Heterogeneous configuration robot neural network migration method, device, equipment and medium

Abstract

The invention discloses a heterogeneous configuration robot neural network migration method and device, equipment and a medium, and belongs to the technical field of robot control. The method comprises the steps that configuration parameters of a target robot are extracted and coded, and configuration codes are obtained; collecting a small amount of target domain data to obtain a target domain data set; based on a general skill kernel network, the target domain data set and the configuration code, training a configuration adapter to obtain a trained configuration adapter; and obtaining a control instruction of the target robot based on the configuration code, the trained configuration adapter and the general motor skill kernel. By means of the method and device, the technical problems that in the prior art, data dependence is heavy, model reuse is difficult, and the development period is long are solved.

Description

Technical Field

[0001] This application belongs to the field of robot control technology, and specifically relates to a method, device, equipment and medium for neural network transfer in heterogeneous robots. Background Technology

[0002] Currently, robots exhibit significant diversity in their mechanical structures, including robotic arms with varying numbers of joints, legged robots with different leg structures, and various wheeled or tracked mobile platforms. Developing efficient and precise motion control models for these robots (often likened to the function of the "cerebellum") is fundamental to ensuring their ability to complete complex tasks. However, for each new robot configuration or model, collecting massive amounts of data from scratch and training a deep neural network controller faces severe challenges, including long development cycles, high costs, and difficulties in data acquisition (especially in the early stages of physical machine debugging). Therefore, the industry urgently needs a technology that can quickly transfer and adapt existing key motion control capabilities to robots with different configurations.

[0003] In existing technologies, methods for achieving cross-configuration motion control model transfer or rapid deployment can be mainly divided into three categories: (1) Direct transfer and full fine-tuning: The neural network controller trained on a certain configuration robot (source robot) is directly used as the initial parameters of the model of the new configuration robot (target robot), and then the entire network is retrained (fine-tuned) using the new dataset of the target robot. This is equivalent to treating the source model as only a "pre-trained weight". Although it can improve the starting point of training, it is still necessary to collect a dataset sufficient to train the entire large network for the target robot. It fails to fundamentally reduce data requirements and training time, has heavy data dependence, and low transfer efficiency.

[0004] (2) Controller design based on analytical dynamics model: Ignoring neural network transfer, this method focuses on the target robot and directly designs and tunes controller parameters (such as PID, computational torque controller) based on its kinematic and dynamic equations (such as Lagrange equations). This method heavily relies on accurate and difficult-to-obtain robot dynamics model parameters (such as friction, inertia tensor). For robots with complex structures, model mismatch can lead to a sharp decline in control performance, and this method cannot directly reuse learned complex skills (such as dexterous manipulation strategies), making model reuse difficult.

[0005] (3) Traditional adaptive control: Design a controller with a fixed structure (such as a model-based controller) and integrate an online parameter estimator to adjust the controller parameters in real time to approximate the dynamic characteristics of the target robot. However, its online learning capacity is limited, and it can usually only adapt to slow parameter changes or linear dynamics. For the fundamental nonlinear dynamic differences brought about by robots with completely different configurations, its adaptive process is slow and unstable, and it cannot achieve "fast" adaptation, resulting in a long development cycle. Summary of the Invention

[0006] To address the aforementioned technical problems, this application provides a method, apparatus, device, and storage medium for transferring neural networks in heterogeneous robots.

[0007] To achieve the above objectives, this application provides the following technical solution: A method for neural network transfer in heterogeneous robots includes: Extract and encode the configuration parameters of the target robot to obtain the configuration code; Collect a small amount of data from the target domain to obtain the target domain dataset; Based on the general skill kernel network, the target domain dataset, and the configuration encoding, the configuration adapter is trained to obtain the trained configuration adapter; Based on the configuration encoding, the trained configuration adapter, and the general motion skill kernel, control commands for the target robot are obtained.

[0008] Further, in one embodiment, the step of extracting and encoding the configuration parameters of the target robot to obtain configuration coding includes: Obtain the key physical parameters of the target robot and generate a configuration description vector; The configuration description vector is input into the configuration encoder to obtain the configuration code.

[0009] Further, in one embodiment, training the configuration adapter based on the general skill kernel network, the target domain dataset, and the configuration encoding to obtain the trained configuration adapter includes: The current initial state of the target robot is determined based on the target domain dataset; The current original state of the target robot is concatenated with the configuration code to obtain the first concatenation vector; Normalize the current original state of the target robot to obtain the normalized current original state; The normalized current original state is concatenated with the task instructions to obtain the second concatenation vector; Input the second concatenated vector into the general skill kernel network to obtain the normalized basic action output by the general skill kernel network; Input the first splicing vector into the configuration adapter to obtain the motion correction amount output by the configuration adapter; The normalized base action is combined with the action correction amount to obtain the predicted normalized action; The predicted normalized action is denormalized to obtain the predicted original action; The loss function is calculated based on the predicted original action and the actual action performed by the target robot in the target domain dataset to obtain the loss value of the configuration adapter; If the loss value is greater than the preset threshold, continue iteratively training the configuration adapter; When the loss value is less than or equal to a preset threshold, the trained configuration adapter is obtained.

[0010] Furthermore, in one embodiment, obtaining control commands for the target robot based on the configuration encoding, the trained configuration adapter, and the general motion skill kernel includes: Read the target robot's original state and task instructions at any time within each control cycle; Normalize the original state at that moment to obtain the normalized original state at that moment; The normalized original state at that moment is concatenated with the task instruction and then input into the general motor skill kernel to obtain the normalized basic action at that moment output by the general skill kernel network. After concatenating the original state at that moment with the configuration code, the result is input into the trained configuration adapter to obtain the action correction amount at that moment output by the trained configuration adapter. Based on the normalized basic motion and motion correction amount at this moment, the control command of the target robot is obtained through the control command synthesis formula and sent to the actuator. The formula for synthesizing control commands is as follows: a_final(t) = (a_base(t) + δ(t))×σ_a + μ_a; In the formula, σ_a represents the scaling factor for motion normalization, and μ_a represents the mean offset for motion normalization.

[0011] Further, in one embodiment, after obtaining the control commands for the target robot based on the trained configuration adapter and general motion skill kernel, the process includes: Continuously monitor the control performance indicators of the target robot during operation; When the average error of the control performance index exceeds the threshold, the latest running data of the target robot within a preset time period is added to the real-time data buffer, and the parameters of the trained configuration adapter are updated online based on this data.

