Robot meta-skill management method, device and equipment and storage medium

Abstract

The invention discloses a robot meta-skill management method, device and equipment and a storage medium, and relates to the technical field of robots. The method comprises the following steps: acquiring task execution data of a robot, and slicing the task execution data to obtain a meta-skill training sample; training an initial model by using the meta-skill training sample to obtain a target model, and generating a meta-skill data packet according to the meta-skill training sample and model parameters of the target model; a user task is generated according to meta-skill editing information input by a user and the meta-skill data packet, the user task is operated, and the user task comprises subtasks corresponding to the edited meta-skill data packet. According to the technical scheme of the embodiment of the invention, a complete closed-loop system integrating multi-source data acquisition, model training, meta-skill packaging and task self-definition is provided.

Description

Technical Field

[0001] This invention relates to the field of robotics, and particularly to a method, apparatus, device, and storage medium for managing robot meta-skills. Background Technology

[0002] In the industrial manufacturing sector, with the maturation of humanoid robots and embodied intelligence technologies, companies hope to use robots to undertake more dexterous, hand-coordinated, and compliant tasks, such as tightening screws, sorting packages, or moving irregular objects.

[0003] As humanoid robot technology matures, higher demands are being placed on it, with expectations that robots can perform more dexterous, complex tasks requiring hand coordination and compliant manipulation, such as tightening screws in precision assembly, identifying and grasping packages in logistics sorting, and safely handling irregularly shaped objects. These tasks often rely on multimodal perception and decision-making, including vision and force sensing. Traditional control methods based on fixed rules or single-modal programming are difficult to adapt to dynamically changing work scenarios.

[0004] However, the current generation efficiency of meta-skills is low, relying on manual labor and lacking an automatic generation system for meta-skills (such as basic action units like grasping, twisting, and placing). This results in long development cycles, low reusability, and difficulty in quickly responding to new task requirements. Consequently, the deployment efficiency and reusability of industrial robots in complex tasks are both low. Summary of the Invention

[0005] This invention provides a robot meta-skill management method, apparatus, device, and storage medium to address the problem of low deployment efficiency and low reuse value of robots in complex tasks.

[0006] In a first aspect, the present invention provides a robot meta-skill management method, comprising:

[0007] Obtain the robot's task execution data and slice the task execution data to obtain meta-skill training samples;

[0008] An initial model is trained using the meta-skill training samples to obtain a target model, and a meta-skill data package is generated based on the model parameters of the meta-skill training samples and the target model.

[0009] A user task is generated based on the user-input meta-skill editing information and the meta-skill data package, and the user task is run. The user task includes a sub-task corresponding to the edited meta-skill data package.

[0010] In a second aspect, the present invention provides a robot meta-skill management device, comprising:

[0011] The training sample determination module is used to acquire the robot's task execution data and slice the task execution data to obtain meta-skill training samples.

[0012] The meta-skill data package generation module is used to train an initial model using the meta-skill training samples to obtain a target model, and to generate a meta-skill data package based on the model parameters of the meta-skill training samples and the target model.

[0013] The task execution module is used to generate user tasks based on the meta-skill editing information input by the user and the meta-skill data package, and to run the user tasks, wherein the user tasks include sub-tasks corresponding to the edited meta-skill data package.

[0014] Thirdly, the present invention provides an electronic device comprising:

[0015] At least one processor;

[0016] and memory that is communicatively connected to at least one processor;

[0017] The memory stores a computer program that can be executed by at least one processor, which enables the at least one processor to perform the robot meta-skill management method of the first aspect described above.

[0018] Fourthly, the present invention provides a computer-readable storage medium storing computer instructions for causing a processor to execute the robot meta-skill management method of the first aspect described above.

[0019] The robot meta-skill management solution provided by this invention realizes a complete system for generating robot training samples, training models, generating and editing robot meta-skills, and customizing user tasks. It improves the generation efficiency, reliability, and reusability of industrial robot meta-skills, enables the rapid construction of industrial robot meta-skills, reduces reliance on professional personnel, and significantly improves the deployment efficiency and reuse value of industrial robots in complex tasks.

