Humanoid robot scheduling method in humanoid robot training ground

Abstract

The invention relates to the technical field of robot scheduling, in particular to a humanoid robot scheduling method in a humanoid robot training ground, and the method comprises the steps: collecting global environment data in the training ground in real time through a multi-modal sensor array, and fusing the global environment data with the obtained real-time pose information of a robot; analyzing the obtained training task queue into atomic operation data, and associating the atomic operation data with the constructed training ground resource table; based on the dynamic passage map and the training ground resource table, the resource constraint operation chain is dynamically allocated, and a scheduling path is generated after space-time joint path planning; and the generated allocation task is transmitted to the corresponding humanoid robot, and the dynamic passage map and the training ground resource table are updated based on the obtained execution data, so that the efficient and conflict-free real-time scheduling method capable of adapting to the motion characteristics of the humanoid robot is realized, the task and path conflict between the robots is avoided, and the real-time scheduling efficiency of the humanoid robot is improved. The overall utilization rate and the training efficiency of the training ground are improved.

Description

Technical Field

[0001] This invention relates to the field of robot scheduling technology, and in particular to a method for scheduling humanoid robots in a humanoid robot training field. Background Technology

[0002] In the context of the rapid development of robotics technology, humanoid robots, with their anthropomorphic form and maneuvering capabilities, are expected to perform tasks in complex human environments in the future, such as factories, warehouses, disaster relief sites, or homes. To equip humanoid robots with the ability to cope with complex environments, it is necessary to construct specialized "training grounds" for extensive, parallel algorithm training and task simulation.

[0003] In existing technologies, the scheduling and management methods for such training grounds are typically quite rudimentary. A common approach is to use a central control system to pre-assign fixed task areas and static task sequences to each robot. This method has significant drawbacks: First, the training ground environment is dynamically changing, and there may be temporary obstacles or other robots. Fixed path planning cannot handle these unexpected situations, easily leading to path conflicts or even deadlocks between robots, thus interrupting training. Second, when multiple robots compete for the same resource (such as a training prop or a charging station), the lack of an effective global coordination mechanism may cause some robots to wait for extended periods while other resources remain idle, resulting in low overall training efficiency. Finally, existing scheduling methods usually only focus on "reaching the target point," ignoring the unique dynamic balance and gait adjustment characteristics of humanoid robots during task execution, and failing to consider the risks to their movement safety and stability under different task states. Therefore, existing technologies cannot achieve an efficient, conflict-free, and adaptable real-time scheduling method for a dynamic, multi-task parallel humanoid robot training ground to avoid task and path conflicts between robots and improve the overall utilization and training efficiency of the training ground. Summary of the Invention

[0004] The purpose of this invention is to provide a humanoid robot scheduling method in a humanoid robot training field. This method enables a real-time scheduling method that is efficient, conflict-free, and adaptable to the motion characteristics of humanoid robots in a dynamic, multi-task parallel humanoid robot training field. This avoids task and path conflicts between robots and improves the overall utilization and training efficiency of the training field.

[0005] To achieve the above objectives, the present invention provides a method for scheduling humanoid robots in a humanoid robot training field, comprising the following steps: A multimodal sensor array is used to collect global environmental data in the training field in real time, and the data is fused with the real-time pose information of the robot to generate a dynamic access map. The acquired training task queue is parsed into atomic operation data and associated with the constructed training field resource table to obtain the resource constraint operation chain; Based on the dynamic access map and the training ground resource table, the resource constraint operation chain is dynamically allocated, and a scheduling path is generated after spatiotemporal joint path planning. The generated assignment task and the scheduling path are transmitted to the corresponding humanoid robot, and the dynamic access map and the training ground resource table are updated based on the acquired execution data.

[0006] This involves using a multimodal sensor array to collect real-time global environmental data within the training field, and fusing this data with the acquired real-time robot pose information to generate a dynamic access map, including: A multimodal sensor array is used to collect 3D point cloud data and depth image data in the training field in real time; The robot's acceleration, angular velocity, and the angles of its joints are obtained to calculate the estimated pose. At the same time, the absolute pose is calculated by identifying visual markers deployed at known locations on the training field. The estimated pose and the absolute pose are fused using a Kalman filter to obtain the robot's real-time pose information. The robot's real-time pose information, the 3D point cloud data, and the depth image data are fused to generate a dynamic access map.

[0007] The process of fusing the robot's real-time pose information, the 3D point cloud data, and the depth image data to generate a dynamic access map includes: The coordinates of the 3D point cloud data and the depth image data are unified and fused, and the time of the multimodal sensor array and the humanoid robot are synchronized. Statistical analysis is performed on the three-dimensional point cloud data of multiple consecutive frames to obtain a static background model, and dynamic objects are obtained based on the static background model. The dynamic object is associated with the robot's real-time pose information to obtain the robot's identification code and generate a predicted trajectory band. A dynamic access map is generated based on the static background model, the predicted trajectory band, and the robot's real-time location information.

