:

[0126] ;

[0127] in: Total length of all cable paths;

[0128] Total number of bends or total bend radius penalty (encourages straightness);

[0129] Undesirable proximity penalty between high-voltage and low-voltage cables;

[0130] Penalty for the required number of fixing clamps (encourages the use of existing support structures);

[0131] Penalties for violating various hard constraints (safety clearance, minimum bending radius, etc.) have extremely high weights to ensure that the constraints are satisfied;

[0132] , , , , Indicates the weight.

[0133] Constraint Integration: The various rules defined in 1.1.1 are transformed into mathematical constraints and integrated into the optimization model. For example, the electrical clearance constraint is expressed as: for any two points... They belong to high-voltage cables. and low voltage cables .

[0134] Require This is a complex spatial constraint that needs to be judged based on path discretization.

[0135] Then, using the hierarchical ant colony optimization algorithm and the digital twin model of the electrical cabinet, based on the undirected connection graph, the comprehensive cost function corresponding to different cable bundle planning schemes can be solved, and the cable bundle planning scheme with the minimum comprehensive cost function can be used as the final global topology layout.

[0136] For example, an improved hierarchical ant colony optimization (ACO) algorithm can be used to solve the comprehensive cost function.

[0137] Path discretization: The three-dimensional space of the electrical cabinet is discretized into a fine-grained grid or graph structure, where nodes represent feasible spatial points and edges represent passable directions.

[0138] Pheromone Design and Initialization: Pheromone exists not only on the grid edges but also on the "wiring patterns" (such as "wiring along the top plate" and "wiring along the right-side column"). During initialization, based on the reachability analysis results, good wiring patterns are assigned higher initial pheromone levels.

[0139] Ant-based solution construction: Each "ant" represents a cable bundle planning scheme. Ants sequentially select paths for each cable. The selection probability is determined by heuristic information (such as the reciprocal of the straight-line distance to the destination) and the overall pheromone intensity. When selecting paths, ants perform "forward-looking" conflict detection to avoid serious conflicts between new paths and already planned paths, and calculate their contribution to the global cost function.

[0140] Local pheromone update: After each ant completes the construction of a solution, it immediately updates the pheromone content based on that solution. The quality is improved by locally updating pheromones (evaporation and enhancement), guiding subsequent ants to search for better areas.

[0141] Global pheromone update: After all ants have completed one round of iteration, they select the best solution in this round and the best solution in the global history, and perform a global pheromone update to strengthen excellent paths and patterns.

[0142] Bundle merging and splitting: Integrate heuristic rules into the algorithm to dynamically determine which cables can be merged into the same cable bundle for planning (shared path), and when it is necessary to split the bundle to bypass obstacles.

[0143] Output: The algorithm finally outputs the globally optimal cable bundle topology, the detailed spatial path point sequence of each cable, and the recommended fixed point locations.

[0144] 1.1.4 Dual-arm task sequence and spatial allocation pre-planning

[0145] For S30, routing instructions are generated.

[0146] Specifically, task decomposition: Based on the topology planning results, the operation for each cable is decomposed into a sequence of atomic tasks:

[0147]

[0148]

[0149] Dual-arm collaborative timing planning: Based on task priority relationships (e.g., deployment must precede fixation, one end must be fixed before the other can be plugged in) and resource constraints (each task requires a specific tool), a priority-based scheduling algorithm or heuristic search is used to generate a preliminary dual-arm collaborative Gantt chart. This chart defines the start time, duration, and assigned robotic arm (left arm L or right arm R) for each task.

[0150] Workspace pre-segmentation: To avoid frequent collisions between the two arms in a shared, confined space, a dynamic workspace segmentation strategy is predefined during the offline phase based on the cabinet structure and task distribution. For example, the cabinet can be divided into two main work areas along the vertical centerline or based on component layout, and these areas can be preferentially assigned to the left and right arms respectively. When a task needs to cross boundaries, an "interlocking" protocol is introduced to ensure that only one arm is working in the critical area at any given time.

[0151] Output: The final output of the offline planning layer is a structured "routing process package," which includes: cable path data, fixed point data, dual-arm task sequence, initial spatial partitioning strategy, and estimated operation time. This process package will be loaded into the online execution layer.

[0152] 1.2 Online Sensing and Execution Control Layer

[0153] The online perception and execution control layer is the "cerebellum" and "limbs" of the system, responsible for executing the results of offline planning in real time in the physical world and handling various uncertainties. It adopts a strategy of "global guidance, local optimization, multi-modal perception, and dynamic collaboration".

[0154] 1.2.1 Multimodal Real-time Sensing Fusion

[0155] Specifically, global environment perception: 3D vision sensors (such as binocular stereo cameras or structured light cameras) are installed on the robot's base or top to acquire point cloud data of the work area in real time and compare it with the digital twin model to detect whether there are unmodeled obstacles (such as temporarily placed tools) or slight deviations between the component positions and the model, so as to achieve "model-reality" registration and update.

[0156] Fine-grained local perception: The end effector uses hand vision, with a high-resolution macro camera mounted on the robot's wrist for precise identification and positioning of terminal blocks before wiring. Deep learning object detection algorithms (such as YOLO) are used to identify terminal hole positions, combined with sub-pixel edge extraction, to achieve a positioning accuracy better than 0.05mm.

[0157] Six-dimensional force / torque sensor: Installed between the robot's end effector and the gripper, it measures the contact force / torque F_ext in real time during the insertion process. Used for force control, anomaly detection, and human-machine interaction safety.

