An artificial intelligence-based multi-robot process scheduling and collaborative control method

Abstract

This invention discloses an artificial intelligence-based method for multi-robot process scheduling and collaborative control, comprising: process structure and robot resource modeling to generate a unified task information queue; environmental perception and multi-source point cloud acquisition to generate multi-source point cloud data; Poisson surface reconstruction to obtain 3D reconstruction results and an occupied grid map; selecting projection viewpoint and grid resolution to generate a multi-channel 2D environmental grid map; constructing an improved HarDNet network to generate environmental semantics and risk maps; task allocation and path planning to form a collaborative control plan; and collaborative execution, controlling the physical robot to perform tasks and controlling the software robot to handle business. This invention achieves unified scheduling and collaborative control of multi-robot tasks by combining Poisson surface reconstruction and an improved HarDNet network.

Description

Technical Field

[0001] This invention relates to the field of robot intelligent scheduling technology, and in particular to a multi-robot process scheduling and collaborative control method based on artificial intelligence. Background Technology

[0002] In existing technologies, enterprises commonly employ business process management systems (BPMs) to model and execute cross-departmental and cross-system business processes in manufacturing, warehousing, logistics, and security inspection scenarios. They also introduce various types of robotic equipment, including mobile robots, robotic arms, inspection robots, and software robots. However, these robots are typically provided by different manufacturers, resulting in differences in control protocols, scheduling systems, and status feedback methods. The integration of BPMs with these heterogeneous robots often relies on point-to-point interfaces or customized integration methods, lacking a unified task description and resource modeling mechanism. This leads to unclear mapping relationships between process nodes and robot tasks, complex system integration, and difficulties in expansion. When business processes or robot resources change, extensive manual modifications and secondary development are required, making it difficult to achieve unified modeling and collaborative scheduling of heterogeneous robots.

[0003] In multi-robot collaborative operations, existing technologies often employ single-source or limited sensor mapping to construct environmental maps. A common approach is to use a single or a small number of robots to collect point cloud data, generating coarse-grained occupancy grids and topology maps for path planning and obstacle avoidance decisions. Environmental modeling and semantic analysis are typically independent of business processes and task types, lacking comprehensive utilization of historical multi-robot operational behavior, task distribution patterns, and risk event distribution. This makes it difficult to perform refined modeling and dynamic risk assessment for high-frequency operational areas, critical passages, and workstations. Multi-robot task allocation and path planning are largely based on static rules and simple heuristic algorithms, failing to unify the modeling and coordinated control of business process states, robot real-time states, and environmental semantic information. There is a lack of closed-loop feedback mechanisms between environmental reconstruction accuracy, traversable area identification results, and scheduling strategies, making it difficult to dynamically optimize and adjust multi-robot tasks and processes in a timely manner based on changes in business needs, robot resources, and the environment. Overall, this approach struggles to meet the comprehensive requirements for collaborative efficiency, safety, flexibility, and resource utilization in complex scenarios.

[0004] Therefore, how to provide a method for scheduling and collaborative control of multi-robot processes based on artificial intelligence is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] One objective of this invention is to propose an artificial intelligence-based method for multi-robot process scheduling and collaborative control. This invention targets business process management platforms and multi-source heterogeneous robot clusters, providing a complete methodological chain from process modeling, robot resource registration, environmental perception and reconstruction, semantic analysis, to task allocation, path planning, and collaborative execution. In environmental modeling, this invention achieves a balance between geometric accuracy and modeling accuracy of point clouds collected by multiple robots by reconstructing Poisson surfaces, introducing multi-source weighted vector field construction and task-driven octree subdivision. In feature extraction and semantic analysis, it improves the HarDNet network by introducing adaptive harmonic connections, task perception channel configuration, and frequency domain feature processing, making environmental semantics and risk maps more aligned with the needs of different task types. Compared with existing technologies, this invention achieves unified modeling and coordinated control of business processes, environmental semantics, and multi-robot states, possessing advantages such as high scheduling accuracy, strong environmental perception and decision-making closure, and good adaptability to complex scenarios and heterogeneous robot clusters, facilitating engineering implementation in enterprise-level multi-robot applications.

[0006] According to an embodiment of the present invention, a multi-robot process scheduling and collaborative control method based on artificial intelligence includes:

[0007] In business process management, establish a process structure that spans business scenarios, unify physical robot task nodes and software robot task nodes as robot task nodes, establish resource description information for various robots and complete robot resource registration, and generate a unified task information queue.

[0008] Collect environmental sensor data based on a unified task information queue to obtain multi-source point cloud data carrying spatial coordinates, normal information and source identifiers;

[0009] Three-dimensional environment reconstruction is performed on multi-source point cloud data by Poisson surface reconstruction. Multi-source weights are assigned to each point cloud data, vector field divergence is calculated, granular subdivision is performed, and three-dimensional surface reconstruction results and occupied grid map are obtained.

[0010] Based on the viewpoint and raster resolution, the 3D surface reconstruction results and the occupied raster map are projected into a multi-channel 2D environment raster map;

[0011] An improved HarDNet network is constructed to extract features and perform semantic analysis on a multi-channel two-dimensional environmental raster map. An adaptive harmonic connection method is used to stitch features together. The features are then subjected to frequency domain transformation, frequency band weighting, and inverse transformation to obtain an environmental semantic and risk map.

[0012] By combining a unified task information queue with environmental semantics and risk maps, multi-robot task allocation and path planning are performed to generate task sequences, running paths and operation sequences for each physical robot, thus forming a collaborative control plan.

[0013] The collaborative control plan is sent to the control platform to control the physical robot to perform handling, inspection or assembly tasks, and to control the software robot to perform data entry or status update operations.

[0014] Optionally, generating a unified task information queue includes:

[0015] In business process management, create process definitions and model business steps across business scenarios as directed process structures, including multiple nodes and directed connections between nodes. Nodes involving physical robot execution and nodes involving software robot execution are uniformly defined as robot task nodes. Each robot task node includes a task identifier, task type, target location identifier, and time constraint parameters. The time constraint parameters include the earliest start time and the latest finish time.

[0016] Generate robot resource description information for various types of robots, including resource identifier, resource type and set of executable task types. Register the robot resource description information in the robot resource library of business process management to complete robot resource registration. Generate a unified task information queue based on robot task nodes. Each task record in the unified task information queue includes task identifier, task type and target location identifier.