[0008] The multimodal sensor array includes a three-dimensional lidar cluster and panoramic vision sensor nodes; The three-dimensional lidar cluster is used to perform a rotating scan of the entire site from a top-down angle to obtain three-dimensional point cloud data. The panoramic vision sensor node is used to calculate the distance between each pixel in the image and the camera while acquiring a conventional RGB color image using stereo binocular vision or structured light technology, thereby obtaining depth image data.

[0009] The acquired training task queue is parsed into atomic operation data and associated with the constructed training field resource table to obtain a resource constraint operation chain, including: The acquired training task queue is subjected to syntax and semantic validation, and after passing the validation, it is marked as pending parsing and cached in the server. Keyword extraction and intent recognition are performed on the training task queue. Based on the recognized intent, the corresponding atomic operation template is matched, and a resource constraint operation chain is generated by combining it with the constructed training field resource table.

[0010] The method further includes: Label each atomic operation in the resource constraint operation chain with the execution prerequisites and the execution result.

[0011] Specifically, based on the dynamic access map and the training ground resource table, the resource constraint operation chain is dynamically allocated, and a scheduling path is generated after spatiotemporal joint path planning, including: When the set trigger event occurs, the constructed candidate matrix is ​​used to traverse the resource constraint operation chain and the training field resource table, and the dynamic communication map is referenced to determine whether the path is unobstructed. The comprehensive cost score of each scheduling scheme is calculated using a multi-objective optimization algorithm, and the scheduling and task allocation scheme with the minimum total comprehensive cost is obtained by weighted summation. Based on the scheduling and allocation of tasks, a spatiotemporal trajectory band is generated, and after conflict resolution based on the dynamic priority of tasks, a scheduling path is obtained.

[0012] The comprehensive cost score includes the time cost score, energy consumption cost score, system equilibrium cost score, and stability cost score.

[0013] The process includes transmitting the generated assignment task and the scheduling path to the corresponding humanoid robot, and updating the dynamic access map and the training ground resource table based on the acquired execution data, including: The generated allocation task and the scheduling path are assembled into a synchronization instruction package and transmitted through the constructed communication registry. After receiving the synchronization command packet, the humanoid robot uploads the actual execution status and execution location in real time. The dynamic access map and the training ground resource table are updated based on the uploaded actual execution status and execution location.

[0014] The method further includes: Cross-validate and monitor the actual execution status and execution location of the uploaded data; The execution location is compared with the spatiotemporal trajectory in real time, and the comparison result is judged against the set distance threshold to obtain the corresponding deviation result; A graded early warning system will be implemented based on the deviation results.

[0015] This invention discloses a humanoid robot scheduling method in a humanoid robot training field. It utilizes a multimodal sensor array to collect real-time global environmental data within the training field and fuses this data with real-time robot pose information to generate a dynamic access map. The acquired training task queue is parsed into atomic operation data and associated with a constructed training field resource table to obtain a resource constraint operation chain. Based on the dynamic access map and the training field resource table, the resource constraint operation chain is dynamically allocated, and a scheduling path is generated after spatiotemporal joint path planning. The generated allocated tasks and the scheduling path are transmitted to the corresponding humanoid robots. Simultaneously, the dynamic access map and the training field resource table are updated based on the acquired execution data. This method enables a highly efficient, conflict-free, and adaptable real-time scheduling method in a dynamic, multi-task parallel humanoid robot training field, avoiding task and path conflicts between robots and improving the overall utilization and training efficiency of the training field. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below.

[0017] Figure 1 This is a schematic diagram illustrating the steps of a humanoid robot scheduling method in a humanoid robot training field according to the first embodiment of the present invention.

[0018] Figure 2 This is a flowchart illustrating a humanoid robot scheduling method in a humanoid robot training field provided by the present invention. Detailed Implementation

[0019] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numbers in different drawings represent the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this application.

[0020] The terminology used in this application is for the purpose of describing particular embodiments only and is not intended to be limiting of the application. The singular forms “a,” “the,” and “the” used in this application and the appended claims are also intended to include the plural forms unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used herein refers to and includes any or all possible combinations of one or more of the associated listed items.

[0021] It should be understood that although the terms first, second, third, etc., may be used in this application to describe various information, such information should not be limited to these terms. These terms are only used to distinguish information of the same type from one another. For example, without departing from the scope of this application, first information may also be referred to as second information, and similarly, second information may also be referred to as first information. Depending on the context, the word "if" as used herein may be interpreted as "when," "when," or "in response to determination."

[0022] Please see Figures 1-2 This invention provides a method for scheduling humanoid robots in a humanoid robot training field, comprising the following steps: S101. Use a multimodal sensor array to collect global environmental data in the training field in real time, and fuse it with the acquired real-time robot pose information to generate a dynamic access map.