[0158] Proximity sensor: An optoelectronic or ultrasonic proximity sensor is integrated at the front end of the gripper to detect the distance to the terminal surface before contact occurs, thereby enabling decelerated approach.

[0159] Acoustic emission sensor: Integrated into the clamp, it monitors the high-frequency sound signals generated by the friction and collision between the terminal and the socket during insertion, and is used to determine the insertion status (such as whether it is in place or whether it is stuck).

[0160] 1.2.2 Hierarchical Hybrid Motion Planning and Controller

[0161] Global trajectory guide: Receives the path point sequence output from offline planning. Each point contains both position and orientation. These points are then interpolated using spline interpolation to generate a smooth reference trajectory. This trajectory serves as high-level guidance.

[0162] Local Real-Time Motion Optimizer: A planner based on Deep Reinforcement Learning (DRL), its implementation process is as follows.

[0163] 1.2.2.1 Problem Modeling (MDP)

[0164] state space High-dimensional continuous state, including: the robot's current joint angle Joint velocity End position / attitude Local environmental point clouds (Represented in the end coordinate system); Current target path point Relative pose; cable tension estimate (Indirectly estimated via joint torque sensors or force sensors); Current mission phase indicator.

[0165] Action space The robot's end effector's "acceleration command" in Cartesian space. Or acceleration commands in the joint space

[0166] Choosing acceleration as the action facilitates the integration of dynamic models.

[0167] reward function The reward function is carefully designed to guide the agent in learning safe, efficient, and smooth movements. It is a weighted sum of multiple sub-rewards.

[0168]

[0169] in, The current actual position vector of the robot's end effector;

[0170] : The reference desired position vector generated by global path planning;

[0171] The current actual posture of the robot's end effector;

[0172] : The desired pose in the reference trajectory;

[0173] Weighting coefficients for position and attitude errors;

[0174] : ;

[0175] The preset absolute safety distance threshold; a distance less than this means an extremely high risk of collision or that a collision has already occurred.

[0176] The preset warning distance threshold means that the robot will be penalized when it enters this range, but will not have collided yet.

[0177] Collision constant penalty, a very large negative value, used in extremely dangerous situations. Forcibly terminate or severely punish actions;

[0178] : The weighting coefficient of the obstacle avoidance term, used to adjust the penalty intensity of the squared distance term;

[0179] The acceleration command output at the current time step is penalized to prevent sudden changes in motion and protect the motor.

[0180] The current actual speed of the robot's end effector;

[0181] The velocity of the robot's end effector at the previous time step;

[0182] : Weighting coefficients corresponding to acceleration amplitude and velocity jitter (jumpiness), respectively;

[0183] : Estimated cable tension (estimated by force sensor or inverse kinematics of joint torque);

[0184] : The maximum allowable tension threshold for the cable; exceeding this value may break the cable or damage the terminals.

[0185] : Weighting coefficient for tension overload. The rectification function indicates that only when the tension exceeds The penalty is calculated only when tension is within the safe range; the penalty is 0.

[0186] The next discrete path point currently being tracked;

[0187] Distance tolerance threshold (radius for determining whether a waypoint has been "reached");

[0188] The indicator function takes a value of 1 when the condition is met, and 0 otherwise. In other words, a reward is only given when the user successfully reaches the vicinity of a waypoint.

[0189] : Sparse reward weight for reaching waypoints;

[0190] : Joint torque vector of the robot;

[0191] : The joint angular velocity vector of the robot;

[0192] The dot product of torque and velocity, physically represented as instantaneous power;

[0193] : Weighting coefficient for energy consumption penalty, encouraging low-power movement.

[0194] 1.2.2.1 Algorithm and Training

[0195] The use of state-of-the-art offline reinforcement learning algorithms (such as Conservative Q-Learning, CQL) or model-based reinforcement learning (such as PETS) is chosen because interactive sampling with real robots is costly and risky. First, massive, parallel training is performed in a high-fidelity digital twin simulation environment (integrating robot dynamics, cable flexibility models, and sensor noise models) to teach the agent to plan trajectories under complex constraints. The trained policy network... .

[0196] 1.2.2.2 Deployment to a real system

[0197] In real-world systems, a "simulation-based strategy supplemented by online fine-tuning" approach is adopted. Online fine-tuning utilizes online learning methods under safety constraints (such as Safe RL) to make small, safe adjustments to the strategy using real-world interaction data to adapt to the differences between simulation and reality (Sim2Real).

[0198] Output and Execution: In each control cycle (e.g., 10ms), the DRL planner calculates based on the current state. Through policy network Output Action .

[0199] Acceleration command: This command is combined with the robot's underlying impedance controller or torque controller to be converted into actual joint torque commands, driving the robot to move.

[0200] 1.2.3 High-precision compliant wiring controller

[0201] Wiring and connection control are achieved through a compliant wiring controller.

[0202] State machine design: The process of connecting a single terminal is divided into a finite state machine (FSM). The states of the state machine include:

[0203] Coarse approach: Based on offline planning and visual coarse positioning, move to a safe distance above the target point;

[0204] Visual alignment: Activate the end-effector macro camera, identify the target terminal hole, and calculate the precise offset. The robot moves to the target point ;

[0205] Force-controlled compliant insertion: This is the core stage, employing a variable impedance control model, which is as follows:

[0206]

[0207] The virtual mass matrix defines the inertial characteristics exhibited by the robot's end effector.

