Mechanical arm multi-target motion path planning method and system

Abstract

The invention relates to the technical field of industrial robot and mechanical arm motion control, in particular to a mechanical arm multi-target motion path planning method and system. The method comprises the following steps: firstly, constructing a scene graph state structure containing geometric, topological and real-time states by integrating multi-source data; constructing a multi-objective cost function fusing time, energy consumption, risk and collaborative constraint based on the structure, and coding by using a graph neural network; carrying out state construction and strategy training on the coding space by applying reinforcement learning, and generating a candidate path set of cooperation of multiple mechanical arms; and finally, performing collision detection, optimization screening and smoothing processing on the candidate paths through digital twinborn simulation, configuring a local re-planning mechanism, and finally generating a control instruction. According to the method, the path planning collaboration, safety and comprehensive efficiency of the multi-mechanical-arm system in a complex scene are effectively improved.

Description

Technical Field

[0001] This invention relates to the field of motion control technology for industrial robots and robotic arms, and in particular to a method and system for multi-objective motion path planning for robotic arms. Background Technology

[0002] In the field of industrial robot and robotic arm motion control technology, existing solutions for multi-robotic arm motion path planning in robotic arm operation scenarios typically revolve around discrete workstations and fixed process flows. They construct simplified workspace models for single or a small number of robotic arms, generate motion paths through preset rules or single-objective optimization strategies, and issue joint trajectory and posture commands to the robotic arm controller from independent control logic. However, these solutions suffer from limitations such as difficulty in expressing robotic arm structural parameters, cargo location, pallet location, and obstacle contour data in a unified data structure, difficulty in forming a scene graph state structure that covers the global environment and task relationships, and lack of path planning constraints that are closely coupled with the operation cycle requirements. Existing methods mostly focus on path planning for a single optimization objective, often favoring time or energy costs. They lack systematic modeling of collision risk and collaborative constraints, making it difficult to uniformly construct, normalize, and set weight parameters for time, energy, collision risk, and collaborative constraints within the same multi-objective path planning cost function. Graph neural network input channel mapping processing also largely remains at the level of geometric features and a few state features. Under the demanding cycle time requirements of complex warehousing operations, problems such as path interference, frequent task switching, and uneven load distribution can easily occur between multiple robotic arms. For the numerous node attributes and edge relationships in the scene graph state structure, existing technologies generally lack a linkage design oriented towards multi-objective costs in the construction of graph encoding state representation. The state construction and action space partitioning of reinforcement learning are often disconnected from the cost function of multi-objective path planning. The generation of candidate multi-robot path sets often fails to fully reflect the work cycle requirements and collaborative constraints. Digital twin warehouse environment models in existing systems are mostly used for static verification or offline simulation, making it difficult to form a closed loop with the candidate multi-robot path sets. There is a lack of close correspondence between path simulation execution, collision detection, path selection and trajectory smoothing and actual control commands. The mechanism for writing control commands and work process data structures back to the scene graph state structure is also imperfect. This results in a significant disconnect between warehouse production cycle control, resource utilization in stacking and sorting processes, accumulation of work process data structures and path planning updates, which adversely affects the continuous and stable operation of multi-robot collaborative operations and production organization.

[0003] Furthermore, existing methods mostly focus on path planning for a single optimization objective, and the combination of graph neural networks (GNNs) and reinforcement learning (RL) in robotic arm path planning lacks a coordinated design for the costs of multiple objectives (for example, the oral robotic arm solution in CN120953489A only uses GNNs to process geometric data, without addressing the trade-offs of energy consumption, risk, and other multiple objectives). This invention solves the problem of path selection imbalance in complex scenarios for multiple robotic arms by integrating a cost function that incorporates time, energy consumption, risk, and collaborative constraints. Summary of the Invention

[0004] This invention provides a method for multi-objective motion path planning for a robotic arm, comprising:

[0005] Acquire the robotic arm's structural parameters, cargo location, pallet location, and obstacle contour data; perform coordinate system calibration, topology analysis, node attribute and edge relationship configuration, sensor channel binding, and state field initialization to generate a scene graph state structure.

[0006] Based on the scene graph state structure, the cost function for multi-objective path planning is constructed, the cost term is normalized, the weight parameters are set, and the graph neural network input channel is mapped to generate the graph neural network encoding configuration structure.

[0007] Based on the graph neural network encoding configuration structure, reinforcement learning state construction, action space partitioning, policy network training and candidate path search processing are performed to generate a candidate set of multi-robot paths.

[0008] Based on the candidate set of multiple robotic arm paths, path simulation execution, collision detection, path selection, trajectory smoothing, and local replanning trigger configuration processing are performed to generate control commands and operation process data structures.

[0009] Preferably, the robotic arm structural parameters, cargo location, pallet location, and obstacle contour data include:

[0010] Robotic arm structural parameters refer to a set of parameters that describe the geometry and kinematic boundaries of each robotic arm, including the initial posture of each joint, the range of joint rotation, the length of the link, the posture and installation position of the end effector, and the installation position of the robotic arm base in the field.

[0011] The location of a storage location refers to the coordinates of the center point, orientation information, and size range of each storage unit on the shelf in the global coordinate system of the site.

[0012] Pallet location refers to the pallet's external dimensions, center coordinates of the bearing surface, and boundaries of the permitted parking area used for stacking or transporting materials;

[0013] Obstacle contour data refers to the external boundary, multi-segment polyline description, and height range information of fixed equipment, temporary storage objects, or safety isolation facilities that affect the movement trajectory of the robotic arm and mobile chassis.

[0014] Preferably, the process of coordinate system calibration and topological relation analysis also includes:

[0015] Coordinate system calibration involves transforming the robot arm's body coordinate system, sensor mounting coordinate system, local coordinate system of the storage location, and local coordinate system of the pallet to the global coordinate system of the site.

[0016] Topological relation resolution involves identifying reachability, interference, and constraint relationships between nodes based on geometric data in a unified coordinate system to form graph structure connection information.

[0017] Preferably, the process of configuring node attributes and edge relationships, binding sensor channels, and initializing state fields further includes:

[0018] The node attribute configuration includes creating attribute field sets for robotic arm nodes, storage location nodes, pallet nodes, and obstacle nodes respectively;

[0019] Edge relationship configuration includes setting direction attributes, reachability tags, and initial values ​​for association costs for connected edges;

[0020] Sensor channel binding involves establishing a mapping relationship between visual sensors, LiDAR sensors, depth cameras, and environmental perception devices and relevant nodes or edges in the scene graph, and recording the perception channel number.

[0021] State field initialization involves creating fields for each node and edge to describe the current and historical states and writing initial values ​​based on sensor observations.

[0022] Preferably, the process of constructing the multi-objective path planning cost function, normalizing the cost term, setting weight parameters, and mapping the input channels of the graph neural network further includes:

[0023] The construction of the cost function for multi-objective path planning includes creating time cost terms, energy cost terms, collision risk terms, and collaborative constraint terms for connecting edges;

[0024] Cost term normalization involves mapping the original cost terms of time cost, energy cost, collision risk, and collaborative constraint to dimensionless values ​​within a specific range.

[0025] The weighting parameter settings include configuring weighting coefficients for various normalized cost items based on task priority strategies and equipment health status, according to job cycle requirements.

[0026] Input channel mapping involves mapping node attributes, edge attributes, and global features to the node feature channels, edge feature channels, and global feature channels of the graph neural network.

[0027] Preferably, the process of constructing reinforcement learning states and dividing the action space further includes:

[0028] State construction includes constructing reinforcement learning state vectors based on graph-encoded state representations generated by graph neural network encoding; action space partitioning includes dividing the action space into task-level actions and trajectory segment-level actions and generating a set of candidate actions for each robotic arm.

[0029] Preferably, the process of training the policy network and searching candidate paths further includes:

[0030] Policy network training involves updating network parameters during virtual interaction by outputting the action probability distribution based on the state vector and accumulating the reward based on the multi-objective cost function; candidate path search involves generating candidate paths through the trained policy network by expanding multiple trajectories in parallel.

[0031] Preferably, the process of performing path simulation execution, collision detection, and path selection also includes:

[0032] Path simulation execution includes discretizing path sequences in a digital twin warehouse environment model and performing kinematic and dynamic calculations; collision detection includes checking the geometric overlap between the robotic arm and obstacles and the mutual interference between robotic arms at each simulation time step; path filtering includes eliminating paths that do not meet the requirements and retaining a representative subset of paths based on collision event statistics, cumulative time consumption, and estimated energy consumption in the simulation evaluation results data.

[0033] Preferably, the process of trajectory smoothing and local replanning trigger configuration processing further includes:

[0034] Trajectory smoothing involves reconstructing the trajectory curve through spline interpolation or piecewise polynomial fitting to meet velocity and acceleration constraints; local replanning trigger configuration involves setting trigger thresholds for critical road segments based on trajectory tracking error, obstacle distance, and time delay.

[0035] Preferably, a multi-objective motion path planning system for a robotic arm, applied to any of the methods described above, includes:

[0036] The scene graph state structure construction module is used to obtain the robot arm structural parameters, cargo location, pallet location and obstacle contour data based on the basic data of the robot arm operation scene, perform coordinate system calibration, topology relationship parsing, node attribute and edge relationship configuration, sensor channel binding and state field initialization processing, and output the scene graph state structure for downstream modules to call.

[0037] The multi-objective path planning cost function configuration module is used to obtain multiple cost items from the scene graph state structure and operation cycle requirements, construct the multi-objective path planning cost function, normalize the cost items and set weight parameters, and output the multi-objective path planning cost function object for the graph neural network encoding configuration module to call.

[0038] The graph neural network encoding configuration module is used to divide node feature channels, edge feature channels and global feature channels according to the aforementioned multi-objective path planning cost function object and scene graph state structure, complete the mapping between fields and channels and generate the graph neural network encoding configuration structure, and provide encoding configuration to the graph encoding and reinforcement learning state construction module.

[0039] The graph coding and reinforcement learning state construction module is used to obtain graph coding state representations based on the graph neural network coding configuration structure and scene graph state structure, construct reinforcement learning states, and generate a candidate multi-robot path set by combining action space partitioning and policy network training. The candidate multi-robot path set is then sent to the digital twin warehouse environment simulation and evaluation module.