[0017] Optionally, the environmental sensing data includes laser point cloud data, depth map data from a depth camera, RGB image data, and the robot's own pose data.

[0018] Optionally, acquiring multi-source point cloud data carrying spatial coordinates, normal information, and source identifiers includes:

[0019] Control multiple physical robots to move according to the target position in a unified task information queue, and collect environmental sensing data using LiDAR, depth camera and vision camera respectively during the movement;

[0020] Based on the pose data, the environmental sensing data collected by different robots are unified into the same spatial coordinate system. The laser point cloud data and depth map data are fused to generate point cloud data carrying spatial coordinates and normal information. A source identifier is added to each laser point cloud data to obtain multi-source point cloud data.

[0021] Optionally, obtaining the three-dimensional surface reconstruction result and the occupied grid map includes:

[0022] For each point cloud data, the sensor type, acquisition distance, incident angle and timestamp parameters are read. According to the weight calculation rules, a corresponding multi-source weight value is generated for each point cloud data. The normal information of each point cloud data is weighted and accumulated with the multi-source weight value. A weighted vector field is constructed on the three-dimensional spatial discrete grid. The divergence value of the weighted vector field is calculated at each node to obtain the corresponding divergence discrete data.

[0023] An octree root node is initialized within the three-dimensional space of multi-source point cloud data. The octree nodes are initially subdivided according to the point cloud distribution. The task importance of each node is calculated based on the historical trajectory and task records of multiple robots. A first threshold and a second threshold are set. Nodes with task importance greater than the first threshold are marked as high-frequency operation nodes, and nodes with task importance less than the second threshold are marked as low-frequency operation nodes. High-frequency operation nodes are further subdivided according to fine-grained rules until the maximum level is reached or the number of point clouds in the node is lower than the first quantity threshold. Low-frequency operation nodes maintain a coarse-grained level or are subdivided a limited number of times only when the number of point clouds in the node exceeds the second quantity threshold, forming a task-driven octree spatial discrete structure.

[0024] Divergence discrete data is introduced at each node of the octree spatial discrete structure to construct the Poisson discrete equation. The Poisson discrete equation is expressed as a sparse linear system of equations and numerically solved to obtain the scalar function values ​​at each node of the octree.

[0025] In the octree spatial discrete structure, a preset isovalue is selected, and the point set whose scalar function value is equal to the preset isovalue is extracted as the 3D surface reconstruction result. The spatial voxels are occupied according to the 3D surface reconstruction result and the scalar function value at the node, and the occupied raster map corresponding to the 3D surface reconstruction result is generated.

[0026] Optionally, projecting the three-dimensional surface reconstruction result and the occupied grid map into a multi-channel two-dimensional environment grid map includes:

[0027] Select a projection reference plane from the 3D spatial coordinate system corresponding to the 3D surface reconstruction result and the occupied raster map, determine the projection view direction and the 2D environment raster coordinate system, set the raster resolution, number of rows and columns and raster boundary range, and map the spatial coordinates of each 3D point to the row index and column index in the 2D environment raster map according to the projection rules.

[0028] For each 2D environment grid cell, point data from the 3D surface reconstruction result mapped to the current grid cell and voxel data from the occupied grid map are collected. The depth value, height value, normal component and occupation probability value of the grid cell are calculated by statistical method. The depth value is the distance statistical result along the projection view direction, the height value is the coordinate statistical result along the vertical direction, the normal component is the component statistical result of normal information, and the occupation probability value is the probability statistical result of the occupation grid voxel state.

[0029] The depth, height, normal components, and occupancy probability values ​​of each two-dimensional environment raster cell are combined in channel order to form a multi-channel two-dimensional environment raster map with row and column indices as position indices and various statistical values ​​as channel data.

[0030] Optionally, obtaining the environmental semantics and risk map includes:

[0031] An improved HarDNet network is constructed, consisting of an input layer, an improved block feature layer, channel units, frequency domain units, and an output layer. Through the input layer, the multi-channel two-dimensional environment raster map is convolved and pooled to obtain an intermediate feature map.

[0032] The improved block feature layer predetermines a set of candidate previous convolutional layers for the current convolutional layer, calculates the feature importance of the output feature maps of the candidate previous convolutional layers, and selects several feature maps from the output of the candidate previous convolutional layers using an adaptive harmonic connection method based on the feature importance. The feature maps are then concatenated along the channel dimension to form the input feature map of the current convolutional layer.

[0033] The channel unit receives the process task type parameters and environment complexity parameters, and allocates the number of channels to the convolutional layers in each improved block in the improved block feature layer according to the mapping relationship. The number of convolutional layer channels is adjusted under different task types and environment complexities to form task-related channel configurations.

[0034] The frequency domain unit performs a frequency domain transformation on the feature map output by the improved block to obtain a frequency domain coefficient map. Frequency band weights are applied to the coefficients of the frequency bands in the frequency domain coefficient map for weighting. Then, an inverse frequency domain transformation is performed on the weighted frequency domain coefficient map to obtain the updated feature map after frequency domain processing.

[0035] The output layer processes the input feature map, channel configuration, and updated feature map to obtain the environmental semantics and risk map.

[0036] Optionally, the formation of the collaborative control plan includes:

[0037] The task records in the unified task information queue are associated with the target location grid and risk level grid in the environmental semantics and risk map to form multi-robot scheduling status data that includes task set, robot set, target location set and passable grid set;

[0038] The tasks in the task set are sorted according to task priority and time constraints. Candidate robots that meet the task type and workspace conditions are selected from the robot set. For each task and candidate robot, a task allocation candidate pair is constructed. The task allocation candidate pair is evaluated according to the scheduling cost function to determine the task allocation result of the task and robot. Path planning is performed for each robot with an assigned task within the range of the passable grid set to generate the corresponding running path.

[0039] Based on the task allocation results and the running path, a task sequence and operation sequence arranged in execution order are generated for each physical robot, and corresponding operation instructions are generated for the software robot associated with the physical robot's task. The task sequence, running path, operation sequence of each physical robot and the operation instructions of the software robot are combined to form a collaborative control plan.

[0040] Optionally, the step of sending the collaborative control plan to the control platform to control the physical robot to perform handling, inspection, or assembly tasks, and to control the software robot to perform data entry or status update operations, includes:

[0041] The collaborative control plan is distributed to the physical robot control platform and the software robot execution platform. The task sequence, running path and operation sequence for each physical robot are converted into motion control instructions. The operation instructions for each software robot are converted into business operation instructions and sent for execution respectively.