[0208] The virtual damping matrix is ​​used to dissipate energy, ensure system stability, and prevent oscillations.

[0209] The adaptive stiffness matrix is ​​the core of this model. It is not a constant, but a function that changes dynamically according to the insertion state.

[0210] The actual position, velocity, and acceleration vector of the robot's end effector.

[0211] : The desired (reference) position, velocity, and acceleration vector of the robot's end effector.

[0212] : The vector of external contact force / torque measured by the robot's end effector sensor.

[0213] Current insertion depth refers to the axial distance from the terminal into the socket or terminal block.

[0214] Alignment Status is a scalar or vector calculated by fusing visual recognition bias and force feedback (such as the magnitude of lateral force), representing the confidence level of hole alignment.

[0215] Adaptive stiffness matrix Stiffness varies with insertion depth h and perceived alignment state change:

[0216] ;

[0217] in, Indicates virtual quality; Indicates virtual damping; Represents the adaptive stiffness matrix; This represents the actual position vector of the robot's end effector. Represents the desired position vector of the robot's end effector; This represents the actual velocity vector of the robot's end effector. This represents the desired velocity vector at the robot's end effector. This represents the actual acceleration vector at the robot's end effector. This represents the desired acceleration vector at the robot's end effector. This represents the external contact force sensed by the robot's end effector; Indicates the stiffness during the search phase; Indicates the stiffness during the insertion phase; Indicates the insertion depth; Indicates the alignment status; Indicates depth Weighting coefficients; Indicates alignment status Weighting coefficients;

[0218] During the search phase Small, Low), Set to smaller The end is allowed to shift laterally under contact force to "slide" into the guide groove. Alignment is achieved when the force sensor detects a contact force pattern that matches the guide groove characteristics (e.g., decreased lateral force, stable axial force), or when visual confirmation indicates alignment. As the stiffness increases, the controller smoothly transitions to a higher level. To ensure precise straight insertion depth.

[0219] Position determination and holding: monitoring axial force Harmony emission signal. When A sudden increase in pressure (indicating contact at the bottom of the terminal) followed by stabilization within a preset range, coupled with an acoustic emission signal characteristic matching a "click" sound indicating proper insertion, indicates successful insertion. Maintain pressure for a certain time and then release.

[0220] Error handling: If during insertion... If the safety threshold is exceeded, or if a persistent jam occurs (displacement does not increase), a retreat strategy is triggered (retreating a short distance in the opposite direction of the force), and the system may re-enter the "visual precision alignment" state.

[0221] Multi-sensor fusion state estimation: Using a Kalman filter to fuse visual positioning results, force sensor readings, and joint encoder information, a more robust estimate of the relative pose and contact state of the end effector and terminal is obtained and used as input to the controller.

[0222] 1.2.4 Dynamic Collaborative Task Arbitrator

[0223] The following information is monitored in real time through a dynamic collaborative task arbitrator: the current task status of each robotic arm, the estimated remaining time, the joint torque / power load, the end effector status, the distance between the end effects of the two arms, and the short-term trajectory predicted by the local planner.

[0224] Constructing a Conflict Prediction and Performance Evaluation Model: Establish a lightweight model to predict the probability of a collision occurring in the future if the current task allocation scheme continues. Total task completion time Dual-arm load balance (e.g., the ratio of maximum load to average load).

[0225] Decision triggers and strategy library: When key metrics are monitored to exceed thresholds (such as...) The arbitrator is triggered when a preset decision point is reached (such as when a cable is laid).

[0226] Decision optimization: The arbitrator selects or combines response strategies from a predefined strategy library. Strategies include: Task swapping: swapping a future task originally assigned to arm A with a task in arm B; Timing fine-tuning: inserting waiting time for a task without changing its assignment, to stagger the working time of the two arms in the conflict zone; Spatial re-segmentation: temporarily adjusting the boundaries of the dynamic workspace; Inserting avoidance actions: commanding one arm to execute a predefined avoidance posture, allowing the other arm to pass first.

[0227] Decision Execution: After making a decision, the arbitrator sends new task instructions or constraints to the local planners of each arm. The local planners then replan the subsequent trajectories based on the new instructions. The entire process is transparent to the higher-level task flow, ensuring the continuity of operations.

[0228] 1.3 Data Management and Learning Optimization Layer

[0229] The data management and learning optimization layer is the system's "memory" and "evolution engine," responsible for collecting and managing data throughout the entire process, and using this data for offline model optimization and knowledge mining.

[0230] 1.3.1 Full Lifecycle Data Lake

[0231] All relevant data for each wiring task is stored in a structured format: offline planning input (drawings, models), planning output (process packages), all sensor timing data executed online (joint encoders, force sensors, vision images, acoustic emissions), control commands, event logs (state switching, abnormal alarms), and final results (success / failure, total time, quality inspection results).

[0232] The data has a unified spatiotemporal label, which facilitates correlation analysis and tracing.

[0233] 1.3.2 Offline Learning and Engine Optimization

[0234] Digital twin model calibration: By comparing the precise pose data collected during actual execution (such as the key point positions calibrated by high-precision measuring equipment) with the twin model, the installation position errors of the components in the model are automatically corrected, making the digital twin increasingly closer to reality.

[0235] Planning algorithm parameter tuning: Analyze the relationship between planning parameters (such as pheromone evaporation coefficient and heuristic weight in ACO algorithm) and final performance indicators (such as total length, number of bends, and operation time) in a large number of successful cases, and use Bayesian optimization or meta-learning methods to automatically find the optimal combination of algorithm parameters for a certain type of cabinet or cable.

