Multi-object arrangement method and system for mobile operation robot

Abstract

The invention discloses a multi-object arrangement method and system for a mobile operation robot, and the method comprises the steps: obtaining an object image of the robot at a current visual angle, and extracting the attribute information of an object; constructing a bearing relation matrix and a semantic scene graph; the bearing relation matrix is converted into a directed acyclic graph, an edge set serves as a state space in reinforcement learning, an action space is defined to select one of unselected edges, the depth of the edge set serves as the basis of current grouping, and objects are grouped based on a high-level reinforcement learning model; obtaining a stacking execution sequence of the objects in the group based on the semantic scene graph; and the grabbing posture of the current to-be-operated object is obtained through prediction of the grabbing network, the object is grabbed according to the grabbing posture, and the grabbed object is stacked according to the planned stacking sequence. According to the robot, efficient action decision making is achieved in the grabbing, stacking and placing processes, and the robot has the capacity of carrying multiple objects in different places and orderly placing the multiple objects in a complex environment.

Description

Technical Field

[0001] This invention relates to the field of service robot task planning technology, and in particular to a method and system for organizing multiple objects in a mobile operating robot. Background Technology

[0002] The statements in this section are merely background information related to the present invention and do not necessarily constitute prior art.

[0003] Existing methods for object sorting in mobile manipulators mainly include three types: physical modeling-based, pose estimation-based, and large model-driven. Among them, physical modeling-based methods use known geometric and mechanical models for simulation or planning operations, but they are highly dependent on the object model and have poor versatility. Pose estimation-based methods rely on visual perception to obtain the 3D pose of the object and then sort it through grasping or trajectory planning, but require high-precision perception capabilities. Large model-driven methods use multimodal large models to understand images and instructions and directly generate operation strategies, which have strong generalization capabilities, but require high computational resources and lack fine process control capabilities.

[0004] Current research on object organization primarily focuses on scene understanding and motion planning methods based on visual perception, especially relying on deep learning-driven perception models trained on large amounts of manually labeled data. However, these methods are often limited to enclosed spaces, such as desktops or containers, failing to consider the movement costs and multi-objective decision-making processes of robots in large-scale spaces. Furthermore, existing methods often employ a per-object processing strategy, lacking optimization for overall task efficiency. In real-world complex environments, especially open scenarios, these methods are inefficient when facing varying distributions or multi-objective tasks, and lack adaptability to environmental changes, limiting their widespread application in real-world scenarios such as home services and warehousing logistics. Summary of the Invention

[0005] To address the aforementioned issues, this invention proposes a method and system for organizing multiple objects in a mobile robot. By introducing a large language model to achieve semantic understanding of objects and support relationship modeling, and combining a hierarchical reinforcement learning structure of high-level grouping and low-level stacking planning, it enables efficient movement and rearrangement of multiple objects in complex environments, thereby effectively improving the robot's operational efficiency and autonomy in open scenarios.

[0006] In some implementations, the following technical solutions are adopted: A method for organizing multiple objects using a mobile robot includes: Acquire an image of an object from the robot's current viewpoint and extract the object's attribute information; Using a large language model, we obtain the support relationships and semantic associations between objects, and construct a support relationship matrix and a semantic scene graph. The support relationship matrix is ​​transformed into a directed acyclic graph, where nodes represent objects and edges represent support relationships. The set of edges is used as the state space in reinforcement learning, and the action space is defined as selecting an unselected edge and using its depth as the basis for the current grouping. Objects are grouped based on a high-level reinforcement learning model. Based on the semantic scene graph, a low-level reinforcement learning model is used to plan the stacking order of each group of objects, thereby obtaining the stacking execution order of objects within the group. The grasping network predicts the grasping posture of the object to be manipulated, the object is grasped according to the grasping posture, and the grasped objects are stacked in the planned stacking order.

[0007] As a further embodiment, the support relationship matrix is ​​specifically configured such that diagonal elements are always 0, and off-diagonal elements are 1 to represent objects. Can support objects Otherwise, the value is 0, indicating that the support relationship is a one-way relationship.

