A forked mobile robot scheduling method based on deep learning

Abstract

The application discloses a kind of based on deep learning's fork type mobile robot scheduling method, it is related to intelligent logistics scheduling technical field, including: to unified state package is filtered with adjacent constraint, constructs candidate combination set, and carries out road network and resource context feature coding, forms candidate subgraph feature package;Candidate subgraph feature package is inferred to deep learning, and output scheduling inference result package, and scheduling inference result package is assembled as space-time plan draft;According to space-time plan draft, resource mutual exclusion and road section capacity constraint correction are carried out by differentiable projection, and output executable scheduling instruction package.The application is further through to the time alignment and topological mapping of robot state, task state and resource occupation state, further through to inference result assembly forms space-time plan draft, realizes the smooth transition from learning result to executable scheduling plan, improves execution safety, environmental adaptability and task implementation effect.

Description

Technical Field

[0001] This invention relates to the field of intelligent logistics scheduling technology, and in particular to a deep learning-based scheduling method for forklift mobile robots. Background Technology

[0002] Against the backdrop of the rapid development of smart logistics and flexible manufacturing systems, automated warehousing and production line scheduling technology based on multi-mobile robot (AMR) collaborative operations has become a core driving force for improving logistics efficiency. Existing forklift scheduling systems mainly rely on classic operations research optimization algorithms, such as mixed integer programming (MIP), heuristic rules (such as genetic algorithms and ant colony algorithms), and time window-based dynamic programming methods, which can effectively calculate the theoretically optimal or suboptimal solution in a known topology. With the popularization of IoT sensing technology, existing systems have the ability to collect robot position, power, and task status in real time, and centrally collect data and issue instructions through a central control server, thus initially realizing the digital management of the work process.

[0003] Although existing rule-based scheduling methods have constructed a rigorous mathematical framework at the theoretical level, in practical engineering applications, traditional methods often treat road network topology, resource occupancy status, and robot dynamics characteristics separately. It is difficult to complete the deep integration and global extrapolation of the multi-dimensional coupling characteristics of "task-resource-road network" within milliseconds. When faced with dynamic disturbances such as sudden order insertion, local road congestion, or instantaneous mutual exclusion of resources, existing technical solutions often lead to scheduling delays due to excessive computational load, or get stuck in local optima due to a lack of in-depth mining of historical evolution patterns, and cannot generate a smooth scheduling strategy that takes into account both global efficiency and local feasibility. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, this invention provides a deep learning-based scheduling method for fork-type mobile robots to solve the problems of difficulty in real-time fusion of multi-dimensional coupled information under dynamic disturbances and the tendency of scheduling results to get trapped in local optima.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0007] This invention provides a deep learning-based scheduling method for fork-type mobile robots, comprising: collecting robot state, task, and resource occupancy data, performing timestamp alignment and topology mapping to generate a unified state package; performing feasibility filtering and proximity constraints on the unified state package to construct a candidate combination set, and encoding road network and resource context features to form a candidate subgraph feature package; performing deep learning inference on the candidate subgraph feature package to output a scheduling inference result package, and assembling the scheduling inference result package into a spatiotemporal plan draft; performing resource mutual exclusion and road segment capacity constraint correction through differentiable projection based on the spatiotemporal plan draft to output an executable scheduling instruction package; parsing the executable scheduling instructions to generate robot speed and heading control, and using a neural control barrier function for online correction to output an execution baseline package.

[0008] As a preferred embodiment of the deep learning-based fork-type mobile robot scheduling method of the present invention, the steps of collecting robot state task and resource usage data, performing timestamp alignment and topology mapping, and generating a unified state packet are as follows:

[0009] Collect robot status data, task data, and resource usage data of the forklift mobile robot, and convert the time information attached to each data to the same scheduling clock reference to form a standardized data set;

[0010] Time slice snapshots are constructed based on standardized data sets, and robot state, task state, and resource occupancy state within the same time slice are aggregated to obtain a set of time slice snapshots;

[0011] Perform topology mapping on the time slice snapshot set, project the robot pose onto the road network topology nodes, associate the task start and end points with corresponding workstation nodes, associate resource usage data with mutual exclusion domain identifiers, output a topological snapshot set, perform consistency verification on the topological snapshot set, and generate a unified state package.

[0012] As a preferred embodiment of the deep learning-based forklift mobile robot scheduling method of the present invention, the robot status, task and resource occupancy data includes the forklift mobile robot's pose information, cargo status, power status, current execution stage, workstation occupancy status and narrow domain occupancy status, and the task data includes task identifier, task start point and task end point.

[0013] As a preferred embodiment of the deep learning-based fork-type mobile robot scheduling method of the present invention, the specific steps for constructing a candidate combination set by performing feasibility filtering and proximity constraints on the unified state packet are as follows:

[0014] Extract the set of tasks to be assigned and the set of single-fork mobile robots that can be connected from the unified state packet, and collect the capability parameters of the fork mobile robots and the loading and unloading constraint parameters of the tasks to form an initial matching set;

[0015] Perform feasibility filtering on the initial matching set and output a feasible matching set. Perform proximity constraint filtering on the feasible matching set and calculate the reachable distance based on the topological position of the forklift mobile robot and the topological node of the task starting point, and output a candidate combination set.

[0016] As a preferred embodiment of the deep learning-based fork-type mobile robot scheduling method of the present invention, the specific steps of encoding road network and resource context features to form candidate subgraph feature packages are as follows:

[0017] The candidate combination set is encoded with road network and resource context features. Temporal features are constructed based on road network edge attributes and resource mutual exclusion domain states to form an encoded feature set.

