Model sharing method and system for heterogeneous robot

Abstract

The invention discloses a model sharing method and system for a heterogeneous robot, and relates to the technical field of artificial intelligence, and the method comprises the steps: obtaining a trained target model, and packaging the trained target model into a corresponding meta-skill; obtaining a target task and a corresponding heterogeneous robot; reading the skill description file of the meta-skill based on the target task to obtain the lowest capability configuration; the capacity of the heterogeneous robot is judged on the basis of the minimum capacity configuration, and a meta-action sequence is obtained on the basis of the skill description file; converting into a native control instruction which can be executed by the heterogeneous robot based on the meta-action sequence; and inputting the native control instruction to the heterogeneous robot to execute the target task. The shackle binding between an algorithm model and robot hardware is broken through, and lossless migration and sharing of the intelligent ability are achieved.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and more specifically to a model sharing method and system for heterogeneous robots. Background Technology

[0002] Currently, artificial intelligence technologies, represented by deep learning, have achieved a series of breakthroughs in fields such as visual recognition and natural language processing. Cutting-edge research is focused on training general-purpose algorithmic models capable of understanding the world, making decisions, and controlling robots to complete complex tasks, which is considered key to unlocking the next generation of autonomous robots. However, when deploying these advanced intelligent models from the digital world to physical robots of various forms, a fundamental technical bottleneck is becoming increasingly apparent: the deep coupling between the algorithmic model and the robot hardware. Existing technological approaches typically bind the model's inference output directly to the motion control system of a specific robot. This tight integration causes the intelligent model to become a "proprietary asset" of a specific robot platform, forming serious "model silos."

[0003] Specifically, existing technologies suffer from the following drawbacks and limitations: 1. High model porting costs hinder technology reuse: When a successfully trained object grasping model for robotic arm A needs to be deployed to robotic arm B or mobile robot C, developers often have to invest considerable effort in model adjustment or retraining. This is because different robots have different kinematic structures, drive characteristics, and sensor configurations, making the low-level control parameters output by the model unsuitable for direct application. This not only wastes R&D resources but also slows down the deployment of advanced algorithms. 2. The "black box" nature of end-to-end models exacerbates adaptation difficulties: Especially for end-to-end models that directly use raw perception data as input and low-level control commands as output, their internal representation is highly bound to the motion space of a specific robot. Cross-platform migration is almost a complete redevelopment, making these most promising general-purpose models extremely "specialized" in practice. 3. Suppressing the synergistic development of the algorithm ecosystem and hardware ecosystem: Due to the strong binding between models and hardware, excellent algorithm models are difficult to verify and iterate on a wider range of robot hardware; at the same time, robot manufacturers cannot easily adapt their hardware to a wealth of third-party intelligent applications. This disrupts the positive cycle between algorithmic and hardware innovation, limiting the prosperity of the entire embodied intelligence industry ecosystem.

[0004] To address the coupling problem, academia has attempted several approaches, such as designing more abstract, robot-independent task representations for training or using simulators for domain randomization to enhance generalization. However, most of these methods remain at the simulation or limited experimental stage, lacking a standardized, engineered solution for "train once, deploy everywhere" in real robotic systems. Another approach is to use robot middleware (such as ROS) for unified communication, but this only standardizes message passing and does not solve the core problem of dynamic mapping between model output semantics and different robot actuators.

[0005] Therefore, how to break through the binding shackles between algorithm models and robot hardware and achieve lossless transfer and sharing of intelligent capabilities is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0006] In view of the above problems, this invention is proposed to provide a model sharing method and system for heterogeneous robots that overcomes or at least partially solves the above problems, breaks through the binding shackles between algorithm models and robot hardware, and realizes lossless transfer and sharing of intelligent capabilities.