[0008] As a further approach, objects are grouped based on a high-level reinforcement learning model, specifically as follows: With edge set Constructing the reinforcement learning state space Each edge Includes parent node child nodes Depth of the parent node and a flag bit , =0 indicates that the edge has been selected. =1 indicates that the edge was not selected; Using the selected and unselected edges as the action space The depth of the edge is used as the basis for the current grouping, and only one edge of the same depth can exist in each group. Grouping based on the depth of the edge helps to limit the selection of actions to the set of edges with similar depths, reduce the complexity of the search space, and improve the stability and rationality of the stacking strategy. Then, the unselected edges are used as the action space, and the depth difference between two consecutive selected edges is set to not exceed a preset threshold. The task is terminated when the edge with the maximum depth is selected.

[0009] Finally obtained Grouping objects into groups based on their support capabilities, denoted as... ,satisfy ,in, Indicates the first Group, This is the final number of groups. Indicates the first The first assigned group Strip edge, It is the number of edges in the graph structure.

[0010] As a further approach, a low-level reinforcement learning model is used to plan the stacking order of each group of objects, specifically as follows: The low-level reinforcement learning model includes a graph capsule convolutional network layer, a regularization layer, and a softmax layer. The graph capsule convolutional layer receives the position and size information of the objects to be rearranged, as well as the grouping results output by the high-level reinforcement learning model, and outputs intermediate features. The intermediate features are processed by regularization and softmax, and each object gets a score. The object with the highest score is selected, and combined with the grouping results, objects not in the same group as the selected object are selected and stacked together with it. This process is repeated until all object stacking is planned, and the optimal decision result for the current stacking order is obtained.

[0011] As a further solution, it also includes: using the negative of the time required for the robot to execute the entire scheme as the reward function to optimize the low-level reinforcement learning model and the high-level reinforcement learning model; and optimizing the stacking order strategy with the goal of minimizing the execution time.

[0012] As a further step, the specific time required for the robot to execute the entire plan is as follows:

[0013] in, Indicates the execution time of the plan. The time it takes for a robot to grasp and stack an object. and These are the robotic arm's moving speed and the chassis's moving speed, respectively. To collect groups The minimum distance that all objects within the area must move.

[0014] As a further solution, a two-stage training strategy is adopted to optimize the low-level reinforcement learning model and the high-level reinforcement learning model, specifically as follows: In the first stage, objects are initially grouped using a heuristic strategy to train a high-level reinforcement learning model. In the second stage, the grouping output of the trained high-level reinforcement learning model is used as input to optimize and train the low-level reinforcement learning model, thereby achieving collaborative decision-making between high- and low-level strategies.

[0015] In other embodiments, the following technical solutions are adopted: A mobile robotic multi-object sorting system includes: The data acquisition module is used to acquire images of objects from the robot's current perspective and extract the object's attribute information; The semantic awareness module is used to obtain the support relationships and semantic associations between objects using a large language model, and to construct a support relationship matrix and a semantic scene graph. The grouping decision module is used to transform the support relationship matrix into a directed acyclic graph, where the nodes of the directed acyclic graph represent objects and the edges represent support relationships. The set of edges is used as the state space in reinforcement learning, and the action space is defined as selecting an unselected edge and using its depth as the basis for the current grouping. The objects are grouped based on a high-level reinforcement learning model. The stacking planning module is used to plan the stacking order of each group of objects based on the semantic scene graph through a low-level reinforcement learning model, so as to obtain the stacking execution order of objects within the group. The execution control module is used to predict the grasping posture of the object to be operated through the grasping network, grasp the object according to the grasping posture, and stack the grasped objects according to the planned stacking order.

[0016] In some implementations, the following technical solutions are adopted: A robot includes a robot body, a multi-object grasping and rearranging actuator mounted on the robot body, a hierarchical reinforcement learning controller, and a computer program stored on and executed by the controller; when the hierarchical reinforcement learning controller executes the computer program, it can realize the above-mentioned multi-object sorting method for mobile operation robots.

[0017] In some implementations, the following technical solutions are adopted: A computer-readable storage medium storing a plurality of instructions adapted for loading and execution by a processor of a terminal device of the above-described method for sorting multiple objects in a mobile robot.