[0018] The fork-type mobile robot nodes, task nodes, and resource nodes are treated as a set of nodes, and feasible matching relationships and occupancy associations are treated as a set of edges. The encoded feature set is written into the node set and the edge set to form a candidate subgraph feature package.

[0019] As a preferred embodiment of the deep learning-based fork-type mobile robot scheduling method of the present invention, the specific steps of performing deep learning inference on the candidate subgraph feature packets and outputting the scheduling inference result packet are as follows:

[0020] Tensor quantization transformation is performed on the candidate subgraph feature package to form an inference tensor package. The inference tensor package is then input into the deep learning inference model to obtain the joint representation.

[0021] The deep learning inference model includes a message passing layer, an attention aggregation layer, a joint representation construction layer, and an output mapping layer; the message passing layer propagates and merges neighborhood information between nodes and edges along the edge structure of the graph to form intermediate node representations;

[0022] The attention aggregation layer assigns weights to different neighborhood information in the intermediate representation of a node and performs weighted fusion to generate a node embedding representation. The joint representation construction layer extracts the two node embeddings corresponding to each candidate matching edge from the node embedding representation to construct a joint representation. The output mapping layer maps the joint representation to generate scheduling scores and time window parameter information and forms an edge-level inference set through mapping and slicing.

[0023] Align and encapsulate the node and edge identifiers in the edge-level inference set and the candidate subgraph feature package, and output the scheduling inference result package.

[0024] As a preferred embodiment of the deep learning-based fork-type mobile robot scheduling method of the present invention, the specific steps for assembling the scheduling inference results into a spatiotemporal plan draft are as follows:

[0025] Read the scheduling score and time window parameter information from the scheduling inference result package and associate them with the corresponding candidate combinations to form a weighted candidate set. Based on the weighted candidate set, construct a candidate execution sequence for each task and a candidate order-taking sequence for each forklift mobile robot to form a sequence draft set.

[0026] Based on the sequence draft set, the time window parameter information is written into the occupied time window corresponding to the task start position and the task end position to form the reservation draft set. The reservation draft set is associated with the corresponding road network traffic path information and encapsulated into a spatiotemporal plan draft.

[0027] As a preferred embodiment of the deep learning-based forklift mobile robot scheduling method of the present invention, the specific steps of performing resource mutual exclusion and road segment capacity constraint correction through differentiable projection based on the spatiotemporal plan draft, and outputting an executable scheduling instruction package are as follows:

[0028] Read the occupancy time window, mutual exclusion domain reservation information and road network travel path information from the draft spatiotemporal plan, and convert them into a set of constraint representations through constraint feature vectorization and constraint tensor quantization;

[0029] Based on the constraint representation set, resource mutual exclusion constraints and road segment capacity constraints are constructed. The occupancy time window and mutual exclusion domain reservation information are mapped to the resource mutual exclusion constraints, and the road network travel path information is mapped to the road segment capacity constraints, forming a conflict representation set.

[0030] Perform a differentiable projection on the set of conflict representations and encapsulate the set of corrected plans into an executable scheduling instruction package.

[0031] As a preferred embodiment of the deep learning-based fork-type mobile robot scheduling method of the present invention, the steps of parsing the executable scheduling instructions to generate robot speed and heading control, and using a neural control barrier function for online correction to output an execution benchmark packet are as follows:

[0032] Read the modified occupancy time window, modified mutual exclusion domain reservation information and modified passage arrangement from the executable scheduling instruction package, and parse the modified occupancy time window into arrival time constraints and docking duration constraints to form a set of control constraints;

[0033] Based on the control constraint set and the current pose and cargo status of the forklift mobile robot, the robot speed and heading control are generated. At the same time, the relative pose, relative speed and mutual exclusion domain occupancy status of neighboring robots are collected to form a barrier constraint set.

[0034] The set of barrier constraints is modified online by using a neural control barrier function to obtain a set of safety controls. The set of safety controls, the modified occupancy time window, and the modified mutual exclusion domain reservation information are then encapsulated into an execution baseline package.

[0035] As a preferred embodiment of the deep learning-based fork-type mobile robot scheduling method of the present invention, the online correction of the barrier constraint set through the neural control barrier function refers to calculating the barrier function value and change trend based on the relative pose, relative velocity and mutual exclusion domain occupancy status in the barrier constraint set in each control update cycle, and making projection-type adjustments to the robot speed and heading control.

[0036] The beneficial effects of this invention are as follows: by performing time alignment and topological mapping on robot state, task state and resource occupancy state, a unified expression of scheduling input information is achieved. Furthermore, by assembling the inference results to form a spatiotemporal plan draft, a smooth transition from learning results to an executable scheduling plan is achieved. This invention is suitable for safe handling and collaborative operations in complex dynamic environments, improving execution safety, environmental adaptability and task implementation effectiveness. Attached Figure Description

[0037] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0038] Figure 1 This is a flowchart of a deep learning-based scheduling method for fork-type mobile robots.

[0039] Figure 2 A flowchart for generating a unified state packet.

[0040] Figure 3 The flowchart describes the process of constructing the candidate combination set and the candidate subgraph feature package.

[0041] Figure 4 This is a flowchart of a draft plan for deep learning inference and generative spatiotemporal planning.

[0042] Figure 5 The graph shows the relationship between the number of forklift mobile robots and the time alignment error under different perturbation ratios.

[0043] Figure 6 This is a time-series comparison diagram of the minimum safe distance between the present invention and the control scheme. Detailed Implementation

[0044] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0045] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0046] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0047] Reference Figures 1-6 This is one embodiment of the present invention, which provides a deep learning-based scheduling method for fork-type mobile robots, comprising the following steps:

[0048] S1. Collect robot status, task, and resource usage data, perform timestamp alignment and topology mapping, and generate a unified status package.