[0007] To achieve the above objectives, the present invention adopts the following technical solution:

[0008] In a first aspect, embodiments of the present invention provide a model sharing method for heterogeneous robots, comprising: Obtain the trained target model and encapsulate it into corresponding meta-skills; Obtain the target task and the corresponding heterogeneous robot; Based on the target task, the skill description file of the meta-skill is read to obtain the minimum capability configuration; The capabilities of the heterogeneous robot are determined based on the minimum capability configuration, and the meta-action sequence is obtained based on the skill description file; Based on the meta-action sequence, it is converted into native control instructions that the heterogeneous robot can execute; The target task is executed by the heterogeneous robot based on the input of the native control commands.

[0009] In another embodiment, the target model is a pre-trained, hardware-decoupled original algorithm model file that takes a high-level task objective as input and outputs abstract task space parameters. The original algorithm model file is then standardized, encapsulated, and lightweighted before being stored in the model library.

[0010] In another embodiment, it is encapsulated as a corresponding meta-skill, specifically including: Based on the target model, parameter verification logic, and preprocessing / postprocessing steps, a complete meta-skill plugin is woven together and registered in the meta-skill library. The skill description file is then edited to realize the combination and arrangement between the target model, adapter, and meta-actions, forming the meta-skill with clear input and output.

[0011] In another embodiment, the meta-skill includes: the adapter and a skill description file; The adapter is used to call a specific model in the model library. The output of the specific model is only related to the input result of the model algorithm itself, and is not related to the hardware and actions of the specific robot. The skill description file includes the input interface of the meta-skill, the output semantics, the minimum capability configuration of the required robot, and the judgment conditions for successful execution. It also records the model to be invoked, the adapter, and the meta-action to be executed based on the adapter result.

[0012] In another embodiment, the meta-action is the lowest-level, indivisible robotic atomic action; Each of the aforementioned meta-actions is a well-encapsulated function or service, corresponding to a fundamental capability that can be reliably executed at the robot's underlying level, and its implementation is related to the robot's hardware.

[0013] In another embodiment, the target task uses a task template obtained from a task library and is used to invoke the meta-skill.

[0014] In another embodiment, the task library includes a complete task flow oriented towards business scenarios; The task flow consists of the meta-skills and skill flow templates from the composite action library.

[0015] In another embodiment, the composite action library stores reusable skill flow templates composed of multiple meta-actions or meta-skills arranged and combined according to specific logic.

[0016] In another embodiment, determining the capabilities of the heterogeneous robot specifically includes: Obtain the capability configuration file of the heterogeneous robot; The capability configuration file includes a list of all meta-actions supported by the heterogeneous robot and component information; Based on the component information, determine whether the minimum capability configuration is met; If satisfied, then obtain the meta-action sequence based on the skill description file; Otherwise, it means that the target task cannot be completed and the requirements are not met.

[0017] Secondly, embodiments of the present invention provide a model sharing system for heterogeneous robots, including: a meta-skill acquisition module, a target task acquisition module, a minimum configuration acquisition module, a meta-action sequence output module, and a heterogeneous execution module; The meta-skill acquisition module is used to acquire the trained target model and encapsulate it into a corresponding meta-skill; The target task acquisition module is used to acquire the target task and the corresponding heterogeneous robot; The minimum configuration acquisition module is used to read the skill description file of the meta-skill based on the target task to obtain the minimum capability configuration; The meta-action sequence output module is used to determine the capabilities of the heterogeneous robot based on the minimum capability configuration, and to obtain the meta-action sequence based on the skill description file. The heterogeneous execution module is used to convert the meta-action sequence into native control instructions that the heterogeneous robot can execute; and input the native control instructions to the heterogeneous robot to execute the target task.

[0018] As can be seen from the above technical solution, compared with the prior art, the present invention discloses a model sharing method and system for heterogeneous robots, which has the following beneficial effects: 1. This invention does not directly use low-level parameters (such as joint angles and motor torque) output by the algorithm model that are specific to a particular robot. Instead, it mandates or uses a lightweight adapter to convert the model output into a set of predefined, general commands described by task space or motion primitives. This command set is completely abstracted from the mechanical configuration and drive details of the specific robot, thereby completely separating the "decision logic" of the intelligent model from the "execution details." This makes the model itself a hardware-independent, shareable asset, promoting the separation and collaborative development of the algorithm ecosystem and the hardware ecosystem, and enhancing the system's robustness and flexibility in the face of hardware iteration and diversity.