[0020] It should be understood that the description in this section is not intended to identify key or essential features of the invention, nor is it intended to limit the scope of the invention. Other features of the invention will become readily apparent from the following description. Attached Figure Description

[0021] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0022] Figure 1 This is a flowchart of a robot meta-skill management method according to Embodiment 1 of the present invention;

[0023] Figure 2 This is a flowchart of a robot meta-skill management method according to Embodiment 2 of the present invention;

[0024] Figure 3 This is a schematic diagram of the structure of a robot meta-skill management device according to Embodiment 3 of the present invention;

[0025] Figure 4 This is a schematic diagram of the structure of an electronic device provided according to Embodiment 4 of the present invention. Detailed Implementation

[0026] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. 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 should fall within the scope of protection of the present invention.

[0027] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that embodiments of the invention described herein can be implemented in orders other than those illustrated or described herein. In the description of this invention, unless otherwise stated, "a plurality of" means two or more. "And / or" describes the relationship between related objects, indicating that three relationships can exist; for example, A and / or B can represent: A alone, A and B simultaneously, and B alone. The character " / " generally indicates that the preceding and following related objects are in an "or" relationship. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or device that includes a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or devices.

[0028] Example 1

[0029] Figure 1 The flowchart of a robot meta-skill management method provided in Embodiment 1 of the present invention is applicable to the management of robot meta-skills. The method can be executed by a robot meta-skill management device, which can be implemented in hardware and / or software. The robot meta-skill management device can be configured in an electronic device, which can be composed of two or more physical entities or a single physical entity.

[0030] like Figure 1 As shown, the robot meta-skill management method provided in Embodiment 1 of the present invention specifically includes the following steps:

[0031] S101. Obtain the robot's task execution data and slice the task execution data to obtain meta-skill training samples.

[0032] In this embodiment, task execution data during robot task execution can be acquired through an end-effector layer and a data acquisition layer deployed on a local terminal. The end-effector layer may include a dual-arm humanoid robot with full-body force control capabilities, an end effector, a force or tactile sensor array, multimodal sensors (such as RGB-D cameras, IMUs, torque sensors, tactile arrays, and joint encoders), and an edge control box. The end-effector layer can provide a high-speed local bus, supporting millisecond-level force control closed-loop. The edge control box is responsible for performing real-time control, safety constraints, data caching, and local preprocessing, mapping the robot control interface to cloud standard protocols. Task execution data typically includes multiple meta-skill (motion) data. Meta-skill training samples can be obtained by slicing the task execution data using a pre-built meta-skill factory. Each meta-skill training sample typically includes task execution data for one meta-skill.

[0033] S102. Train an initial model using the meta-skill training samples to obtain a target model, and generate a meta-skill data package based on the model parameters of the meta-skill training samples and the target model.

[0034] In this embodiment, the initial model can be trained using meta-skill training samples through the meta-skill factory to obtain the target model, and a meta-skill data package containing the model parameters of the target model and the corresponding meta-skill training samples can be generated. The meta-skill data package is then saved to the meta-skill factory.

[0035] S103. Generate a user task based on the user-input meta-skill editing information and the meta-skill data package, and run the user task, wherein the user task includes a sub-task corresponding to the edited meta-skill data package.

[0036] In this embodiment, the meta-skill factory can receive meta-skill editing information input by the user and adjust the meta-skill data package according to the editing information to obtain the adjusted meta-skill data package. For example, sorting and combining multiple meta-skill data packages yields the adjusted meta-skill data package. A corresponding user task can be generated based on this adjusted meta-skill data package. Running this user task enables the user to manage the meta-skills of the industrial robot, such as rapid orchestration and testing.

[0037] The technical solution of this invention realizes a complete system for generating robot training samples, training models, generating robot meta-skills, editing robot meta-skills, and customizing user tasks. It improves the generation efficiency, reliability, and reusability of industrial robot meta-skills, enables the rapid construction of industrial robot meta-skills, reduces reliance on professional personnel, and significantly improves the deployment efficiency and reuse value of industrial robots in complex tasks.

[0038] Optionally, the method is applied in the cloud.

[0039] Specifically, the meta-skill factory described above can be deployed in the cloud. Once deployed in the cloud, this method can achieve rapid construction, testing, and simulation of industrial robot meta-skills through interaction with local terminals.