[0040] The digital twin warehouse environment simulation and evaluation module is used to perform path simulation, collision detection and path selection based on the candidate multi-robotic arm path set and the digital twin warehouse environment model. It performs trajectory smoothing on the selected paths and generates simulation evaluation result data, and outputs path-level evaluation information to the control command and operation process data generation module.

[0041] The control command and operation process data generation module is used to generate control commands and operation process data structures based on simulation evaluation results and filtered, smooth paths. It sends control commands to the robotic arm and mobile chassis controller, and writes the operation process data structure back to the scene graph state structure for subsequent path planning cycles to update the basic data.

[0042] Beneficial effects:

[0043] To address the issue of inconsistent path planning inputs caused by the scattered storage and lack of unified representation of robotic arm structural parameters, cargo location, pallet location, and obstacle contour data in existing technologies, this paper proposes a continuous modeling chain that integrates coordinate system calibration, topological relationship parsing, node attribute and edge relationship configuration, sensor channel binding, and state field initialization. This chain converges the basic data of the robotic arm operation scenario into a single scene graph state structure, providing a unified environmental description foundation for subsequent multi-objective path planning cost function construction, graph neural network encoding, and candidate multi-robotic arm path set generation. This reduces the impact of incomplete environmental modeling and fragmented multi-source data on robotic arm motion path planning.

[0044] To address the issue that existing path planning technologies often only set single objectives around time or energy costs, lacking unified quantification of collision risk and collaborative constraints, and making them difficult to handle within the same optimization framework, this paper proposes a method that simultaneously constructs time, energy, collision risk, and collaborative constraints based on the scenario graph state structure and job cycle requirements. This involves constructing a multi-objective path planning cost function, normalizing cost terms, and setting weight parameters. Then, a graph neural network encoding configuration structure is generated through graph neural network input channel mapping. This ensures that the multi-objective path planning cost function objects meet unified requirements in terms of numerical scale and channel allocation. This facilitates the comprehensive consideration of time, energy consumption, collision risk, and collaborative constraints in graph encoding state representation and reinforcement learning state construction, mitigating the path selection imbalance problem caused by the difficulty in representing multi-objective trade-offs within the model in existing solutions.

[0045] To address the issue that existing technologies often validate path planning results offline in simplified simulation environments, lacking a closed-loop path evaluation and control generation mechanism that coordinates with graph-encoded state representation, reinforcement learning state construction, and digital twin warehouse environment models, this paper generates graph-encoded state representations based on graph neural network-encoded configuration structures and scene graph state structures. This is combined with reinforcement learning state construction, action space partitioning, and policy network training to generate a candidate set of multi-robot paths. Path simulation execution, collision detection, path selection, and trajectory smoothing are then performed within the digital twin warehouse environment model. After local replanning triggers configuration processing, control commands and operational process data structures are generated. This achieves tight coupling between the candidate multi-robot path set and the digital twin warehouse environment model. The control commands and operational process data structures have already undergone simulation verification for collision risk and trajectory smoothing during generation, which helps alleviate the problems of disconnect between path planning and actual control, and the difficulty of the operational process data structure influencing subsequent path planning in existing technologies. Attached Figure Description

[0046] Figure 1 A flowchart illustrating a multi-objective motion path planning method for a robotic arm provided in an embodiment of this application;

[0047] Figure 2 This is a structural block diagram of a multi-objective motion path planning system for a robotic arm, provided in an embodiment of this application. Detailed Implementation

[0048] Example 1: Refer to Figure 1 In the field of industrial robot and robotic arm motion control technology, Figure 1 This is a flowchart illustrating a multi-objective motion path planning method for a robotic arm according to an embodiment of the present invention. The process may include at least steps S1-S4:

[0049] Step S1: Obtain the structural parameters of the robotic arm, the location of the cargo, the location of the pallet, and the contour data of the obstacles. Perform coordinate system calibration, topology analysis, node attribute and edge relationship configuration, sensor channel binding, and state field initialization to generate the scene graph state structure.

[0050] Step S2: Based on the scene graph state structure, construct the multi-objective path planning cost function, normalize the cost terms, set the weight parameters, and perform graph neural network input channel mapping processing to generate the graph neural network encoding configuration structure.

[0051] Step S3: Based on the graph neural network encoding configuration structure, perform reinforcement learning state construction, action space partitioning, policy network training and candidate path search processing to generate a candidate multi-robot path set;

[0052] Step S4: Based on the candidate multi-robot path set, perform path simulation execution, collision detection, path selection, trajectory smoothing and local replanning trigger configuration processing to generate control commands and operation process data structures.

[0053] To address the problem in existing technologies where the structural parameters of robotic arms, location positions, pallet positions, and obstacle contours are stored in a scattered manner and lack a unified representation, leading to inconsistent path planning inputs and incomplete environment modeling, this invention constructs a scene graph state structure in step S1, providing a unified data foundation for subsequent planning. Specifically, this includes:

[0054] Step S1: Obtain the structural parameters of the robotic arm, the location of the cargo, the location of the pallet, and the contour data of the obstacles. Perform coordinate system calibration, topology analysis, node attribute and edge relationship configuration, sensor channel binding, and state field initialization to generate the scene graph state structure.

[0055] In this embodiment, the basic data for the robotic arm operation scenario comes from the equipment configuration database, environmental modeling files, and data cache synchronously uploaded by online sensing devices in the warehousing and logistics system. This basic data includes robotic arm structural parameters, storage location, pallet location, and obstacle contour data. Robotic arm structural parameters refer to a set of parameters describing the geometry and kinematic boundaries of each robotic arm, including the initial posture of each joint, joint rotation range, link length, end effector posture and installation position, and the installation position of the robotic arm base in the site. Storage location refers to the center point coordinates, orientation information, and storage location size range of each storage unit on the shelf in the global coordinate system of the site. Pallet location refers to the pallet's external dimensions, center coordinates of the bearing surface, and the boundary of the allowed parking area for stacking or handling materials. Obstacle contour data refers to the external boundaries, multi-segment polyline descriptions, and height range information of fixed equipment, temporary storage, or safety isolation facilities that affect the movement trajectory of the robotic arm and the mobile chassis. The system first obtains the equipment list and area number involved in the current task from the task scheduling module. Then, based on this, it retrieves the corresponding robotic arm structural parameters, cargo location, pallet location, and obstacle outline data from the basic data storage. The system then summarizes these data according to a unified site global coordinate code to form the input data set for this step.

[0056] Specifically, after acquiring the aforementioned basic data, the coordinate descriptions used by different data sources need to be uniformly processed; this process is called coordinate system calibration. Coordinate system calibration refers to the process of transforming the coordinate systems of each robotic arm body, sensor installation coordinate system, local coordinate system of the storage location, and local coordinate system of the pallet into a single global coordinate system for the site. The system reads the installation position and attitude configuration of the robotic arm base, obtains the pose of the vision and laser sensors in the global coordinate system through calibration boards or landmark point cloud recognition, and then converts the storage location and pallet position from local coordinate descriptions to global coordinate descriptions based on the rack installation diagram and pallet parking line positions. For cases with calibration errors, the system iteratively corrects the transformation relationship between the coordinate systems by collecting multiple frames of vision or laser observation data, and writes the correction results into the equipment configuration library for subsequent scene construction calls. The robotic arm structural parameters, storage location positions, pallet positions, and obstacle contour data after coordinate system calibration are organized into a geometric description in a unified coordinate system, providing input for subsequent topology analysis.

[0057] After coordinate system calibration, the system enters the topology relation resolution process. Topology relation resolution refers to the process of identifying the reachability, interference, and constraint relationships between nodes based on geometric data in a unified coordinate system, and forming graph structure connection information. First, based on the robot arm base position and joint range of motion, the system performs meshing and bounding volume estimation of the workspace of each robot arm in the global space, obtaining the reachable area volume of the robot arm. Then, based on the location of the storage location and the pallet location, spatial intersection calculations are performed between these target points and the reachable areas of each robot arm, marking which storage locations or pallets are reachable by a single robot arm and which belong to the shared reachable area of ​​multiple robots. Based on this, the system also calculates the distance distribution between the robot arm workspace and the obstacle boundary based on obstacle contour data, generating a description of potential interference areas for subsequent collision risk assessment. The aforementioned reachability and interference relationships are recorded as topological edges between nodes, and then combined with task differentiation and device attributes to generate edge type labels, thus forming a topology sketch containing the basic relationships between nodes and edges.

[0058] Further, after obtaining the topology sketch, detailed attribute information needs to be configured for each node and edge to complete the node attribute and edge relationship configuration. Node attribute configuration refers to establishing attribute field sets for robotic arm nodes, storage location nodes, pallet nodes, and obstacle nodes respectively. For robotic arm nodes, attribute fields include robotic arm identifier, number of degrees of freedom, joint type combination, maximum load weight, current online status, and communication channel identifier with the controller, etc. For storage location nodes, attribute fields include storage location identifier, shelf number, type of materials that can be placed, maximum load capacity, and current occupancy status, etc. For pallet nodes, attribute fields include pallet identifier, pallet type, current loading status, and allowed parking posture range, etc. For obstacle nodes, attribute fields include obstacle category, whether it is movable, outer contour polygon description, and safety buffer distance, etc. Edge relationship configuration refers to setting direction attributes, reachability markers, and initial values ​​of associated costs for connecting edges. The reachability marker is used to describe whether there is a direct movement channel between the robotic arm and the storage location or pallet, and the initial value of associated costs is used to provide an initial time cost and energy consumption estimation reference for the subsequent multi-objective path planning cost function. The system traverses all nodes and edges in the topology sketch, fills in the above attributes for each entity, and verifies and records any missing or abnormal data. Nodes and edges that cannot be parsed or have conflicts are summarized separately for manual review by the upper-level monitoring module.