[0019] 2. For each target robot, this invention is equipped with a runtime command mapper. This mapper loads the robot's hardware capability configuration file (including kinematics, dynamics, workspace, and other constraints). When a general command is received, the mapper automatically uses parameter mapping to map the general command parameters to the input of the robot's specific command interface, thereby realizing the call and execution between the general command parameters and the specific interface.

[0020] 3. This invention achieves standardized output at the algorithm layer through a "general command set" and automated adaptation at the execution layer through a "runtime mapper." This architecture transforms the core work of cross-robot model porting from expensive, expert-experience-dependent "model reconstruction" to automated, engineered "configuration and mapping," providing a crucial foundational framework for building an open, embodied intelligence ecosystem where models can circulate and hardware is interchangeable, thereby reducing the time cost and technical barriers to model porting and system integration.

[0021] 4. This invention constructs a "meta-skill abstraction layer" and a "runtime dynamic orchestration and adaptation" mechanism between the algorithm model and the physical robot. The high-level intent output by the model is converted and encapsulated into standardized, hardware-independent "meta-skills," which are then dynamically instantiated into executable low-level control instructions at runtime based on the specific capabilities of the target robot. This achieves "train once, deploy everywhere," enhancing the value and reusability of AI assets. Attached Figure Description

[0022] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0023] Figure 1 This is a flowchart of a model sharing method for heterogeneous robots provided in an embodiment of the present invention.

[0024] Figure 2 This is a flowchart of the heterogeneous robot capability determination method provided in an embodiment of the present invention.

[0025] Figure 3 This is a schematic diagram of a model sharing system for heterogeneous robots provided in an embodiment of the present invention. Detailed Implementation

[0026] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0027] Example 1 like Figure 1As shown, this embodiment of the invention discloses a model sharing method for heterogeneous robots, including the following steps. For ease of description, numbers S1 to S6 are assigned; these numbers are not intended to limit the sequential relationship between the various steps of this invention: S1 acquires the trained target model and encapsulates it into the corresponding meta-skill.

[0028] Furthermore, the target model is a pre-trained, hardware-decoupled original algorithm model file that takes a high-level task objective as input and outputs abstract task space parameters. The original algorithm model file is standardized, encapsulated, and lightweighted before being stored in the model library.

[0029] Furthermore, the model library is used to store and manage pre-trained, hardware-decoupled raw algorithm model files, such as PyTorch's .pt files or TensorFlow's SavedModel. These models typically take high-level task objectives (such as images or point clouds) as input and output abstract task space parameters (such as grasping pose, navigation target point, and the force vector required to push an object).

[0030] Furthermore, it is encapsulated into corresponding meta-skills, specifically including: Based on the target model, parameter validation logic, and preprocessing / postprocessing steps, a complete meta-skill plugin is woven together and registered in the meta-skill library. The skill description file is then edited to realize the combination and arrangement between the target model, adapter, and meta-actions, forming a meta-skill with clear input and output.

[0031] Furthermore, meta-skills include: adapters and skill description files; An adapter is used to call a specific model from the model library. The output of the specific model is only related to the input result of the model algorithm itself, and is not related to the hardware and actions of the specific robot. For example, the recognition of objects and the coordinate position of objects in a visual recognition algorithm. Skill description files (such as XML or JSON) include the input interface of the meta-skill, output semantics, the minimum capability configuration required for the robot (e.g., "requires a six-DOF robotic arm and a binocular camera"), and the criteria for successful execution. They also record the models and adapters that need to be invoked, as well as the meta-actions that need to be executed based on the adapter results. For example, for a vision-based navigation meta-skill, the model analyzes an input image to obtain the relative distance and angle relative to the current robot position. Based on the output of this model, the adapter can convert it into the angular velocity and linear velocity that the robot needs to move. In addition to describing the models and adapters that need to be invoked, the skill description file also describes the robot meta-actions that need to be executed based on the adapter results. In this example, the meta-action of BODY_MOVE is the body moving according to the calculated angular velocity and linear velocity.