[0023] Specifically, multimodal sensor arrays mainly include two types of sensors: A 3D lidar cluster is deployed at key locations within the training ground (such as the four corners of the ceiling or the central pillar), performing a high-speed rotating scan of the entire area from a top-down perspective. The core data it collects is high-precision 3D point cloud data. Each data point contains its 3D spatial coordinates (X, Y, Z) relative to the lidar itself. Acquisition method: The lidar directly obtains distance information to object surfaces by emitting laser beams and calculating their return time. Tens of thousands of point clouds constitute the "skeleton" or "outline" of the training ground's physical environment.

[0024] Panoramic vision sensor nodes: Deployed above key areas of the training field, complementing the location of the LiDAR. The core data they acquire consists of high-resolution color video streams and depth images. Acquisition method: Employing stereo binocular vision or structured light technology, while acquiring a regular RGB color image, the distance between each pixel in the image and the camera is calculated, thereby generating a depth map that is pixel-by-pixel aligned with the color image.

[0025] LiDAR provides precise, illumination-independent geometric information, but lacks color and texture. Visual sensors offer rich texture, color, and pattern recognition capabilities, but depth information can be affected by lighting and object surface materials. Combining the two allows for complementary advantages, providing a richer data foundation for subsequent environmental understanding.

[0026] Real-time robot pose information is generated by a robot state reporting module deployed on each humanoid robot body. This module integrates multiple sensors and has data processing and wireless communication capabilities. Pose information acquisition includes: Internal state perception: The robot uses its own inertial measurement unit and joint encoders to read the body's acceleration, angular velocity, and the angles of each joint in real time. The body state calculation unit (a built-in algorithm) fuses this data to calculate the robot's estimated pose relative to its initial pose, including 3D coordinates and body orientation. External absolute positioning correction: Relying solely on internal sensors will result in accumulated errors. Therefore, the robot uses a vision sensor on its head to identify visual markers deployed at known locations on the walls or pillars around the training area. By observing these markers and combining them with a camera model, the robot's absolute pose relative to the fixed coordinate system of the training area can be accurately calculated.

[0027] The internal predicted pose is fused with external absolute positioning information using Kalman filtering to obtain an optimal, drift-free real-time pose data packet. This data packet contains the three-dimensional coordinates and orientation of the robot's chassis center point, as well as a unique robot identification code and a timestamp. This data packet is then uploaded to the central dispatch system in real-time via a high-speed wireless communication network covering the entire training area at a fixed high frequency (e.g., 50Hz).

[0028] Upon receiving the data packet, the system first unifies all data into the same global coordinate system of the training field. The system pre-stores the precise position and orientation of each multimodal sensor array node, allowing the conversion of their acquired local coordinate data (such as point clouds from LiDAR and images from cameras) into the global coordinate system, achieving coordinate system unification. Simultaneously, the clocks of all sensors and the robot are strictly synchronized with the master clock of the central scheduling system, ensuring that each frame of data has a precise timestamp and aligning data from different sources on the timeline. The point cloud generated by the 3D LiDAR cluster is then registered with the color depth images generated by the panoramic vision sensor nodes. Specifically, the 3D points in the point cloud are projected onto the image plane of the vision sensor, assigning each 3D point corresponding color and texture information to form a color point cloud.

[0029] Statistical analysis is performed on multiple consecutive frames of color point clouds to establish a static background model for the training field. All continuous, stable, and stationary points (such as the ground, walls, and fixed training facilities) are categorized into the static environment layer. The color point cloud of the current frame is compared with the static background model, and point cloud clusters whose positions have changed or which have newly appeared are identified as dynamic objects. These dynamic point cloud clusters are segmented into independent individuals using a clustering algorithm. The extracted dynamic point cloud clusters are then associated with the real-time pose data packets uploaded by the robot status reporting module. By comparing the geometric center and contour size of the dynamic point cloud clusters with their coordinate positions in the pose data packets, the system can assign a unique identity—a robot identification code—to each moving point cloud cluster. For example, the system confirms that "moving object A" is "robot 001".

[0030] The system acquires the historical movement trajectory of each identified humanoid robot and uses a kinematic model to predict its possible spatial location in the near future (e.g., 2-3 seconds) based on its current speed and direction, forming a predicted trajectory band.

[0031] All the above information is integrated into a unified, dynamic access map. This map has a multi-layered data structure: Basic static layer: Composed of static environment layers, clearly marking fixed obstacles and permanent passable areas.

[0032] Real-time occupancy layer: In the form of a raster or vector map, the current position contours of all robots are marked in real time (not just the center point, but the entire envelope of its chassis or arms after deployment).

[0033] Predicted trajectory layer: This layer displays the predicted trajectory band for each robot overlaid with semi-transparent or differently colored areas. This trajectory band represents the space that the robot "may traverse" in the future and is crucial for collision detection.

[0034] S102. Parse the acquired training task queue into atomic operation data and associate it with the constructed training field resource table to obtain the resource constraint operation chain.