[0059] After the nodes and edges are configured, this implementation performs sensor channel binding. Sensor channel binding refers to establishing a mapping relationship between the visual sensors, LiDAR sensors, depth cameras, and other environmental perception devices deployed in the scene and the relevant nodes or edges in the scene graph. The system reads the sensor installation pose configuration and acquisition channel number, determines the spatial area currently covered by each sensor, and performs spatial overlap determination between the coverage area and the robotic arm workspace, the location of the storage location, and the outline of obstacles. For areas stably covered by a certain sensor, the system adds a perception channel field to the corresponding node or edge attribute, writing the perception channel number, sampling frequency, and data type into the field, so that real-time observation data can be directly accessed through this field when updating the scene state. For areas with overlapping coverage by multiple sensors, the system records the priority and fusion strategy identifier of the perception channels, which is used to select the primary observation source or perform weighted fusion when performing multi-source perception fusion. Through sensor channel binding, static geometric entities in the scene graph are associated with dynamic perception data channels, enabling the scene to move from the static modeling stage to the stage where the state can be updated in real time.

[0060] After channel binding is completed, the system performs state field initialization. State field initialization refers to the process of creating fields in the scene graph to describe the current and historical states for each node and edge, and writing initial values. For robotic arm nodes, state fields include the current joint angle, end effector posture, joint speed limit, current task occupancy flag, and safety mode flag; for storage location nodes, state fields include occupancy flag, current material identifier, locked status, and planned inbound / outbound time window; for pallet nodes, state fields include current position coordinates, current load weight, whether in transit, and available task queue number; for obstacle nodes, state fields include whether they are currently in an effective obstacle state, whether movement has been detected, and update timestamp; for edges connecting robotic arm nodes to storage location nodes or pallet nodes, state fields include estimated motion time, estimated energy consumption, collision risk level, and the flag of the most recent planning result. During initialization, the system rapidly updates the node status within the visible area based on a frame of environmental observation data acquired from the sensor channels at the current moment. For example, it updates the occupancy status of cargo locations based on visual recognition results and the effective status of obstacles based on laser ranging results, writing the update time into the corresponding fields. For areas that are temporarily invisible, the status fields retain static configuration values ​​and are refreshed by the real-time sensing thread at a set period in subsequent steps.

[0061] After the state fields are initialized, this implementation defines the overall data structure, including geometric descriptions, topological relationships, node attributes, edge relationships, sensor binding fields, and state fields, as the Scene Graph State Structure. The Scene Graph State Structure is a graph model data format for multi-objective path planning of robotic arms, where each node and each edge carries complete structural information, constraint information, and real-time state information. The system submits this Scene Graph State Structure to the multi-objective path planning cost function construction and graph neural network encoding module in subsequent steps, using it as the input basis for step S2 to obtain the time cost term and other content from the Scene Graph State Structure and the operation cycle requirements. Simultaneously, the Scene Graph State Structure is periodically cached in system memory and log files for subsequent offline training and source analysis.

[0062] In summary, the technical effects of this step are as follows: By constructing a scene graph state structure, the geometric layout, topological connections, perception channels, and real-time status in the robotic arm's operation scene are organized into a unified data carrier, providing a complete and consistent input basis for the subsequent construction of multi-objective path planning cost functions and graph neural network encoding, and supporting real-time response to environmental changes during the robotic arm's path planning process.

[0063] To address the problem that existing path planning technologies often focus solely on single objectives such as time or energy consumption, lack unified quantification of collision risk and collaborative constraints, and are difficult to handle within the same optimization framework, this invention constructs a multi-objective cost function and completes the graph neural network encoding configuration in step S2 to achieve multi-objective collaborative optimization. Specifically, it includes:

[0064] Step S2: Based on the scene graph state structure, construct the multi-objective path planning cost function, normalize the cost terms, set the weight parameters, and perform graph neural network input channel mapping processing to generate the graph neural network encoding configuration structure.

[0065] In this embodiment, the input for step S2 comes from the scene graph state structure generated in the preceding step S1 and the operation cycle requirements given by the upper-level scheduling module. The scene graph state structure already includes the geometric descriptions, topological connections, sensor channel binding fields, and current status fields of robotic arm nodes, storage location nodes, pallet nodes, and obstacle nodes. The operation cycle requirements are generated by the warehouse control system based on order priority and production plan, and cover information such as the planned start time, planned end time, task priority, maximum allowable delay range, and continuous operation time limits for a single robotic arm or combination of robotic arms within the current time window. At the beginning of this step, the system first reads the scene graph state structure and operation cycle requirements from the data cache, aligns them on a unified time axis, marks the robotic arm nodes, storage location nodes, and pallet nodes involved in the current round of tasks with the planning participation mark, and establishes corresponding task index fields in the scene graph state structure to provide complete input for the subsequent calculation of time cost items, energy consumption cost items, collision risk items, and collaborative constraint items.

[0066] Specifically, the time cost term is a quantified field describing the time consumption required to transition from the current state to the target state. The system traverses all feasible edges connecting the robotic arm node to the storage location node and pallet node in the scene graph state structure, reads the robotic arm structural parameters, current joint state, and target node pose, generates candidate motion trajectories for the corresponding edges through the kinematics solving module, and then estimates the action time required to execute the trajectory based on preset joint speed limits and joint acceleration / deceleration capabilities. When processing edges involving complex actions such as grasping and placing, the system also adds the time occupied by the end effector opening / closing and gripper adjustment. For situations where multiple robotic arms share the same storage location or cross-operation areas, the system derives waiting time and queuing time based on the task sequence and time window in the operation cycle requirements, and writes these time quantities into the time cost term field corresponding to the edge. After traversing all edges, each edge participating in this round of planning in the scene graph state structure has a time cost term bound to the task index, providing basic data for the time part of the multi-objective path planning cost function.

[0067] The energy consumption cost term describes the estimated energy consumption of a robotic arm executing a certain trajectory. The system first reads the structural parameter fields, current load field, and motion distance and height change information of each robotic arm node from the scene graph state structure. Then, it combines this information with the energy consumption characteristic table and historical energy consumption statistics stored in the device management module to construct an energy consumption estimation relationship. For each connection, the system queries the energy consumption estimation relationship based on the robotic arm joint motion amplitude, load weight, motion distance, and operating speed mode to obtain the estimated electrical or other energy consumption value when executing that connection, and writes this value into the energy consumption cost term field of the connection. For connections involving frequent acceleration / deceleration actions or requiring long-term maintenance of load posture, the system adds a correction coefficient to increase the energy consumption estimate, making the energy consumption cost term closer to the actual energy consumption level. During the energy consumption estimation process, if the system detects missing load information associated with a connection, an incomplete device energy consumption characteristic table, or insufficient historical statistical samples, it writes a default flag to the energy consumption cost term field for that connection and records the connection number and the reason for the missing information in the log module, facilitating subsequent manual correction or completion through offline learning.

[0068] The collision risk item describes the degree of risk of the robotic arm colliding with obstacles or other robotic arms while moving along a connecting edge. The system reads the contour field, safety buffer distance field, and robotic arm workspace partitioning field of the obstacle node from the scene graph state structure, and combines this with the sensor channel binding field to extract the most recent frames of environmental observation data from the real-time perception thread cache to correct the obstacle's position and shape. Then, the system discretizes the candidate motion trajectory corresponding to each connecting edge into several key attitude points in the time dimension. For each attitude point, the collision detection module is invoked to calculate the minimum distance between each link of the robotic arm and the obstacle. Through statistical analysis of these distance values, the system obtains the distance distribution characteristics corresponding to the connecting edge. Based on the safety buffer distance and preset risk classification rules, the collision risk of the connecting edge is divided into several levels and mapped to the numerical collision risk item field. For areas involving multiple robotic arms working collaboratively, the system also needs to incorporate the current path cache and timeline information of other robotic arm nodes when calculating collision risk. By analyzing time overlap, potential interference between robotic arms can be identified, and this risk can also be accumulated in the collision risk field, so that the field can reflect both the risks between the equipment and the environment and between the equipment and the equipment.

[0069] The collaborative constraint field describes the constraint relationships between multiple robotic arms in terms of task allocation, cycle time coordination, and resource usage. The system first parses the task dependency description in the job cycle time requirement, identifying the sequential constraints between multiple tasks in the same order, the reuse relationships of tasks on storage locations and pallet nodes, and whether the same robotic arm is required to complete tasks consecutively. Then, the system combines the task occupancy markers and node type fields in the scenario graph state structure to convert these constraints into collaborative constraint descriptions for connections. For example, for a task requiring the same robotic arm to complete the gripping at a storage location node and then place the task at a designated pallet node, the system writes the same task identifier on the connections between the storage location node and the pallet node, and records the constraint in the collaborative constraint field that the two connections must appear consecutively in the same path. For scenarios where multiple tasks contend for the same storage location node or pallet node in adjacent time windows, the system, based on task priority and time window overlap, writes waiting penalties and resource exclusivity flags in the collaborative constraint field for connections related to lower-priority tasks. In collaborative tasks that require multiple robotic arms to simultaneously grasp or move the same large material, the system records time alignment requirements, attitude synchronization range, and allowable time errors in the connection and collaboration constraint field of the relevant robotic arms. This information is used to guide the generation of path sequences that meet the requirements of collaborative actions during subsequent path planning and strategy generation.

[0070] After all time cost, energy cost, collision risk, and collaborative constraint items are written into the scene graph state structure, the system begins to construct the multi-objective path planning cost function. The multi-objective path planning cost function organizes the aforementioned cost items into a comprehensive evaluation structure that can be invoked by the path search process for each edge in the scene graph. During construction, the system creates a cost record unit for each edge participating in the planning. This record unit stores the time cost, energy cost, collision risk, and collaborative constraint items for that edge, along with the associated robotic arm identifier and task identifier fields. Subsequently, the system defines path-level cost aggregation rules in the path evaluation module. These rules describe how to sequentially read the cost record units of each edge when a path consists of multiple edges, accumulating or superimposing the various cost items, and adjusting the aggregation method when encountering special splicing relationships described in the collaborative constraint items. Through the combination of cost record units and aggregation rules, the system forms a multi-objective path planning cost function object and records its unique identifier in the metadata area of ​​the scene graph state structure, enabling subsequent steps to access the complete multi-objective cost description through this identifier.