[0049] S1.1. Collect robot status data, task data, and resource usage data of the forklift mobile robot, and convert the time information attached to each data to the same scheduling clock reference to form a standardized data set.

[0050] Specifically, the system acquires the pose information, loading status, battery status, and current execution stage of the forklift mobile robot; it also acquires the task identifier, task start point, task end point, task priority, and task status; it acquires the workstation occupancy status and narrow domain occupancy status, and retains the timestamp field of each record; it converts the timestamps to the same scheduling clock reference and unifies the time format; it performs unified processing on field names, units of measurement, and identification methods; and it merges and encapsulates the processed robot status data, task data, and resource occupancy data into a standardized data set.

[0051] It should be noted that the robot status, task, and resource usage data include the forklift robot's pose information, cargo status, battery status, current execution stage, workstation occupancy status, and narrow domain occupancy status. The task data includes the task identifier, task start point, and task end point.

[0052] S1.2. Construct time-slice snapshots based on standardized data sets, and aggregate robot state, task state, and resource occupancy state within the same time slice to obtain a time-slice snapshot set.

[0053] Specifically, using the same scheduling clock as a unified time axis, continuous time slice intervals are divided, and start and end time markers are set for each time slice interval. Robot status data, task data, and resource usage data in the standardized dataset are assigned to the corresponding time slice intervals according to the timestamp field, forming a candidate record set indexed by the time slice interval. Data within each time slice interval is aggregated, and multiple robot status data of the same forklift mobile robot are merged, retaining pose information, cargo status, power status, and the end or representative value of the current execution stage of the time slice. For task data, the current values ​​of task status, task start, task end, and task priority corresponding to each task identifier are retained. For resource usage data, the occupant identifier and occupancy start and end time information corresponding to the workstation occupancy status and narrow domain occupancy status are retained. The aggregation results are encapsulated as time slice snapshots and arranged in chronological order to obtain a time slice snapshot set.

[0054] S1.3. Perform topology mapping on the time slice snapshot set, project the robot pose onto the road network topology nodes, associate the task start and end points with corresponding workstation nodes, associate resource occupancy data with mutual exclusion domain identifiers, output the topology snapshot set, perform consistency verification on the topology snapshot set, and generate a unified state package.

[0055] Specifically, snapshots of each time slice are read sequentially, pose information is extracted, and node matching is performed in the road network topology to obtain the corresponding road network topology node identifiers, thereby completing the pose projection of the forklift mobile robot; the task start and end points are read, the corresponding workstation node identifiers are found in the road network topology, and associated with the task identifier, task status, and task priority; the workstation occupancy status and narrow domain occupancy status are read, converted into mutual exclusion domain identifiers, and associated with the occupant identifier and occupancy start and end time information to form a topological snapshot set; consistency verification is performed on the topological snapshot set, including mutual exclusion domain occupancy conflict checking, adjacent time slice node jump checking, and task status reverse change checking; the verified topological snapshot set is encapsulated to generate a unified state package.

[0056] It should be noted that, as Figure 5The figure shows the impact of changes in the number of forklift mobile robots on time alignment error under different perturbation ratios. The horizontal axis represents the number of forklift mobile robots, and the vertical axis represents the time alignment error. The curves correspond to different perturbation ratios. As can be seen from the figure, as the number of forklift mobile robots increases, the time alignment error under each perturbation ratio generally shows a downward trend. This indicates that the data time base conversion, time slice merging, and unified encapsulation processing methods adopted can maintain a good time alignment effect under the scaling of multiple robots. At the same time, the higher the perturbation ratio, the larger the overall time alignment error, indicating that external perturbations will affect the accuracy of time synchronization and state alignment. However, the solution of this invention still maintains a relatively stable error control capability under different perturbation conditions.

[0057] S2. Perform feasibility filtering and proximity constraints on the unified state package, construct a candidate combination set, and encode the road network and resource context features to form a candidate subgraph feature package.

[0058] S2.1. Extract the set of tasks to be assigned and the set of single-fork mobile robots that can be connected from the unified state packet, and collect the capability parameters of the fork mobile robots and the loading and unloading constraint parameters of the tasks to form an initial matching set.

[0059] Specifically, in the unified status package, task identifiers in the pending assignment state are filtered according to task status, and the task start point, task end point, task priority, and task status are summarized to form a set of tasks to be assigned; forklift mobile robot identifiers in the order-accepting state are filtered according to the current execution stage and cargo status, and the road network topology node identifiers, power status, cargo status, and current execution stage are summarized to form a set of forklift mobile robots that can accept orders; the load capacity, fork size range, and lifting height range of the order-accepting forklift mobile robots are associated with the load capacity, fork size range, and lifting height range of the order-accepting forklift mobile robots, and the cargo type, loading and unloading height requirements, and workstation docking method of the tasks to be assigned are associated with the task identifiers; each task identifier is combined with each forklift mobile robot identifier, and the corresponding forklift mobile robot capability parameters and loading and unloading constraint parameters are written to form an initial matching set.

[0060] It should be noted that the capability parameters of the forklift mobile robot can be formed from equipment specifications and calibration records. The equipment specifications are derived from factory technical parameters or acceptance configuration items, while the calibration records are derived from the calibration results of designated workstations. During operation, the capability parameters can also be updated based on maintenance records or component replacement records. The loading and unloading constraint parameters for a task can be generated during task creation based on task source information. This information includes the workstation attributes corresponding to the task start and end points, as well as the cargo attributes. The workstation attributes characterize the loading and unloading height requirements and docking methods, while the cargo attributes characterize the cargo type, size, and weight class.