[0040] Optionally, acquiring the robot's task execution data and slicing the task execution data to obtain meta-skill training samples includes: acquiring multimodal task execution data of the robot through VR teleoperation technology and / or UMI robot data acquisition kit, wherein the task execution data includes at least visual data, motion data, force control data, and task semantic data; using an anomaly detector based on self-supervised contrastive learning to remove abnormal trajectories in the task execution data to obtain data to be processed, and performing semantic slicing on the data to be processed to obtain meta-skill training samples.

[0041] Specifically, the VR teleoperation system comprises a head-mounted display, two-hand controllers or exoskeleton gloves, a force feedback device, external motion capture (IMU array or optical marker system), and the robot's built-in multimodal sensors (such as RGB-D, haptic, and joint encoders). The sensors are connected to the ROS2 network via a high-precision time synchronization module, uniformly publishing timestamped messages. Force control data is broadcast in real-time by joint drives and written to the same ROS2 session, ensuring millisecond-level alignment. The workflow includes: the operator viewing the robot's first-person perspective in real-time on the VR device, controlling hand movements, and simultaneously recording multimodal data such as joint angle q(t), Cartesian trajectory x(t), torque τ(t), haptic signal h(t), and semantic commands L(t). Image data can be compressed and stored using JPEG, H.265, or learnable compression encoders. The data acquisition process only ensures smooth playback, without requiring extremely low latency. Raw data can be written to a local cache with low latency and then uploaded to the cloud in batches during idle periods.

[0042] The UMI robot data acquisition kit includes a camera, a multi-DOF master controller, a gripper, and force or haptic feedback modules. The kit's sampling rate and signal interface are consistent with the sensors used in the aforementioned VR teleoperation. To avoid space constraints at the user's location, the camera in the UMI robot data acquisition kit can be a fixed, high-resolution head-mounted camera (optional 4K wide-angle + depth), used to capture the global scene and operator hand movements, generating additional perspectives. The operator's recorded trajectory, visual stream, and voice commands can be first written to local secure storage media, encrypted, packaged, and then uploaded to the cloud when a network is available. A real-time preview mode can also be selected, transmitting only low-bitrate images and status information.

[0043] Furthermore, for tasks requiring high-precision trajectories (such as screw fastening or fitting assembly), external optical motion capture systems or inertial motion capture suits can be deployed to output the human skeleton and tool trajectories. The external motion capture timestamp is connected to a unified clock via a hardware synchronization interface and aligned with the robot data using an interpolation algorithm to achieve data fusion.

[0044] The task execution data can include visual data, force control data, motion data, and semantic data. After acquisition, the task execution data can be encoded using JPEG, PNG, or learnable super-resolution compression algorithms. To ensure sensor synchronization, precise timestamps and frame IDs are recorded when saving data, and the compressed data is stored along with metadata. The real-time monitoring interface obtains a low frame rate (e.g., 10-15 FPS) preview stream via ROS2 topics, without pursuing extremely low latency. The data storage channel uses a high frame rate (≥60 FPS) and high synchronization configuration, and data integrity is ensured by fragmenting during upload. The real-time visual data stream uses standard encoding (e.g., H.264 / H.265) to meet remote monitoring requirements and ensure a usable frame rate. The original frames used for training are reversibly compressed locally using JPEG / PNG, retaining high fidelity and recording the compression ratio. Non-visual data (such as force control data, motion data, and semantic data) is serialized using binary serialization (e.g., FlatBuffers / Protobuf) and further compressed using differential encoding. Integrity can be ensured by using a low-latency real-time command channel (TCP / UDP hybrid) and a batch data channel with breakpoint resumption and CRC check.

[0045] Next, a unified timeline T can be constructed using the data governance module through the meta-skill factory, mapping visual data, motion data, force control data, and task semantic data to synchronization frames D(t)={I(t),q_r(t),τ(t),L_struct(t)}. Anomaly detectors based on self-supervised contrastive learning are used to filter out abnormal trajectories. When the detector's output score(t) < θ, it is marked as an anomaly and the synchronization frame is removed, yielding the data to be processed. Then, a semantic slicing algorithm is used to segment the data to be processed, generating meta-skill training samples. For example, a complete process of "tightening a screw" can be segmented into segments such as "grabbing the screwdriver," "aligning the screw," "rotating and tightening," and "placing the screwdriver." For complex data to be processed, it can first be segmented by semantic nodes, and then further segmented within each node according to action patterns, forming reusable and combinable meta-skill segments, i.e., meta-skill training samples.