[0018] Compared with the prior art, the beneficial effects of the present invention are: (1) This invention integrates a large language model to model support and semantic relationships, replacing manual rule setting, enabling the robot to automatically determine the stackable relationships and classification associations between objects using its perception capabilities, and complete the construction of support graphs and task strategy generation. Combined with a hierarchical reinforcement learning structure, it achieves efficient action decision-making in the grasping, stacking and placement process, enabling the robot to perform multi-object remote transportation and orderly placement in complex environments.

[0019] (2) The present invention enables the robot to autonomously complete the analysis and grouping of object support capabilities based on high and low layer control strategies. At the same time, based on the spatial distribution, semantic information and physical attributes of objects in the current scene, it plans a reasonable stacking order and placement path. Through quantitative modeling and reinforcement learning optimization of operation time, it effectively reduces the robot's task execution time, improves the rearrangement efficiency, and enhances the robot's adaptability and stability in the actual open environment.

[0020] (3) During the execution process, the robot of the present invention can dynamically evaluate the rationality of grasping and stacking based on its own motion ability, gripper structure parameters and task scenario, so as to achieve the shortest task path and the most simplified action process, and finally complete the entire rearrangement task process efficiently without human intervention.

[0021] After completing each round of object stacking tasks, the robot can optimize the strategy network structure based on the task execution time feedback, and achieve better grouping and stacking action planning in the next round of tasks.

[0022] Through self-learning and structural optimization, the robot gradually improves its rearrangement model, enhances its generalization ability in diverse environments, and achieves task autonomy and long-term execution stability.

[0023] (4) The robot of the present invention can not only perceive and understand the structural and semantic information between multiple objects in the environment, but also actively select the optimal execution strategy based on the task objective and time cost function to maximize the rearrangement efficiency; the robot thus has a highly robust object manipulation capability for complex real-world scenarios and is suitable for a variety of practical application fields such as home service, warehouse sorting, and industrial sorting.

[0024] Other features and advantages of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0025] Figure 1 This is a schematic diagram of the process of a mobile robot organizing multiple objects in an embodiment of the present invention; Figure 2 The following are the results of the visual experimental comparison in the embodiments of the present invention; wherein, (a) to (c) show the training results of capsule networks with different layers when the number of objects is 30 and the robot performs task rearrangement; (d) to (f) show the training results of capsule networks with different layers when the number of objects is 40 and the robot performs task rearrangement; and (g) to (i) show the training results of capsule networks with different layers when the number of objects is 50 and the robot performs task rearrangement. Detailed Implementation

[0026] It should be noted that the following detailed descriptions are illustrative and intended to provide further explanation of this application. Unless otherwise specified, all technical and scientific terms used in this invention have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains.

[0027] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the exemplary embodiments according to this application. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0028] Example 1 In one or more embodiments, a method for organizing multiple objects using a mobile robot is disclosed, combining... Figure 1 Specifically, it includes the following process: S101: Acquire an image of the object from the robot's current perspective and extract the object's attribute information. This attribute information includes the object's position, category, size, and other attributes.

[0029] S102: Use the Large Language Model (LLM) to obtain the support relationships and semantic associations between objects, and construct the support relationship matrix and semantic scene graph.

[0030] Specifically, construct a support relationship matrix. The diagonal elements are always 0, and the off-diagonal elements are... Represents objects Can support objects Otherwise, it is 0. The support relationship is unidirectional to prevent the objects from tipping over when stacked.

[0031] Simultaneously, construct a semantic scene graph. ,in For a set of object nodes, This is a set of edges, representing spatial or semantic relationships between objects. These semantic relationships are based on a pre-defined set of human knowledge. It provides and selects appropriate semantic edge attributes through a large language model to facilitate input to the subsequent policy model.

[0032] In this embodiment, the input of the large language model is the category of the item, such as common kitchen utensils (bowls, plates, spoons, etc.). The model outputs the pairwise support relationship between the objects, such as a bowl can be placed on a plate, a spoon can be placed on a bowl, and a spoon can be placed on a plate. We directly use the Chatgpt-4o model.