[0071] The construction of the cost function for multi-objective path planning includes creating time cost, energy cost, collision risk, and collaborative constraint terms for connecting edges. The specific calculation formula for the cost function is as follows:

[0072]

[0073] in, Represents nodes in the scene graph To the node The edges connecting the points; That is The comprehensive value of connecting edges; This represents the time cost of normalization. This represents the energy cost of normalization. This represents the normalized collision risk cost; This represents the cost of normalized collaborative constraints. Represents the weighting coefficients, satisfying Furthermore, it can be dynamically adjusted according to the operational cycle requirements of different operational stages (such as high-speed handling and precision assembly); i and j are node indices in the scene graph state structure; This represents the i-th source node in the scene graph; This represents the j-th target node in the scene graph.

[0074] The cost function This is one of the core innovations of this method. In step S2, it assigns a comprehensive cost function to each edge in the scene graph. Subsequently, in step S3, the graph neural network (GNN) combines the learned node and edge features with this cost function to provide a state value assessment for the reinforcement learning (RL) agent. The agent continuously attempts to minimize the accumulated cost function. The optimal strategy is learned, and the globally optimal path is generated in step S4. This design successfully unifies the four objectives of time, energy consumption, safety, and collaboration within an optimizable framework, and achieves efficient solution through GNN+RL, thus solving the technical problem of difficult collaborative optimization of multiple objectives in the background technology.

[0075] To ensure that different types of cost items participate in path comparison under a unified numerical scale, this implementation performs cost item normalization processing on time cost items, energy consumption cost items, collision risk items, and collaborative constraint items after constructing the multi-objective path planning cost function. The system first traverses the cost record units of all edges in the current planning round, statistically analyzes the numerical range and distribution characteristics of various cost items, removes outliers that significantly exceed the normal range, and writes the outlier edge number and corresponding cost item type into the log for subsequent manual inspection. Then, based on the statistically obtained maximum, minimum, and typical values, the system sets a mapping interval and mapping method for each type of cost item, converting the original cost value into a dimensionless value within a specific interval, and writing the conversion result into the normalized cost field of the corresponding edge. For cost items lacking effective samples in the current scenario, such as a certain type of risk not being triggered in this round of task, the system sets the normalized result of this type of cost item to a fixed neutral value and records a flag indicating that the cost item is in a weak participation state in this round of planning in the multi-objective path planning cost function object, avoiding misjudgment of its importance in subsequent processing stages.

[0076] After cost term normalization is completed, the system enters the weight parameter setting phase. Weight parameter setting involves assigning weight coefficients to various normalized cost terms during comprehensive cost calculation, reflecting the emphasis between different planning objectives. The system first obtains priority strategy information for the current task set from the work cycle requirements, including the degree of emphasis on completion timeliness, sensitivity to energy costs, and the strategy level of safety bias. Then, it combines the robotic arm health status provided by the equipment maintenance module and experience adjustment parameters accumulated from historical planning results to calculate the initial weight values ​​for various cost terms. For situations with high completion timeliness requirements, the system increases the weight of the time cost term, while also considering the safety strategy level and increasing the weight of the collision risk term; for situations with high energy cost pressure, the system increases the weight of the energy consumption cost term. To incorporate learning feedback, the system also reads weight correction suggestions from the reinforcement learning strategy network results generated in the previous planning cycle. These suggestions are summarized by the reinforcement learning model based on reward signals during long-term operation. The system adds these correction suggestions to the current weight coefficients according to the set safety boundaries and adjustment step size, allowing the weight parameters to be slowly adjusted during operation. After the above processing, the weight coefficients corresponding to the time cost, energy consumption cost, collision risk, and collaborative constraint are written into the multi-objective path planning cost function object for subsequent comprehensive cost calculation.

[0077] After configuring the multi-objective path planning cost function and its weight parameters, this implementation continues with the graph neural network input channel mapping process. Graph Neural Network (GNN) is a type of neural network model for feature extraction from graph-structured data. This invention uses it to encode the state structure of the scene graph and multi-objective cost information. Input channel mapping refers to the process of determining the input positions and dimensional allocation of node attributes, edge attributes, normalized cost fields, and weight fields in the scene graph within the graph neural network. Based on the graph neural network model configuration, the system divides the input into three types: node feature channels, edge feature channels, and global feature channels. The structural parameters of the robotic arm nodes, the current state field, and the time cost and energy cost related to the node are mapped to the node feature channels. The geometric descriptions of the edges, sensor coverage indicators, normalized time cost terms, normalized energy cost terms, normalized collision risk terms, and normalized collaborative constraint terms, along with their corresponding weight fields, are mapped to the edge feature channels. The summary information of the work cycle requirements and the global weight configuration for this round of planning are mapped to the global feature channels. During the mapping process, for missing or invalid fields, the system uses a pre-defined mask value to fill them in and records the mask flag at the corresponding position, enabling the graph neural network to distinguish between real observations and missing data during the encoding stage.

[0078] After completing the input channel mapping, the system summarizes the number of channels, channel dimensions, field-to-channel mapping rules, multi-objective path planning cost function object identifiers, and association information with the scene graph state structure to construct a graph neural network encoding configuration structure. The graph neural network encoding configuration structure is a data set describing how to extract node features, edge features, and global features from the scene graph state structure and organize them into a graph neural network input tensor. In subsequent step S3, it will be used as input along with the scene graph state structure to generate the graph-encoded state representation. At the end of this step, the system registers the graph neural network encoding configuration structure in memory and writes its associated identifier into the metadata area of ​​the scene graph state structure. This allows subsequent processing modules to complete cross-step calls through this identifier, thus forming a continuous link from scene modeling to cost encoding to policy generation during the multi-objective path planning process.

[0079] The technical effects of this step can be summarized as follows: Based on the scenario graph state structure and operation cycle requirements, time cost items, energy consumption cost items, collision risk items, and collaborative constraint items are constructed. The cost function for multi-objective path planning is constructed, normalized, and weight parameters are set. Then, the relevant fields are mapped to the input channel configuration of the graph neural network and a graph neural network encoding configuration structure is formed. This provides a structured, scalable, and unified cost description foundation that can reflect multiple planning objectives for the subsequent graph encoding and strategy search stages.

[0080] To address the problem in existing technologies where path planning and graph encoding state representation are disconnected, and reinforcement learning state construction and multi-objective cost functions lack linkage, resulting in candidate paths failing to reflect collaborative constraints, this invention integrates reinforcement learning in the graph encoding space through step S3 to generate a set of paths with multi-objective constraints. Specifically, this includes:

[0081] Step S3: Based on the graph neural network encoding configuration structure, perform reinforcement learning state construction, action space partitioning, policy network training and candidate path search processing to generate a candidate multi-robot path set;

[0082] In this embodiment, the input source for step S3 is the graph neural network encoding configuration structure generated in the preceding step S2 and the scene graph state structure generated in step S1. The graph neural network encoding configuration structure provides the mapping relationship between each field in the node feature channel, edge feature channel, and global feature channel and the scene graph state structure, clarifying the positions of the robotic arm node, storage location node, pallet node, obstacle node, and the normalized cost field and weight field on the connecting edges in the graph neural network. The scene graph state structure records the geometric attributes, state fields, cost fields, and collaborative constraint information of all relevant nodes and connecting edges within the current planning cycle. At the beginning of step S3, the system reads the graph neural network encoding configuration structure and the scene graph state structure from memory, constructs a graph neural network input batch according to the mapping rules given in the configuration structure, fills in the feature vectors of each node, each connecting edge, and the global feature vector in the input batch, and simultaneously constructs a topology description to drive the subsequent graph neural network encoding process.

[0083] Specifically, the system first traverses all nodes in the scene graph state structure. Based on the field mapping recorded in the graph neural network encoding configuration structure, it combines the structural parameter fields, joint state fields, task occupancy fields, and local cost-related fields of the robotic arm nodes into node feature vectors. It also combines the spatial location, available material types, occupancy status, and task requirement summary fields of the storage location nodes and pallet nodes into node feature vectors. Finally, it combines the contour description, safety buffer distance, and latest observation timestamp fields of the obstacle nodes into node feature vectors. The graph neural network (GNN) forms the basis of the node input channels in the GNN. Subsequently, according to the edge feature mapping rules given in the encoding configuration structure, the system reads the geometric relationship fields, connected node identifier fields, normalized time cost terms, normalized energy cost terms, normalized collision risk terms, and normalized collaborative constraint terms, as well as the corresponding weight fields, of each edge from the scene graph state structure. These contents are then combined into edge feature vectors and filled into the edge input channels of the GNN. The global feature channel is constructed by reading the job cycle requirement summary field, the multi-objective cost weight configuration field, and the global task load statistics field, enabling the GNN to simultaneously perceive the current task set and cost preferences during encoding. If a field is found to be missing or in an abnormal state during the above process, the system writes a specific mask value according to the masking rules given in the encoding configuration structure and marks the position as missing in the internal state. When such a mark is encountered, the GNN's internal aggregation unit will automatically adopt an ignore or substitution strategy.

[0084] After loading node features, edge features, and global features, the system constructs the graph structure input required for the GNN based on the topological description in the scene graph state structure. It transcribes the edge relationships between nodes into an adjacency index set and integrates multiple regions or task scenarios into a batch of graph data according to the batch strategy set in the encoding configuration structure. Subsequently, the system initiates the GNN forward computation process. First, it performs linear transformations and nonlinear activation operations on each node, mapping the original node features to the internal representation space. Then, in each message passing unit, it collects the feature information of adjacent nodes and edges along the edge direction, gradually updating the node's internal representation through weighted aggregation, normalization operations, and residual superposition. Edge features participate in weight calculation and channel selection during message passing. Collision risk terms and collaborative constraint terms can guide the direction of message aggregation through edge features, ensuring that the node representation after multiple layers of transmission simultaneously contains structural topology, cost distribution, and collaborative relationships. In some implementations, a global readout unit can be introduced into the graph structure to perform weighted summarization of the encoding results of all nodes and edges, thereby forming a global encoding vector describing the overall state of the entire graph. After the forward computation is completed, the system obtains a set of results including node encoding, edge encoding, and global encoding. In this embodiment, these results are collectively referred to as graph encoding state representation.