[0035] Specifically, this is done through a standardized task description language interface. Upper-level management systems (such as training simulation platforms and task orchestration systems) submit tasks in a structured data format (such as specific schemas based on JSON or XML) through this interface.

[0036] A complete task request package must contain at least: Task unique identifier: used to track the task throughout the entire system lifecycle.

[0037] Task type: Select from the predefined type library, such as "item handling", "equipment operation", "complex process drill", etc.

[0038] Task objective description: Describe it using natural language combined with key parameters, for example: "Transport the red toolbox from storage area A1 to maintenance workbench B2".

[0039] Task constraints include priority (e.g., "high", "medium", "low"), latest completion time, and allowed number of retries.

[0040] Upon receiving a task request, the system first performs syntax and semantic validation to ensure that the task type is known, all necessary parameters are present, and the format conforms to specifications. If validation passes, a server is created for the task, its status is marked as "pending parsing," and it is then placed in the task parsing queue for processing.

[0041] The task objective description is read from the server, and natural language processing technology is used to extract keywords and identify intents in the task objective description. For example, for the task "transport the red toolbox from storage area A1 to maintenance workbench B2", the core verb "transport" and key entities "red toolbox", "storage area A1" and "maintenance workbench B2" are identified. Based on the identified intent, the most suitable composite task template is matched, where the template is selected from the constructed atomic operation template library.

[0042] When generating the sequence, a "simulated association" is performed beforehand with the training field resource table. It discovers that the "grab toolbox" operation not only requires the robot to reach the "storage area A1" location, but more specifically, it requires occupying the physical resource of the "red toolbox" itself. Similarly, "put down the toolbox" requires occupying an empty slot on the "maintenance workbench B2," thus initially binding atomic operations to specific resource IDs, ultimately resulting in a resource-constrained operation chain. Each atomic operation clearly indicates the target resource it requires: Atomic operation 1 (navigation): target location = storage area A1 (associated resource: spatial area A1).

[0043] Atomic operation 2 (fetch): target resource = red toolbox (associated resource: physical object "red toolbox").

[0044] Atomic Operation 3 (Navigation): Target Location = Maintenance Workbench B2 (Associated Resource: Spatial Region B2).

[0045] Atomic Operation 4 (Placement): Target Resource = Placement Area of ​​Repair Workbench B2 (Associated Resource: Physical Location "Workbench B2 - Placement Area").

[0046] To further ensure logical rigor, each atomic operation in the resource constraint operation chain is labeled with its execution preconditions and execution results.

[0047] The prerequisites for atomic operation 2 (grab) are: "The robot is within the operable range of the red toolbox" AND "The red toolbox is in the 'idle' state".

[0048] The result after execution is: "The red toolbox status changes to 'occupied'", and it is bound to the robot ID that executed the task.

[0049] These conditions constitute a closed loop of internal logical consistency within the task, and are also the core basis for subsequent conflict detection and resource status management.

[0050] The training ground resource table is a core dynamic database within the central scheduling system, maintained by the resource status management module. This table abstracts all key entities in the training ground as schedulable "resources".

[0051] During system initialization, all training facilities must be registered in the resource table. Each resource record contains: Unique identifier for resources: such as Tool_Red_01, Workstation_B2.

[0052] Resource types: such as "tools", "workbench", "charging station", "access gate".

[0053] Resource space attributes: its location coordinates and the space it occupies (a three-dimensional cube).

[0054] Resource status: The initial status is "idle".

[0055] When a task begins to execute atomic operations, it sends a "reservation" or "occupancy" request to the training ground resource table and updates the resource status accordingly.

[0056] A multimodal sensor array continuously monitors critical resources. For example, a vision sensor detects that the "red toolbox" is not in its proper location A1, but has been accidentally placed on the ground. In this case, the resource table will forcibly update the status of the resource's unique identifier to "abnormal location" and issue an alert. This prevents the system from making scheduling decisions based on erroneous status information.

[0057] If all resources are "idle" or "soon to be idle" (about to be released), the task is considered "assignable".

[0058] If a critical resource (such as the "red toolbox") is in a "occupied" or "abnormal" state, the system will: a) wait: if the task has a high priority, wait for the resource to be released. b) replan: if the task allows, the task parsing engine may be triggered to perform dynamic re-parsing. For example, after detecting an aberration in the "red toolbox," the engine may automatically replace the task objective with "complete the repair using the blue toolbox" according to rules, and generate a new resource constraint operation chain associated with the resource's unique identifier.

[0059] Once a task is assigned to a robot and begins execution, the system "resers" the resources to be used in the resource table (to prevent them from being preempted by other tasks). When the atomic operation is actually executed (such as a successful grab), the robot will send a "operation completed" signal, and the resource status management module will then officially update the resource status to "occupied" and record the occupant.