[0046] Optionally, after slicing the task execution data to obtain meta-skill training samples, the method further includes: using a video generation model to supplement the visual images of the robot from a non-eye perspective in the meta-skill training samples to obtain meta-skill training samples from the robot's eye perspective; evaluating the smoothness of robot movements, stability of robot movements, safety of robot torque, and task success rate in the meta-skill training samples from the robot's eye perspective to obtain a trajectory quality score, and determining the updated meta-skill training samples based on the trajectory quality score.

[0047] Specifically, after slicing the task execution data to obtain meta-skill training samples, an automatic label generator can be used to generate initial labels for the meta-skill training samples by combining language instructions and action features. Then, key meta-skill training samples are corrected using manual or semi-automatic review tools. The obtained meta-skill training samples are then uniformly converted into a standardized format (e.g., JSON + binary features) and metadata (e.g., acquisition device, operator, and environmental parameters) is recorded for easy traceability and retraining. The samples are then augmented using augmentation strategies such as perspective transformation, lighting perturbation, background replacement, tool geometry replacement, force control offset, and / or action time scaling. A video generation model (e.g., video diffusion model) is used, with human operation video, synchronized motion capture data, and the robot's URDF (Unified Robot Description Format) model as conditions, to generate simulated videos of the robot performing the same operations from the robot's first or third perspective, supplementing uncaptured angles and obtaining meta-skill training samples from the robot's eye perspective. This significantly expands the training dataset and solves the cross-subject (human to robot) domain adaptation problem.

[0048] Finally, the quality of the meta-skill training samples from the robot's eye perspective can be evaluated. Specific indicators include robot motion smoothness, robot motion stability, robot torque safety, and task success rate, to obtain a trajectory quality score Q. Samples with Q values ​​greater than a preset threshold are identified as updated meta-skill training samples.

[0049] Example 2

[0050] Figure 2 This is a flowchart of a robot meta-skill management method provided in Embodiment 2 of the present invention. The technical solution of the present invention is further optimized based on the above optional technical solutions, and provides a specific way to manage robot meta-skills.

[0051] Optionally, the step of training an initial model using the meta-skill training samples to obtain a target model includes: training an initial VLA model using the meta-skill training samples to obtain a model to be determined; and fine-tuning and optimizing the model to be determined using a digital twin simulation platform and a reinforcement learning platform to obtain a trained target model.

[0052] Optionally, generating the meta-skill data package based on the meta-skill training samples and the model parameters of the target model includes: packaging the model parameters of the target model, the input parameters of the target model, the threshold parameters of the meta-skills corresponding to the model parameters, the semantic description of the meta-skills corresponding to the model parameters, the meta-skill training sample parameters, and the version parameters of the meta-skills corresponding to the model parameters to obtain the meta-skill data package.

[0053] Optionally, before generating the user task based on the user-inputted meta-skill editing information and the meta-skill data package, the method further includes: receiving the user-inputted meta-skill editing information through a preset visual meta-skill editor, wherein the preset visual meta-skill editor can output graphical meta-skill nodes and meta-skill parameter configuration panels through a display interface, and the meta-skill editing information includes editing the meta-skill parameters corresponding to the meta-skill data package and / or combining and sorting the meta-skills corresponding to multiple meta-skill data packages.

[0054] like Figure 2 As shown in Embodiment 2 of the present invention, a robot meta-skill management method specifically includes the following steps:

[0055] S201. Obtain multimodal task execution data of the robot through VR teleoperation technology and / or UMI robot data acquisition kit; use an anomaly detector based on self-supervised contrastive learning to remove abnormal trajectories in the task execution data to obtain data to be processed, and perform semantic slicing on the data to be processed to obtain meta-skill training samples.

[0056] The task execution data includes at least visual data, motion data, force control data, and task semantic data.

[0057] S202. Use a video generation model to supplement the visual images of the robot from a non-eye perspective in the meta-skill training samples to obtain meta-skill training samples from the robot's eye perspective; evaluate the smoothness of robot movements, stability of robot movements, safety of robot torque, and success rate of the meta-skill training samples from the robot's eye perspective to obtain a trajectory quality score, and determine the updated meta-skill training samples based on the trajectory quality score.