[0061] S2.2. Perform feasibility filtering on the initial matching set, output a feasible matching set, perform proximity constraint filtering on the feasible matching set, calculate the reachable distance based on the topological position of the forklift mobile robot and the topological node of the task starting point, and output a candidate combination set.

[0062] Specifically, each combination record in the initial matching set is read and compared with the correspondence between the forklift mobile robot's capability parameters and the task's loading and unloading constraint parameters. Combination records whose capability parameters do not match the task's loading and unloading constraint parameters are removed from the initial matching set. At the same time, the current execution stage and loading status in the unified state packet are combined to exclude combination records that do not match the current execution stage, forming a feasible matching set. Proximity constraint filtering is performed on the feasible matching set. The topological position of each forklift mobile robot in the unified state packet and the topological node of the task's starting point are read. The shortest path length or shortest path cost from the topological position to the topological node of the task's starting point is used as the reachable distance on the road network topology, and the reachable distance is written into the corresponding combination record in the feasible matching set. The feasible matching set is filtered and sorted according to the reachable distance, and the combination records of subsequent road network and resource context feature encoding are retained, outputting a candidate combination set.

[0063] It should be noted that the expression for calculating the reachable distance is:

[0064] ;

[0065] in, The identifier for the forklift mobile robot is: To the task identifier is The reachability distance of the task's starting topology node. This represents a candidate travel path, consisting of several road network edges connected end-to-end, from the starting point to the destination. Indicates starting from the node To the destination node The set of all feasible paths, The identifier for the forklift mobile robot is: The topological location corresponds to the road network topology node identifier. The task identifier is The road network topology node identifier of the task's starting topology node. It represents the length of the centerline of a road segment (used to characterize the geometric length of the corresponding road segment along the centerline direction in the road network). This represents a road network edge in the road network topology. This represents the index variable for the fork-type mobile robot identifier. This represents the task identifier index variable.

[0066] S2.3. Encode the candidate combination set with road network and resource context features, construct temporal features based on road network edge attributes and resource mutual exclusion domain states, and form a set of encoded features.

[0067] Specifically, the system reads each combination record in the candidate combination set to obtain the topological location, task start-point topological node, and task end-point topological node of the forklift mobile robot. It then calculates the corresponding travel path on the road network topology and extracts the road network edge attributes, which include at least segment length, travel direction, travel capacity, and travel restriction information. Next, it reads the resource mutex domain status from the unified state packet, associates the occupant identifier corresponding to the mutex domain identifier with the occupancy start and end time information, matches the mutex domain identifiers involved in the travel path with the resource mutex domain status, and obtains mutex domain waiting association information. Finally, it constructs temporal features based on the road network edge attributes and mutex domain waiting association information, and encapsulates these features with reachability distance, task priority, cargo status, and battery status to form a set of encoded features.

[0068] S2.4. The fork-type mobile robot nodes, task nodes, and resource nodes are taken as a set of nodes, the feasible matching relationships and occupancy associations are taken as a set of edges, and the encoded feature set is written into the node set and the edge set to form a candidate subgraph feature package.

[0069] Specifically, within the entity scope corresponding to the candidate combination set, a node set consisting of forklift mobile robot nodes, task nodes, and resource nodes is established, along with an edge set consisting of feasible matching relationships, access associations, and occupancy associations. Forklift mobile robot nodes are identified by identifiers of single-forklift mobile robots that can be connected, task nodes by identifiers of tasks to be assigned, and resource nodes by identifiers of mutual exclusion domains and workstation nodes. The encoded feature set is written into the node set and edge set. Node attributes include the forklift mobile robot's topological location, loading status, battery status, and current execution stage; the task's start point, end point, priority, and status; and the resource's mutual exclusion domain identifier, occupant identifier, and occupancy start and end times. Edge attributes include reachability distance, road network edge attributes, temporal characteristics, and mutual exclusion domain waiting association information. The node set, edge set, node attributes, and edge attributes are encapsulated to form a candidate subgraph feature package.

[0070] S3. Perform deep learning inference on the candidate subgraph feature package, output the scheduling inference result package, and assemble the scheduling inference result package into a spatiotemporal plan draft.

[0071] S3.1. Perform tensor quantization transformation on the candidate subgraph feature package to form an inference tensor package. Input the inference tensor package into the deep learning inference model to obtain the joint representation.

[0072] Specifically, based on the node set, edge set, and the correspondence between node attributes and edge attributes in the candidate subgraph feature package, the node attributes and edge attributes of the fork-type mobile robot nodes, task nodes, and resource nodes are encoded with fixed length, arranged in order, and stacked as tensors to form an inference tensor package. Then, the node attributes and edge attributes in the inference tensor package are aligned and combined according to the edge connection relationship, and input into the deep learning inference model to obtain a joint representation.

[0073] S3.2. The deep learning inference model includes a message passing layer, an attention aggregation layer, a joint representation construction layer, and an output mapping layer. The message passing layer propagates and merges neighborhood information between nodes and edges along the edge structure of the graph to form intermediate representations of nodes.

[0074] Specifically, the deep learning inference model consists of a message passing layer, an attention aggregation layer, a joint representation construction layer, and an output mapping layer connected sequentially. The message passing layer receives node and edge attributes from the inference tensor packet and propagates and fuses neighborhood information between nodes and edges along the connection relationships corresponding to the edge set, forming an intermediate node representation. The attention aggregation layer assigns weights to different neighborhood information in the intermediate node representation and performs weighted fusion to generate a node embedding representation. The joint representation construction layer extracts the two-end node embeddings corresponding to each candidate matching edge from the node embedding representation and constructs a joint representation by combining the corresponding edge attributes. The output mapping layer maps the joint representation and outputs the scheduling score and time window parameter information corresponding to the candidate matching edge. Based on the connection relationships recorded in the edge set, the information content corresponding to the joint representation is transmitted edge by edge along the edge structure of the graph between connected nodes and edges, and the association information from the neighborhood is summarized and fused to form an intermediate node representation reflecting adjacency relationships, path constraint relationships, and resource occupancy relationships.