[0032] Furthermore, meta-actions are the most basic, indivisible atomic actions of robots; Each meta-action is a pre-packaged function or service, corresponding to a fundamental capability that the robot can reliably execute at its underlying level. For example, Move_ARM(position): indicates that the robotic arm moves to a certain spatial position; SetGripper(open_width): indicates that the end effector opens to a certain angle. Its implementation is related to the robot's hardware.

[0033] S2 acquires the target task and the corresponding heterogeneous robot.

[0034] Furthermore, the target task uses a task template, obtained from the task library, to invoke meta-skills.

[0035] Furthermore, the task library includes complete task flows oriented towards business scenarios; The mission flow consists of meta-skills and skill flow templates from the composite action library.

[0036] Furthermore, based on a visual or domain-specific language (DSL) editing environment, components can be selected from the meta-skill library and composite action library by dragging and dropping, combined into a complete task flow oriented towards business scenarios, and saved to the task library.

[0037] Furthermore, the composite action library stores reusable skill flow templates composed of multiple meta-actions or meta-skills arranged and combined according to specific logic (sequence, branching, looping), such as the "assembling screws" flow.

[0038] S3 reads the skill description file of the meta-skill based on the target task to obtain the minimum capability configuration.

[0039] S4 determines the capabilities of heterogeneous robots based on the minimum capability configuration and obtains the meta-action sequence based on the skill description file.

[0040] Furthermore, such as Figure 2 As shown, the assessment of heterogeneous robot capabilities specifically includes: Obtain the capability configuration files of heterogeneous robots; The capability profile includes a list of all meta-actions supported by the heterogeneous robot, as well as component information; Determine whether the minimum capability configuration is met based on component information; If satisfied, the meta-action sequence is obtained based on the skill description file; Otherwise, it means that the target task cannot be completed and the requirements are not met.

[0041] Furthermore, assuming the minimum capability configuration requires "one 6DOF robotic arm" and "one RGB-D camera," the system searches the heterogeneous robot's component information to determine if a camera and a 6DOF robotic arm are available. If they are, the capability is met; otherwise, the meta-skill cannot be performed. Additionally, for each component, specific technical parameters are configured. For example, for a camera component, this includes the camera component ID and the camera-hand-eye calibration matrix.

[0042] Furthermore, when the requirements are met, the skill description file of the meta-skill is loaded, and the specific model calls and robot meta-action execution semantics are extracted based on the skill description file, thereby dynamically generating the abstract skill into a series of specific, executable meta-action sequences.

[0043] S5 converts meta-action sequences into native control commands that heterogeneous robots can execute.

[0044] Furthermore, in this embodiment, standard meta-action commands (such as Move_ARM(position)) are converted into native control instructions (such as URScript, ROS Control messages) for specific robot brands through a robot access and control adapter.

[0045] Furthermore, the robot access and control adapter is a set of drive adapters for different robots.

[0046] S6 inputs native control commands to heterogeneous robots to execute target tasks.

[0047] Furthermore, the heterogeneous robot executes relevant actions sequentially based on native control commands to complete the target task.

[0048] Example 2 Example of the overall execution flow of the method of this invention: Suppose a deep learning-trained "visual grasping model" is stored in a model library and has been packaged into a GeneralVisionGraspSkill meta-skill and stored in the meta-skill library. Now, a grasping task needs to be performed on a new model of robot, Robot_B: Task Trigger: The user selects a "Grab an Apple from the Table" task template from the task library. The core of this template is to call the GeneralVisionGraspSkill meta-skill.

[0049] Skill Loading and Matching: Load the GeneralVisionGraspSkill meta-skill, read its skill description file, and obtain the minimum capability configuration required for the skill: "one 6DOF robotic arm" and "one RGB-D camera". Simultaneously, obtain the Robot_B capability configuration file and confirm that it meets the requirements.