[0042] The physical robot control platform controls the physical robot to perform handling, inspection or assembly tasks along the running path according to motion control instructions, while the software robot execution platform performs data entry or status update operations according to business operation instructions.

[0043] The beneficial effects of this invention are:

[0044] This invention enhances the automatic scheduling and collaborative control capabilities of multi-source heterogeneous robots within a unified business process framework by deeply integrating business process management, Poisson surface reconstruction, and an improved HarDNet network. Compared to existing solutions that rely on manually configured interfaces, static rules, and coarse-grained environmental maps, this invention achieves consistent modeling of business process nodes and physical / software robot tasks through a unified task information queue and robot resource description. It improves the 3D reconstruction accuracy of multi-robot point clouds in key operational areas through multi-source weighted vector fields and task-driven octree subdivision. Furthermore, it obtains environmental semantics and risk maps closely related to task types through adaptive harmonic connections, task perception channel configuration, and frequency domain feature processing, making task allocation and path planning more aligned with real-time business needs and changes in the on-site environment. This results in significant improvements in multi-robot collaborative efficiency, scheduling accuracy, and system flexibility.

[0045] The integrated closed-loop method proposed by the present invention, from process modeling, environment reconstruction, semantic analysis to task allocation, path planning and collaborative execution, realizes the linkage control and dynamic optimization among the business process state, environmental semantic information and the running states of multiple robots. With the help of the occupancy grid map and the environmental semantic and risk map, the present invention can conduct fine safety assessments on high-frequency operation areas, key channels and operation workstations, effectively reducing the risks of collision and congestion; through the unified distribution and execution feedback of the collaborative control plan, the present invention overcomes the technical bottlenecks in the prior art, such as complex integration, fragmentation of environmental perception and scheduling, and difficulty in online optimization, and is applicable to various scenarios of production manufacturing, warehousing logistics, inspection and security, which is beneficial to the engineering implementation and intelligent operation and maintenance of enterprise-level multi-robot systems. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] The drawings are used to provide a further understanding of the present invention, and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention, and do not constitute a limitation to the present invention. In the drawings:

[0047] Figure 1 is a flowchart of a multi-robot process scheduling and collaborative control method based on artificial intelligence proposed by the present invention;

[0048] Figure 2 is a structural block diagram of Poisson surface reconstruction of a multi-robot process scheduling and collaborative control method based on artificial intelligence proposed by the present invention;

[0049] Figure 3 is a functional schematic diagram of an improved HarDNet network of a multi-robot process scheduling and collaborative control method based on artificial intelligence proposed by the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0050] Now, the present invention will be further described in detail with reference to the drawings. These drawings are all simplified schematic diagrams, only showing the basic structure of the present invention in a schematic way, so they only show the components related to the present invention.

[0051] Refer to Figure 1 、 Figure 2 and Figure 3 , a multi-robot process scheduling and collaborative control method based on artificial intelligence, includes:

[0052] Establish a process structure across business scenarios in business process management, uniformly set physical robot task nodes and software robot task nodes as robot task nodes, establish resource description information for various robots and complete robot resource registration, and generate a unified task information queue;

[0053] Collect environmental sensing data according to the unified task information queue, and obtain multi-source point cloud data carrying spatial coordinates, normal information and source identifiers;

[0054] Three-dimensional environment reconstruction is performed on multi-source point cloud data by Poisson surface reconstruction. Multi-source weights are assigned to each point cloud data, vector field divergence is calculated, granular subdivision is performed, and three-dimensional surface reconstruction results and occupied grid map are obtained.

[0055] Based on the viewpoint and raster resolution, the 3D surface reconstruction results and the occupied raster map are projected into a multi-channel 2D environment raster map;

[0056] An improved HarDNet network is constructed to extract features and perform semantic analysis on a multi-channel two-dimensional environmental raster map. An adaptive harmonic connection method is used to stitch features together. The features are then subjected to frequency domain transformation, frequency band weighting, and inverse transformation to obtain an environmental semantic and risk map.

[0057] By combining a unified task information queue with environmental semantics and risk maps, multi-robot task allocation and path planning are performed to generate task sequences, running paths and operation sequences for each physical robot, thus forming a collaborative control plan.

[0058] The collaborative control plan is sent to the control platform to control the physical robot to perform handling, inspection or assembly tasks, and to control the software robot to perform data entry or status update operations.

[0059] In this embodiment, generating a unified task information queue includes:

[0060] In business process management, create process definitions and model business steps across business scenarios as directed process structures, including multiple nodes and directed connections between nodes. Nodes involving physical robot execution and nodes involving software robot execution are uniformly defined as robot task nodes. Each robot task node includes a task identifier, task type, target location identifier, and time constraint parameters. The time constraint parameters include the earliest start time and the latest finish time.

[0061] Generate robot resource description information for various types of robots, including resource identifier, resource type and set of executable task types. Register the robot resource description information in the robot resource library of business process management to complete robot resource registration. Generate a unified task information queue based on robot task nodes. Each task record in the unified task information queue includes task identifier, task type and target location identifier.

[0062] In this embodiment, the environmental sensing data includes laser point cloud data, depth map data from a depth camera, RGB image data, and the robot's own pose data.

[0063] In this embodiment, acquiring multi-source point cloud data carrying spatial coordinates, normal information, and source identifiers includes:

[0064] Control multiple physical robots to move according to the target position in a unified task information queue, and collect environmental sensing data using LiDAR, depth camera and vision camera respectively during the movement;

[0065] Based on the pose data, the environmental sensing data collected by different robots are unified into the same spatial coordinate system. The laser point cloud data and depth map data are fused to generate point cloud data carrying spatial coordinates and normal information. A source identifier is added to each laser point cloud data to obtain multi-source point cloud data.

[0066] In this embodiment, obtaining the three-dimensional surface reconstruction result and the occupied grid map includes:

[0067] For each point cloud data point, the sensor type, acquisition distance, incident angle, and timestamp parameters are read. A corresponding multi-source weight value is generated for each point cloud data point according to the weight calculation rules. The normal information of each point cloud data point is weighted and accumulated with the multi-source weight value to construct a weighted vector field on a three-dimensional discrete grid. The divergence value of the weighted vector field is calculated at each node to obtain the corresponding divergence discrete data, where:

[0068] The process of generating corresponding multi-source weight values ​​for each point cloud data point according to the weight calculation rules is as follows:

[0069] During the initialization phase, sensor weight tables, distance weight tables, angle weight tables, and time weight tables are preset, and corresponding weight coefficients are configured for different sensor types, different distance ranges, different incident angle ranges, and different time ranges, respectively.