[0012] Further, in one embodiment, before extracting and encoding the configuration parameters of the target robot to obtain the configuration code, the following steps are included: Construct a universal motor skill kernel, which includes: a heterogeneous robot simulation training set, a unified normalized interaction interface, and a universal skill kernel network.

[0013] Furthermore, in one embodiment, constructing a general skill kernel network includes: A deep neural network with shared weights is trained using a hybrid configuration experience replay pool, and the trained network is used as a general skill kernel network. The general skill kernel network is a multilayer perceptron, which contains multiple fully connected layers and uses the ReLU activation function; The training objective is to minimize the loss function L_core: L_core = (1 / N)×Σ||G_θ(s_norm, g) a_target||² + λ×R(s, a) In the formula: L_core represents the training loss of the general skill kernel network; N represents the total number of training samples; G_θ(·) represents the general skill kernel network with parameter θ; s_norm represents the normalized robot state vector; g represents the current task instruction; a_target represents the target action to be output; λ represents the weight coefficient of the reward term; R(s, a) represents the task completion reward function; s represents the robot's original state; a represents the robot's action.

[0014] Based on the same inventive concept, this application also provides a heterogeneous robot neural network transfer device, the device comprising: The parameter encoding module is configured to extract and encode the configuration parameters of the target robot to obtain the configuration code; The data acquisition module is configured to collect a small amount of target domain data to obtain the target domain dataset; The adapter training module is configured to train the configuration adapter based on the general skill kernel network, the target domain dataset, and the configuration encoding to obtain the trained configuration adapter. The control command generation module is configured to obtain control commands for the target robot based on the configuration encoding, the trained configuration adapter, and the general motion skill kernel.

[0015] Furthermore, in one embodiment, the parameter encoding module is configured to: Obtain the key physical parameters of the target robot and generate a configuration description vector; The configuration description vector is input into the configuration encoder to obtain the configuration code.

[0016] Furthermore, in one embodiment, the adapter training module is configured to: The current initial state of the target robot is determined based on the target domain dataset; The current original state of the target robot is concatenated with the configuration code to obtain the first concatenation vector; Normalize the current original state of the target robot to obtain the normalized current original state; The normalized current original state is concatenated with the task instructions to obtain the second concatenation vector; Input the second concatenated vector into the general skill kernel network to obtain the normalized basic action output by the general skill kernel network; Input the first splicing vector into the configuration adapter to obtain the motion correction amount output by the configuration adapter; The normalized base action is combined with the action correction amount to obtain the predicted normalized action; The predicted normalized action is denormalized to obtain the predicted original action; The loss function is calculated based on the predicted original action and the actual action performed by the target robot in the target domain dataset to obtain the loss value of the configuration adapter; If the loss value is greater than the preset threshold, continue iteratively training the configuration adapter; When the loss value is less than or equal to a preset threshold, the trained configuration adapter is obtained.

[0017] Furthermore, in one embodiment, the control command generation module is configured to: Read the target robot's original state and task instructions at any time within each control cycle; Normalize the original state at that moment to obtain the normalized original state at that moment; The normalized original state at that moment is concatenated with the task instruction and then input into the general motor skill kernel to obtain the normalized basic action at that moment output by the general skill kernel network. After concatenating the original state at that moment with the configuration code, the result is input into the trained configuration adapter to obtain the action correction amount at that moment output by the trained configuration adapter. Based on the normalized basic motion and motion correction amount at this moment, the control command of the target robot is obtained through the control command synthesis formula and sent to the actuator. The formula for synthesizing control commands is as follows: a_final(t) = (a_base(t) + δ(t))×σ_a + μ_a; In the formula, σ_a represents the scaling factor for motion normalization, and μ_a represents the mean offset for motion normalization.

[0018] Furthermore, in one embodiment, the device further includes a monitoring module configured to: Continuously monitor the control performance indicators of the target robot during operation; When the average error of the control performance index exceeds the threshold, the latest running data of the target robot within a preset time period is added to the real-time data buffer, and the parameters of the trained configuration adapter are updated online based on this data.

[0019] Furthermore, in one embodiment, a general motor skill kernel module is constructed and configured to: Construct a general motor skill kernel, which includes: a heterogeneous robot simulation training set, a unified normalized interaction interface, and a general skill kernel network; The construction of a universal skill core network includes: A deep neural network with shared weights is trained using a hybrid configuration experience replay pool, and the trained network is used as a general skill kernel network. The general skill kernel network is a multilayer perceptron, which contains multiple fully connected layers and uses the ReLU activation function; The training objective is to minimize the loss function L_core: L_core = (1 / N)×Σ||G_θ(s_norm, g) a_target||² + λ×R(s, a) In the formula: L_core represents the training loss of the general skill kernel network; N represents the total number of training samples; G_θ(·) represents the general skill kernel network with parameter θ; s_norm represents the normalized robot state vector; g represents the current task instruction; a_target represents the target action to be output; λ represents the weight coefficient of the reward term; R(s, a) represents the task completion reward function; s represents the robot's original state; a represents the robot's action.

[0020] Based on the same inventive concept, this application also provides an electronic device, including: a memory and a processor; the processor is used to read and execute a computer program stored in the memory to implement the aforementioned method for neural network transfer in heterogeneous robots.

[0021] Based on the same inventive concept, this application also provides a computer storage medium storing computer-executable instructions, which, when executed, implement the aforementioned method for neural network transfer in heterogeneous robots.

[0022] Compared with the prior art, this application has the following advantages: 1. This significantly shortens the development and deployment cycle of heterogeneous robot controllers, enabling "rapid deployment." Traditional methods require collecting massive amounts of data from scratch and training a complete model for each new robot configuration, taking weeks or even months. This application employs a two-stage architecture of "offline solidification of a general-purpose skill kernel and rapid online training of a dedicated configuration adapter," pre-positioning the core development work and solidifying it within a general-purpose kernel. When faced with a new robot, only a small amount (typically hundreds to thousands of sets) of non-task-specific exploratory data needs to be collected, and a very lightweight adapter network (typically containing only a few thousand parameters) needs to be trained. This adaptation process can be completed within hours on ordinary computing devices, thereby shortening the cycle from "zero" to "usable" for a new robot controller from months in traditional methods to days or even hours, significantly improving the response speed of product iteration and field deployment.

[0023] 2. This significantly reduces the reliance on data and computing resources for model development, achieving "low-cost adaptation." Since it eliminates the need to train massive deep neural networks from scratch for new robots, this application effectively avoids the expensive costs of collecting, cleaning, and labeling massive amounts of dedicated data, as well as the long-term occupation of high-performance computing clusters when training large-scale networks. The core general-purpose skill kernel only requires a one-time large-scale offline training, the results of which can be reused by all subsequent robots. For the adaptation process of new robots, only simple and safe random exploration in the target domain is needed to obtain a small amount of data and train an adapter with a very small parameter scale. The requirements for data quality and quantity are extremely low, and the computing power requirement is also significantly reduced (a regular CPU is sufficient), making rapid adaptation possible at resource-constrained edge computing or on-site environments, greatly saving economic and time costs.