[0033] S103: Transform the support relationship matrix into a directed acyclic graph, where nodes of the directed acyclic graph represent objects and edges represent support relationships; use the set of edges as the state space in reinforcement learning, define the action space as selecting an unselected edge, and use its depth as the basis for the current grouping, and group objects based on a high-level reinforcement learning model.

[0034] In this embodiment, the support relationship matrix obtained in step S102 is used as a basis. Construct a directed acyclic graph ,in For a collection of objects, It is a set of support relationships.

[0035] The set of edges in a directed acyclic graph As the state space in reinforcement learning Each edge Includes parent node child nodes Depth of the parent node and a flag bit The flag indicates whether the edge is selected (0 indicates selected, 1 indicates not selected).

[0036] Action space The task is defined as selecting one edge from the unselected edges and using its depth as the basis for the current grouping; specifically, each edge has a depth attribute, and only one edge of the same depth can exist in each group; the grouping task terminates when all edges in the directed acyclic graph have been selected once.

[0037] The depth difference of the selected edges between consecutive time steps is set to not exceed a preset threshold to ensure the structural rationality of the grouping results; when the edge with the maximum depth is selected, the grouping process terminates, and the final result is obtained. Grouping objects into groups based on their support capabilities, denoted as... ,satisfy ,in, Indicates the first Group, This is the final number of groups. Indicates the first The first assigned group Strip edge, It is the number of edges in the graph structure.

[0038] In this embodiment, the high-level reinforcement learning model mainly consists of a convolutional network layer, a regularization layer, a multi-head attention network layer, and a softmax layer connected in sequence. The model's input is the support relation matrix and semantic scene graph obtained from the large language model, and the model output is the grouping of objects according to their support relations. For example, for an environment where the objects to be rearranged include plates, bowls, and chopsticks, the output may be two groups: the first group contains plates and bowls, and the second group contains chopsticks; all objects in the second group can be placed on top of any object in the first group.

[0039] This embodiment groups edges based on their depth in the directed acyclic graph (DAG). Within each group, only one edge of the same depth can exist. This grouping helps to clarify the structural hierarchy in the stack and avoids support conflicts or resource duplication. Through grouping, subsequent stacking sequence planning can achieve clearer and more efficient path selection while ensuring structural stability, thus improving the overall feasibility and planning quality of the stack.

[0040] S104: In this embodiment, the semantic scene graph extracts semantic relationships and spatial context features between objects through a convolutional network, providing structured perceptual input to the low-level reinforcement learning model and guiding it to identify a reasonable stacking order. The low-level model uses a graph capsule convolutional network to construct an object relationship graph, fusing geometric, positional, and semantic information of objects to output a stacking priority score for each object. The stacking order is selected through Softmax sorting, and stacking decisions within groups are completed by combining high-level grouping information, achieving a structurally reasonable and semantically consistent stacking plan.

[0041] In this embodiment, in each group In the process, a stacked state space is constructed. state space Including object properties Supporting structure Group tags semantic information etc.; among them, Including the position coordinates of each object ,weight Access frequency Allocation status Group tags are denoted as Semantic relationships are derived from edge attributes in the semantic graph.

[0042] Action space Defined as selecting one from unassigned objects and adding it to a stackable target combination.

[0043] In this embodiment, the low-level reinforcement learning model consists of a graph capsule convolutional network layer, a regularization layer, and a softmax layer connected in sequence. First, the graph capsule convolutional layer receives the position and size information of the objects to be rearranged, as well as the grouping results output by the high-level reinforcement learning model, and outputs intermediate features. The intermediate features are then processed by regularization and softmax, and each object corresponds to a score.

[0044] First, select the object with the highest score (highest operation priority). Then, based on the grouping results of the high-level model, select objects not in the same group as the first object and stack them together (objects with higher scores in other groups). Repeat this process until all object stacking is planned. Output the optimal decision result for the current stacking order.