[0070] The sensor weight coefficient is found in the sensor weight table according to the sensor type; the distance weight coefficient is found in the distance weight table according to the distance interval to which the acquisition distance belongs; the angle weight coefficient is found in the angle weight table according to the angle interval to which the incident angle belongs; and the time weight coefficient is found in the time weight table according to the time interval to which the time difference between the timestamp and the current time belongs.

[0071] The four types of weight coefficients are combined to obtain the multi-source weight values ​​of the point cloud data. The combination method is the product of the weight coefficients.

[0072] The weighted vector field is constructed on the three-dimensional discrete grid as follows:

[0073] Discrete grids are divided within the three-dimensional space covered by multi-source point cloud data. Each grid cell corresponds to a grid node. For each grid node, a set of point cloud data within the neighborhood is determined. For each point cloud data in the set, the normal information is multiplied by the corresponding multi-source weight value to obtain a weighted normal vector. Then, all weighted normal vectors in the current neighborhood are accumulated at the grid node. The accumulated result is divided by the sum of the multi-source weight values ​​of all point cloud data in the neighborhood to obtain the normalized node vector, forming a weighted vector field.

[0074] The step of calculating the divergence value of the weighted vector field at each node to obtain the corresponding discrete divergence data is as follows:

[0075] For each grid node, adjacent nodes in three coordinate directions are selected, and the component differences of the weighted vector in each coordinate direction are calculated respectively. The differences in each direction are summed according to preset coefficients to obtain the divergence approximation at the current node. Through discrete divergence calculation, the divergence discrete data corresponding to the weighted vector field is obtained at each node of the three-dimensional discrete grid.

[0076] An octree root node is initialized within the three-dimensional space of multi-source point cloud data. The octree nodes are initially subdivided according to the point cloud distribution. Task importance is calculated for each node based on multi-robot historical trajectories and task records. A first threshold and a second threshold are set. Nodes with task importance greater than the first threshold are marked as high-frequency operation nodes, and nodes with task importance less than the second threshold are marked as low-frequency operation nodes. High-frequency operation nodes are further subdivided according to fine-grained rules until the maximum level is reached or the number of point clouds within the node is lower than the first threshold. Low-frequency operation nodes maintain a coarse-grained level or are subdivided a limited number of times only when the number of point clouds within the node exceeds the second threshold, forming a task-driven octree spatial discrete structure. Specifically, forming the task-driven octree spatial discrete structure involves:

[0077] The three-dimensional spatial boundary is determined based on the minimum outer cube of the multi-source point cloud data. The three-dimensional spatial boundary is used as the spatial range of the octree root node. The octree root node is initially subdivided into eight child nodes according to the overall distribution of the point cloud, so that the spatial range of each child node contains point cloud data, thus obtaining an initial octree structure covering the multi-source point cloud data.

[0078] Record the motion trajectory, task triggering position and work dwell time of each robot in three-dimensional space, map the historical trajectory points and task events to octree nodes according to spatial coordinates, count the number of times the robot trajectory passes through, the number of times the task is triggered and the work dwell time for each node, and combine the indicators into the task importance value of the current node according to a weighted linear combination;

[0079] Set a first threshold and a second threshold. Nodes with task importance values ​​greater than the first threshold are marked as high-frequency operation nodes, nodes with task importance values ​​less than the second threshold are marked as low-frequency operation nodes, and nodes with task importance values ​​between the first and second thresholds are marked as ordinary nodes. For nodes marked as high-frequency operation nodes, spatial subdivision is performed according to fine-grained rules to divide the corresponding spatial range into smaller sub-nodes. The subdivision process stops when the maximum level limit is met or the number of point clouds in the node is lower than the first number threshold.

[0080] The first threshold and the second threshold are set by sorting the task importance values ​​of all octree nodes, taking the top 20% of task importance values ​​in the sorted sequence as the first threshold, and taking the bottom 10% of task importance values ​​in the sorted sequence as the second threshold.

[0081] The fine-grained rule, for nodes marked as high-frequency operation nodes, under the condition that the current level of the node is less than the maximum level and the number of point clouds in the node is not less than the first threshold, divides the cube space corresponding to the node into eight sub-nodes along each of the three coordinate axes, and repeats the condition judgment and subdivision operation for the newly generated sub-nodes until the subdivision stop condition is met.

[0082] For nodes marked as low-frequency operation nodes, the coarser level is maintained and not further subdivided. By applying different subdivision granularities to nodes with different task importance, a task-driven octree spatial discrete structure is formed in the three-dimensional space of multi-source point cloud data.

[0083] Divergence discrete data is introduced at each node of the octree spatial discrete structure to construct a Poisson discrete equation. This Poisson discrete equation is then expressed as a sparse linear system of equations and numerically solved to obtain the scalar function values ​​at each node of the octree. Specifically, obtaining the scalar function values ​​at each node of the octree involves:

[0084] For each node, the difference form of the discrete Laplace operator is used to write the difference relationship between the scalar function value at the node and the scalar function values ​​of the adjacent nodes in the three coordinate directions into an algebraic equation. The result of the discrete Laplace operator at the node is equal to the divergence discrete data corresponding to the current node, and the Poisson discrete equation of the current node is obtained.

[0085] After constructing the discrete Poisson equations for all nodes, the equations corresponding to each node are arranged in order of node number. The scalar unknowns of each node are used as the unknowns of the equation system. The coefficients before the scalar unknowns of each node are used to form a coefficient matrix. The discrete divergence data of each node are used to form a constant term vector, resulting in a sparse linear equation system. The conjugate gradient iteration method is used to numerically solve the sparse linear equation system, and the scalar unknowns corresponding to all nodes are obtained. The scalar unknowns of each node are used as the scalar function values ​​at each node of the octree.

[0086] In the octree spatial discrete structure, a preset isovalue is selected, and the point set whose scalar function value is equal to the preset isovalue is extracted as the 3D surface reconstruction result. Based on the 3D surface reconstruction result and the scalar function value at the node, the spatial voxels are occupancy determined, and an occupancy raster map corresponding to the 3D surface reconstruction result is generated. Specifically, the selection of the preset isovalue in the octree spatial discrete structure involves:

[0087] After obtaining the scalar function values ​​at each node of the octree, each point cloud data in the multi-source point cloud data is interpolated in the discrete structure of the octree space to obtain the scalar function value corresponding to the current point cloud data. The scalar function values ​​corresponding to all point cloud data are added together and divided by the total number of point cloud data to obtain the average value of the scalar function. The average value is used as the preset equal value.