[0024] 3. Effectively balancing general robustness and individual precision in control performance, achieving "high-performance transfer." This application does not simply transfer the source model parameters, but rather employs a collaborative mechanism of "general kernel outputting basic strategies + dedicated adapters for dynamic fine-tuning." The general kernel inherits rich prior knowledge from multi-configuration, multi-task training, possessing excellent generalization ability and robustness to complex dynamics (such as water flow disturbances and uneven ground). Simultaneously, the lightweight adapter, by learning the unique configurational physical parameters and dynamic responses of the target robot, can perform refined and personalized corrections to the general strategy, thereby accurately compensating for the specific effects caused by configurational differences, manufacturing tolerances, wear, etc. This allows the transferred controller to maintain the stability of general intelligence while achieving a high degree of matching with specific hardware, ensuring final control accuracy and task success rate.

[0025] 4. Significantly improved maintainability and scalability of the robot controller system, achieving "sustainable evolution." The layered decoupled architecture of this application brings excellent system modularity. The general skill kernel can be continuously upgraded and expanded as an independent module without modifying the deployed adapters; conversely, hardware changes or performance optimizations for a single robot can be quickly completed by retraining or fine-tuning its dedicated adapter without affecting other robots. In addition, the online adaptive fine-tuning capability enables the robot to continuously optimize in actual operation, adapting to slow environmental changes or hardware wear. This concept greatly facilitates the collaborative management, unified software upgrades, and full lifecycle performance maintenance of large-scale robot swarms, strongly supporting the large-scale and sustainable application of robot swarms in scenarios such as energy inspection and power plant operation and maintenance.

[0026] This application achieves a breakthrough in efficiency for the new robot controller, moving from "months of training and massive amounts of data" to "minutes of adaptation and hundreds of sets of data." It also ensures superior control performance after migration, possesses strong engineering application value and broad industrial application potential, and solves the technical problems of heavy data dependence, difficulty in model reuse, and long development cycle in existing technologies.

[0027] Other features and advantages of this application will be set forth in the description which follows, and will be apparent in part from the description, or may be learned by practicing the application. The objectives and other advantages of this application may be realized and obtained by means of the structures pointed out in the description, claims and drawings. Attached Figure Description

[0028] To more clearly illustrate the technical solutions in the embodiments of this application 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 some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0029] Figure 1 A flowchart illustrating the method provided in the embodiments of this application; Figure 2 This is a schematic diagram of the overall process of an embodiment of this application; Figure 3 This is a schematic diagram of a real-time control architecture according to an embodiment of this application; Figure 4 This is a schematic diagram of the configuration adapter training process according to an embodiment of this application; Figure 5 This is a functional module diagram of an embodiment of the heterogeneous robot neural network transfer device of this application; Figure 6 This is a schematic diagram of the structure of an electronic device according to an embodiment of this application. Detailed Implementation

[0030] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0031] To address the shortcomings of existing technologies, refer to Figure 1 This application discloses a neural network transfer method for heterogeneous robot configurations, comprising: Step S10: Extract and encode the configuration parameters of the target robot to obtain the configuration code; In some specific embodiments, step S10 includes: Obtain the key physical parameters of the target robot and generate a configuration description vector; The configuration description vector is input into the configuration encoder to obtain the configuration code.

[0032] In this embodiment, refer to Figure 2 Extract the target configuration parameters, i.e., the key physical parameters of the target robot, to form the configuration description vector P: P = [n, L_1, L_2, ..., L_n, m_total, τ_max_1, τ_max_2, ..., τ_max_n]; Where n is the number of joints, L_i is the length of the i-th link, and m_total is the total mass of the target robot. For the first The maximum torque of each joint; Input the configuration description vector P into a lightweight configuration encoder to obtain the configuration code c_p: c_p = Encoder_φ(P), where Encoder_φ(P) ) represents a lightweight encoder with parameter φ.

[0033] Furthermore, in one embodiment, before step S10, the method further includes: Construct a universal motor skill kernel, which includes: a heterogeneous robot simulation training set, a unified normalized interaction interface, and a universal skill kernel network.

[0034] In this embodiment, we continue to refer to... Figure 2 To construct a heterogeneous robot simulation training set, specifically, in a physical simulation environment (such as MuJoCo or PyBullet), a model library containing various robot configurations is established, including: a 6-DOF serial robotic arm, a 4-DOF SCARA robot, a 3-DOF Delta parallel robot, a quadruped robot, and a wheeled mobile platform. A common set of basic motion tasks is designed for each robot configuration model.

[0035] To construct a unified, normalized interaction interface, specifically, to decouple the model input / output from the specific robot dimensions and units, the following standardized interfaces are established: State normalization: Normalize the original state vector s_raw (which includes joint position, velocity, end pose, etc.).

[0036] s_norm = (s_raw - μ_s) / σ_s; Where μ_s and σ_s are the mean vector and standard deviation vector pre-calculated based on the typical working range of each robot configuration.

[0037] Action normalization: The action commands output by the controller are normalized to the range [-1, 1].

[0038] a_norm = 2×(a_raw / a_max) - 1; Where a_max is the maximum capability vector of each executor.

[0039] A general skill core network (G-Core Net) is constructed, specifically using a deep neural network with shared weights as the G-Core Net. Its structure is a multilayer perceptron (MLP) containing 3 to 5 fully connected layers, using the ReLU activation function.

[0040] Input layer: Receives the concatenated vector I = [s_norm, g], where g is the task instruction.

[0041] Output layer: The basic action for output normalization is a_base = G_θ(I).

[0042] A deep neural network with shared weights is trained using a hybrid configuration experience replay pool, and the trained network is used as a general skill kernel network. The training objective is to minimize the following loss function L_core: L_core = (1 / N)хΣ || G_θ(s_norm, g) - a_target ||^2 + λхR(s, a); Where L_core represents the training loss of the general skill kernel network; N represents the total number of training samples; G_θ(·) represents the general skill kernel network with parameter θ; s_norm represents the normalized robot state vector; g represents the current task instruction; a_target represents the target action to be output; λ represents the weight coefficient of the reward term; R(s, a) represents the task completion reward function; s represents the robot's original state; a represents the robot's action, the first term (1 / N)хΣ || G_θ(s_norm, g) - a_target ||^2 is the mean square error of the action, and the second term λхR(s, a) is the task completion reward.