[0045] In this embodiment, the stacking planning is guided by grouping results, combining geometric, positional, and semantic information of objects for progressive selection and stacking. The system first clarifies the support hierarchy between objects, grouping objects with similar support roles together to provide structural constraints for subsequent selections. During stacking, the highest-scoring operable object is always prioritized as the stacking benchmark. Then, structurally compatible objects are selected from different groups to combine with it, ensuring that each stack satisfies both support stability and semantic common sense. This progressive stacking method effectively avoids invalid combinations and conflicting layouts, improving the overall rationality and success rate of stacking. Simultaneously, hierarchical planning reduces the action space dimension, improving decision-making efficiency and generalization ability, making it particularly suitable for efficient stacking tasks in diverse and complex object environments.

[0046] S105: The grasping posture of the object to be operated is predicted by the grasping network, the object is grasped according to the grasping posture, and the grasped objects are stacked according to the planned stacking order.

[0047] In this embodiment, the execution phase of the action includes: grasping, carrying and placing actions.

[0048] In the grasping phase, the GraspNet model is used to predict the 6-DOF grasping pose of the object to be grasped. The GraspNet model takes the RGB-D image of the scene as input, combines it with camera intrinsics to generate a 3D point cloud, and performs a grasping feasibility analysis for each point in the scene. The output consists of multiple candidate 6-DOF grasping poses (including position and orientation) and their corresponding grasping scores. The model samples candidate grasping points on the 3D point cloud and uses a convolutional neural network to encode local geometric features, ultimately regressing the grasping pose and success probability of each point, achieving accurate grasping prediction of the target object.

[0049] During the stacking phase, the objects to be grabbed are stacked according to the stacking order output by the low-level reinforcement learning model, and the center of gravity is kept aligned as much as possible to improve the stability of the stacking.

[0050] During the placement phase, the robot prioritizes moving objects to a temporary buffer zone to avoid placement failures due to limited space in the target area or obstruction by existing objects. By adjusting the order and avoiding obstacles through the buffer zone, the system can flexibly arrange the placement order of objects. In implementation, the robot first determines whether the target area is placeable; if not, it uses point-to-point path planning to send the object to the buffer zone. Once the conditions are met, the object is then sequentially transferred from the buffer zone to the target area, improving the stability and efficiency of placement.

[0051] In this embodiment, a total task execution time model is constructed for reward function design. The task execution time includes: ; in, Indicates the time for a single grab and the time for stacking. This indicates the robot's speed of movement during the stacking process. Indicates the first The minimum distance of the traversal path for grabbing a group of objects.

[0052] ; in, For the robot's handling speed, This is the one-way distance from the desktop to the target storage area.

[0053] ; ; To minimize total execution time To achieve the goal, the reward function is designed as follows: ; This embodiment employs a two-stage training strategy to collaboratively optimize different modules within a hierarchical structure. The two-stage strategy first optimizes the high-level and low-level strategies separately, making each module learn more stably and efficiently on its respective task. Subsequently, joint training further enhances inter-module collaboration, improving the overall strategy's adaptability and execution performance; specifically: The first stage uses a heuristic approach to group objects based on their load-bearing capacity, generating preliminary strategies for pre-training high-level reinforcement learning models. The second stage uses the output of the high-level reinforcement learning model as input to train the low-level reinforcement learning model, gradually replacing the heuristic grouping scheme. The REINFORCE algorithm (Monte Carlo policy gradient algorithm) is used to update the high- and low-level policy networks separately. The learning rate of the two-layer policy is coordinated through a time scale control mechanism, thereby enabling the robot to perform efficiently in the task of rearranging multiple objects in different locations.

[0054] After completing each round of object stacking tasks, the robot can optimize the strategy network structure based on the task execution time feedback, and achieve better grouping and stacking action planning in the next round of tasks.

[0055] During execution, the robot can dynamically evaluate the rationality of grasping and stacking based on its own mobility, gripper structure parameters, and task scenario, thereby minimizing the task path and simplifying the action flow. Ultimately, it can efficiently complete the entire rearrangement task without human intervention. Through self-learning and structural optimization, the robot gradually improves its rearrangement model, enhances its generalization ability in diverse environments, and achieves task autonomy and long-term execution stability.