[0236] DRL policy network refinement: This involves refining new data collected from real-world environments. Data is added to the experience pool, and the deployed policy network is periodically trained using offline reinforcement learning algorithms. .

[0237] Fine-tuning is performed to better adapt it to the dynamic characteristics of the real physical world.

[0238] Wiring strategy optimization: Collecting successful and failed insertion force curves Combined with acoustic signals, a classifier or anomaly detection model is trained. This model can be used to predict connection failures earlier and more accurately in the online phase. Simultaneously, the variable impedance parameters in successful cases are analyzed. This allows us to recommend the optimal control parameters for different types of terminals and cables.

[0239] Process rule mining: Data mining is performed on a large number of successful routing schemes. By using association rule learning or graph neural networks, hidden and effective routing patterns or rules can be automatically discovered. These new rules can supplement or optimize the existing process rule library.

[0240] Through the close collaboration of these three layers, this invention achieves a complete closed loop, from high-quality pre-planning in virtual space to safe, efficient, and flexible execution in physical space, and then to continuous self-optimization based on execution feedback, thus forming a truly intelligent automated cabling solution for electrical cabinets.

[0241] 2.1 Mathematical Model and Improved ACO Algorithm for Topology Optimization of Multi-Constraint Cable Harnesses

[0242] 2.1.1 Formalization of Mathematical Model

[0243] Let the set of cables to be laid in the electrical cabinet be denoted as Each cable The attributes include: starting point ,end Minimum bending radius Electrical type (such as high voltage, low voltage, signal).

[0244] Define decision variables: for each cable Its path It is a continuous curve in three-dimensional space, using a series of ordered path points. To approximate, where

[0245] Objective function:

[0246] ;

[0247] in, :

[0248] Defined as Or the curvature at the bend exceeds The punishment.

[0249] Cable and Electrical separation penalty between them. If and If isolation is required (such as between high-voltage and low-voltage electricity), the penalty is proportional to the length of their parallel proximity: ,in It is the inverse function of distance.

[0250] path The estimated number of fixing clamps required depends on the path length and the spacing rules of the support points.

[0251] Total penalty term for violating hard constraints, weight great.

[0252] Hard constraints: Spatial obstacle avoidance: (Safe distance from the robot body) and regarding the cables themselves,

[0253] Minimum bending radius: Approximate curvature at any point on the path .

[0254] Electrical safety clearance: ,,if If the type of pair requires isolation, then... .

[0255] Maximum spacing between fixed points: , in the path Above, the distance between any two adjacent planned fixed points (or support points). .

[0256] 2.1.2 Improved Hierarchical Ant Colony Optimization (HACO) Algorithm

[0257] Since it is very difficult to search directly in a high-dimensional continuous space, a hierarchical strategy and a discretized search space are adopted.

[0258] Step 1: Spatial Discretization and Graph Construction

[0259] Analyzing the digital twin model of the electrical cabinet reveals “channels” that can be used for wiring: such as U-shaped grooves along the edge of the cabinet, gaps between component rows, and dedicated cable tray areas.

[0260] Discretize these channels into a coarse topological graph. node Represents key locations near the intersections, endpoints, and connection points of the channel. (Edge) The representative channel segment has the following attributes: length, direction, region, and default cable mounting method.

[0261] At the same time, a finer-grained "local fine grid" is established within each channel segment to plan the specific route of the cable within the channel.

[0262] Step 2: Ant Colony Optimization - Performing Bundle-Level Planning on a Coarse Topology Graph

[0263] Each ant The task is to plan a shared backbone path for a group (bundle) of cables. Ants start from the set of origin of the bundle.

[0264] State transition rule: The probability that an ant chooses the next node j from node i is:

[0265] ;

[0266] in: It is the edge The concentration of pheromones on the surface.

[0267] It is heuristic information, usually taken

[0268] This is the "forward-looking compatibility factor," a key improvement. It estimates the compatibility factor if edges are chosen... As part of the current cable bundle path, consider the ease with which all cables within the bundle can be subsequently connected to their respective endpoints, and whether there will be serious conflicts (such as electrical isolation conflicts) with other planned cable bundles. Estimation is achieved through rapid local search and collision detection. These are weight parameters.

[0269] Constraint checking: When selecting a node, a real-time check is performed to see if any hard constraints are violated (e.g., bending radius is determined by estimating the path's turning angle). If a violation is found, the node is removed from the candidate list. Remove from the list.

[0270] Pheromones Update:

[0271] Local update: After each ant traverses an edge, it immediately releases local pheromones. To prevent premature convergence.

[0272] Global update: After all ants have completed one iteration, select the optimal solution for that iteration. and the global historical optimal solution Update the edges on their paths according to the following rules:

[0273] ;

[0274] in, It is the global pheromone evaporation coefficient. It is a constant. It is the objective function value of the optimal solution. The smaller the size, the greater the pheromone enhancement.

[0275] Step 3: Fine-mesh path adjustment based on coarse path

[0276] After obtaining the coarse bundle backbone path, the precise position of each individual cable within the bundle is planned in the corresponding local fine mesh. This is a relatively simple local planning problem and can be solved using... The algorithm takes into account the minute spacing between cables and the location of fixed points. Ultimately, a detailed discrete path point sequence is obtained for each cable.

[0277] Step 4: Path smoothing and fixed point determination

[0278] Apply B-spline curve fitting to discrete path points to obtain a smooth, continuous path, and ensure that the fitted curve satisfies the minimum bending radius constraint.