[0043] Step S20: Collect a small amount of target domain data to obtain the target domain dataset; In this embodiment, we continue to refer to... Figure 2 On the simulation model or physical entity of the target robot, a random action strategy is executed. K sets of interaction data are collected (typically K=500~2000) to form a target domain dataset D_target={(s_raw_j, a_executed_j)}. s_raw_j represents the original state of the target robot in the j-th sample in the dataset, and a_executed_j represents the actual action performed by the target robot in the j-th sample in the dataset.

[0044] Step S30: Based on the general skill kernel network, the target domain dataset, and the configuration encoding, the configuration adapter is trained to obtain the trained configuration adapter; In some specific embodiments, step S30 includes: The current initial state of the target robot is determined based on the target domain dataset; The current original state of the target robot is concatenated with the configuration code to obtain the first concatenation vector; Normalize the current original state of the target robot to obtain the normalized current original state; The normalized current original state is concatenated with the task instructions to obtain the second concatenation vector; Input the second concatenated vector into the general skill kernel network to obtain the normalized basic action output by the general skill kernel network; Input the first splicing vector into the configuration adapter to obtain the motion correction amount output by the configuration adapter; The normalized base action is combined with the action correction amount to obtain the predicted normalized action; The predicted normalized action is denormalized to obtain the predicted original action; The loss function is calculated based on the predicted original action and the actual action performed by the target robot in the target domain dataset to obtain the loss value of the configuration adapter; If the loss value is greater than the preset threshold, continue iteratively training the configuration adapter; When the loss value is less than or equal to a preset threshold, the trained configuration adapter is obtained.

[0045] In this embodiment, refer to Figure 4 The configuration adapter A_ψ is a very lightweight neural network (e.g., 2-3 layers, 64-128 hidden units).

[0046] Adapter input: The concatenated vector of the current raw state s_raw and the configuration code c_p.

[0047] Adapter output: Motion correction δ = A_ψ(s_raw, c_p).

[0048] Training process: Fix the parameters θ of the general skill kernel network (G-Core Net) and the parameters φ of the configuration encoder, and optimize only the adapter parameters ψ.

[0049] Specifically, the current raw state s_raw of the target robot is determined based on the target domain dataset; The current raw state s_raw of the target robot is concatenated with the configuration code c_p to obtain the first concatenation vector; Normalize the current raw state s_raw of the target robot to obtain the normalized current raw state s_norm: s_norm = (s_raw - μ_s) / σ_s; The normalized current original state is concatenated with the task instruction g to obtain the second concatenation vector I=[s_norm,g]. The second concatenation vector is input into the general skill kernel network to obtain the normalized basic action a_base output by the general skill kernel network. Input the first splicing vector into the configuration adapter to obtain the motion correction amount δ output by the configuration adapter; The base action is combined with the action correction amount to obtain the predicted normalized action a_pred_norm: a_pred_norm = a_base +δ; The predicted normalized action is denormalized to obtain the predicted original action a_pred_raw: a_pred_raw = a_pred_norm × σ_a + μ_a; The loss function L_adapt = || a_pred_raw - a_executed ||² is calculated based on the predicted original action and the actual action executed by the target robot in the target domain dataset to obtain the loss value of the configuration adapter; where: L_adapt is the loss value of the configuration adapter; a_pred_raw is the predicted original action; a_executed is the actual action executed by the target robot; μ_s is the state mean vector of the typical working range of each configuration robot; σ_s is the state standard deviation vector of the typical working range of each configuration robot; μ_a is the mean offset of the action normalization; σ_a is the scaling factor of the action normalization. If the loss value is greater than the preset threshold, continue iteratively training the configuration adapter; When the loss value is less than or equal to a preset threshold, the trained configuration adapter is obtained.

[0050] It should be noted that training the adapter using mini-batch stochastic gradient descent typically converges within 10 minutes.

[0051] Step S40: Based on the configuration encoding, the trained configuration adapter, and the general motion skill kernel, control commands for the target robot are obtained.

[0052] In some specific embodiments, step S40 includes: Read the target robot's original state and task instructions at any time within each control cycle; The sensor data at that moment is normalized to obtain the normalized sensor data at that moment; The normalized sensor data at that moment is concatenated with the task instructions and then input into the general motion skill kernel to obtain the normalized basic motion at that moment output by the general skill kernel network. After concatenating the sensor data at that moment with the configuration code, the data is input into the trained configuration adapter to obtain the motion correction amount at that moment output by the trained configuration adapter. Based on the normalized basic motion and motion correction amount at this moment, the control command of the target robot is obtained through the control command synthesis formula and sent to the actuator. The formula for synthesizing control commands is as follows: a_final(t) = (a_base(t) + δ(t))×σ_a + μ_a; In the formula, σ_a represents the scaling factor for motion normalization, and μ_a represents the mean offset for motion normalization.

[0053] In this embodiment, refer to Figure 3 In each control cycle (e.g., 1ms): first, sensing is performed: the target robot's original state s_raw(t) and task instruction g(t) at time t are read; Normalize the original state s_raw(t) of the target robot at time t to obtain the normalized original state s_norm(t) of the target robot at time t; After concatenating the normalized original state s_norm(t) of the target robot at time t with the task instruction g(t), the concatenation is input into the general motion skill kernel to obtain the normalized basic action a_base(t) at time t output by the general skill kernel network, a_base(t) = G_θ(s_norm(t), g(t)). The original state s_raw(t) of the target robot at time t is concatenated with the configuration code c_p and then input into the trained configuration adapter to obtain the motion correction amount δ at time t output by the trained configuration adapter. Based on the normalized basic motion and motion correction amount at time t, the control command of the target robot is obtained through the control command synthesis formula and sent to the driver. The formula for synthesizing control commands is as follows: a_final(t) = (a_base(t) + δ(t))×σ_a + μ_a; In the formula, σ_a represents the scaling factor for motion normalization, and μ_a represents the mean offset for motion normalization.

[0054] Further, in one embodiment, after step S40, the following is included: Continuously monitor the control performance indicators of the target robot during operation; When the average error of the control performance index exceeds the threshold, the latest running data of the target robot within a preset time period is added to the real-time data buffer, and the parameters of the trained configuration adapter are updated online based on this data.

[0055] In this embodiment, the control performance indicators are continuously monitored during operation. When the average error exceeds a set threshold, the latest operating data of the target robot within a preset time period is added to the real-time data buffer, and the parameters of the trained configuration adapter are updated online based on the latest operating data of the target robot within the preset time period to achieve adaptive optimization during operation.