[0060] S103. Based on the dynamic access map and the training ground resource table, the resource constraint operation chain is dynamically allocated, and a scheduling path is generated after spatiotemporal joint path planning.

[0061] Specifically, when a trigger event occurs, the central scheduling system's dynamic allocation engine creates a candidate matrix. It iterates through all pending resource constraint operation chains and all robots in the "idle" or "about to be idle" state. For each (task, robot) combination, the engine performs a rapid feasibility assessment: Check the training ground resource table to see if all the resources required for the task can be reserved by the robot.

[0062] Referring to the dynamic traffic map, it can be preliminarily determined whether the path from the robot's current location to the task starting point is basically unobstructed (without long-term blockages).

[0063] The triggering events include: A new task has arrived in the task parsing queue.

[0064] Some robots have completed their current tasks ahead of schedule and entered an idle state.

[0065] High-priority tasks are being inserted.

[0066] The system detected that the current allocation plan is not feasible due to an unexpected event (such as robot failure or resource occupation).

[0067] For combinations of tasks and robots that pass the feasibility assessment, the dynamic allocation engine activates its core multi-objective optimization algorithm to calculate a comprehensive cost score for each combination. A lower score indicates a better allocation scheme. This algorithm considers the following objectives: Objective 1: The Cost of Time Estimation content: Estimating the total time required to sequentially execute all atomic operations in the resource constraint operation chain of this task, starting from the robot's current pose.

[0068] Calculation methods include: ① travel time to the task starting point (based on the current congestion situation in the dynamic traffic map); ② estimated execution time of each operation (such as the fixed time for grabbing and placing); ③ possible waiting time at resource points (based on the prediction of resource release).

[0069] Travel time prediction based on spatiotemporal trajectory. Unlike simply dividing the straight-line distance by the speed, this method combines generated scheduling paths to quickly generate a preliminary spatiotemporal trajectory and calculates the accurate travel time accordingly. This trajectory also preliminarily avoids known static obstacles.

[0070] Objective Two: Energy Consumption Cost Estimate content: A comprehensive assessment of the energy consumption of the robot in performing this task.

[0071] Calculation method: Consider: ① Total distance traveled; ② Task operation type (e.g., grasping heavy objects consumes more power than walking without a load); ③ Robot's current battery level. Assigning a high-energy-consuming task to a robot with low battery will incur a high penalty, encouraging it to recharge first.

[0072] Objective 3: System Equilibrium Cost Estimate the impact of this allocation on the overall system load balancing and resource utilization.

[0073] Calculation method: Resource contention: If the resources required for this task are also needed by other candidate tasks, the allocation cost will increase to avoid all robots rushing to the same hot resource.

[0074] Regional congestion: Analyze the dynamic traffic map. If the target area already has multiple robots' predicted trajectory zones, the cost will increase in order to alleviate traffic pressure.

[0075] Objective 4: Stability Cost Estimation Content: This is a unique optimization objective introduced specifically for humanoid robots. It assesses the impact of task paths and operations on the stability of the humanoid robot.

[0076] Calculation method: Terrain stability margin: The system maintains a dynamic access map, marking the ground flatness, slope, friction coefficient, etc., of different areas. For tasks requiring traversal of highly complex terrain, the system queries the robot's stability model library to assess the risk level of traversing such terrain and increases the cost score accordingly.

[0077] Operational stability margin: For resource-constrained operation chains involving lifting, carrying, etc., the system considers the impact of load weight and robot arm span on center of gravity stability. Assigning a small humanoid robot a task of carrying a large box would incur a very high stability cost.

[0078] Each of the four cost objectives is assigned a configurable weight (e.g., time has the highest weight, followed by stability), and the weighted comprehensive cost score for each candidate combination (task, robot) is calculated. Using algorithms such as the Hungarian algorithm or more advanced metaheuristics, an allocation scheme is selected from all candidate combinations that minimizes the total comprehensive cost of all assigned tasks. Once the optimal allocation scheme is found, the central scheduling system formally executes the assignment. It sends a "task ready" instruction to the selected robot, containing the corresponding resource constraint operation chain. Simultaneously, it sends an instruction to the resource state management module to pre-lock the critical resources required in the task chain to prevent them from being preempted by other tasks before the task begins.

[0079] Based on the robot's initial pose (from the dynamic navigation map) and the position sequence of each navigation sub-objective in its resource constraint operation chain, planning is performed on the latest environment model provided by the dynamic navigation map. Static obstacles, the current contours of other robots, and the predicted trajectory bands in this map together constitute the spatial constraints of the planning. The generation process is as follows: Step A: Spatial Path Search: First, in the spatial dimension, a fast randomized exploratory tree algorithm is used to search for a collision-free (relative to static obstacles and other robot outlines at the current moment) geometric path from the starting point to the ending point. This path consists of a series of waypoints.