[0050] The Robot_B configuration file contains a list of all the meta-actions the robot can support, along with component information. The system checks if a camera and a 6DOF robotic arm are present; if so, the robot possesses the capability; otherwise, it cannot perform the meta-skill. Furthermore, each component is configured with its specific technical parameters. For example, for a camera component, this includes the camera component ID and the camera-hand-eye calibration matrix.

[0051] Meanwhile, after confirming that Robot_B meets the requirements, the following action sequence is obtained by reading the skill description file of the GeneralVisionGraspSkill meta-skill: Take_Photo(camera_id)→Get_Position(image)→Move_ARM(position)→SetGripper(open)→Grasp(); respectively represent triggering the camera to take a picture (meta-action), calling the visual model to analyze and obtain the object's position (visual model), moving the robotic arm to the target position (meta-action), opening the end effector (meta-action), and grasping (meta-action); Dynamic reasoning and transformation: During execution, the Take_Photo(camera_id) meta-action is executed first to drive the robot camera with id_camera_id to take a picture and obtain the current RGB-D image. After taking the picture, the model inference request Get_Position(image) is actively sent to the meta-skill adapter, so that it calls the "visual grasping model" through the algorithm model service and inputs the current RGB-D image. The model outputs a grasping pose suggestion that is independent of the robot, such as a 6D pose of the object relative to the camera coordinate system.

[0052] The adapter module combines this abstract pose with information such as the specific camera-hand-eye calibration matrix of Robot_B, and calculates the specific target pose that the Robot_B end effector needs to reach through coordinate transformation.

[0053] After obtaining the target pose, the subsequent Move_ARM(position) → SetGripper(open) → Grasp() meta-actions are executed in sequence to complete the full meta-skill execution.

[0054] Execution and Adaptation: The generated sequence of meta-actions is handed over to the scheduling and execution engine. The engine converts each meta-action into a native instruction through the driver corresponding to Robot_B in the robot access and control adapter, and sends it to Robot_B for execution to complete the grasping.

[0055] Example 3 like Figure 3 As shown, based on the same inventive concept, this embodiment of the invention also provides a model sharing system for heterogeneous robots, including: a meta-skill acquisition module, a target task acquisition module, a minimum configuration acquisition module, a meta-action sequence output module, and a heterogeneous execution module; The meta-skill acquisition module is used to acquire the trained target model and encapsulate it into the corresponding meta-skill; The target task acquisition module is used to acquire target tasks and corresponding heterogeneous robots; The minimum configuration acquisition module is used to read the skill description file of the meta-skill based on the target task and obtain the minimum capability configuration; The meta-action sequence output module is used to determine the capabilities of heterogeneous robots based on the minimum capability configuration and to obtain meta-action sequences based on the skill description file. The heterogeneous execution module is used to convert the meta-action sequence into native control commands that the heterogeneous robot can execute; and input the native control commands to the heterogeneous robot to execute the target task.

[0056] Furthermore, in this embodiment, the functional implementation methods of each functional module correspond one-to-one with the methods described above, and will not be repeated here.

[0057] Example 4 Based on the same inventive concept, the present invention also provides an electronic device, which includes a processor and a memory, wherein the memory stores instructions, characterized in that the instructions are loaded and executed by the processor to implement a model sharing method for heterogeneous robots as in Embodiment 1.

[0058] Based on the same inventive concept, the present invention also provides a computer device, including a processor, a communication interface, a memory, and a communication bus, wherein the processor, the communication interface, and the memory communicate with each other through the communication bus; Memory, used to store computer programs; When the processor executes a program stored in memory, it can implement a model sharing method for heterogeneous robots as shown in Example 1.

[0059] The electronic device may include a processor, a communications interface, a memory, and a communication bus, wherein the processor, communications interface, and memory communicate with each other via the communication bus. The processor can invoke logical instructions in the memory to execute a model sharing method for heterogeneous robots as described in Embodiment 1.

[0060] Furthermore, the logical instructions in the aforementioned memory can be implemented as software functional units and sold or used as independent products, and can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0061] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the apparatus disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple; relevant parts can be referred to the method section.