[0056] In this application, a general motion skill network (G-CoreNet) decoupled from specific physical parameters is first trained offline to form a reusable intelligent base; then, for any new robot configuration, only a small amount of data needs to be collected and a lightweight adaptation network needs to be trained to achieve high-precision and rapid adaptation of general skills.

[0057] The decoupled "general kernel-dedicated adapter" two-level architecture clearly separates the complex robot motion control intelligence into a reusable general skill base and a lightweight adaptation module for specific hardware. This fundamentally changes the traditional development model that is strongly tied to hardware, and lays a structural foundation for achieving rapid and low-cost knowledge transfer.

[0058] A universal interaction interface based on a normalized state / action space: This decouples the control model from the robot's specific physical dimensions and units, and constructs a unified semantic space that can be "understood" and "used" by all robot configurations. This is a key prerequisite for the effective transmission and function of the universal skill core between different physical entities.

[0059] Lightweight configuration adapter and its rapid training mechanism: Instead of retraining the entire network, it learns a dynamic mapping from general output to precise control of the target robot through a dedicated network with a very small number of parameters. Its training only relies on a small amount of readily available non-task data and has low computational requirements, thus achieving a practical breakthrough in engineering with "minute-level adaptation and hour-level deployment".

[0060] A complete engineering path from offline pre-training to online rapid adaptation: It not only proposes an algorithm model, but also plans a clear and operable implementation process, covering all aspects such as general kernel construction, configuration parameter encoding, data acquisition, adapter training and integrated deployment, ensuring the high feasibility and repeatability of the method in real industrial scenarios.

[0061] This application achieves a breakthrough in efficiency for the new robot controller by using core technologies such as normalized interface decoupling, configuration parameter encoding, and lightweight correction networks. This breakthrough is achieved from "months of training and massive amounts of data" to "minutes of adaptation and hundreds of sets of data". At the same time, it ensures the superior control performance after migration, has strong engineering application value and wide industrial application potential, and solves the technical problems of heavy data dependence, difficulty in model reuse and long development cycle in the existing technology.

[0062] Furthermore, Example 1: Rapid deployment of a robot for inspecting turbine blades in a pumped storage power station.

[0063] Application background and problems: A pumped-storage power station has multiple large hydroelectric generator units, requiring regular defect inspections of the underwater blades. Traditionally, the station uses a custom-developed six-legged underwater robot (Model A) for this task. Its onboard "cerebellar neural network controller," trained over months with massive amounts of data, is capable of stably clinging to the blade surface and performing visual scans even under water current disturbances. The station has now purchased a new, higher-performance, heavier-load eight-legged underwater robot (Model B), with a completely different number of joints, leg configuration, and mass distribution compared to Model A. Re-collecting underwater data and training a new controller for Model B would be time-consuming and costly, disrupting the normal inspection schedule.

[0064] Solutions employing the technology of this application: Phase 1: Leveraging existing foundations (universal skills core network): A general-purpose skill kernel network (G-Core Net) was pre-trained. This network was trained in a simulation environment using data from a hexapod robot A, a four-wheeled underwater ROV, and various biomimetic swimming robots. It mastered general skills such as maintaining stability in turbulent environments, resisting water flow impacts, and achieving precise foot landings.

[0065] Phase Two: Rapid Adaptation for Target Robot B Step 1.1: Configuration parameter extraction and encoding: Measure the physical parameters of robot B to form the configuration vector P_B: P_B = [Number of joints: 8, Leg length: [0.8, 0.8, ...], Total mass: 120kg, Maximum torque per joint: 350Nm, ...]; Input this configuration vector P_B into the fixed configuration encoder to obtain its unique configuration code c_B.

[0066] Step 1.2: Collect a small amount of data from the target domain: In the test tank of the power plant (simulating a real water flow environment), robot B was operated to freely explore for approximately 2 hours. The operator remotely controlled it to walk with a random gait, climb slopes, and resist simulated water flow. This process did not perform any specific inspection tasks, but only collected approximately 1500 data pairs (raw joint angles / velocities / torques, and executed control commands).

[0067] Step 1.3: Training the lightweight configuration adapter: Data preprocessing: The collected raw state data is normalized according to the same rules used when training G-Core Net.

[0068] Fixed kernel inference: The normalized state data, along with the simple task instruction to "keep the current position", is input into G-Core Net to obtain a series of general basic actions a_base.

[0069] Adapter Training: Construct a lightweight adapter network A_B with only three layers and approximately 5000 parameters. Its inputs are the raw state data and configuration code c_B of robot B, and its output is the motion correction δ. The training objective is to minimize the following loss: Loss = || (a_base + δ) - Action after denormalization - Actual control commands acquired ||^2.

[0070] Step 1.4: System Integration and Field Testing The trained "General Skill Core Network (G-Core Net), encoder, and adapter A_B" were integrated into the embedded control unit of robot B. Testing was conducted inside a real turbine tailrace pipe. Perception: Robot B's sensors read the body posture and joint status in real time under the impact of water flow.

[0071] Control: The general kernel calculates the general gait basic motion adapted to the water flow based on the normalized state and the instruction to "move along the blade surface"; at the same time, the adapter fine-tunes the basic motion based on the real-time state of B and its own configuration encoding to compensate for the difference between its eight-legged structure and A, as well as its greater mass inertia.

[0072] Execution: The synthesized high-precision control commands drive the eight joints of B, enabling it to stably and flexibly complete the full-coverage inspection of the blades.

[0073] Furthermore, Example 2: Rapid expansion and adaptation of large-scale photovoltaic power plant panel cleaning robot clusters.

[0074] 1. Application Background and Problems: A large desert photovoltaic power station has deployed 20 tracked photovoltaic panel cleaning robots (model C). Their controllers have been optimized, enabling them to autonomously navigate and clean tilted, dusty panels. Due to capacity expansion, the power station needs to add 30 new wheel-legged hybrid cleaning robots (model D). The D-type robot combines the speed of wheeled movement with the stability of legged obstacle crossing, but its configuration and control dynamics are completely different from the C-type. Developing controllers for each of the 30 D-type robots would be an enormous undertaking.

[0075] 2. The solution adopted in this application: Phase 1: Preparation of the core general skills: Using historical operational data (tens of thousands of hours) from 20 existing C-type robots, combined with simulation data from other ground mobile robots, a general-purpose photovoltaic power plant terrain movement and cleaning general-skill kernel network (G-CoreNet) was trained. This general-skill kernel network has mastered general skills such as navigation between photovoltaic arrays, identification of panel boundaries, adjustment of brush head pressure, and adaptation to different slopes.