[0080] Step B: Time Axis Introduction and Velocity Planning: Instead of assigning a constant speed to the robot, the system assigns a suitable speed curve to each segment of the above geometric path based on the global traffic conditions. In wide, uninhabited areas, allocate higher speeds.

[0081] Distribute a lower speed when turning, approaching obstacles, or when anticipating entering a traffic-heavy area.

[0082] The result of this speed planning is that each point on the geometric path is assigned an expected arrival timestamp. Thus, a static path is elevated to a four-dimensional spacetime trajectory. This trajectory can be precisely described as: "Robot 001 will be located at coordinates (x1, y1) in a certain posture at t=10.5 seconds."

[0083] Step C: Generating the Spatiotemporal Trajectory Band: For safety reasons, not only is the spatiotemporal trajectory of the robot's center of mass planned, but also a spatiotemporal trajectory band is generated to wrap around this trajectory based on the robot's physical shape (chassis size, arm span). This "band" represents the actual spatial volume occupied by the robot over time during the entire movement. It serves as the direct basis for subsequent accurate collision detection.

[0084] All pre-planned spatiotemporal trajectory bands of robots, along with predicted trajectory bands of other robots not yet replanned from the dynamic traffic map, are loaded into a unified virtual spatiotemporal simulation environment. Within this environment, the simulation progresses progressively along the time axis. In each simulation time slice (e.g., starting at t=0, one slice every 0.1 seconds), it checks for intersections between the cross-sections of all robot spatiotemporal trajectory bands at that moment (i.e., the spatial regions they occupy at that moment). For efficient computation, the entire space is discretized into extremely small cubes (voxels), and the occupancy of each trajectory band at each moment is mapped to a set of these voxels. Conflict detection is then transformed into quickly determining whether different sets intersect at the same moment. This method is computationally efficient and meets real-time requirements. Once an intersection is detected, the engine identifies the conflict type: Cross-path conflict: The trajectories of two robots will intersect at a point in the future.

[0085] Rear-end collision: One robot is moving faster than the other, and their paths are the same or partially overlap.

[0086] Opposing collision: Two robots are moving towards each other in the same passage.

[0087] When conflict is anticipated, a smart negotiation mechanism should be activated, rather than abruptly stopping one party completely.

[0088] Decision-making basis: Dynamic priority mechanism: The system calculates a dynamic priority score for each robot's task. This score is not fixed, but is calculated in real time based on the following factors: Task inherent priority: "high", "medium", and "low" levels assigned by the upper-level system.

[0089] Task urgency: Based on the urgency of the remaining execution time and the final deadline of the task.

[0090] Robot state weights: Robots with low battery levels may be given higher priority so that they can complete their tasks as soon as possible and then recharge.

[0091] Impact on overall system efficiency: Which robot's path has the least impact on the overall system throughput?

[0092] The system selects the most appropriate strategy from a range of methods to resolve conflicts based on a comparison of conflict types and priorities. 1. Priority-driven path replanning: In the event of a conflicting path, the lower-priority robot is typically instructed to detour. The spatiotemporal path planner will then replan a path for it, possibly by making a detour before the conflict point, ensuring that its spatiotemporal trajectory completely avoids the path of the higher-priority robot.

[0093] In the event of a conflict between robots traveling in opposite directions on a narrow passage, the low-priority robot needs to wait at the entrance until the high-priority robot passes through.

[0094] 2. Non-invasive speed adjustment: For less urgent conflicts (such as a minor risk of rear-end collision), the system may not instruct the low-priority robot to change its path, but instead instruct it to slow down slightly, thus staggering its passage time with the robot ahead. This usually has a smaller impact on overall efficiency than detouring.

[0095] Similarly, high-priority robots can be accelerated appropriately to pass through conflict zones more quickly.

[0096] 3. Agreement reached through negotiation: In complex scenarios where both parties have similar priorities, the system may perform a simplified "negotiation": for example, generating two options for each robot, namely "decelerate-pause-pass" and "accelerate-pass first", estimating their respective time losses, and selecting the option with the minimum total loss to execute.

[0097] Regardless of the conflict resolution strategy employed, the generated new solution must be sent back to the conflict prediction engine for verification to ensure that the newly generated spatiotemporal trajectory band no longer conflicts with any other trajectory bands in the system. This process may iterate several times until a conflict-free solution is found. Ultimately, all these coordinated and compatible scheduling paths are issued to the corresponding robots for execution. In this way, the entire system maintains efficient, smooth, and safe operation in a dynamically changing multi-robot training environment through a continuous cycle of "planning-prediction-detection-resolution".

[0098] S104. Transmit the generated assignment task and the scheduling path to the corresponding humanoid robot, and update the dynamic access map and the training ground resource table based on the acquired execution data.