[0279] Following the smoothed path, based on the maximum spacing constraint of fixed points It automatically determines the location of fixing points based on available support structures (such as tying holes).

[0280] This algorithm significantly improves the efficiency and success rate of finding high-quality cable bundle topologies in complex, multi-constraint environments by hierarchical search and incorporating forward-looking factors.

[0281] 2.2 Local Motion Planning (DRL-MP) Algorithm Based on Deep Reinforcement Learning

[0282] 2.2.1 State Representation and Feature Extraction

[0283] state Sufficient information is needed for the policy network to make decisions; therefore, an efficient feature extraction network was designed.

[0284] Robot body state: Then, normalization is performed.

[0285] Target and progress status: Current target path point Relative position in the terminal coordinate system (3D) and relative attitude (represented by quaternion differences, 3D), and the proportion of remaining path length to the destination. (1-dimensional).

[0286] Environmental perception status The key to processing high-dimensional point clouds is to employ a lightweight variant of PointNet to handle end-centric local point clouds. (For example, N points, each with 3-dimensional coordinates).

[0287] right Perform random downsampling or farthest point sampling, and fix it to One point.

[0288] Features for each point are extracted using several multilayer perceptrons (MLPs) with shared weights. .

[0289] To preserve spatial structure information, the "Voxel" method can be introduced to divide the local space into a voxel grid, calculate whether there are obstacles within each voxel, forming a 3D occupancy grid, and then extract features through a 3D convolutional network. This invention uses a combination of PointNet and shallow voxel features.

[0290] Cable status Scalar value of cable tension estimated by a joint torque sensor or end-effector force sensor. (1D), and a flag indicating whether the cable is being dragged.

[0291] Historical Information : the past Actions at each time step Alternatively, state changes can be used as input to help the network understand motion trends. This can be achieved using a small recurrent neural network (RNN) layer or by directly concatenating features from the past few frames.

[0292] Ultimately, state It is a concatenation of all the above feature vectors:

[0293]

[0294] 2.2.2 Policy Network and Value Network Structure

[0295] The Actor-Critic architecture is adopted, specifically the Soft Actor-Critic (SAC) algorithm, which has the advantages of high sample efficiency and good stability in continuous control tasks.

[0296] Actor Network (Policy Network)

[0297] Input: Status

[0298] Network structure: Multilayer perceptron (MLP), with each layer followed by a ReLU activation function;

[0299] Output: Action mean Log-standard deviation of the diagonal covariance matrix

[0300] action From Gaussian distribution Medium sampling. During deployment, the mean μ is typically used directly to eliminate randomness.

[0301] Output layer use The activation function limits the output to [-1, 1], and then scales it according to the robot's actual acceleration limit.

[0302] Critic Network (Value Network) ):

[0303] SAC uses two Critic networks And their target networks, to reduce overestimation.

[0304] Input: Status and actions splicing.

[0305] Network structure: MLP, similar to Actor but possibly deeper or wider.

[0306] Output: Scalar Q value, representing the state at which the output is Q. Next action The long-term expected return.

[0307] 2.2.3 Core Update Formula of SAC Algorithm

[0308] SAC adds an entropy regularization term to the standard maximum expected return objective to encourage exploration:

[0309] ;

[0310] in, It's a temperature parameter that controls the balance between exploration and utilization. It is the entropy of the strategy.

[0311] The objective function of the policy network $\pi$ is to maximize the cumulative reward and policy entropy.

[0312] Entropy of a policy. It measures the randomness of a policy; the higher the entropy, the stronger the exploratory nature and the better it can prevent getting trapped in local optima.

[0313] Critic Update: Updating the parameters of the two Critic networks by minimizing the Bellman error loss.

[0314] ;

[0315] ;

[0316] in, It is an experience replay buffer. It is a discount factor. It is the target Critic network (parameters are periodically copied from the main network).

[0317] : The loss function of the Critic network. Critic Network, with the following parameters: Used to assess the state Take action The value of.

[0318] Actor Update: Update the policy network parameters by minimizing the following loss.

[0319] ;

[0320] This encourages strategies to choose actions that yield high Q values ​​and have high entropy (i.e., are more exploratory).

[0321] Temperature parameters renew: It can also adjust automatically to maintain a target entropy value. (Usually set to a negative number of action dimensions).

[0322] The loss function of the Actor network (policy network) has the following parameters: The policy is updated by minimizing this loss, making it more inclined to choose actions with high Q-values ​​and high entropy.

[0323] 2.2.4 Simulation Training and Sim2Real Transfer

[0324] High-fidelity simulation environment construction: The simulation environment is built using physics engines such as NVIDIA Isaac Sim, MuJoCo, and PyBullet. The environment includes:

[0325] A precise URDF model of a dual-arm robot;

[0326] Simplified collision geometry of the electrical cabinet and its internal components;

[0327] Physical modeling of the cable: The "Discrete Elastic Rods" model or a series of rigid body-spring-damper models are used to simulate the bending, torsion and tension dynamics of the cable, as well as its contact friction with the environment.

[0328] Sensor model: Add noise, distortion, and random occlusion to the camera; add Gaussian noise and drift to the force sensor.

[0329] Curriculum Learning: Training begins with simple scenarios (such as an open space or a single cable) and gradually increases in difficulty (more obstacles, narrower passages, multiple cables, and coordinated arm movements), allowing the agent to learn complex skills step by step.