[0076] Phase Two: Rapid Mass Adaptation of New D-Type Robots Step 2.1: Unified coding of configuration parameters; Extract the configuration parameter vector P_D of the D-type robot to highlight its wheel-leg composite characteristics: P_D = [Number of driving wheels: 4, Number of swing legs: 2, Wheel diameter: 0.25m, Leg length: 0.5m, Mode switching mechanism parameters...]; The configuration code c_D is obtained through the configuration encoder.

[0077] Step 2.2: Data Acquisition and Adapter Training for a Single Prototype Unit: One D-type robot was selected as a prototype and allowed to autonomously explore and operate in the power plant's test area for half a day. It attempted to cross simulated cable trenches and ramps, and performed cleaning actions on dirty panels. Approximately 800 sets of state-action data were collected.

[0078] Using the same process as in Example 1, a lightweight adapter A_D specifically for the D-type architecture is trained using prototype data with the general-purpose skill core network G-Core Net fixed.

[0079] Step 2.3: Clustered Deployment and Individual Fine-tuning: Batch deployment: The trained A_D adapter, along with the general kernel and encoder, is batch-programmed into the controllers of the remaining 29 D-type robots.

[0080] Self-learning upon startup: Each D-type robot performs a 15-minute rapid self-calibration process upon its first startup and operation. During this process, the robot performs small-amplitude wheel rotations and leg swings in place, and collects approximately 100 sets of unique sensor data (minor manufacturing differences, tire wear, etc.) using feedback from its built-in IMU and encoders.

[0081] Individual fine-tuning: Based on these 100 sets of data, each robot performs online incremental fine-tuning of its shared copy of the A_D adapter over several minutes, generating a fine-tuned individual adaptation parameter ψ_i. This process is fully automated and requires no human intervention.

[0082] Step 2.4: Cluster Collaborative Operations: All D-type robots join the existing C-type robot cleaning cluster. The central dispatch system issues task instructions (such as "clean array 7"). When the D-type robots are performing tasks, the General Skill Core Network (G-Core Net) outputs basic movement instructions based on the map and their own normalized position information.

[0083] Configuration adaptation: The A_D adapter (combined with the personal fine-tuning parameter ψ_i) dynamically adjusts the basic movement commands based on real-time attitude, wheel slippage signals, etc., to ensure optimal wheel-foot coordination. For example, it can switch to wheel mode for high-speed travel on flat roads and automatically switch to foot mode for stable obstacle crossing when encountering ditches and bumps.

[0084] Task execution: The system accurately reached the designated panel and invoked the cleaning skill module in the general kernel to complete the task.

[0085] This application significantly shortens the development and deployment cycle of heterogeneous robot controllers, achieving "rapid deployment." Traditional methods require collecting massive amounts of data from scratch and training a complete model for each new robot configuration, taking weeks or even months. This application, through a two-stage architecture of "offline solidification of a general-purpose skill kernel and rapid online training of a dedicated configuration adapter," pre-positions the core development work and solidifies it within a general-purpose kernel. When faced with a new robot, only a small amount (typically hundreds to thousands of sets) of non-task-specific exploratory data needs to be collected, and a very lightweight adapter network (typically containing only a few thousand parameters) needs to be trained. This adaptation process can be completed within hours on ordinary computing devices, thereby shortening the cycle from "zero" to "usable" for a new robot controller from months in traditional methods to days or even hours, significantly improving the response speed of product iteration and field deployment.

[0086] 2. This significantly reduces the reliance on data and computing resources for model development, achieving "low-cost adaptation." Since it eliminates the need to train massive deep neural networks from scratch for new robots, this application effectively avoids the expensive costs of collecting, cleaning, and labeling massive amounts of dedicated data, as well as the long-term occupation of high-performance computing clusters when training large-scale networks. The core general-purpose skill kernel only requires a one-time large-scale offline training, the results of which can be reused by all subsequent robots. For the adaptation process of new robots, only simple and safe random exploration in the target domain is needed to obtain a small amount of data and train an adapter with a very small parameter scale. The requirements for data quality and quantity are extremely low, and the computing power requirement is also significantly reduced (a regular CPU is sufficient), making rapid adaptation possible at resource-constrained edge computing or on-site environments, greatly saving economic and time costs.

[0087] 3. Effectively balancing general robustness and individual precision in control performance, achieving "high-performance transfer." This application does not simply transfer the source model parameters, but rather employs a collaborative mechanism of "general kernel outputting basic strategies + dedicated adapters for dynamic fine-tuning." The general kernel inherits rich prior knowledge from multi-configuration, multi-task training, possessing excellent generalization ability and robustness to complex dynamics (such as water flow disturbances and uneven ground). Simultaneously, the lightweight adapter, by learning the unique configurational physical parameters and dynamic responses of the target robot, can perform refined and personalized corrections to the general strategy, thereby accurately compensating for the specific effects caused by configurational differences, manufacturing tolerances, wear, etc. This allows the transferred controller to maintain the stability of general intelligence while achieving a high degree of matching with specific hardware, ensuring final control accuracy and task success rate.

[0088] 4. Significantly improved maintainability and scalability of the robot controller system, achieving "sustainable evolution." The layered decoupled architecture of this application brings excellent system modularity. The general skill kernel can be continuously upgraded and expanded as an independent module without modifying the deployed adapters; conversely, hardware changes or performance optimizations for a single robot can be quickly completed by retraining or fine-tuning its dedicated adapter without affecting other robots. In addition, the online adaptive fine-tuning capability enables the robot to continuously optimize in actual operation, adapting to slow environmental changes or hardware wear. This concept greatly facilitates the collaborative management, unified software upgrades, and full lifecycle performance maintenance of large-scale robot swarms, strongly supporting the large-scale and sustainable application of robot swarms in scenarios such as energy inspection and power plant operation and maintenance.

[0089] Based on the same inventive concept, this application also provides a heterogeneous robot neural network transfer device.

[0090] In one embodiment, reference is made to Figure 5 , Figure 5 This is a functional module diagram of an embodiment of the heterogeneous robot neural network transfer device of this application. Figure 5 As shown, the heterogeneous robot neural network transfer device includes: The parameter encoding module 10 is configured to extract and encode the configuration parameters of the target robot to obtain the configuration code. Data acquisition module 20 is configured to collect a small amount of target domain data to obtain a target domain dataset; The adapter training module 30 is configured to train the configuration adapter based on the general skill kernel network, the target domain dataset, and the configuration encoding to obtain the trained configuration adapter. The control instruction generation module 40 is configured to obtain control instructions for the target robot based on the configuration encoding, the trained configuration adapter, and the general motion skill kernel.