[0088] In this embodiment, projecting the three-dimensional surface reconstruction result and the occupied grid map into a multi-channel two-dimensional environment grid map includes:

[0089] A projection reference plane is selected from the 3D spatial coordinate system corresponding to the 3D surface reconstruction result and the occupied raster map. The projection viewpoint direction and the 2D environment raster coordinate system are determined. The raster resolution, number of rows and columns, and raster boundary range are set. The spatial coordinates of each 3D point are mapped to the row index and column index in the 2D environment raster map according to the projection rules. The projection rules are as follows:

[0090] A horizontal plane in three-dimensional space is selected as the projection reference plane. Orthogonal projection is performed on the three-dimensional points along the vertical direction. The coordinates of the three-dimensional points in the column direction are subtracted from the minimum coordinate value in the column direction, divided by the raster resolution, and the integer part is used as the column index. The coordinates of the three-dimensional points in the row direction are subtracted from the minimum coordinate value in the row direction, divided by the raster resolution, and the integer part is used as the row index. Points that exceed the range of the number of rows and columns do not participate in the projection mapping of the current two-dimensional environment raster map.

[0091] For each 2D environment grid cell, point data from the 3D surface reconstruction result mapped to the current grid cell and voxel data from the occupied grid map are collected. The depth value, height value, normal component and occupation probability value of the grid cell are calculated by statistical method. The depth value is the distance statistical result along the projection view direction, the height value is the coordinate statistical result along the vertical direction, the normal component is the component statistical result of normal information, and the occupation probability value is the probability statistical result of the occupation grid voxel state.

[0092] The depth, height, normal components, and occupancy probability values ​​of each two-dimensional environment raster cell are combined in channel order to form a multi-channel two-dimensional environment raster map with row and column indices as position indices and various statistical values ​​as channel data.

[0093] In this embodiment, obtaining the environmental semantics and risk map includes:

[0094] An improved HarDNet network is constructed, comprising an input layer, an improved block feature layer, channel units, frequency domain units, and an output layer. The input layer performs convolution and pooling on the multi-channel two-dimensional environment raster map to obtain intermediate feature maps, where:

[0095] The improved HarDNet network takes the input layer as the entry point. While maintaining the original convolution and pooling structure, it replaces the ordinary feature extraction layer in the original network with an improved block feature layer formed by sequentially connecting multiple improved blocks. After the improved block feature layer, channel units and frequency units are embedded in sequence. The channel units adjust the channel configuration of each convolutional layer according to the task-related parameters. The feature map output by the channel units then enters the frequency units to perform frequency domain transformation, weighting and inverse transformation. Finally, the output layer processes the feature map jointly improved by the channel units and frequency units to generate environmental semantics and risk maps.

[0096] The improved block feature layer predetermines a set of candidate previous convolutional layers for the current convolutional layer, calculates the feature importance of the output feature maps of the candidate previous convolutional layers, and selects several feature maps from the output of the candidate previous convolutional layers using an adaptive harmonic connection method based on the feature importance. These feature maps are then concatenated along the channel dimension to form the input feature map of the current convolutional layer. Specifically, forming the input feature map of the current convolutional layer involves:

[0097] Based on the current convolutional layer's layer number, select preceding convolutional layers with integer intervals of one, two, four, and eight indices. Add preceding convolutional layers that satisfy the index difference condition to the candidate preceding convolutional layer set, so that each preceding convolutional layer in the candidate set satisfies the harmonic interval relationship in terms of layer number.

[0098] The values ​​of the current feature map in the spatial and channel dimensions are statistically analyzed to obtain the average response value, maximum response value, and variance. The multiple statistics are combined to form the feature importance of the feature map, and the candidate previous convolutional layers are sorted according to the feature importance.

[0099] Based on the feature importance ranking results, an adaptive harmonic connection method is adopted to select several feature maps from the output of the candidate previous convolutional layer. Under the premise of satisfying the harmonic interval relationship, the output of the previous convolutional layer with higher feature importance is selected first. The selected multiple feature maps are then concatenated sequentially in the channel dimension to form the input feature map of the current convolutional layer.

[0100] The channel unit receives the process task type parameters and environment complexity parameters, and allocates the number of channels to the convolutional layers in each improved block in the improved block feature layer according to the mapping relationship. The number of convolutional layer channels is adjusted under different task types and environment complexities to form task-related channel configurations.

[0101] The frequency domain unit performs a frequency domain transformation on the feature map output by the improved block to obtain a frequency domain coefficient map. Frequency band weights are then applied to the coefficients of each frequency band in the frequency domain coefficient map for weighting. Finally, an inverse frequency domain transformation is performed on the weighted frequency domain coefficient map to obtain an updated feature map processed in the frequency domain. Specifically, obtaining the updated feature map processed in the frequency domain involves:

[0102] The frequency domain unit selects the feature map to be processed from the output of the improved block feature layer, and performs frequency domain transformation processing on the feature map one by one according to the channel. For the two-dimensional feature map of any channel, a two-dimensional discrete cosine transform is performed on the feature map of the current channel while keeping the spatial size unchanged, converting the pixel values ​​in the spatial domain into the coefficient distribution in the frequency domain, and obtaining the frequency domain coefficient map corresponding to the channel. Each position in the frequency domain coefficient map corresponds to the coefficient of a spatial frequency component.

[0103] Multiple frequency bands are divided according to the frequency index. The low-frequency region near the origin, the middle region, and the high-frequency region far from the origin are defined as low-frequency band, mid-frequency band, and high-frequency band, respectively. An independent frequency band weight parameter is configured for each frequency band. The coefficients belonging to the same frequency band in the frequency domain coefficient diagram are multiplied by the corresponding frequency band weight parameter for weighted processing. This completes the differential adjustment of low-frequency structural information and high-frequency detail information. After processing all frequency bands of the current channel, the weighted frequency domain coefficient diagram is obtained.