[0330] Domain randomization: Randomizing various parameters in the simulation, such as the friction coefficient of an object, the stiffness of a cable, the extrinsic parameters of a camera, lighting conditions, and minute positional shifts of obstacles. This makes the trained policy more robust to parameter variations in the real world.

[0331] Offline training and online fine-tuning: After large-scale training in simulation, the strategy is deployed to the real robot. Human supervision is used during the initial operation phase of the real system.

[0332] 3. Online Optimization and Adaptive Plugging Control of Impedance Model Based on Learning Mechanism

[0333] To achieve a higher success rate and more adaptable terminal insertion operation, this invention introduces a data-driven learning mechanism on the basis of traditional variable impedance control, enabling the impedance model to learn autonomously from historical insertion data and continuously optimize, forming an intelligent closed loop of "perception-execution-learning".

[0334] 3.1 Learning-based impedance control framework

[0335] The core concept lies in building a learnable variable impedance controller. Its impedance parameters no longer rely entirely on preset empirical values, but are dynamically generated through a machine learning model and adaptively adjusted according to the type of connection task, terminal type, cable specification, and environmental conditions. The learnable variable impedance controller consists of the following components: a data acquisition and feature engineering unit; an offline supervised / reinforcement learning training module; an online adaptive and meta-learning module; and safety constraints and anomaly handlers.

[0336] 3.2 Data Acquisition and Feature Representation

[0337] The collected data includes:

[0338] Process signal: Six-dimensional force / torque sequence terminal pose Joint torque

[0339] Visual features: Feature vector of terminal image captured by the front-end camera. (Extracted via CNN);

[0340] Context information: Terminal model Cable specifications Robot configuration

[0341] Result label: Connection success / failure indicator If successful, record the insertion depth. and time consumption

[0342] Feature Construction: Construct the above data into a state feature vector. :

[0343] ;

[0344] in, The difference feature represents the trend of change.

[0345] 3.3 Offline Learning: Impedance Policy Training Based on Deep Reinforcement Learning

[0346] Objective: To train a policy network in a simulation environment. The network is based on the current state Directly output impedance parameter adjustment command 'a' to maximize the success rate and efficiency of the connection.

[0347] Algorithm Design

[0348] 3.3.1. Motion Space Design

[0349] Action 'a' is an incremental adjustment of the impedance parameter:

[0350] ;

[0351] in, These are the main diagonal elements of the stiffness and damping matrices. These are the parameters for variable impedance switching.

[0352] 3.3.2. Reward Function Design

[0353] ;

[0354] in, A significant positive reward is given when the connection is made correctly;

[0355] Increased reward insertion depth and penalty time consumption;

[0356] A significant negative reward is given when the contact force exceeds the safety threshold;

[0357] : The penalty is the drastic fluctuation of the impedance parameter;

[0358] 3.3.3. Course Learning Strategies

[0359] Phase 1: Simple environment (large-diameter terminals, unobstructed access) to learn basic mating strategies.

[0360] Phase Two: Increase environmental disturbances (simulate visual positioning errors and terminal position offsets).

[0361] Phase 3: Complex environment (small aperture, high precision requirements, dynamic obstacles).

[0362] 3.3.4. Domain Randomization

[0363] Randomization in simulations includes terminal friction coefficients, robot joint clearances, visual noise models, and force sensor noise, enhancing the generalization ability of the strategy.

[0364] Training output: A well-trained policy network is obtained. It can output the optimal impedance adjustment command based on the real-time status.

[0365] 3.4 Online Adaptive Learning: Rapid Optimization Based on Online Learning and Meta-Learning

[0366] After the offline training strategy is deployed in a real environment, it is continuously fine-tuned through an online learning mechanism to adapt to the unmodeled dynamics of the real world.

[0367] 3.4.1 Online Adaptive Learning (using a model-based online adaptive algorithm)

[0368] Collect real data after each plug-in task. Store in buffer

[0369] Employing a mini-batch online gradient update strategy for network parameters Make fine adjustments:

[0370] ;

[0371] Update step size Smaller size ensures a smooth transition of strategies.

[0372] 3.4.2 Meta-learning framework

[0373] Construct a meta-learner that enables the policy network to be quickly adjusted based on a small number of new terminal plugging samples.

[0374] Algorithm flow:

[0375] 3.4.2.1. Collect a small amount of insertion data for the new type of terminal. .

[0376] 3.4.2.2. Utilizing meta-learners (such as MAML, Reptile) in Based on this, a small number of intra-step gradient updates are performed to obtain a strategy adapted to the new task. .

[0377] This process can be completed in minutes, enabling rapid adaptation for small samples.

[0378] 3.5 Impedance Parameter Generation and Safety Monitoring

[0379] 3.5.1 Real-time generation of impedance parameters

[0380] During the actual connection process, the controller executes once every 0.01 seconds:

[0381] Obtain current state features

[0382] Input Policy Network Get adjustment action

[0383] Update current impedance parameters:

[0384]

[0385]

[0386]

[0387]

[0388] Substitute the updated parameters into the variable impedance control law (see the original scheme formula) to calculate the current control output.

[0389] 3.5.2 Security Monitoring and Anomaly Handling

[0390] Set safety boundaries: limits on the rate of change of impedance parameters, upper and lower limits on stiffness, and maximum permissible contact force.

[0391] When a network output command may cause an out-of-bounds error, a security filter is triggered to cut the command to a safe range.

[0392] If multiple connection failures occur consecutively, the system automatically switches to conservative impedance mode (high damping, low stiffness) and reports a request for manual intervention or relearning.