[0091] Optionally, in one embodiment, the parameter encoding module 10 is configured to: Obtain the key physical parameters of the target robot and generate a configuration description vector; The configuration description vector is input into the configuration encoder to obtain the configuration code.

[0092] Optionally, in one embodiment, the adapter training module 30 is configured to: The current initial state of the target robot is determined based on the target domain dataset; The current original state of the target robot is concatenated with the configuration code to obtain the first concatenation vector; Normalize the current original state of the target robot to obtain the normalized current original state; The normalized current original state is concatenated with the task instructions to obtain the second concatenation vector; Input the second concatenated vector into the general skill kernel network to obtain the normalized basic action output by the general skill kernel network; Input the first splicing vector into the configuration adapter to obtain the motion correction amount output by the configuration adapter; The normalized base action is combined with the action correction amount to obtain the predicted normalized action; The predicted normalized action is denormalized to obtain the predicted original action; The loss function is calculated based on the predicted original action and the actual action performed by the target robot in the target domain dataset to obtain the loss value of the configuration adapter; If the loss value is greater than the preset threshold, continue iteratively training the configuration adapter; When the loss value is less than or equal to a preset threshold, the trained configuration adapter is obtained.

[0093] Optionally, in one embodiment, the control command generation module 40 is configured to: Read the target robot's original state and task instructions at any time within each control cycle; Normalize the original state at that moment to obtain the normalized original state at that moment; The normalized original state at that moment is concatenated with the task instruction and then input into the general motor skill kernel to obtain the normalized basic action at that moment output by the general skill kernel network. After concatenating the original state at that moment with the configuration code, the result is input into the trained configuration adapter to obtain the action correction amount at that moment output by the trained configuration adapter. Based on the normalized basic motion and motion correction amount at this moment, the control command of the target robot is obtained through the control command synthesis formula and sent to the actuator. The formula for synthesizing control commands is as follows: a_final(t) = (a_base(t) + δ(t))×σ_a + μ_a; In the formula, σ_a represents the scaling factor for motion normalization, and μ_a represents the mean offset for motion normalization.

[0094] Optionally, in one embodiment, the device further includes a monitoring module configured to: Continuously monitor the control performance indicators of the target robot during operation; When the average error of the control performance index exceeds the threshold, the latest running data of the target robot within a preset time period is added to the real-time data buffer, and the parameters of the trained configuration adapter are updated online based on this data.

[0095] Optionally, in one embodiment, the device further includes a module for constructing a general motor skill kernel, configured to: Construct a general motor skill kernel, which includes: a heterogeneous robot simulation training set, a unified normalized interaction interface, and a general skill kernel network; The construction of a universal skill core network includes: A deep neural network with shared weights is trained using a hybrid configuration experience replay pool, and the trained network is used as a general skill kernel network. The general skill kernel network is a multilayer perceptron, which contains multiple fully connected layers and uses the ReLU activation function; The training objective is to minimize the loss function L_core: L_core = (1 / N)×Σ||G_θ(s_norm, g) a_target||² + λ×R(s, a) In the formula: L_core represents the training loss of the general skill kernel network; N represents the total number of training samples; G_θ(·) represents the general skill kernel network with parameter θ; s_norm represents the normalized robot state vector; g represents the current task instruction; a_target represents the target action to be output; λ represents the weight coefficient of the reward term; R(s, a) represents the task completion reward function; s represents the robot's original state; a represents the robot's action.

[0096] The functions of each module in the above-mentioned heterogeneous robot neural network transfer device correspond to the steps in the above-mentioned heterogeneous robot neural network transfer method embodiment, and their functions and implementation processes will not be described in detail here.

[0097] Based on the same inventive concept, embodiments of this application also provide an electronic device, the structure of which is as follows: Figure 6 As shown, it includes: a memory and a processor, wherein the processor is used to read and execute the computer program stored in the memory to implement the aforementioned method for neural network transfer in a heterogeneous robot.

[0098] Based on the same inventive concept, this application also provides a computer storage medium storing computer-executable instructions, which, when executed, implement the aforementioned method for neural network transfer in heterogeneous robots.

[0099] Finally, it should be noted that while some processes described in the embodiments of this application include multiple operations or steps that appear in a specific order, it should be understood that these operations or steps may not be executed in the order they appear in the embodiments of this application, or may be executed in parallel. The sequence number of the operation is only used to distinguish different operations, and the sequence number itself does not represent any execution order. In addition, these processes may include more or fewer operations, and these operations or steps may be executed sequentially or in parallel, and these operations or steps may be combined.

[0100] Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application.

Claims

1. A neural network transfer method for heterogeneous robot configurations, characterized in that, include: Extract and encode the configuration parameters of the target robot to obtain the configuration code; Collect a small amount of data from the target domain to obtain the target domain dataset; Based on the general skill kernel network, the target domain dataset, and the configuration encoding, the configuration adapter is trained to obtain the trained configuration adapter; Based on the configuration encoding, the trained configuration adapter, and the general motion skill kernel, control commands for the target robot are obtained.

2. The method according to claim 1, characterized in that, The process of extracting and encoding the configuration parameters of the target robot to obtain configuration coding includes: Obtain the key physical parameters of the target robot and generate a configuration description vector; The configuration description vector is input into the configuration encoder to obtain the configuration code.

3. The method according to claim 1, characterized in that, The training of the configuration adapter based on the general skill kernel network, the target domain dataset, and the configuration encoding to obtain the trained configuration adapter includes: The current initial state of the target robot is determined based on the target domain dataset; The current original state of the target robot is concatenated with the configuration code to obtain the first concatenation vector; Normalize the current original state of the target robot to obtain the normalized current original state; The normalized current original state is concatenated with the task instructions to obtain the second concatenation vector; Input the second concatenated vector into the general skill kernel network to obtain the normalized basic action output by the general skill kernel network; Input the first splicing vector into the configuration adapter to obtain the motion correction amount output by the configuration adapter; The normalized base action is combined with the action correction amount to obtain the predicted normalized action; The predicted normalized action is denormalized to obtain the predicted original action; The loss function is calculated based on the predicted original action and the actual action performed by the target robot in the target domain dataset to obtain the loss value of the configuration adapter; If the loss value is greater than the preset threshold, continue iteratively training the configuration adapter; When the loss value is less than or equal to a preset threshold, the trained configuration adapter is obtained.