[0104] A two-dimensional discrete cosine inverse transform is performed on the weighted frequency domain coefficient map to restore the frequency domain coefficients back to the spatial domain, resulting in a channel update feature map after frequency domain processing. The frequency domain transformation, frequency band weighting, and inverse transform process are repeated for all channels of the same feature map to finally obtain an update feature map that is consistent with the original feature map in terms of the number of channels and spatial size.

[0105] The output layer processes the input feature map, channel configuration, and updated feature map to obtain an environmental semantic and risk map. Specifically, obtaining the environmental semantic and risk map involves:

[0106] The input feature map, channel configuration, and updated feature map are concatenated along the channel dimension to form a fused feature map containing multiple semantic and structural information. Convolution and upsampling operations are then performed on the fused feature map sequentially to restore the feature map to the same spatial resolution as the multi-channel two-dimensional environment raster map. An environment semantic segmentation map is output through several convolutional channels for semantic classification, a accessibility raster for each raster is output through a convolutional channel for accessibility determination, and a risk level raster for each raster is output through a convolutional channel for risk level estimation. The environment semantic segmentation map, accessibility raster, and risk level raster are then combined to form an environment semantic and risk map.

[0107] In this embodiment, the formation of a collaborative control plan includes:

[0108] The task records in the unified task information queue are associated with the target location grid and risk level grid in the environmental semantics and risk map to form multi-robot scheduling status data that includes task set, robot set, target location set and passable grid set;

[0109] The tasks in the task set are sorted according to task priority and time constraints. Candidate robots that meet the task type and workspace conditions are selected from the robot set. A task allocation candidate pair is constructed for each task and candidate robot combination. The task allocation candidate pairs are evaluated according to the scheduling cost function to determine the task allocation result between the task and the robot. Path planning is then performed for each robot with an assigned task within the accessible grid set to generate the corresponding running path. Specifically, the evaluation of the task allocation candidate pairs according to the scheduling cost function to determine the task allocation result between the task and the robot involves:

[0110] The scheduling cost function gives a cost value for each task allocation candidate pair, which is used to measure the comprehensive cost of a robot performing a task. First, based on the environmental semantics and risk map and the set of traversable grids, the shortest feasible path between the current position of the candidate robot and the target position of the task is calculated, and the running time corresponding to the path length is used as the travel time index.

[0111] Combining the robot's current task queue length and task time constraints, the difference between the estimated task completion time and the latest task completion time is calculated as a time urgency index. While calculating the shortest feasible path, the risk level of each grid cell through which the path passes is read. The risk level of each grid cell is multiplied by the corresponding path step length and then summed to obtain the cumulative risk value of the path as a safety risk index. Furthermore, a priority matching index is obtained based on the task priority and whether the robot's capabilities include the task type and the covered task space area.

[0112] The total value of the candidate pair is obtained by weighted summation of the driving time index, time urgency index, safety risk index and priority matching index. The smaller the total value, the better the candidate pair. After combining all tasks and candidate robots to form a cost matrix, the task allocation candidate pair with the smallest total value and not occupied is selected step by step to determine the final execution robot corresponding to each task, thus forming the task and robot allocation result.

[0113] Based on the task allocation results and the running path, a task sequence and operation sequence arranged in execution order are generated for each physical robot, and corresponding operation instructions are generated for the software robot associated with the physical robot's task. The task sequence, running path, operation sequence of each physical robot and the operation instructions of the software robot are combined to form a collaborative control plan.

[0114] In this embodiment, the step of sending the collaborative control plan to the control platform to control the physical robot to perform handling, inspection, or assembly tasks, and controlling the software robot to perform data entry or status update operations, includes:

[0115] The collaborative control plan is distributed to the physical robot control platform and the software robot execution platform. The task sequence, running path and operation sequence for each physical robot are converted into motion control instructions. The operation instructions for each software robot are converted into business operation instructions and sent for execution respectively.

[0116] The physical robot control platform controls the physical robot to perform handling, inspection or assembly tasks along the running path according to motion control instructions, while the software robot execution platform performs data entry or status update operations according to business operation instructions.

[0117] Example 1:

[0118] To verify the feasibility of this invention in practice, it was applied to an electronics manufacturing company with an automated production line. The automated warehouse for finished products covers an area of ​​approximately 15,000 square meters and is equipped with six mobile robots, two stacker cranes, one inspection robot, and multiple software robots deployed on the business system. The existing business processes are linked by the company's original business process management platform, connecting steps such as "receiving registration, quality inspection, shelving, replenishment, and inventory." The original system can only drive manual operation and a small number of fixed devices. The mobile robots use their respective manufacturers' independent scheduling systems, and the software robots are configured separately by an RPA platform. This results in a "receiving-shelving-warehousing" process requiring manual switching between the three systems. The robot's operating status cannot be uniformly visualized at the process level, leading to significant task conflicts, wasted runs, and waiting. During peak periods of concentrated warehousing and inventory at the end of the month, robot queues and congestion occur in the warehouse, requiring emergency manual intervention. This makes it difficult to meet the company's needs for unified scheduling and process linkage control of multiple robots.

[0119] The method of this invention is deployed as a multi-robot process scheduling and collaborative control platform, which interfaces with existing business process management systems, mobile robot scheduling systems, and RPA platforms. In business process management, existing manual nodes, equipment nodes, mobile robot tasks, and software robot tasks are uniformly modeled as robot task nodes, driven by a unified task information queue for multiple batches of warehousing and inventory processes. All mobile robots and stacker cranes on-site are connected to a unified robot resource library, while multiple mobile robots and inspection robots continuously collect laser point clouds, depth maps, and their own pose information to construct multi-source point clouds. Internally, the platform uses Poisson surface reconstruction to perform 3D reconstruction of warehouse aisles, shelves, intersections, and other areas, inputting an improved HarDNet network to generate environmental semantics and risk maps. The scheduling unit simultaneously performs task allocation and path planning for mobile robots and software robots within the same platform. Mobile robots are responsible for goods handling, inventory inspection, and software robots automatically complete receiving slip entry, quality inspection result backfilling, and inventory status updates in ERP and WMS, achieving unified collaboration between business processes, physical robots, and software robots.