[0085] The graph-encoded state representation is the core intermediate product in this step, used to carry the state expression of scene structure and multi-objective cost information in the subsequent reinforcement learning stage. When constructing the graph-encoded state representation, the system assigns a node encoding vector to each node and an edge encoding vector to each connection, and generates a graph-level encoding vector based on the readout unit. The node encoding vector is used to describe the local features of a single robotic arm, a single storage location, a single pallet, or a single obstacle in the current planning state; the edge encoding vector is used to describe the relationship features between the robotic arm and task points or nodes; and the graph-level encoding vector is used to describe the overall situation of the entire multi-robotic arm operation scene at the current moment. All encoding vectors are incorporated into the graph-encoded state representation structure, which simultaneously records the mapping relationship between node indices, edge indices, and the original scene graph state structure, ensuring a clear correspondence for subsequent modules when backtracking from the encoding space to the physical space. After the graph-encoded state representation structure is generated, the system registers the structure's identifier in memory, which serves as the unified state input source for subsequent sub-processes in step S3.

[0086] After obtaining the graph-encoded state representation structure, the system enters the reinforcement learning state construction phase. Reinforcement learning (RL) is used in this implementation to construct RL states in the graph encoding space, driving the policy network to make decisions on multi-robot path planning problems. Specifically, the system uses the graph-level encoding vector as the global state basis, selecting the robot node codes, task node codes, and key edge codes related to the current decision time step. These are then concatenated in a predetermined order or aggregated using attention weights to generate the reinforcement learning state vector for that time step. Fields such as remaining task time, robot load level, and historical action summary are appended to the state vector. These fields are obtained by reading the state fields, cost fields, and historical trajectory cache fields from the scene graph state structure. Normalization and truncation are performed during state vector construction to avoid excessively large numerical ranges or too much historical information. For multi-robot task scenarios, the system adopts a centralized state construction strategy, unifying the node codes of multiple robots and related task node codes into the same state vector. Simultaneously, masks or flags are used in the state vector to distinguish the identities of different robots, providing support for subsequent action space partitioning and policy output. Once constructed, the reinforcement learning state vector is stored in the reinforcement learning state cache, which can be accessed repeatedly by the policy network and the value evaluation network at the same time step.

[0087] After constructing the reinforcement learning state, the system performs motion space partitioning according to the coordination and control requirements of multiple robotic arms. Motion space partitioning refers to dividing the overall set of selectable actions into several levels or categories for a multi-robotic arm operation scenario, and determining the set of executable actions for each robotic arm or each task type. Based on the robotic arm node information and task node information in the scenario graph state structure, the system divides the motion space into two main categories: task-level actions and trajectory segment-level actions. Task-level actions are used to indicate the selection of a certain task or target node as the next service object, while trajectory segment-level actions are used to indicate which local trajectory segment or type of motion mode to adopt after selecting a task. For task-level actions, the system selects task nodes with valid states and unclosed time windows from the current incomplete task queue, generates a candidate task set for each robotic arm based on the cooperative constraint terms and the robotic arm capability field, and writes the task set index into the motion space description field. For trajectory segment-level actions, the system selects edges that satisfy basic safety and reachability constraints from the edges related to the current state based on the edge set in the scene graph state structure and the GNN encoding results. These edges are then added to the trajectory candidate set, and several priority layers are defined according to the cost distribution. At the end of the action space partitioning, the system establishes a mask field for each type of action to mark actions that are not selectable at the current time step, such as actions that conflict with cooperative constraints, actions that trigger high-risk collisions, or actions that exceed the robot arm's workspace. Action space encoding is then appended to the reinforcement learning state, enabling the policy network to automatically skip masked actions when generating action probabilities.

[0088] After completing the reinforcement learning state construction and action space partitioning, this implementation enters the policy network training phase. The policy network is the core decision-making module in the RL framework, used to output action sampling probabilities or directly output action selection results based on the current state and action space distribution. The system first constructs a virtual interaction process during the training phase, using the graph-encoded state representation structure and multi-objective path planning cost function as the basis for environment evaluation. The path evaluation module simulates the new state, immediate cost, and task progress updates obtained after performing an action from the current state. Each virtual interaction process consists of several time steps. In each time step, the system reads the current state vector from the reinforcement learning state cache, feeds it into the policy network for forward computation to obtain the probability distribution of each action, and then performs action sampling based on this distribution or selects actions using a greedy sampling strategy. Subsequently, the system calls the environment evaluation module to update the states of the robotic arm nodes and task nodes in the scene graph state structure based on the selected action, and re-triggers GNN encoding to generate a new graph-encoded state representation, thereby obtaining the reinforcement learning state for the next time step. The entire process loops until the end of the current task sequence or the set step size limit is reached.

[0089] During the aforementioned interaction, the system synchronously constructs experience samples, recording the state vector, action selection, comprehensive cost, and task completion marker for each time step in the experience cache. The comprehensive cost is obtained through the multi-objective path planning cost function and weight parameters, with time cost, energy cost, collision risk, and collaborative constraint terms each holding a fixed weight, forming the reward signal in RL. Depending on the selected RL algorithm, the system employs different loss function construction methods to update the parameters of the policy network using gradients. During training, if a certain type of action or state is detected to have a low frequency in the experience set, the system can increase its access probability by adjusting the sampling strategy or introducing exploration noise, while simultaneously recording key parameter changes and training round information during the adjustment process for subsequent analysis and retraining. After multiple training iterations, the policy network converges to a stable parameter set, enabling it to output a consistent action distribution that conforms to multi-objective cost preferences when faced with the same graph-encoded state representation. At this point, the system marks the current policy network parameter version as an online version suitable for candidate path search.

[0090] After the policy network training reaches the set conditions, the system initiates candidate path search processing. By running the trained policy network within the graph-encoded state representation space, several candidate paths are generated for each robotic arm. During the candidate path search process, the system starts from the graph-encoded state representation structure corresponding to the current actual operation scenario, using it as the initial state. It then calls the reinforcement learning state construction module and action space partitioning module to generate the state vector and action space description for the initial time step. Subsequently, the system adopts a parallel unfolding approach for multiple trajectories, running the policy network forward computation and action sampling several times under the same initial state. Each unfolded trajectory corresponds to one candidate path construction process. At each time step, the system selects actions based on the action probability distribution output by the policy network, updates the scene graph state structure and graph-encoded state representation through the environment evaluation module, accumulates the path cost according to the multi-objective cost function, and records the traversed node sequence, edge sequence, and action sequence. When a trajectory meets the task completion conditions or reaches the maximum step limit, the node sequence and edge sequence corresponding to that trajectory are solidified into a candidate multi-robotic arm path. The path contains the target nodes, selected edges, and estimated costs for each robotic arm at each key time point.

[0091] To improve the applicability of the candidate path set for subsequent simulation evaluation and optimization, the system performs filtering and deduplication on the generated path set after the candidate path search is completed. Specifically, several paths with lower costs but significant differences in path structure are selected from all candidate paths as primary candidate paths. For paths with highly similar structures and small cost differences, the system can group them into the same cluster, retaining only representative paths within each cluster. During the filtering process, the system considers task completion status, time cost, energy cost, and collision risk level distribution, retaining path samples that exhibit different trade-offs in these indicators according to a preset strategy. Finally, all paths that pass the filtering are organized into a candidate multi-robot path set, with each path having a unique identifier in the set, and its path structure, cost vector, task coverage, and key time node information are recorded. At the end of step S3, the candidate multi-robot path set is written into the output data structure. The field name in this structure is the candidate multi-robot path set, which will be used as input in the subsequent step S4 to perform path simulation execution, collision detection, path selection and trajectory smoothing in the digital twin warehouse environment model, thereby completing the cross-step connection from graph encoding, strategy generation to physical simulation.

[0092] The technical effects of this step can be summarized as follows: Step S3 constructs reinforcement learning states based on graph neural network encoding, divides the action space of multiple robotic arms, trains the policy network and performs candidate path search to form a set of candidate multi-robotic arm paths. This provides a structured path candidate set for subsequent simulation evaluation and online replanning process, and achieves tight coupling between the path planning search strategy and graph encoding representation under the constraint of multi-objective cost function.

[0093] To address the problems in existing technologies where path planning results are mostly verified offline in simplified simulation environments, resulting in a disconnect from actual control commands, and where operational process data is difficult to use for reverse optimization, this invention utilizes a digital twin model for closed-loop simulation and optimization in step S4 to ensure path feasibility and control effectiveness. Specifically, this includes:

[0094] Step S4: Based on the candidate multi-robot path set, perform path simulation execution, collision detection, path selection, trajectory smoothing and local replanning trigger configuration processing to generate control commands and operation process data structure;

[0095] In this embodiment, the input source for step S4 is the candidate multi-robot path set output from the preceding step S3 and the digital twin warehouse environment model pre-built and continuously updated by the system. The candidate multi-robot path set records multiple path schemes searched in the graph-encoded state representation space by a reinforcement learning policy network for the current task scenario. Each path scheme includes a path identifier, the robot arm numbers involved, a sequence of key nodes, a sequence of edges, a multi-objective cost vector, and task coverage information. The digital twin warehouse environment model is a virtual replica of the real warehouse scenario, including warehouse geometry, shelf layout, ground attributes, pallet parking areas, rigid body parameters of the robot arm and mobile chassis, joint dynamic parameters, drive constraints, control cycle configuration, and a virtual sensor model. During operation, it is periodically aligned with the scene graph state structure and real-time sensing data. When step S4 begins, the system first reads the candidate multi-robotic arm path set and the digital twin warehouse environment model from the storage area. By comparing the node status and task progress recorded in the current scene graph state structure, the system filters out the path identifiers in the path set that are irrelevant to the current task round or have expired. The remaining paths are marked as simulation evaluation paths, and these paths are bound to the digital twin warehouse environment model to form the input basis for path simulation execution.