[0099] Specifically, the planned resource constraint operation chain and spatiotemporal trajectory are safely, reliably, and accurately sent to the corresponding robot. The generated allocation task and the scheduling path are assembled into a synchronization instruction package, which combines the resource constraint operation chain and the spatiotemporal trajectory into a single synchronization instruction package. When an instruction needs to be sent, a temporary, encrypted point-to-point communication channel is established with the target robot through the high-speed wireless communication network within the training ground. Point-to-point communication is primarily based on the unique identification code and corresponding network address (such as IP address) of each robot recorded in the constructed robot communication registry. Before sending, the communication registry is queried to obtain the corresponding identification code and network address, thereby enabling point-to-point communication transmission.

[0100] After sending the synchronization command packet, a three-way handshake confirmation mechanism will be initiated: Send and request confirmation: Send a command packet and wait for the receiver's confirmation.

[0101] Reception and Verification Feedback: After receiving the instruction packet, the robot performs an integrity check (such as CRC check). Upon successful verification, it does not execute the instruction immediately, but first sends a confirmation signal indicating that the instruction has been received and is valid.

[0102] Execution Authorization: After receiving confirmation signals from all target robots, a unified "Global Execution Authorization" command is broadcast. This command carries a unified future timestamp (e.g., "Execute at the start of the next global clock cycle").

[0103] This mechanism ensures that all relevant robots begin executing the new plan at almost the same time, avoiding new timing conflicts caused by time differences in instruction reception.

[0104] The robot sends status heartbeat packets to the execution monitor at a fixed frequency (e.g., 10Hz), containing information such as real-time pose, battery level, and joint torque. Whenever an atomic operation is completed (e.g., "successful grasp"), the robot proactively reports an "atomic operation completion event." Based on this, the internal progress status of the task is updated.

[0105] Cross-validate the "successful capture" event reported by the robot with the resource's status changed to "occupied" in the resource's unique identifier. If the information is inconsistent, immediately trigger exception handling.

[0106] The robot's reported real-time pose is compared with its expected spatiotemporal trajectory in real time. The comparison result is then compared with a set distance threshold to obtain the corresponding deviation result.

[0107] Level 1 deviation (slight deviation): If the deviation between the robot's actual position and the expected trajectory is within the allowable threshold (e.g., 5 cm), the system only records it and assumes that the robot's local controller can correct it on its own.

[0108] Secondary deviation (significant deviation): If the deviation exceeds the threshold, but the robot is still within the envelope of its predicted trajectory, a "path fine-tuning cue" is sent to the robot to remind its local controller to make corrections.

[0109] Level 3 Deviation (Severe Deviation / Abnormality): A severe abnormality is identified if the following conditions are detected: The robot completely detached from its track and was unable to recover on its own.

[0110] The robot has not reported a status heartbeat for an extended period of time (communication interruption).

[0111] The robot reported an "operation failure" signal (such as "object dropped").

[0112] The multimodal sensor array detected that the robot had experienced an actual physical collision (even though it did not report it itself).

[0113] If a Level 3 deviation occurs, immediately set the robot's task status to "abnormal interruption" and initiate the emergency response procedure: First, send an "emergency pause" command to the malfunctioning robot (if communication is normal), and simultaneously mark it as a "dynamic obstacle" on the dynamic traffic map. Then, immediately perform local path replanning for all other robots that might be colliding with the malfunctioning robot, directing them to bypass this new "obstacle." Release all resources reserved or occupied by the malfunctioning task. Return the task (or its unfinished portion) to the task resolution queue. Based on the latest global state, reassign a suitable robot to the task and replan the path.

[0114] By dynamically allocating tasks and predicting conflicts, the robot avoids stagnation caused by waiting for resources or path deadlock, thus maximizing the utilization of hardware resources (such as training props and charging stations) and robot time resources in the training field, thereby completing more training tasks per unit of time.

[0115] By establishing real-time dynamic maps and conflict detection and resolution mechanisms, physical conflicts between robots can be proactively anticipated and avoided, ensuring the safety of expensive humanoid robot hardware and guaranteeing the continuity and stability of the training process.

[0116] The scheduling strategy fully considers the kinematic characteristics of the humanoid robot, such as stability margin, making path planning and task allocation more consistent with its actual movement capabilities and improving the success rate of task execution. The system can adapt to dynamic changes in the training environment, exhibiting stronger robustness and practicality.

[0117] Other embodiments of this application will readily occur to those skilled in the art upon consideration of the specification and practice of the disclosure herein. This application is intended to cover any variations, uses, or adaptations of this application that follow the general principles of this application and include common knowledge or customary techniques in the art not disclosed herein.

[0118] It should be understood that this application is not limited to the precise structure described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope.

Claims

1. A humanoid robot scheduling method in a humanoid robot training field, characterized by, The method comprises the following steps: Real-time acquisition of global environment data in the training field by a multi-modal sensor array, fusion with acquired real-time pose information of the robot, and generation of a dynamic passable map; Parsing of the acquired training task queue into atomic operation data, association with the constructed training field resource table, and obtaining of a resource-constrained operation chain; Dynamic allocation of the resource-constrained operation chain based on the dynamic passable map and the training field resource table, and generation of a scheduling path after spatio-temporal joint path planning; Transmission of the generated allocated tasks and the scheduling path to the corresponding humanoid robot, and updating of the dynamic passable map and the training field resource table based on acquired execution data.