[0120] The continuous trial operation from May to July 2025 showed that in the typical warehouse "receiving-quality inspection-shelving" process, the overall time from receiving and registration to shelving completion for a single batch of goods was significantly shortened. Especially during peak inbound periods, the waiting time between multiple mobile robots was greatly reduced, and task conflicts and scheduling alarms were significantly decreased. In the monthly full-warehouse inventory count, the platform automatically planned inspection paths and scanning poses based on the semantic risk map generated by Poisson surface reconstruction and the improved HarDNet network. The accuracy and consistency of the inventory results were significantly improved, and no significant congestion or prolonged downtime occurred during the operation. The overall trial performance demonstrates that this invention effectively alleviates the contradiction between the complex integration and difficulty in unified modeling between business process management systems and multi-source heterogeneous robots in practical applications, significantly improving the efficiency of multi-robot collaborative operations and the dynamic scheduling capability based on environmental perception.

[0121] Table 1 Comparative Experimental Results of Multi-Robot Process Scheduling Methods

[0122] index manual solution Local scheduling method Static scheduling method Method of the present invention Average time (minutes) for a single batch of goods receiving process 96.2 82.7 71.5 59.1 AMR average no-load rate (%) 29.4 24.8 21.1 17.3 Congestion / conflict events (times / day) 7.8 5.2 3.4 1.9 Task human intervention rate (%) 21.6 15.3 9.7 5.1 Inventory discrepancy rate (%) 2.4 1.9 1.2 0.6 RPA document error rate (%) 4.1 3.0 1.6 0.8 Number of tasks completed per shift (per shift) 268 305 362 425 Robot utilization rate (%) 54.7 63.2 71.8 79.2

[0123] As shown in Table 1, the average time for a single batch of warehousing processes gradually decreased from 96.2 minutes with the manual method to 82.7 minutes with the partial scheduling method and 71.5 minutes with the static scheduling method. After adopting the method of this invention, it was further shortened to 59.1 minutes, showing a continuous downward trend as the intelligence of scheduling increases. The number of tasks completed per shift increased progressively from 268 with the manual method, 305 with the partial scheduling method, and 362 with the static scheduling method to 425 with the method of this invention. This indicates that, under the same shift length and number of devices, this invention is significantly more efficient in task organization and path arrangement, and can more fully release the capacity of multi-robot clusters within a unified process framework.

[0124] From the perspective of robot operation status and collaborative effect, the average idle rate of AMR decreased from 29.4% in the manual scheme to 24.8% and 21.1% through local scheduling and static scheduling, respectively, and further decreased to 17.3% in this invention, indicating a significant reduction in ineffective movement and that robot movement is more concentrated on effective tasks. The number of congestion and conflict events decreased from 7.8 times / day in the manual scheme to 5.2 times / day and 3.4 times / day through local scheduling and static scheduling, respectively, and further decreased to 1.9 times / day under the method of this invention. The rate of human intervention in conjunction with tasks decreased step by step from 21.6% to 5.1%, indicating that global scheduling based on environmental semantics and risk maps is more effective in avoidance planning, bottleneck diversion, and multi-robot collaboration, resulting in smoother system operation and a significant reduction in reliance on human intervention.

[0125] From the perspectives of business accuracy and system efficiency, the inventory discrepancy rate decreased from 2.4% in the manual method to 1.9% and 1.2% through partial scheduling and static scheduling, respectively, and further to 0.6% using the method of this invention. The RPA document error rate gradually decreased from 4.1% to 0.8%, indicating that under a unified process, physical robot operations and software robot data processing can form a reliable closed loop, reducing errors caused by manual data entry and system fragmentation. Robot utilization increased from 54.7% in the manual method to 63.2% and 71.8% through partial scheduling and static scheduling, respectively, reaching 79.2% in this invention. This demonstrates that the environmental semantics and risk information obtained under Poisson reconstruction and improved HarDNet network support effectively helps the scheduling layer to utilize multi-robot clusters more fully and stably, achieving a comprehensive improvement in efficiency, accuracy, and resource utilization.

[0126] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for multi-robot process scheduling and collaborative control based on artificial intelligence, characterized in that, include: In business process management, establish a process structure that spans business scenarios, unify physical robot task nodes and software robot task nodes as robot task nodes, establish resource description information for various robots and complete robot resource registration, and generate a unified task information queue. Collect environmental sensor data based on a unified task information queue to obtain multi-source point cloud data carrying spatial coordinates, normal information and source identifiers; Three-dimensional environment reconstruction is performed on multi-source point cloud data by Poisson surface reconstruction. Multi-source weights are assigned to each point cloud data, vector field divergence is calculated, granular subdivision is performed, and three-dimensional surface reconstruction results and occupied grid map are obtained. Based on the viewpoint and raster resolution, the 3D surface reconstruction results and the occupied raster map are projected into a multi-channel 2D environment raster map; An improved HarDNet network is constructed to extract features and perform semantic analysis on a multi-channel two-dimensional environmental raster map. An adaptive harmonic connection method is used to stitch features together. The features are then subjected to frequency domain transformation, frequency band weighting, and inverse transformation to obtain an environmental semantic and risk map. By combining a unified task information queue with environmental semantics and risk maps, multi-robot task allocation and path planning are performed to generate task sequences, running paths and operation sequences for each physical robot, thus forming a collaborative control plan. The collaborative control plan is sent to the control platform to control the physical robot to perform handling, inspection or assembly tasks, and to control the software robot to perform data entry or status update operations.

2. The multi-robot process scheduling and collaborative control method based on artificial intelligence according to claim 1, characterized in that, The generation of the unified task information queue includes: In business process management, create process definitions and model business steps across business scenarios as directed process structures, including multiple nodes and directed connections between nodes. Nodes involving physical robot execution and nodes involving software robot execution are uniformly defined as robot task nodes. Each robot task node includes a task identifier, task type, target location identifier, and time constraint parameters. The time constraint parameters include the earliest start time and the latest finish time. Generate robot resource description information for various types of robots, including resource identifier, resource type and set of executable task types. Register the robot resource description information in the robot resource library of business process management to complete robot resource registration. Generate a unified task information queue based on robot task nodes. Each task record in the unified task information queue includes task identifier, task type and target location identifier.

3. The multi-robot process scheduling and collaborative control method based on artificial intelligence according to claim 1, characterized in that, The environmental sensing data includes laser point cloud data, depth map data from depth cameras, RGB image data, and the robot's own pose data.

4. The multi-robot process scheduling and collaborative control method based on artificial intelligence according to claim 1, characterized in that, The acquisition of multi-source point cloud data carrying spatial coordinates, normal information, and source identifiers includes: Control multiple physical robots to move according to the target position in a unified task information queue, and collect environmental sensing data using LiDAR, depth camera and vision camera respectively during the movement; Based on the pose data, the environmental sensing data collected by different robots are unified into the same spatial coordinate system. The laser point cloud data and depth map data are fused to generate point cloud data carrying spatial coordinates and normal information. A source identifier is added to each laser point cloud data to obtain multi-source point cloud data.