[0096] Specifically, to perform path simulation, the system performs a simulation initialization configuration on the digital twin warehouse environment model. This initialization configuration includes reading the current joint angles of each robotic arm, the attitude of the end effector, the grasping state, the pose of the moving chassis, and the latest states of the storage location nodes, pallet nodes, and obstacle nodes from the scene graph state structure. These states are then written into the corresponding virtual entities in the digital twin warehouse environment model, ensuring that the simulation environment remains consistent with the actual environment at the current moment. Subsequently, the system discretizes the key node sequence and edge sequence of each candidate path into several simulation time points according to the simulation time step setting. At each simulation time point, a reference pose or reference joint state is generated for the corresponding robotic arm and moving chassis, and then submitted to the simulation solver for kinematic and dynamic calculations. At each time step, the simulation solver updates the joint angles, joint velocities, and joint torques of the virtual robotic arm, as well as the velocity and acceleration of the moving chassis, while simultaneously recording the virtual sensor outputs for subsequent collision detection and trajectory evaluation. During the simulation, if the virtual robotic arm or mobile chassis is detected to enter the preset soft limit area in the model, such as approaching the mechanical limit of the joint or the boundary of the safety fence, the system will add a corresponding mark to the simulation record to prompt the subsequent path selection process to consider the risk of the action approaching the boundary.

[0097] While the path simulation is running, the system operates a collision detection module within the digital twin warehouse environment model, performing time- and space-based collision detection for each candidate path. The collision detection module first obtains the robotic arm linkage mesh model, the polyhedral models of shelves and obstacles, ground height information, and pallet geometry from the digital twin warehouse environment model. Then, based on the virtual joint states and virtual poses generated during the simulation execution phase, it updates bounding boxes and determines spatial overlap for all relevant entities at each simulation time step. For each candidate path, the system sequentially checks along the time axis for geometric overlap between the robotic arm and obstacles, interference between robotic arms, and unreasonable interweaving between the robotic arm and the moving chassis, pallet, or goods. It records the event time, involved entity identifiers, closest distance, contact depth, and relative speed level for each detected contact or proximity event. Furthermore, the system, combining the collision risk term configuration in the multi-objective path planning cost function, converts the number, severity, and duration of collision events detected in the simulation into a path collision score field, writing it into the corresponding path record item in the simulation evaluation result data, providing a quantitative basis for subsequent path selection.

[0098] Building upon collision detection, this implementation method aggregates various statistical data generated during simulation to generate simulation evaluation result data. The simulation evaluation result data is a comprehensive record of the execution status of each candidate multi-robotic arm path in the digital twin warehouse environment model, including fields such as path identifier, task completion marker, simulation end time, cumulative time consumption, estimated energy consumption, collision event statistics, proximity to safety boundary statistics, joint saturation statistics, and anomaly flags encountered during execution. The system iterates through the simulation log, summarizing the motion quantities and sensor outputs recorded at each time step, accumulating and integrating time consumption and energy consumption, classifying and statistically analyzing collision and proximity events, and synthesizing a multi-dimensional evaluation vector according to preset rules, storing it in the record item corresponding to the path identifier in the simulation evaluation result data. Understandably, the simulation evaluation result data is considered the fundamental data source for candidate path quality evaluation in this step, and also an important basis for judging whether the path performance meets expectations during subsequent local replanning trigger configuration.

[0099] After obtaining the simulation evaluation results data, the system performs path filtering. The goal of path filtering is to eliminate paths from the candidate multi-robot path set whose simulation performance clearly does not meet the operational requirements, retaining a representative subset of paths in terms of multi-objective cost and balance. Specifically, the system first performs a preliminary safety screening of all paths based on the collision event statistics field recorded in the simulation evaluation results data, eliminating paths with serious collision events or multiple intersections with obstacles, as well as paths whose joint saturation time exceeds a threshold. Subsequently, the system sorts and clusters the remaining paths based on the cumulative time consumption field and the estimated energy consumption field, grouping paths with similar performance in time and energy consumption into the same cluster. From each cluster, several paths with better performance in terms of collision risk, waiting time, and compliance with cooperative constraints are selected as representative paths to form the filtered path set. Furthermore, for paths belonging to multi-robotic arm collaborative tasks, the system will also check the time alignment, spacing, and task coverage order of key actions of multiple robotic arms in the path, eliminate paths with obvious collaborative action conflicts, and record the collaborative consistency level of each retained path in the screening results for reference when configuring subsequent trajectory smoothing and local replanning triggers.

[0100] After obtaining the filtered path set, this implementation method performs trajectory smoothing processing on each retained path. Trajectory smoothing refers to reconstructing and resampling the trajectory of the robotic arm joints and the mobile chassis at each discrete time point in the path while maintaining the path topology and task sequence. This makes the trajectory more continuous in terms of velocity and acceleration changes, reducing mechanical structural impact and abrupt changes in control commands. The system first reads the key node sequence and edge sequence of each path in the filtered path set, converts them into the corresponding joint space trajectory of the robotic arm and the pose trajectory of the mobile chassis, and inserts additional intermediate points on the time axis. The trajectory curve is reconstructed through spline interpolation or piecewise polynomial fitting. During the reconstruction process, the system sets upper limits for velocity and acceleration for each joint and each translational movement. If the fitted trajectory exceeds the upper limit in a certain time period, the system automatically extends the length of that time period or adjusts the position of the control points so that the trajectory meets the constraints again. For road segments involving critical actions such as grasping and placing, the system retains the key attitude points unchanged during trajectory smoothing and only smooths the transition segments before and after them to maintain the accuracy of the task operation posture. After trajectory smoothing is completed, the joint trajectory and pose trajectory corresponding to each path will be re-discretized into a time series consistent with the control cycle. At the same time, the reference state and control mode identifier of each time step will be recorded as the direct basis for subsequent generation of control commands.

[0101] After trajectory smoothing, the system configures local replanning triggers based on simulation evaluation data and the characteristics of the filtered paths. Local replanning trigger configuration refers to pre-setting a set of trigger conditions for each path to identify situations requiring local path adjustments during online runtime, ensuring dynamic linkage between the planning process and actual execution. The system first extracts indicators such as collision approach distance statistics, time buffer margin, energy consumption fluctuations, and cooperative action sensitivity for each path in key road segments from the simulation evaluation data. These indicators are then compared and analyzed with the environmental change frequency, equipment failure rate, and sensor noise level recorded in the scene graph state structure to form a risk level classification for different road segments. Subsequently, the system configures several local replanning trigger thresholds for each key road segment based on the risk level. Examples include trajectory tracking error exceeding a certain range, the actual distance to an obstacle being reduced by a certain proportion compared to the simulation-estimated distance, task execution time delay exceeding the time buffer, and the cooperative robotic arm spacing being lower than the cooperative safety distance. When the actual observed state during online runtime meets one of these thresholds, the local replanning module can regenerate candidate paths for that road segment by calling the graph neural network encoding configuration structure and graph encoding state representation from the preceding steps according to preset logic. Therefore, in this embodiment, the relevant parameters for local replanning trigger configuration are written into the path description structure and summarized into the corresponding fields in the final control command and operation process data structure.

[0102] After path simulation execution, collision detection, path selection, trajectory smoothing, and local replanning trigger configuration are completed, the system generates control commands and operation process data structures according to the interface specifications of the multi-manipulator control system. The control commands and operation process data structures are the output field names of this step, used to pass the planning results and operation configuration to the lower-level control system and the upper-level scheduling and learning module. Specifically, the system generates a sequence of control frames arranged in chronological order for each retained path and each manipulator. Each control frame includes the target joint angle or target end-effector pose, velocity and acceleration reference values, end-effector action identifier, and control mode flag, with a path identifier and timestamp appended before the control frame, allowing the lower-level controller to execute frame by frame according to the timeline. Simultaneously, the system also generates corresponding trajectory control frames for the mobile chassis, including position, velocity, and steering commands. The operation process data section records the simulation evaluation result summary, multi-objective cost vector, local replanning trigger threshold set, path cooperative attribute identifier, and mapping relationship with the scene graph state structure obtained in the simulation and selection phases. This allows the upper-level module to monitor the path execution status during execution and feed back data for offline learning. Understandably, the control commands and operation process data structure are sent to the actual robotic arm and mobile chassis controller for execution, and written back to the scene graph state structure as the input basis for updating the basic data of the robotic arm operation scene and constructing the cost function in the next round of steps S1 and S2, thus forming a closed-loop link from planning to simulation to execution and replanning in the overall method.

[0103] The technical effects of this step can be summarized as follows: Step S4 uses a digital twin warehouse environment model to simulate and perform collision detection on the candidate multi-robotic arm path set, combines multi-dimensional evaluation results to complete path selection and trajectory smoothing, configures local replanning trigger conditions, and finally generates control commands and operation process data structures, so that the path planning results are consistent with the actual control and subsequent planning processes at the data and process level.

[0104] Example 2: Figure 2 A structural block diagram of a multi-objective motion path planning system for a robotic arm according to an embodiment of the present invention is shown. Figure 2 As shown, the structure may include:

[0105] The scene graph state structure construction module 01 is used to acquire the robotic arm structural parameters, cargo location, pallet location, and obstacle contour data based on the basic data of the robotic arm operation scene. It performs coordinate system calibration, topology relationship parsing, node attribute and edge relationship configuration, sensor channel binding, and state field initialization, outputting the scene graph state structure for downstream modules to call. Specifically, the scene graph state structure construction module receives the basic data of the robotic arm operation scene from the sensor acquisition layer and the business configuration database, importing the original records of the robotic arm structural parameters, cargo location and pallet layout information, site boundary description, obstacle contour point cloud data, and environmental marker data into its internal cache. The module performs format conversion and coordinate unit unification on the above data through the interface adaptation unit. It establishes an index table for the robotic arm, cargo location, pallet, and obstacles based on the equipment identification index, calls the geometric solution process to generate transformation parameters between the workstation coordinate system, the robotic arm base coordinate system, and the environmental world coordinate system, and writes these parameters into the coordinate system calibration field. The scene graph state structure construction module generates a set of nodes and a set of edges based on the center positions of the cargo locations and pallets, the outer contours of obstacles, and preset passability rules, through adjacency determination. For each node, it writes a type flag, geometric parameters, and a region flag; for each edge, it writes a connection endpoint flag, passability attribute, and an initial cost placeholder field. According to the sensor configuration table, the module establishes binding relationships between the visual acquisition device, distance measurement device, and attitude detection device and the corresponding robotic arm node, cargo location node, pallet node, or obstacle node, writing the binding results into the sensor channel binding field. During the state field initialization phase, it sets a default state, timestamp placeholder, and data quality flag for nodes that have not yet obtained real-time observation data. After completing the above processing, the scene graph state structure construction module encapsulates the node set, edge set, node attributes, edge relationships, sensor channel binding fields, and state fields into a scene graph state structure, and sends it to the multi-objective path planning cost function configuration module, graph neural network encoding configuration module, and graph encoding and reinforcement learning state construction module through the system bus. At the same time, it writes the scene graph state structure summary into persistent storage for the control command and operation process data generation module to index and associate when writing back the operation process data structure.