[0075] S3.3. The attention aggregation layer assigns weights to different neighborhood information in the intermediate representation of a node and performs weighted fusion to generate a node embedding representation. The joint representation construction layer extracts the two-end node embeddings corresponding to each candidate matching edge from the node embedding representation to construct a joint representation. The output mapping layer maps the joint representation to generate scheduling scores and time window parameter information, and forms an edge-level inference set through mapping and slicing.

[0076] Specifically, the attention aggregation layer calculates the strength of the influence of different neighborhood information on the formation of the current node representation based on the correspondence between the neighborhood information and the current node's associated information in the intermediate node representation, and assigns corresponding weights to each neighborhood information. Then, it performs weighted fusion of each neighborhood information to generate a node embedding representation. The joint representation construction layer extracts the two-end node embeddings corresponding to each candidate matching edge from the node embedding representation, and concatenates or combines them with the edge attribute content corresponding to the candidate matching edge to construct a joint representation. The output mapping layer maps the joint representation to obtain the scheduling score and time window parameter information corresponding to each candidate matching edge, and performs result alignment, field segmentation, and sequential arrangement according to the order of candidate matching edges to form an edge-level inference set.

[0077] S3.4. Align and encapsulate the node and edge identifiers in the edge-level inference set and the candidate subgraph feature package, and output the scheduling inference result package.

[0078] Specifically, according to the correspondence between edge identifiers and node identifiers in the candidate subgraph feature package, the reasoning results of each edge in the edge-level reasoning set are read one by one. The position of each edge reasoning result is aligned and the identifier is associated with the corresponding edge identifier. The corresponding node identifier is extracted according to the relationship between the two nodes connected by the edge identifier. Then, the scheduling score, time window parameter information, edge identifiers and node identifiers are encapsulated in a unified record format so that each record corresponds to a unique candidate matching edge and a unique node connection relationship. Finally, all encapsulated records are summarized and arranged in the order of candidate matching edges, and the scheduling reasoning result package is output.

[0079] S3.5. Read the scheduling score and time window parameter information in the scheduling inference result package and associate them with the corresponding candidate combinations to form a weighted candidate set. Based on the weighted candidate set, construct a candidate execution sequence for each task and a candidate order-taking sequence for each forklift mobile robot to form a sequence draft set.

[0080] Specifically, based on the alignment relationship between edge identifiers and node identifiers in the scheduling inference result package, the fork-type mobile robot node and task node corresponding to the candidate matching edge are located, and the corresponding scheduling score and time window parameter information are written back to the combination record in the candidate combination set where the fork-type mobile robot identifier and task identifier are consistent, forming a weighted candidate set; the combination records in the weighted candidate set are summarized using the task identifier and fork-type mobile robot identifier as indices, and sorted by scheduling score to form candidate execution sequence and candidate order receiving sequence, while retaining the corresponding time window parameter information; the candidate execution sequence and candidate order receiving sequence are encapsulated by task identifier and fork-type mobile robot identifier to form a sequence draft set.

[0081] S3.6. Based on the sequence draft set, write the time window parameter information into the occupied time window corresponding to the task start position and the task end position to form a reservation draft set. Associate the reservation draft set with the corresponding road network travel path information and encapsulate it into a spatiotemporal plan draft.

[0082] Specifically, the candidate execution sequences in the sequence draft set are traversed according to the task identifier. The time window parameter information carried by each combination record is read and split into the task start station occupation time window and the task end station occupation time window. The time window is written into the corresponding occupation time window field. Combined with the task identifier, forklift robot identifier, task start station, and task end station, a reservation draft set is formed. The candidate order acceptance sequence is traversed according to the forklift robot identifier. The time window parameter information is associated with the mutual exclusion domain reservation field and written into the relevant records in the reservation draft set. Then, the road network travel path information corresponding to the topology position of the forklift robot to the task start topology node and the task start topology node to the task end topology node is retrieved. The road network travel path information is associated with the occupation time window field and the mutual exclusion domain reservation field to form a plan record set. The plan is then encapsulated to obtain the spatiotemporal plan draft.

[0083] S4. Based on the draft spatiotemporal plan, resource mutual exclusion and segment capacity constraint correction are performed through differentiable projection, and an executable scheduling instruction package is output.

[0084] S4.1. Read the occupancy time window, mutual exclusion domain reservation information and road network travel path information from the draft spatiotemporal plan, and convert them into a constraint representation set through constraint feature vectorization and constraint tensor quantization.

[0085] Specifically, the process involves traversing the plan records in the draft spatiotemporal plan, extracting the occupancy time window field, mutual exclusion domain reservation field, and road network travel path information field, while maintaining the correspondence between task identifiers, forklift robot identifiers, workstation node identifiers, and mutual exclusion domain identifiers. Constraint feature vectorization involves converting each constraint field in the occupancy time window, mutual exclusion domain reservation, and road network travel path information into numerical feature vectors of uniform length, and arranging them in the order of the plan records. Constraint tensorization involves stacking the sequentially arranged occupancy time window constraint feature vector, mutual exclusion domain reservation constraint feature vector, and road network travel path constraint feature vector to form a tensor, while retaining the plan record index and mutual exclusion domain identifier index. The resulting tensors are then encapsulated to form a constraint representation set.