5. The multi-robot process scheduling and collaborative control method based on artificial intelligence according to claim 1, characterized in that, The obtained 3D surface reconstruction results and occupied grid map include: For each point cloud data, the sensor type, acquisition distance, incident angle and timestamp parameters are read. According to the weight calculation rules, a corresponding multi-source weight value is generated for each point cloud data. The normal information of each point cloud data is weighted and accumulated with the multi-source weight value. A weighted vector field is constructed on the three-dimensional spatial discrete grid. The divergence value of the weighted vector field is calculated at each node to obtain the corresponding divergence discrete data. An octree root node is initialized within the three-dimensional space of multi-source point cloud data. The octree nodes are initially subdivided according to the point cloud distribution. The task importance of each node is calculated based on the historical trajectory and task records of multiple robots. A first threshold and a second threshold are set. Nodes with task importance greater than the first threshold are marked as high-frequency operation nodes, and nodes with task importance less than the second threshold are marked as low-frequency operation nodes. High-frequency operation nodes are further subdivided according to fine-grained rules until the maximum level is reached or the number of point clouds in the node is lower than the first quantity threshold. Low-frequency operation nodes maintain a coarse-grained level or are subdivided a limited number of times only when the number of point clouds in the node exceeds the second quantity threshold, forming a task-driven octree spatial discrete structure. Divergence discrete data is introduced at each node of the octree spatial discrete structure to construct the Poisson discrete equation. The Poisson discrete equation is expressed as a sparse linear system of equations and numerically solved to obtain the scalar function values ​​at each node of the octree. In the octree spatial discrete structure, a preset isovalue is selected, and the point set whose scalar function value is equal to the preset isovalue is extracted as the 3D surface reconstruction result. The spatial voxels are occupied according to the 3D surface reconstruction result and the scalar function value at the node, and the occupied raster map corresponding to the 3D surface reconstruction result is generated.

6. The multi-robot process scheduling and collaborative control method based on artificial intelligence according to claim 1, characterized in that, The process of projecting the 3D surface reconstruction result and the occupied grid map into a multi-channel 2D environment grid map includes: Select a projection reference plane from the 3D spatial coordinate system corresponding to the 3D surface reconstruction result and the occupied raster map, determine the projection view direction and the 2D environment raster coordinate system, set the raster resolution, number of rows and columns and raster boundary range, and map the spatial coordinates of each 3D point to the row index and column index in the 2D environment raster map according to the projection rules. For each 2D environment grid cell, point data from the 3D surface reconstruction result mapped to the current grid cell and voxel data from the occupied grid map are collected. The depth value, height value, normal component and occupation probability value of the grid cell are calculated by statistical method. The depth value is the distance statistical result along the projection view direction, the height value is the coordinate statistical result along the vertical direction, the normal component is the component statistical result of normal information, and the occupation probability value is the probability statistical result of the occupation grid voxel state. The depth, height, normal components, and occupancy probability values ​​of each two-dimensional environment raster cell are combined in channel order to form a multi-channel two-dimensional environment raster map with row and column indices as position indices and various statistical values ​​as channel data.

7. The multi-robot process scheduling and collaborative control method based on artificial intelligence according to claim 1, characterized in that, The obtained environmental semantics and risk map includes: An improved HarDNet network is constructed, consisting of an input layer, an improved block feature layer, channel units, frequency domain units, and an output layer. Through the input layer, the multi-channel two-dimensional environment raster map is convolved and pooled to obtain an intermediate feature map. The improved block feature layer predetermines a set of candidate previous convolutional layers for the current convolutional layer, calculates the feature importance of the output feature maps of the candidate previous convolutional layers, and selects several feature maps from the output of the candidate previous convolutional layers using an adaptive harmonic connection method based on the feature importance. The feature maps are then concatenated along the channel dimension to form the input feature map of the current convolutional layer. The channel unit receives the process task type parameters and environment complexity parameters, and allocates the number of channels to the convolutional layers in each improved block in the improved block feature layer according to the mapping relationship. The number of convolutional layer channels is adjusted under different task types and environment complexities to form task-related channel configurations. The frequency domain unit performs a frequency domain transformation on the feature map output by the improved block to obtain a frequency domain coefficient map. Frequency band weights are applied to the coefficients of the frequency bands in the frequency domain coefficient map for weighting. Then, an inverse frequency domain transformation is performed on the weighted frequency domain coefficient map to obtain the updated feature map after frequency domain processing. The output layer processes the input feature map, channel configuration, and updated feature map to obtain the environmental semantics and risk map.

8. The multi-robot process scheduling and collaborative control method based on artificial intelligence according to claim 1, characterized in that, The formation of the collaborative control plan includes: The task records in the unified task information queue are associated with the target location grid and risk level grid in the environmental semantics and risk map to form multi-robot scheduling status data that includes task set, robot set, target location set and passable grid set; The tasks in the task set are sorted according to task priority and time constraints. Candidate robots that meet the task type and workspace conditions are selected from the robot set. For each task and candidate robot, a task allocation candidate pair is constructed. The task allocation candidate pair is evaluated according to the scheduling cost function to determine the task allocation result of the task and robot. Path planning is performed for each robot with an assigned task within the range of the passable grid set to generate the corresponding running path. Based on the task allocation results and the running path, a task sequence and operation sequence arranged in execution order are generated for each physical robot, and corresponding operation instructions are generated for the software robot associated with the physical robot's task. The task sequence, running path, operation sequence of each physical robot and the operation instructions of the software robot are combined to form a collaborative control plan.

9. The multi-robot process scheduling and collaborative control method based on artificial intelligence according to claim 1, characterized in that, The process of distributing the collaborative control plan to the control platform to control the physical robot to perform handling, inspection, or assembly tasks, and controlling the software robot to perform data entry or status update operations includes: The collaborative control plan is distributed to the physical robot control platform and the software robot execution platform. The task sequence, running path and operation sequence for each physical robot are converted into motion control instructions. The operation instructions for each software robot are converted into business operation instructions and sent for execution respectively. The physical robot control platform controls the physical robot to perform handling, inspection or assembly tasks along the running path according to motion control instructions, while the software robot execution platform performs data entry or status update operations according to business operation instructions.

Images