[0106] The multi-objective path planning cost function configuration module 02 is used to obtain multiple cost items from the scene graph state structure and operation cycle requirements, construct a multi-objective path planning cost function, normalize the cost items and set weight parameters, and output a multi-objective path planning cost function object for the graph neural network encoding configuration module to call. Specifically, the multi-objective path planning cost function configuration module reads the geometric description, topological connection relationship, sensor channel binding field and current state field of the robotic arm node, storage location node, pallet node and obstacle node from the scene graph state structure, receives the operation cycle requirements from the scheduling management unit, completes the alignment processing of the task time window and node state timestamp on a unified time axis, and obtains the set of participating nodes and task set for the current planning round. When traversing the participating set, the module generates several candidate trajectory segments through a kinematic solution process based on the robotic arm structural parameters, current joint state and target node posture. It calculates the time consumption field for each candidate trajectory segment, adds the end effector action duration to the trajectory segment containing grasping or placing actions, and writes the result into the time cost item of the corresponding connection. The multi-objective path planning cost function configuration module calls the resource consumption estimation sub-process, reads the robotic arm load field, trajectory length field, and operation mode field from the equipment management unit, obtains the estimated resource consumption value by combining it with the historical statistics table, and writes the value into the resource-type cost item field corresponding to the connection edge, which is used as the resource consumption-related part in the multi-objective path planning cost function. The module retrieves environmental observation data from the real-time observation cache based on the obstacle contour field, robotic arm workspace partitioning field, and sensor channel binding field in the scene graph state structure, updates the obstacle position and shape, arranges several attitude sampling points on the candidate trajectory segment, calculates the minimum distance between the robotic arm link and the obstacle through the collision detection process, and generates a collision risk cost item field according to the preset risk classification rules. The multi-objective path planning cost function configuration module parses the task dependency description in the operation cycle requirement, combines the task occupancy marker and node type field in the scene graph state structure, establishes sequence constraints, continuous operation constraints of the same robotic arm, and collaborative operation time alignment constraints for tasks sharing a storage location or pallet, forms a collaborative constraint cost item field, and writes it into the corresponding connection edge record. After the above cost items are written, the module creates a cost record unit for each participating edge, stores the time cost item, resource cost item, collision risk cost item, and collaborative constraint cost item into the unit, performs statistical calculations on all cost record units to obtain the numerical range and distribution characteristics of various cost items, removes abnormal records and performs normalization processing, writes the normalization result into the normalized cost field, and determines the weight parameters of various normalized cost items based on the priority strategy configuration in the operation cycle requirements, equipment maintenance status, and historical planning experience, and writes the weight parameters into the multi-objective path planning cost function object.The multi-objective path planning cost function configuration module registers the identifier field of the multi-objective path planning cost function object in the scene graph state structure metadata area. At the same time, it outputs the multi-objective path planning cost function object and related statistical parameters to the graph neural network encoding configuration module through an internal interface, providing a basis for subsequent graph neural network input channel partitioning and field mapping.

[0107] The graph neural network encoding configuration module 03 is used to divide node feature channels, edge feature channels, and global feature channels according to the aforementioned multi-objective path planning cost function object and the scene graph state structure, complete the mapping between fields and channels, and generate a graph neural network encoding configuration structure, providing encoding configuration to the graph encoding and reinforcement learning state construction module. Specifically, the graph neural network encoding configuration module receives the multi-objective path planning cost function object from the multi-objective path planning cost function configuration module and the scene graph state structure from the scene graph state structure construction module, and describes the dimensional division scheme of node feature channels, edge feature channels, and global feature channels according to the preset graph neural network structure configuration. In the node feature channel construction stage, the module writes structural parameter fields, current state fields, task occupancy markers, and cost statistics fields related to the node to the robotic arm node; writes spatial location fields, storage attribute fields, and task requirement summary fields to the storage location node and pallet node; and writes contour fields, safety buffer distance fields, and update timestamp fields to the obstacle node, forming a set of node feature vectors. In the edge feature channel construction phase, the graph neural network encoding configuration module traverses the set of edges, combining the geometric description fields, connection endpoint markers, traversability attributes, normalized cost fields, and weight parameter fields of each edge into a set of edge feature vectors. It also adds the corresponding robotic arm marker and task marker to each edge, completing the registration of the correspondence between edge features and edge indices. In the global feature channel construction phase, the module extracts summary fields such as total task volume, overall time window, resource pressure level, and safety policy level from the job cycle requirements, combines them with the global statistical fields in the multi-objective path planning cost function object to form a global feature vector, and registers the mapping relationship between it and the current planning round identifier. The graph neural network encoding configuration module sets masking strategies for node feature channels, edge feature channels, and global feature channels, writing uniform placeholder values ​​for missing or invalid fields, and recording the corresponding node or edge number in the internal mask table. This allows the graph encoding and reinforcement learning state construction module to handle missing information based on the mask table when using the graph neural network encoding configuration structure. After completing the above operations, the graph neural network encoding configuration module encapsulates the dimensional descriptions, field mapping rules, and mask tables of the node feature channels, edge feature channels, and global feature channels, along with the multi-objective path planning cost function object identifiers, into a graph neural network encoding configuration structure. This structure is then sent to the graph encoding and reinforcement learning state construction module via the internal message bus. Simultaneously, the graph neural network encoding configuration structure identifier is registered in the scene graph state structure metadata area for reuse in subsequent planning rounds.

[0108] The graph coding and reinforcement learning state construction module 04 is used to obtain a graph-coded state representation based on the graph neural network coding configuration structure and the scene graph state structure, construct a reinforcement learning state, and generate a candidate multi-robot path set by combining action space partitioning and policy network training. The candidate multi-robot path set is then sent to the digital twin warehouse environment simulation and evaluation module. Specifically, the graph coding and reinforcement learning state construction module receives the graph neural network coding configuration structure from the graph neural network coding configuration module and the scene graph state structure from the scene graph state structure construction module. It assembles node feature tensors, edge feature tensors, and global feature tensors according to the field mapping rules specified in the coding configuration structure, and calls the graph neural network forward computation process to generate a graph-coded state representation. The module constructs a reinforcement learning state based on the node encoding, edge encoding, and global encoding in the graph-coded state representation. It selects and aggregates the robot node, task node, and key edge encoding involved in the current decision moment, and concatenates the remaining task time, robot load level, and historical action summary fields into the state vector to generate a reinforcement learning state description for single-step decision-making. The graph coding and reinforcement learning state construction module divides the action space based on the scene graph state structure and cooperative constraint cost terms. It includes task selection actions, trajectory segment selection actions, and cooperative action adjustment actions in the action set. Actions that do not meet safety or cooperative constraints are masked, and the action mask information is carried in the state description for the policy network to reference during the training and inference phases. During the policy network training phase, the module iteratively reads the reinforcement learning state description, calls the policy network to generate action distributions, samples actions from the action set, and drives the environment evaluation process to update the scene graph state structure and graph coding state representation. Each step's state, action, cost vector, and task completion marker are written to the experience cache, providing training samples for the policy update phase. When the policy network training converges to the preset conditions, the graph coding and reinforcement learning state construction module performs candidate path search based on the current policy network parameters and graph coding state representation. Multiple decision trajectories are developed under the same initial state. For each trajectory, the robot arm node access order, edge sequence, and cumulative cost vector are recorded, forming a candidate multi-robot path set and registering the set identifier. During the output phase, the module sends the candidate multi-robotic arm path set to the digital twin warehouse environment simulation and evaluation module. At the same time, it writes the candidate multi-robotic arm path set identifier and key statistical summary into the scene graph state structure, which supports the control command and operation process data generation module to track the path source during the execution phase.

[0109] The digital twin warehouse environment simulation and evaluation module 05 is used to perform path simulation, collision detection, and path selection based on the candidate multi-robot path set and the digital twin warehouse environment model. It performs trajectory smoothing on the selected paths and generates simulation evaluation result data, outputting path-level evaluation information to the control command and operation process data generation module. Specifically, the digital twin warehouse environment simulation and evaluation module receives the candidate multi-robot path set from the graph coding and reinforcement learning state construction module, loads the digital twin warehouse environment model from the environment modeling unit, and synchronizes the initial states of the virtual robot arm, storage location, pallet, obstacles, and mobile chassis according to the current node state in the scene graph state structure, ensuring that the initial state of the simulation environment is consistent with the actual operation state. The module expands each path into several simulation time steps according to the node sequence and edge sequence recorded in the candidate multi-robot path set. For each time step, it calls the motion solving unit to update the joint angles, end-effector posture, and mobile chassis pose of the virtual robot arm, and drives the virtual sensor model to generate collision-related observation data and state monitoring data. The digital twin warehouse environment simulation and evaluation module runs collision detection logic during the simulation process. Based on the virtual robotic arm linkage model, shelf and obstacle geometry models, and pallet contour model, it performs spatial overlap determination at each time step, recording contact and proximity events between the robotic arm and obstacles, between robotic arms, and between the robotic arm and pallets. The timing of each event, proximity distance, and contact depth are written to the collision event log. After completing the simulation of the entire path, the module performs aggregation operations on the time consumption fields, resource consumption fields, collision event logs, and proximity event statistics recorded at each time step, forming a simulation evaluation result data record for each path. This record includes fields such as task completion marker, cumulative time consumption, estimated resource consumption, number of collision events, and severity level. Based on the simulation evaluation result data, the module executes path filtering logic, eliminating candidate paths with severe collision events or multiple path segments approaching the safety boundary. The remaining paths are sorted and clustered according to time consumption and resource consumption. Several representative paths are selected from each cluster to form a set of filtered paths, which are then processed by the trajectory smoothing process. During the trajectory smoothing phase, the module interpolates and resamples the joint trajectories and moving chassis trajectories of each path in the selected paths, performs smoothing processing on the velocity and acceleration curves, re-discretes the trajectory point sequence over the control cycle, and writes the updated trajectory into the smoothed path set, which serves as the trajectory input source for the control command and operation process data generation module. After the simulation, the digital twin warehouse environment simulation evaluation module organizes the simulation evaluation results data for each path and the corresponding path identifier into path-level evaluation information, which is sent to the control command and operation process data generation module through the inter-module interface. At the same time, the main evaluation summary is written into the scene graph state structure to provide a reference for subsequent planning rounds.