[0086] It should be noted that the mutual exclusion domain reservation information includes a specific mutual exclusion domain identifier (such as a workstation or narrow passage), the reserved entry and exit times on the mutual exclusion domain, and an optional passage sequence identifier, which describes when and in what order a task or robot plans to use this mutual exclusion resource; the road network passage path information includes a sequence of road network topology node identifiers, and the road network edge attributes of each road segment in the sequence (such as length, capacity, and direction restrictions).

[0087] S4.2. Based on the constraint representation set, construct resource mutual exclusion constraints and road segment capacity constraints, map the occupancy time window and mutual exclusion domain reservation information to the resource mutual exclusion constraints, and map the road network traffic path information to the road segment capacity constraints to form a conflict representation set.

[0088] Specifically, the constraint representation set reads the tensor corresponding to the occupancy time window and the tensor corresponding to the mutual exclusion domain reservation information, and constructs resource mutual exclusion constraints using the mutual exclusion domain identifier and the workstation node identifier as grouping indexes. The resource mutual exclusion constraints are used to express the occupancy mutual exclusion relationship of the same mutual exclusion domain identifier or the same workstation node identifier during overlapping time periods. The occupancy time window and the mutual exclusion domain reservation information are mapped to the resource mutual exclusion constraints. In the mapping process, the occupancy start time, occupancy end time, reservation start time, reservation end time, and passage sequence identifier are written into the constraint items of the resource mutual exclusion constraints, and the correspondence between the task identifier and the forklift mobile robot identifier is retained. The constraint representation set reads the tensor corresponding to the road network passage path information, and constructs road segment capacity constraints using the road segment identifier and passage direction information in the road network edge attribute sequence as grouping indexes. The road network passage path information is mapped to the road segment capacity constraints. In the mapping process, the correspondence between the path passage time sequence and the road segment identifier is written into the constraint items of the road segment capacity constraints, and the correspondence between the task identifier and the forklift mobile robot identifier is retained. Based on the resource mutual exclusion constraints and the road segment capacity constraints, the constraint item markers corresponding to overlapping occupancy and overcapacity passage are calculated and encapsulated to form a conflict representation set.

[0089] S4.3. Perform a differentiable projection on the set of conflict representations and encapsulate the set of revised plans into an executable scheduling instruction package.

[0090] Specifically, the overlapping occupancy constraints and overcapacity passage constraints marked in the conflict representation set are read, and a differentiable projection variable representation is established using the occupancy time window mutual exclusion domain reservation information and road network passage path information as the objects to be corrected. Differentiable projection is performed on the overlapping occupancy constraints, and the occupancy start time, occupancy end time, reservation start time, and reservation end time in the occupancy time window and mutual exclusion domain reservation information are continuously differentiable shifted or adjusted in duration, so that the occupancy time window and reservation period corresponding to the same mutual exclusion domain identifier and the same workstation node identifier are aligned on the time axis. The process involves: performing differentiable projection on the overcapacity traffic constraint, continuously adjusting the traffic timing or inserting waiting timing into the path passage time sequence in the road network traffic path information to distribute the traffic volume corresponding to the same road segment identifier within the time range; summarizing the reservation information of the occupancy time window mutual exclusion domain after the differentiable projection is completed with the road network traffic path information into a correction plan set, which maintains the correspondence between task identifiers and forklift mobile robot identifiers, as well as the binding relationship between workstation node identifiers and mutual exclusion domain identifiers; and encapsulating the correction plan set into an executable scheduling instruction package.

[0091] S5. Parse the executable scheduling instructions to generate robot speed and heading control, and use a neural control barrier function for online correction, outputting the execution reference package.

[0092] S5.1. Read the modified occupancy time window, modified mutex domain reservation information, and modified passage arrangement from the executable scheduling instruction package, and parse the modified occupancy time window into arrival time constraints and docking duration constraints to form a set of control constraints.

[0093] Specifically, the plan records in the executable scheduling instruction package are traversed according to the task identifier and the forklift mobile robot identifier. The corrected occupancy time window fields corresponding to the task start and end workstations are extracted, along with the corrected mutual exclusion domain reservation information field corresponding to the mutual exclusion domain identifier and the corrected passage arrangement field. Simultaneously, the association between the workstation node identifier, the mutual exclusion domain identifier, and the road network topology node identifier is maintained. The corrected occupancy time window is parsed into arrival time constraints and docking duration constraints. The arrival time constraint is expressed by the occupancy start time of the corrected occupancy time window and bound to the workstation node identifier. The docking duration constraint is determined by the corrected... The duration difference between the end time and start time of the occupancy window is expressed and bound to the task identifier and workstation node identifier; the modified mutual exclusion domain reservation information is parsed into mutual exclusion domain entry time constraint, mutual exclusion domain exit time constraint and passage order constraint and bound to the mutual exclusion domain identifier; the modified passage arrangement is parsed into path passage order and waiting node order and bound to the road network topology node identifier; the arrival time constraint, stopping time constraint, mutual exclusion domain entry time constraint, mutual exclusion domain exit time constraint, passage order constraint, path passage order and waiting node order are uniformly encapsulated to form a set of control constraints.

[0094] S5.2. Based on the control constraint set and the current pose and cargo status of the forklift mobile robot, generate robot speed and heading control, and at the same time collect the relative pose, relative speed and mutual exclusion domain occupancy status of neighboring robots to form a barrier constraint set.