[0056] To verify the effectiveness of the method in this embodiment, Figure 2 Panels (a)-(i) in the table present the visualized qualitative results and execution efficiency comparisons of multiple sets of simulation and actual physical experiments. The experiments were conducted under different object numbers and stacking heights. Ours represents the method of this embodiment, and the comparison algorithm is Attention Mechanism-based RL (AM-RL). It can be observed that under different experimental conditions of 30, 40, and 50 objects, the method of this embodiment can significantly shorten the task execution time while maintaining a smaller number of stacking groups, and its decision-making efficiency is also superior to other comparative methods. The method of this embodiment can successfully complete grouping and rearrangement operations in different scenarios, demonstrating good cross-scenario generalization performance. These results further verify the effectiveness of the method of this embodiment in improving robot operation efficiency and autonomy in complex environments.

[0057] Example 2 In one or more embodiments, a multi-object sorting system for a mobile robot is disclosed, specifically including: The data acquisition module is used to acquire images of objects from the robot's current perspective and extract the object's attribute information; The semantic awareness module is used to obtain the support relationships and semantic associations between objects using a large language model, and to construct a support relationship matrix and a semantic scene graph. The grouping decision module is used to transform the support relationship matrix into a directed acyclic graph, where the nodes of the directed acyclic graph represent objects and the edges represent support relationships. The set of edges is used as the state space in reinforcement learning, and the action space is defined as selecting an unselected edge and using its depth as the basis for the current grouping. The objects are grouped based on a high-level reinforcement learning model. The stacking planning module is used to plan the stacking order of each group of objects based on the semantic scene graph through a low-level reinforcement learning model, so as to obtain the stacking execution order of objects within the group. The execution control module is used to predict the grasping posture of the object to be operated through the grasping network, grasp the object according to the grasping posture, and stack the grasped objects according to the planned stacking order.

[0058] It should be noted that the specific implementation methods of the above modules are exactly the same as those in Example 1, and will not be described in detail again.

[0059] Example 3 In one or more embodiments, a robot is disclosed, comprising: a robot body and a multi-object grasping and rearranging actuator mounted on the robot body, a hierarchical reinforcement learning controller, and a computer program stored on and executed by the controller; when the hierarchical reinforcement learning controller executes the computer program, it can implement the multi-object sorting method of the mobile operation robot described in Embodiment 1.

[0060] Example 4 In one or more embodiments, a computer-readable storage medium is disclosed, wherein a plurality of instructions are stored, the instructions being adapted to be loaded by a processor of a terminal device and executed as described in Embodiment 1, the method for sorting multiple objects in a mobile operating robot.

[0061] While the specific embodiments of the present invention have been described above in conjunction with the accompanying drawings, this is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art without creative effort based on the technical solutions of the present invention are still within the scope of protection of the present invention.

Claims

1. A method for organizing multiple objects using a mobile robot, characterized in that, include: Acquire an image of an object from the robot's current viewpoint and extract the object's attribute information; Using a large language model, we obtain the support relationships and semantic associations between objects, and construct a support relationship matrix and a semantic scene graph. The support relationship matrix is ​​transformed into a directed acyclic graph, where nodes represent objects and edges represent support relationships. The set of edges is used as the state space in reinforcement learning, and the action space is defined as selecting an unselected edge and using its depth as the basis for the current grouping. Objects are grouped based on a high-level reinforcement learning model. Based on the semantic scene graph, a low-level reinforcement learning model is used to plan the stacking order of each group of objects, thereby obtaining the stacking execution order of objects within the group. The grasping network predicts the grasping posture of the object to be manipulated, the object is grasped according to the grasping posture, and the grasped objects are stacked in the planned stacking order.

2. The method for organizing multiple objects using a mobile robot as described in claim 1, characterized in that, The support relationship matrix is ​​specifically defined as follows: diagonal elements are always 0, and off-diagonal elements are 1 to represent objects. Can support objects Otherwise, the value is 0, indicating that the support relationship is a one-way relationship.