[0110] The control command and operation process data generation module 06 is used to generate control command and operation process data structures based on the simulation evaluation results data and the filtered and smoothed paths. It then sends control commands to the robotic arm and mobile chassis controllers and writes the operation process data structures back to the scene graph state structure for subsequent path planning rounds to update the basic data. Specifically, the control command and operation process data generation module receives path-level evaluation information from the digital twin warehouse environment simulation evaluation module and a smoothed path set from the trajectory smoothing process. Based on the task completion markers, time consumption, and resource consumption statistics recorded in the evaluation information, it selects a target execution path for each path, rearranges and segments the node sequence and trajectory point sequence in the path, and obtains a single-device trajectory segment sequence for each robotic arm and mobile chassis. The module generates continuous control frames for each single-device trajectory segment according to the control cycle. Each control frame contains the target joint angle or end effector pose, velocity reference value, acceleration reference value, and end effector action flag, along with path identifier, timestamp, and device identifier. These continuous control frames are combined into a control command data stream, which is sent to the corresponding robotic arm controller and mobile chassis controller via fieldbus to drive the operation. The control command and operation process data generation module simultaneously constructs the operation process data structure during control frame generation. It extracts action sequences, device status records, and path numbers from the control command data stream and collects execution status, anomaly flags, and alarm records from the underlying controller feedback. The two sets of data are aligned chronologically to form an operation process data structure covering task, device, and time dimensions. The module writes the control command transmission status and operation process data structure to log storage and a historical database. Node identifiers, edge identifiers, and planning cycle identifiers from the scene graph state structure are attached to the operation process data structure, allowing subsequent planning cycles to directly reference the actual execution records when updating the scene graph state structure and multi-objective path planning cost function object. In terms of closed-loop planning, the control command and operation process data generation module extracts the statistical results related to scene updates from the operation process data structure and writes them back to the scene graph state structure construction module and the multi-objective path planning cost function configuration module through the internal interface. The module updates the node state field, cost statistics field and weight adjustment reference field, so that the next round of multi-objective motion path planning of the robotic arm is based on the latest operation process data.

Claims

1. A method for multi-objective motion path planning for a robotic arm, characterized in that, include: Acquire the robotic arm's structural parameters, cargo location, pallet location, and obstacle contour data; perform coordinate system calibration, topology analysis, node attribute and edge relationship configuration, sensor channel binding, and state field initialization to generate a scene graph state structure. Based on the scene graph state structure, the cost function for multi-objective path planning is constructed, the cost term is normalized, the weight parameters are set, and the graph neural network input channel is mapped to generate the graph neural network encoding configuration structure. Based on the graph neural network encoding configuration structure, reinforcement learning state construction, action space partitioning, policy network training and candidate path search processing are performed to generate a candidate set of multi-robot paths. Based on the candidate set of multiple robotic arm paths, path simulation execution, collision detection, path selection, trajectory smoothing, and local replanning trigger configuration processing are performed to generate control commands and operation process data structures.

2. The method according to claim 1, characterized in that, The robotic arm's structural parameters, cargo location, pallet location, and obstacle contour data include: Robotic arm structural parameters refer to a set of parameters that describe the geometry and kinematic boundaries of each robotic arm, including the initial posture of each joint, the range of joint rotation, the length of the link, the posture and installation position of the end effector, and the installation position of the robotic arm base in the field. The location of a storage location refers to the coordinates of the center point, orientation information, and size range of each storage unit on the shelf in the global coordinate system of the site. Pallet location refers to the pallet's external dimensions, center coordinates of the bearing surface, and boundaries of the permitted parking area used for stacking or transporting materials; Obstacle contour data refers to the external boundary, multi-segment polyline description, and height range information of fixed equipment, temporary storage objects, or safety isolation facilities that affect the movement trajectory of the robotic arm and mobile chassis.

3. The method according to claim 1, characterized in that, The process of coordinate system calibration and topological relation resolution also includes: Coordinate system calibration involves transforming the robot arm's body coordinate system, sensor mounting coordinate system, local coordinate system of the storage location, and local coordinate system of the pallet to the global coordinate system of the site. Topological relation resolution involves identifying reachability, interference, and constraint relationships between nodes based on geometric data in a unified coordinate system to form graph structure connection information.

4. The method according to claim 1, characterized in that, The process of configuring node attributes and edge relationships, binding sensor channels, and initializing state fields also includes: The node attribute configuration includes creating attribute field sets for robotic arm nodes, storage location nodes, pallet nodes, and obstacle nodes respectively; Edge relationship configuration includes setting direction attributes, reachability tags, and initial values ​​for association costs for connected edges; Sensor channel binding involves establishing a mapping relationship between visual sensors, LiDAR sensors, depth cameras, and environmental perception devices and relevant nodes or edges in the scene graph, and recording the perception channel number. State field initialization involves creating fields for each node and edge to describe the current and historical states and writing initial values ​​based on sensor observations.

5. The method according to claim 1, characterized in that, The process of constructing the cost function for multi-objective path planning, normalizing the cost term, setting weight parameters, and mapping the input channels of the graph neural network also includes: The construction of the cost function for multi-objective path planning includes creating time cost terms, energy cost terms, collision risk terms, and collaborative constraint terms for connecting edges; Cost term normalization involves mapping the original cost terms of time cost, energy cost, collision risk, and collaborative constraint to dimensionless values ​​within a specific range. The weighting parameter settings include configuring weighting coefficients for various normalized cost items based on task priority strategies and equipment health status, according to job cycle requirements. Input channel mapping involves mapping node attributes, edge attributes, and global features to the node feature channels, edge feature channels, and global feature channels of the graph neural network.

6. The method according to claim 1, characterized in that, The process of constructing states and dividing the action space in reinforcement learning also includes: State construction includes constructing reinforcement learning state vectors based on graph-encoded state representations generated by graph neural network encoding; action space partitioning includes dividing the action space into task-level actions and trajectory segment-level actions and generating a set of candidate actions for each robotic arm.

7. The method according to claim 1, characterized in that, The process of training the policy network and searching candidate paths also includes: Policy network training involves updating network parameters during virtual interaction by outputting the action probability distribution based on the state vector and accumulating the reward based on the multi-objective cost function; candidate path search involves generating candidate paths through the trained policy network by expanding multiple trajectories in parallel.

8. The method according to claim 1, characterized in that, The process of performing path simulation execution, collision detection, and path selection also includes: Path simulation execution includes discretizing path sequences in a digital twin warehouse environment model and performing kinematic and dynamic calculations; collision detection includes checking the geometric overlap between the robotic arm and obstacles and the mutual interference between robotic arms at each simulation time step; path filtering includes eliminating paths that do not meet the requirements and retaining a representative subset of paths based on collision event statistics, cumulative time consumption, and estimated energy consumption in the simulation evaluation results data.

9. The method according to claim 1, characterized in that, The process of trajectory smoothing and local replanning-triggered configuration processing also includes: Trajectory smoothing involves reconstructing the trajectory curve through spline interpolation or piecewise polynomial fitting to meet velocity and acceleration constraints; local replanning trigger configuration involves setting trigger thresholds for critical road segments based on trajectory tracking error, obstacle distance, and time delay.

10. A multi-objective motion path planning system for a robotic arm, applied to the method described in any one of claims 1-9, characterized in that, include: The scene graph state structure construction module is used to obtain the robot arm structural parameters, cargo location, pallet location and obstacle contour data based on the basic data of the robot arm operation scene, perform coordinate system calibration, topology relationship parsing, node attribute and edge relationship configuration, sensor channel binding and state field initialization processing, and output the scene graph state structure for downstream modules to call. The multi-objective path planning cost function configuration module is used to obtain multiple cost items from the scene graph state structure and operation cycle requirements, construct the multi-objective path planning cost function, normalize the cost items and set weight parameters, and output the multi-objective path planning cost function object for the graph neural network encoding configuration module to call. The graph neural network encoding configuration module is used to divide node feature channels, edge feature channels and global feature channels according to the aforementioned multi-objective path planning cost function object and scene graph state structure, complete the mapping between fields and channels and generate the graph neural network encoding configuration structure, and provide encoding configuration to the graph encoding and reinforcement learning state construction module. The graph coding and reinforcement learning state construction module is used to obtain graph coding state representations based on the graph neural network coding configuration structure and scene graph state structure, construct reinforcement learning states, and generate a candidate multi-robot path set by combining action space partitioning and policy network training. The candidate multi-robot path set is then sent to the digital twin warehouse environment simulation and evaluation module. The digital twin warehouse environment simulation and evaluation module is used to perform path simulation, collision detection and path selection based on the candidate multi-robotic arm path set and the digital twin warehouse environment model. It performs trajectory smoothing on the selected paths and generates simulation evaluation result data, and outputs path-level evaluation information to the control command and operation process data generation module. The control command and operation process data generation module is used to generate control commands and operation process data structures based on simulation evaluation results and filtered, smooth paths. It sends control commands to the robotic arm and mobile chassis controller, and writes the operation process data structure back to the scene graph state structure for subsequent path planning cycles to update the basic data.

Images