4. The method according to claim 1, characterized in that, The control commands for the target robot, obtained based on the configuration encoding, the trained configuration adapter, and the general motion skill kernel, include: Read the target robot's original state and task instructions at any time within each control cycle; Normalize the original state at that moment to obtain the normalized original state at that moment; The normalized original state at that moment is concatenated with the task instruction and then input into the general motor skill kernel to obtain the normalized basic action at that moment output by the general skill kernel network. After concatenating the original state at that moment with the configuration code, the result is input into the trained configuration adapter to obtain the action correction amount at that moment output by the trained configuration adapter. Based on the normalized basic motion and motion correction amount at this moment, the control command of the target robot is obtained through the control command synthesis formula and sent to the actuator. The formula for synthesizing control commands is as follows: a_final(t) = (a_base(t) + δ(t))×σ_a + μ_a; In the formula, σ_a represents the scaling factor for motion normalization, and μ_a represents the mean offset for motion normalization.

5. The method according to claim 1, characterized in that, After obtaining control commands for the target robot based on the trained configuration adapter and general motion skill kernel, the following steps are included: Continuously monitor the control performance indicators of the target robot during operation; When the average error of the control performance index exceeds the threshold, the latest running data of the target robot within a preset time period is added to the real-time data buffer, and the parameters of the trained configuration adapter are updated online based on this data.

6. The method according to claim 1, characterized in that, Before extracting and encoding the configuration parameters of the target robot to obtain the configuration code, the process includes: Construct a universal motor skill kernel, which includes: a heterogeneous robot simulation training set, a unified normalized interaction interface, and a universal skill kernel network.

7. The method according to claim 6, characterized in that, Constructing a universal skill core network, including: A deep neural network with shared weights is trained using a hybrid configuration experience replay pool, and the trained network is used as a general skill kernel network. The general skill kernel network is a multilayer perceptron, which contains multiple fully connected layers and uses the ReLU activation function; The training objective is to minimize the loss function L_core: L_core = (1 / N)×Σ||G_θ(s_norm, g) a_target||² + λ×R(s, a) In the formula: L_core represents the training loss of the general skill kernel network; N represents the total number of training samples; G_θ(·) represents the general skill kernel network with parameter θ; s_norm represents the normalized robot state vector; g represents the current task instruction; a_target represents the target action to be output; λ represents the weight coefficient of the reward term; R(s, a) represents the task completion reward function; s represents the robot's original state; a represents the robot's action.

8. A heterogeneous robot neural network transfer device, characterized in that, The device includes: The parameter encoding module is configured to extract and encode the configuration parameters of the target robot to obtain the configuration code; The data acquisition module is configured to collect a small amount of target domain data to obtain the target domain dataset; The adapter training module is configured to train the configuration adapter based on the general skill kernel network, the target domain dataset, and the configuration encoding to obtain the trained configuration adapter. The control command generation module is configured to obtain control commands for the target robot based on the configuration encoding, the trained configuration adapter, and the general motion skill kernel.

9. The apparatus according to claim 8, characterized in that, The parameter encoding module is configured to: Obtain the key physical parameters of the target robot and generate a configuration description vector; The configuration description vector is input into the configuration encoder to obtain the configuration code.

10. The apparatus according to claim 8, characterized in that, The adapter training module is configured to: The current initial state of the target robot is determined based on the target domain dataset; The current original state of the target robot is concatenated with the configuration code to obtain the first concatenation vector; Normalize the current original state of the target robot to obtain the normalized current original state; The normalized current original state is concatenated with the task instructions to obtain the second concatenation vector; Input the second concatenated vector into the general skill kernel network to obtain the normalized basic action output by the general skill kernel network; Input the first splicing vector into the configuration adapter to obtain the motion correction amount output by the configuration adapter; The normalized base action is combined with the action correction amount to obtain the predicted normalized action; The predicted normalized action is denormalized to obtain the predicted original action; The loss function is calculated based on the predicted original action and the actual action performed by the target robot in the target domain dataset to obtain the loss value of the configuration adapter; If the loss value is greater than the preset threshold, continue iteratively training the configuration adapter; When the loss value is less than or equal to a preset threshold, the trained configuration adapter is obtained.

11. The apparatus according to claim 8, characterized in that, The control command generation module is configured to: Read the target robot's original state and task instructions at any time within each control cycle; Normalize the original state at that moment to obtain the normalized original state at that moment; The normalized original state at that moment is concatenated with the task instruction and then input into the general motor skill kernel to obtain the normalized basic action at that moment output by the general skill kernel network. After concatenating the original state at that moment with the configuration code, the result is input into the trained configuration adapter to obtain the action correction amount at that moment output by the trained configuration adapter. Based on the normalized basic motion and motion correction amount at this moment, the control command of the target robot is obtained through the control command synthesis formula and sent to the actuator. The formula for synthesizing control commands is as follows: a_final(t) = (a_base(t) + δ(t))×σ_a + μ_a; In the formula, σ_a represents the scaling factor for motion normalization, and μ_a represents the mean offset for motion normalization.

12. The apparatus according to claim 8, characterized in that, The device also includes a monitoring module configured to: Continuously monitor the control performance indicators of the target robot during operation; When the average error of the control performance index exceeds the threshold, the latest running data of the target robot within a preset time period is added to the real-time data buffer, and the parameters of the trained configuration adapter are updated online based on this data.

13. The apparatus according to claim 8, characterized in that, The device also includes a module for constructing a general motor skill kernel, configured to: Construct a general motor skill kernel, which includes: a heterogeneous robot simulation training set, a unified normalized interaction interface, and a general skill kernel network; The construction of a universal skill core network includes: A deep neural network with shared weights is trained using a hybrid configuration experience replay pool, and the trained network is used as a general skill kernel network. The general skill kernel network is a multilayer perceptron, which contains multiple fully connected layers and uses the ReLU activation function; The training objective is to minimize the loss function L_core: L_core = (1 / N)×Σ||G_θ(s_norm, g) a_target||² + λ×R(s, a) In the formula: L_core represents the training loss of the general skill kernel network; N represents the total number of training samples; G_θ(·) represents the general skill kernel network with parameter θ; s_norm represents the normalized robot state vector; g represents the current task instruction; a_target represents the target action to be output; λ represents the weight coefficient of the reward term; R(s, a) represents the task completion reward function; s represents the robot's original state; a represents the robot's action.

14. An electronic device, characterized in that, include: Memory, processor; The processor is configured to read and execute the computer program stored in the memory to implement the heterogeneous robot neural network transfer method according to any one of claims 1-7.

15. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer-executable instructions, which, when executed, implement the heterogeneous robot neural network transfer method according to any one of claims 1-7.

Images