[0095] Specifically, the system reads the arrival time constraint, docking duration constraint, mutual exclusion domain entry time constraint, mutual exclusion domain exit time constraint, passage order constraint, path traversal order, and waiting node order from the control constraint set according to the forklift mobile robot identifier, and associates them with the current pose and cargo status of the forklift mobile robot in the unified state package. Based on the path traversal order and waiting node order, segmented travel targets are generated on the road network topology node sequence, and the arrival time constraint and mutual exclusion domain entry time constraint are used as the time constraints of the segmented travel targets, and the cargo status is used as the speed limit basis for the segmented travel targets, thus forming robot speed and heading control for each segmented travel target. The system collects the relative pose, relative speed, and mutual exclusion domain occupancy status of neighboring robots. The relative pose and relative speed of neighboring robots are calculated from the current pose and speed information of different forklift mobile robots in the unified state package, and the mutual exclusion domain occupancy status is obtained from the mutual exclusion domain identifier, occupant identifier, and occupancy start and end time information in the unified state package. The robot speed and heading control are bound and encapsulated with the relative pose, relative speed, and mutual exclusion domain occupancy status to form a barrier constraint set.

[0096] S5.3. Based on the neural control barrier function, perform online correction on the barrier constraint set to obtain the safety control set, and encapsulate the safety control set, the correction occupancy time window, and the correction mutual exclusion domain reservation information into an execution baseline package.

[0097] Specifically, the robot speed and heading control, relative pose, relative velocity, and mutual exclusion domain occupancy status in the barrier constraint set are read according to the fork-type mobile robot identifier. The relative pose, relative velocity, and mutual exclusion domain occupancy status are then converted into barrier constraint conditions. These barrier constraint conditions express the constraint relationship between the safety distance constraint and the mutual exclusion domain passage order constraint at the current moment. Within each control update cycle, the robot speed and heading control are adjusted in a projective manner based on the barrier constraint conditions. This projective adjustment replaces the control components in the robot speed and heading control that cause safety distance contraction or mutual exclusion domain passage order conflict with control components that satisfy the barrier constraint conditions, thereby obtaining a safety control set. The safety control set is then consistently associated with the corrected occupancy time window and corrected mutual exclusion domain reservation information in the executable scheduling instruction package and encapsulated according to the task identifier, fork-type mobile robot identifier, and mutual exclusion domain identifier to form an execution baseline package.

[0098] It should be noted that the online correction of the barrier constraint set based on the neural control barrier function is performed in each control update cycle by calculating the barrier function value and its changing trend according to the relative pose, relative velocity and mutual exclusion domain occupancy status in the barrier constraint set, and then making projection adjustments to the robot's speed and heading control.

[0099] Furthermore, such as Figure 6 The figure shows a time-series comparison of the minimum safe distance between the proposed solution and the control solution. The upper figure is an overview, showing the trend of the minimum safe distance changes for both solutions throughout the entire passing process. The lower figure is a magnified view, highlighting the differences between the two solutions within the passing risk zone. In the figure, the black curve represents the control solution, the red curve represents the proposed solution, and the blue dashed line represents the safety distance threshold of 1.2 meters. As can be seen from the figure, the minimum safe distance of the control solution drops significantly in the middle of the passing phase and falls below the safety distance threshold, while the proposed solution remains near or above the safety distance threshold during the same period, with relatively small overall fluctuations. This indicates that... The present invention, by parsing the executable scheduling instruction package and combining it with a neural control barrier function for online correction, can more effectively maintain the safety margin during vehicle passing and improve execution safety. The contrasting scheme is a baseline scheduling scheme that only performs routine task allocation and path execution based on the robot's current position, the task target position, and the currently available resources. It does not perform unified timestamp alignment and topology mapping for robot state, task state, and resource occupancy state, does not generate a unified state package, does not assemble the scheduling results into a spatiotemporal plan draft, does not perform differentiable projection correction, and does not use a neural control barrier function for online safety correction.

[0100] In summary, this invention achieves a unified expression of scheduling input information by performing time alignment and topological mapping on robot state, task state, and resource occupancy state. Furthermore, by assembling the inference results into a spatiotemporal plan draft, it achieves a smooth transition from learning results to an executable scheduling plan. This invention is suitable for safe handling and collaborative operations in complex dynamic environments, improving execution safety, environmental adaptability, and task completion effectiveness.

[0101] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A deep learning-based scheduling method for forklift mobile robots, characterized in that, include: Collect robot status, task, and resource usage data, perform timestamp alignment and topology mapping, and generate a unified status package; Feasibility filtering and proximity constraints are applied to the unified state package to construct a candidate combination set, and road network and resource context features are encoded to form a candidate subgraph feature package. Deep learning inference is performed on the candidate subgraph feature package to output the scheduling inference result package, and the scheduling inference result package is assembled into a spatiotemporal plan draft. Based on the draft spatiotemporal plan, resource mutual exclusion and segment capacity constraint correction are performed through differentiable projection, and an executable scheduling instruction package is output. The executable scheduling instructions are parsed to generate robot speed and heading control, and a neural control barrier function is used for online correction, outputting an execution reference package.

2. The deep learning-based fork-type mobile robot scheduling method as described in claim 1, characterized in that, The process of collecting robot status, task, and resource usage data, performing timestamp alignment and topology mapping, and generating a unified status package involves the following steps: Collect robot status data, task data, and resource usage data of the forklift mobile robot, and convert the time information attached to each data to the same scheduling clock reference to form a standardized data set; Time slice snapshots are constructed based on standardized data sets, and robot state, task state, and resource occupancy state within the same time slice are aggregated to obtain a set of time slice snapshots; Perform topology mapping on the time slice snapshot set, project the robot pose onto the road network topology nodes, associate the task start and end points with corresponding workstation nodes, associate resource usage data with mutual exclusion domain identifiers, output a topological snapshot set, perform consistency verification on the topological snapshot set, and generate a unified state package.