[0393] The following technical effects can be achieved in this invention:

[0394] 1. Significantly improves first-time connection success rate: The learning impedance controller can automatically adapt to different terminal types and cable characteristics, eliminating the need for manual parameter tuning and improving the first-time success rate by ≥15%.

[0395] 2. Possesses continuous evolution capability: As operational data accumulates, the controller is continuously optimized, and the plug-in efficiency (average time) can be gradually reduced.

[0396] 3. Strong environmental adaptability: Through online learning and meta-learning, the system can quickly adapt to unforeseen disturbances (such as terminal manufacturing tolerances and robot wear).

[0397] 4. Reduced reliance on precise models: No need for precise terminal-socket friction models; contact dynamics are learned autonomously through data-driven processes.

[0398] 5. Form a process knowledge base: The impedance strategies learned can be archived as "process knowledge packages" for rapid deployment and inheritance of similar terminals.

[0399] On the other hand, the present invention provides a cable routing device for a dual-arm robot based on an electrical cabinet digital twin. Figure 5 This is a structural block diagram of a dual-arm robot cable routing device based on an electrical cabinet digital twin according to an embodiment of the present invention. (Reference) Figure 5 The dual-arm robot cable routing equipment based on the digital twin of the electrical cabinet includes: a digital twin model construction module, a global topology layout generation module, a routing instruction generation module, a routing and plugging operation module, and an action adjustment module.

[0400] The digital twin model building module is used to: build a digital twin model of the electrical cabinet based on the geometric model of the electrical cabinet.

[0401] The global topology layout generation module is used to: extract electrical connection relationships from cabling files, generate an undirected connection graph, and based on the undirected connection graph, use the digital twin model of the electrical cabinet to perform global topology layout for cable bundles and generate an offline planned path point sequence.

[0402] The cabling instruction generation module is used to: decompose the work of each cable in the cable bundle based on the offline planned path point sequence, perform dual-arm collaborative timing planning, and pre-divide the workspace to form cabling instructions.

[0403] The wiring and splicing module is used to perform wiring and splicing operations for each cable using a dual-arm robot, based on wiring instructions and real-time sensing data.

[0404] The motion adjustment module is used to: determine whether the current cable has completed the wiring and plugging operation, and if the wiring and plugging operation has not been completed, dynamically adjust the task sequence or motion trajectory, and return to the wiring and plugging operation module for execution.

[0405] Other implementation details of the cable routing equipment for the dual-arm robot based on the digital twin of the electrical cabinet can be found in the above description of the cable routing method for the dual-arm robot based on the digital twin of the electrical cabinet, and will not be repeated here.

[0406] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above description is only a specific embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A cable routing method for a dual-arm robot based on an electrical cabinet digital twin, characterized in that, The method includes: Based on the geometric model of the electrical cabinet, a digital twin model of the electrical cabinet is constructed; Electrical connection relationships are extracted from the wiring file to generate an undirected connection graph. Based on the undirected connection graph, the digital twin model of the electrical cabinet is used to perform global topology layout for the cable bundle and generate an offline planned path point sequence. Based on the offline planned path point sequence, the operation of each cable in the cable bundle is decomposed into tasks, the dual-arm collaborative timing is planned, and the workspace is pre-divided to form wiring instructions. Using a dual-arm robot, wiring instructions and real-time sensing data are employed to perform wiring and splicing operations for each cable. Determine whether the current cable has completed the wiring and plugging operation, and if the wiring and plugging operation has not been completed, dynamically adjust the task sequence or movement trajectory, and return to step: using a dual-arm robot, the wiring instructions and real-time sensing data are used to perform wiring and plugging operations for each cable.

2. The method according to claim 1, characterized in that, Based on the geometric model of the electrical cabinet, a digital twin model of the electrical cabinet is constructed, including: An electrical cabinet geometric model is constructed based on the electrical cabinet's geometric structure, wherein the electrical cabinet's geometric structure includes: structural components and electrical elements; Static attributes are added to each electrical component in the electrical cabinet geometric model, wherein the static attributes include semantic tags and associated electrical attributes, and the cabling process specification library is bound to the electrical cabinet geometric model in the form of computable logical rules; Define the dynamic attributes, accessibility, and traversability of the electrical cabinet's geometric model to form a digital twin model of the electrical cabinet.

3. The method according to claim 2, characterized in that, Based on the undirected connectivity graph, and using the digital twin model of the electrical cabinet, a global topology layout is performed for the cable bundles, including: The mathematical constraints derived from the computable logical rules are integrated into the digital twin model of the electrical cabinet, and a comprehensive cost function for all cable path sets is established for the mathematically constrained digital twin model of the electrical cabinet. ; in, Indicates the total length of all cable paths; Indicates the total number of bends or the total bend radius penalty; This represents the penalty for undesirable proximity between high-voltage and low-voltage cables. Penalty indicating the required number of clamps; Indicates penalties for violating hard constraints; , , , , Indicates weight; Using the hierarchical ant colony optimization algorithm and the digital twin model of the electrical cabinet, the comprehensive cost function corresponding to different cable bundle planning schemes is solved based on the undirected connection graph. The cable bundle planning scheme that minimizes the comprehensive cost function is used as the final global topology layout.