2. The humanoid robot scheduling method in a humanoid robot training field according to claim 1, wherein, Real-time acquisition of global environment data in the training field by a multi-modal sensor array, fusion with acquired real-time pose information of the robot, and generation of a dynamic passable map, comprising: Real-time acquisition of three-dimensional point cloud data and depth image data in the training field by a multi-modal sensor array; Acquisition of acceleration, angular velocity, and angles of each joint of the humanoid robot, calculation of an estimated pose, and calculation of an absolute pose by recognizing visual marker codes deployed at known positions in the training field; Kalman filter fusion of the estimated pose and the absolute pose to obtain real-time pose information of the robot; Fusion of the real-time pose information of the robot, the three-dimensional point cloud data, and the depth image data to generate a dynamic passable map.

3. The humanoid robot scheduling method in a humanoid robot training field according to claim 2, wherein, Fusion of the real-time pose information of the robot, the three-dimensional point cloud data, and the depth image data to generate a dynamic passable map, comprising: Coordinate unification and fusion of the three-dimensional point cloud data and the depth image data, and time synchronization of the multi-modal sensor array and the humanoid robot; Statistical analysis of consecutive multiple frames of the three-dimensional point cloud data to obtain a static background model, and obtaining of dynamic objects based on the static background model; Association of the dynamic objects with the real-time pose information of the robot to obtain a robot identity code, and generation of a predicted trajectory band; Generation of a dynamic passable map based on the static background model, the predicted trajectory band, and the real-time position information of the robot.

4. The humanoid robot scheduling method in a humanoid robot training field according to claim 2, wherein, The multi-modal sensor array comprises a three-dimensional laser radar cluster and a panoramic visual sensor node; The three-dimensional laser radar cluster is configured to perform rotational scanning of the entire field from a top-down angle to obtain three-dimensional point cloud data; The panoramic visual sensor node is configured to use stereo binocular vision or structured light technology to calculate the distance between each pixel point in the image and the camera while acquiring conventional RGB color images, thereby obtaining depth image data.

5. The humanoid robot scheduling method in a humanoid robot training field of claim 1, wherein, Parsing of the acquired training task queue into atomic operation data, association with the constructed training field resource table, and obtaining of a resource-constrained operation chain, comprising: Syntax and semantic checking of the acquired training task queue, and marking as to-be-parsed after passing the checking, while being cached in the server; Keyword extraction and intent recognition of the training task queue, matching of corresponding atomic operation templates according to the recognized intent, and generation of a resource-constrained operation chain in combination with the constructed training field resource table.

6. The humanoid robot scheduling method in a humanoid robot training field according to claim 5, wherein, The method further comprises: annotating execution precondition and execution postcondition for each atomic operation in the resource-constrained operation chain.

7. The humanoid robot scheduling method in a humanoid robot training field of claim 1, wherein, Based on the dynamic communication map and the training field resource table, the resource-constrained operation chain is dynamically allocated, and a scheduling path is generated after spatiotemporal joint path planning, comprising: When a set trigger event occurs, the constructed candidate matrix is used to traverse the resource-constrained operation chain and the training field resource table, and it is judged whether the path is unobstructed by referring to the dynamic communication map; The comprehensive cost score of each scheduling scheme is calculated by using a multi-objective optimization algorithm, and the scheduling allocation task with the minimum comprehensive cost total sum is obtained after weighted summation; Based on the scheduling allocation task, a spatiotemporal trajectory band is generated, and after conflict resolution based on task dynamic priority, a scheduling path is obtained.

8. The humanoid robot dispatching method in a humanoid robot training field according to claim 7, wherein, The comprehensive cost score includes time cost score, energy consumption cost score, system balance cost score and stability cost score.

9. The humanoid robot scheduling method in a humanoid robot training field of claim 1, wherein, The generated allocation task and the scheduling path are transmitted to the corresponding humanoid robot, and the dynamic communication map and the training field resource table are updated based on the obtained execution data, comprising: The generated allocation task and the scheduling path are assembled into a synchronization instruction package, which is transmitted through the constructed communication registry; When the humanoid robot receives the synchronization instruction package, the actual execution state and execution position are uploaded in real time; Based on the uploaded actual execution state and execution position, the dynamic communication map and the training field resource table are updated.

10. The humanoid robot scheduling method in a humanoid robot training field of claim 9, wherein, The method further comprises: Cross-validation and monitoring of the uploaded actual execution state and execution position are performed; The execution position is compared with the spatiotemporal trajectory band in real time, and the comparison result is compared with a set distance threshold to obtain a corresponding deviation result; According to the deviation result, a hierarchical early warning is performed.

Images