3. The deep learning-based fork-type mobile robot scheduling method as described in claim 2, characterized in that, The robot status, task, and resource occupancy data include the forklift robot's pose information, cargo status, battery status, current execution stage, workstation occupancy status, and narrow domain occupancy status. The task data includes the task identifier, task start point, and task end point.

4. The deep learning-based fork-type mobile robot scheduling method as described in claim 1, characterized in that, The specific steps for performing feasibility filtering and proximity constraints on the unified state packet to construct a candidate combination set are as follows: Extract the set of tasks to be assigned and the set of single-fork mobile robots that can be connected from the unified state packet, and collect the capability parameters of the fork mobile robots and the loading and unloading constraint parameters of the tasks to form an initial matching set; Perform feasibility filtering on the initial matching set and output a feasible matching set. Perform proximity constraint filtering on the feasible matching set and calculate the reachable distance based on the topological position of the forklift mobile robot and the topological node of the task starting point, and output a candidate combination set.

5. The deep learning-based fork-type mobile robot scheduling method as described in claim 1, characterized in that, The specific steps for encoding road network and resource context features to form candidate subgraph feature packages are as follows: The candidate combination set is encoded with road network and resource context features. Temporal features are constructed based on road network edge attributes and resource mutual exclusion domain states to form an encoded feature set. The fork-type mobile robot nodes, task nodes, and resource nodes are treated as a set of nodes, and feasible matching relationships and occupancy associations are treated as a set of edges. The encoded feature set is written into the node set and the edge set to form a candidate subgraph feature package.

6. The deep learning-based fork-type mobile robot scheduling method as described in claim 1, characterized in that, The specific steps for performing deep learning inference on the candidate subgraph feature package and outputting the scheduling inference result package are as follows: Tensor quantization transformation is performed on the candidate subgraph feature package to form an inference tensor package. The inference tensor package is then input into the deep learning inference model to obtain the joint representation. The deep learning inference model includes a message passing layer, an attention aggregation layer, a joint representation construction layer, and an output mapping layer; the message passing layer propagates and merges neighborhood information between nodes and edges along the edge structure of the graph to form intermediate node representations; The attention aggregation layer assigns weights to different neighborhood information in the intermediate representation of a node and performs weighted fusion to generate a node embedding representation. The joint representation construction layer extracts the two node embeddings corresponding to each candidate matching edge from the node embedding representation to construct a joint representation. The output mapping layer maps the joint representation to generate scheduling scores and time window parameter information and forms an edge-level inference set through mapping and slicing. Align and encapsulate the node and edge identifiers in the edge-level inference set and the candidate subgraph feature package, and output the scheduling inference result package.

7. The deep learning-based fork-type mobile robot scheduling method as described in claim 1, characterized in that, The specific steps for assembling the scheduling inference results into a spatiotemporal plan draft are as follows: Read the scheduling score and time window parameter information from the scheduling inference result package and associate them with the corresponding candidate combinations to form a weighted candidate set. Based on the weighted candidate set, construct a candidate execution sequence for each task and a candidate order-taking sequence for each forklift mobile robot to form a sequence draft set. Based on the sequence draft set, the time window parameter information is written into the occupied time window corresponding to the task start position and the task end position to form the reservation draft set. The reservation draft set is associated with the corresponding road network traffic path information and encapsulated into a spatiotemporal plan draft.

8. The deep learning-based fork-type mobile robot scheduling method as described in claim 7, characterized in that, The process of performing resource mutual exclusion and road segment capacity constraint correction through differentiable projection based on the spatiotemporal plan draft, and outputting an executable scheduling instruction package, is as follows: Read the occupancy time window, mutual exclusion domain reservation information and road network travel path information from the draft spatiotemporal plan, and convert them into a set of constraint representations through constraint feature vectorization and constraint tensor quantization; Based on the constraint representation set, resource mutual exclusion constraints and road segment capacity constraints are constructed. The occupancy time window and mutual exclusion domain reservation information are mapped to the resource mutual exclusion constraints, and the road network travel path information is mapped to the road segment capacity constraints, forming a conflict representation set. Perform a differentiable projection on the set of conflict representations and encapsulate the set of corrected plans into an executable scheduling instruction package.

9. The deep learning-based fork-type mobile robot scheduling method as described in claim 8, characterized in that, The steps for parsing executable scheduling instructions to generate robot speed and heading control, using a neural control barrier function for online correction, and outputting an execution baseline package are as follows: Read the modified occupancy time window, modified mutual exclusion domain reservation information and modified passage arrangement from the executable scheduling instruction package, and parse the modified occupancy time window into arrival time constraints and docking duration constraints to form a set of control constraints; Based on the control constraint set and the current pose and cargo status of the forklift mobile robot, the robot speed and heading control are generated. At the same time, the relative pose, relative speed and mutual exclusion domain occupancy status of neighboring robots are collected to form a barrier constraint set. The set of barrier constraints is modified online by using a neural control barrier function to obtain a set of safety controls. The set of safety controls, the modified occupancy time window, and the modified mutual exclusion domain reservation information are then encapsulated into an execution baseline package.

10. The deep learning-based fork-type mobile robot scheduling method as described in claim 9, characterized in that, The online correction of the barrier constraint set through neural control barrier function refers to calculating the barrier function value and its changing trend based on the relative pose, relative velocity and mutual exclusion domain occupancy status in the barrier constraint set during each control update cycle, and then making projection-based adjustments to the robot's speed and heading control.

Images