[0062] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A model-sharing method for heterogeneous robots, characterized in that, include: Obtain the trained target model and encapsulate it into corresponding meta-skills; Obtain the target task and the corresponding heterogeneous robot; Based on the target task, the skill description file of the meta-skill is read to obtain the minimum capability configuration; The capabilities of the heterogeneous robot are determined based on the minimum capability configuration, and the meta-action sequence is obtained based on the skill description file; Based on the meta-action sequence, it is converted into native control instructions that the heterogeneous robot can execute; The target task is executed by the heterogeneous robot based on the input of the native control commands.

2. The model sharing method for heterogeneous robots as described in claim 1, characterized in that, The target model is a pre-trained, hardware-decoupled original algorithm model file that takes a high-level task objective as input and outputs abstract task space parameters. The original algorithm model file is then standardized, encapsulated, and lightweighted before being stored in the model library.

3. The model sharing method for heterogeneous robots as described in claim 2, characterized in that, Encapsulated as corresponding meta-skills, specifically including: Based on the target model, parameter verification logic, and preprocessing / postprocessing steps, a complete meta-skill plugin is woven together and registered in the meta-skill library. The skill description file is then edited to realize the combination and arrangement between the target model, adapter, and meta-actions, forming the meta-skill with clear input and output.

4. The model sharing method for heterogeneous robots as described in claim 3, characterized in that, The meta-skill includes: the adapter and the skill description file; The adapter is used to call a specific model in the model library. The output of the specific model is only related to the input result of the model algorithm itself, and is not related to the hardware and actions of the specific robot. The skill description file includes the input interface of the meta-skill, the output semantics, the minimum capability configuration of the required robot, and the judgment conditions for successful execution. It also records the model to be invoked, the adapter, and the meta-action to be executed based on the adapter result.

5. A model sharing method for heterogeneous robots as described in claim 4, characterized in that, The meta-actions are the lowest-level, indivisible atomic actions of the robot; Each of the aforementioned meta-actions is a well-encapsulated function or service, corresponding to a fundamental capability that can be reliably executed at the robot's underlying level, and its implementation is related to the robot's hardware.

6. A model sharing method for heterogeneous robots as described in claim 5, characterized in that, The target task uses a task template, obtained from the task library, and is used to invoke the meta-skill.

7. A model sharing method for heterogeneous robots as described in claim 6, characterized in that, The task library includes complete task flows oriented towards business scenarios; The task flow consists of the meta-skills and skill flow templates from the composite action library.

8. A model sharing method for heterogeneous robots as described in claim 7, characterized in that, The composite action library stores reusable skill flow templates composed of multiple meta-actions or meta-skills arranged and combined according to specific logic.

9. A model sharing method for heterogeneous robots as described in claim 8, characterized in that, The determination of the capabilities of the heterogeneous robot specifically includes: Obtain the capability configuration file of the heterogeneous robot; The capability configuration file includes a list of all meta-actions supported by the heterogeneous robot and component information; Based on the component information, determine whether the minimum capability configuration is met; If satisfied, then obtain the meta-action sequence based on the skill description file; Otherwise, it means that the target task cannot be completed and the requirements are not met.

10. A model-sharing system for heterogeneous robots, used to execute a model-sharing method for heterogeneous robots as described in any one of claims 1-9, characterized in that, include: The module includes a meta-skill acquisition module, a target task acquisition module, a minimum configuration acquisition module, a meta-action sequence output module, and a heterogeneous execution module. The meta-skill acquisition module is used to acquire the trained target model and encapsulate it into a corresponding meta-skill; The target task acquisition module is used to acquire the target task and the corresponding heterogeneous robot; The minimum configuration acquisition module is used to read the skill description file of the meta-skill based on the target task to obtain the minimum capability configuration; The meta-action sequence output module is used to determine the capabilities of the heterogeneous robot based on the minimum capability configuration, and to obtain the meta-action sequence based on the skill description file. The heterogeneous execution module is used to convert the meta-action sequence into native control instructions that the heterogeneous robot can execute; The target task is executed by the heterogeneous robot based on the input of the native control commands.

Images