4. The method according to claim 3, characterized in that, Using the hierarchical ant colony optimization algorithm and the digital twin model of the electrical cabinet, based on the undirected connectivity graph, the comprehensive cost function corresponding to different cable bundle planning schemes is solved, including: The channels available for wiring are identified from the digital twin model of the electrical cabinet, and the channels are spatially discretized to form a coarse topology map, and a fine-grained local fine mesh is established within each channel. Using ant colony optimization, cable bundle-level routing planning is performed on the coarse topology map to obtain coarse bundle backbone paths; In the local fine grid of the coarse bundle backbone path, the position of each cable in its respective cable bundle is planned for each cable; Curve fitting is performed on the spatially discretized path points of each cable under the minimum bending radius constraint to obtain a smooth continuous path, and the location of the cable's fixed points is determined based on the maximum spacing constraint of the fixed points and the available support structure.

5. The method according to claim 4, characterized in that, Using ant colony optimization, cable bundle-level routing planning is performed on the coarse topology map to obtain coarse bundle backbone paths, including: Design and initialize pheromones; Let the probability be: An ant starts from the set of starting points of the cable bundle and chooses the next node from the current node. ; in, For path superior The pheromone concentration at any given time For heuristic information, As a forward-looking compatibility factor, , , Indicates the importance coefficient; Check if the selected node violates hard constraints; Select the chosen node without violating the hard constraints; Update the pheromones; After an ant has traversed a path, it releases pheromones locally. After all ants have completed one iteration, select the best solution of this iteration and the best solution in the global history to update the path.

6. The method according to any one of claims 1 to 5, characterized in that, Using a dual-arm robot, and leveraging the wiring instructions and real-time sensing data, wiring and splicing operations are performed for each cable, including: Using the global trajectory guide and the wiring instructions, a path reference trajectory for each cable is generated. A local real-time motion planner based on deep reinforcement learning generates locally optimized action instructions for wiring and splicing based on the path reference trajectory, robot body state, target and progress state, real-time perception data, cable state, and historical information.

7. The method according to claim 6, characterized in that, A local real-time motion planner based on deep reinforcement learning generates locally optimized action commands for wiring and connection based on the path reference trajectory, robot body state, target and progress state, real-time perception data, cable status, and historical information. The insertion process of a single terminal is divided into a finite state machine, and the states of the finite state machine include: Coarse approach: Based on offline wiring planning and visual coarse positioning, move to a safe distance above the target point; Visual alignment: Identify the target terminal hole and calculate the offset; Force-controlled compliant insertion: A variable impedance control model is used to control the insertion force. The variable impedance control model is as follows: ; ; in, Indicates virtual quality; Indicates virtual damping; Represents the adaptive stiffness matrix; This represents the actual position vector of the robot's end effector. Represents the desired position vector of the robot's end effector; This represents the actual velocity vector of the robot's end effector. This represents the desired velocity vector at the robot's end effector. This represents the actual acceleration vector at the robot's end effector. This represents the desired acceleration vector at the robot's end effector. This represents the external contact force sensed by the robot's end effector; Indicates the stiffness during the search phase; Indicates the stiffness during the insertion phase; Indicates the insertion depth; Indicates the alignment status; Indicates depth Weighting coefficients; Indicates alignment status Weighting coefficients; Position determination and holding: monitoring axial force and acoustic emission signals; Abnormal handling: When the axial force exceeds the safety threshold or continuous jamming occurs, the yielding strategy is triggered.

8. The method according to claim 7, characterized in that, The method further includes: Real-time monitoring of impedance parameters during wiring and connection processes; The adaptive stiffness matrix is ​​updated based on the impedance parameters monitored during the wiring and connection process.

9. The method according to claim 8, characterized in that, The method further includes: The following information is monitored in real time: current task status of each robotic arm, estimated remaining time, joint torque / power load, end effector status, distance between the end effects of the two arms, and short-term trajectory predicted by the local planner; Construct a conflict prediction and performance evaluation model, and based on the monitoring information, use the conflict prediction and performance evaluation model to predict the probability of a collision occurring within a predetermined time in the future, the total task completion time, and the load balancing degree of the two arms, continuing the current task allocation scheme. If a key indicator in the monitoring information exceeds a threshold or reaches a preset decision point, the arbitrator is triggered to arbitrate. The arbitrator selects or combines response strategies from a predefined strategy library to make a decision, wherein the response strategies include: task swapping, timing fine-tuning, spatial resegmentation, and inserting avoidance actions. Based on the decision made by the arbitrator, new task instructions or constraints are sent to the local real-time motion planners of each arm of the dual-arm robot, and the local real-time motion planners replan the subsequent trajectory based on the new instructions or constraints.

10. A dual-arm robot cable routing device based on electrical cabinet digital twin, characterized in that, The device includes: The digital twin model building module is used to: build a digital twin model of the electrical cabinet based on the geometric model of the electrical cabinet; The global topology layout generation module is used to: extract electrical connection relationships from the cabling file, generate an undirected connection graph, and based on the undirected connection graph, use the digital twin model of the electrical cabinet to perform global topology layout for the cable bundle and generate an offline planned path point sequence. The cabling instruction generation module is used to: based on the offline planned path point sequence, perform task decomposition, dual-arm collaborative timing planning, and workspace pre-segmentation for the operation of each cable in the cable bundle to form cabling instructions; The wiring and splicing operation module is used to perform wiring and splicing operations for each cable by means of a dual-arm robot, using the wiring instructions and real-time sensing data; The action adjustment module is used to: determine whether the current cable has completed the wiring and plugging operation, and if the wiring and plugging operation has not been completed, dynamically adjust the task sequence or movement trajectory, and return to the wiring and plugging operation module for execution.

Images