3. The method for organizing multiple objects using a mobile robot as described in claim 1, characterized in that, Objects are grouped based on a high-level reinforcement learning model, specifically as follows: With edge set Constructing the reinforcement learning state space Each edge Includes parent node child nodes Depth of the parent node and a flag bit , =0 indicates that the edge has been selected. =1 indicates that the edge was not selected; Using the selected and unselected edges as the action space The depth of the edge is used as the basis for the current grouping, and only one edge of the same depth can exist in each group. Grouping based on the depth of the edge helps to limit the selection of actions to the set of edges with similar depths, reduce the complexity of the search space, and improve the stability and rationality of the stacking strategy. Then, the unselected edges are used as the action space, and the depth difference between two consecutive selected edges is set to not exceed a preset threshold. The task is terminated when the edge with the maximum depth is selected. Finally obtained The support capabilities of an object are grouped together, denoted as... ,satisfy ,in, Indicates the first Group, This is the final number of groups. Indicates the first The first assigned group Edge, It is the number of edges in the graph structure.

4. The method for organizing multiple objects using a mobile robot as described in claim 1, characterized in that, The stacking order of each group of objects is planned using a low-level reinforcement learning model, specifically as follows: The low-level reinforcement learning model includes a graph capsule convolutional network layer, a regularization layer, and a softmax layer. The graph capsule convolutional layer receives the position and size information of the objects to be rearranged, as well as the grouping results output by the high-level reinforcement learning model, and outputs intermediate features. The intermediate features are processed by regularization and softmax, and each object gets a score. Select the object with the highest score, and based on the grouping results, select objects that are not in the same group as the selected object and stack them together. Repeat this process until all object stacking plans are completed, and obtain the optimal decision result for the current stacking order.

5. A method for organizing multiple objects using a mobile robot as described in claim 1, characterized in that, Also includes: The low-level reinforcement learning model and the high-level reinforcement learning model are optimized by using the negative of the time required for the robot to execute the entire scheme as the reward function; The stacking order is optimized with the goal of minimizing execution time.

6. A method for organizing multiple objects using a mobile robot as described in claim 5, characterized in that, The specific time required for the robot to execute the entire plan is as follows: in, Indicates the execution time of the plan. The time it takes for a robot to grasp and stack an object. and These are the robotic arm's moving speed and the chassis's moving speed, respectively. To collect groups The minimum distance that all objects within the area must move.

7. A method for organizing multiple objects using a mobile robot as described in claim 5, characterized in that, A two-stage training strategy is employed to optimize both the low-level and high-level reinforcement learning models, specifically: In the first stage, heuristic strategies are used to initially group objects and train a high-level reinforcement learning model. In the second stage, the grouped output of the trained high-level reinforcement learning model is used as input to optimize and train the low-level reinforcement learning model, thereby achieving collaborative decision-making between high-level and low-level policies.

8. A multi-object sorting system for a mobile robot, characterized in that, include: The data acquisition module is used to acquire images of objects from the robot's current perspective and extract the object's attribute information; The semantic awareness module is used to obtain the support relationships and semantic associations between objects using a large language model, and to construct a support relationship matrix and a semantic scene graph. The grouping decision module is used to transform the support relationship matrix into a directed acyclic graph, where the nodes of the directed acyclic graph represent objects and the edges represent support relationships. The set of edges is used as the state space in reinforcement learning, and the action space is defined as selecting an unselected edge and using its depth as the basis for the current grouping. The objects are grouped based on a high-level reinforcement learning model. The stacking planning module is used to plan the stacking order of each group of objects based on the semantic scene graph through a low-level reinforcement learning model, so as to obtain the stacking execution order of objects within the group. The execution control module is used to predict the grasping posture of the object to be operated through the grasping network, grasp the object according to the grasping posture, and stack the grasped objects according to the planned stacking order.

9. A robot, comprising a robot body, a multi-object grasping and rearranging actuator mounted on the robot body, a hierarchical reinforcement learning controller, and a computer program stored on and executed by the controller; characterized in that, When the hierarchical reinforcement learning controller executes the computer program, it can implement the multi-object sorting method for mobile operation robots as described in any one of claims 1-7.

10. A computer-readable storage medium storing a plurality of instructions, characterized in that, The instructions are adapted to be loaded and executed by the processor of the terminal device using the multi-object sorting method for mobile operating robots according to any one of claims 1-7.

Images