[0058] S203. Train an initial VLA model using the meta-skill training samples to obtain a model to be determined; fine-tune and optimize the model to be determined using a digital twin simulation platform and a reinforcement learning platform to obtain a target model that has been trained.

[0059] Specifically, the VLA model is a Transformer-based multimodal policy model. Its structure mainly includes:

[0060] 1) Multimodal encoder, specifically including:

[0061] Visual encoder: Extracts spatial features of image I(t) using a pre-trained VisionTransformer.

[0062] Proprioceptive encoder: Captures the robot's physical state by encoding joint state q(t), force control information τ(t), and wrist end force F(t) through a small Transformer or MLP.

[0063] Language encoder: Obtains the semantic embedding of instruction L(t) through the encoder part of a pre-trained language model (such as BERT).

[0064] 2) Cross-modal fusion module: Through a multi-layer cross-attention mechanism, features from different modalities are deeply fused. For example, visual features are used as queries, and language and proprioceptive features are used as keys and values, enabling the model to "pay attention" to relevant regions in the image based on instructions and its own state.

[0065] 3) Policy Decoder: A diffusion model is used as the policy decoder. Compared to direct action regression, the diffusion model generates action sequences through an iterative denoising process.

[0066] 4) Loss Function: The training process uses a composite loss function for optimization, specifically including:

[0067] The denoising loss of the diffusion model is usually the mean square error between the predicted noise and the actual noise.

[0068] Cross-modal contrastive learning loss (similar to CLIP): used to bring the embedding representations of matching (image-language) pairs closer together and push away mismatched pairs, enhancing the model's generalization and understanding capabilities.

[0069] Force / torque prediction loss: Allows the model to explicitly learn the relationship between action and contact force, improving the accuracy of force control operations.

[0070] Specifically, a digital twin simulation platform is a crucial bridge connecting cloud algorithms and the physical world. Its core task is to provide a high-fidelity, scalable, and efficient virtual environment for model reinforcement learning, security verification, and policy generalization. The construction and application process of a digital twin simulation platform includes:

[0071] 1) High-fidelity robotics and environmental modeling

[0072] Robot dynamics modeling: Based on the URDF or SDF files provided by the robot manufacturer, a kinematic and dynamic model of the robot is constructed. This model not only includes the geometry and joint structure, but also precisely defines the physical properties of each link, including mass, center of mass position, and inertia tensor. Furthermore, the model includes joint friction models, such as a combination of Coulomb friction and viscous friction, and actuator dynamics models (such as motor torque limits and response delays). These parameters form the basis of the initial simulation model.

[0073] Environment and Object Modeling: Supports importing CAD models of industrial sites (such as STEP or IGES formats) to construct accurate factory layouts, workbenches, and equipment. For interactive objects (such as screws, tools, and packages to be sorted), in addition to the geometric model, their physical material properties must be defined, including static friction coefficient, dynamic friction coefficient, coefficient of restitution (elasticity), and surface stiffness, etc. These parameters are crucial for the simulation of contact mechanics.

[0074] 2) Physics engine and multi-fidelity simulation construction

[0075] Simulation Engine Selection: A variety of industry-leading physics simulation engines can be integrated. These engines were chosen because of their excellent support for GPU parallel computing, which can significantly accelerate the tens of thousands of parallel environment interactions required for reinforcement learning, and provide accurate rigid body dynamics and contact force models.

[0076] Construction of multi-fidelity simulation strategy: To balance training speed and simulation accuracy, a staged multi-fidelity simulation strategy can be adopted. For example:

[0077] Low-fidelity phase: In the early stages of policy learning, simplified physical models (e.g., considering only rigid body dynamics and ignoring complex friction) and geometric models (e.g., using convex hulls instead of fine meshes for collision detection) are used. This phase aims to rapidly explore the policy space, allowing the VLA model to quickly learn the basic workflow of the task.

[0078] High-fidelity phase: Once the strategy has initially converged, it automatically switches to high-fidelity mode. This phase loads complete physical parameters, a fine-grained collision mesh, simulated sensor noise, and introduces force control safety thresholds collected from the real world. This phase is used for fine-tuning the strategy, verifying safety boundaries, and bridging the simulation-to-real gap.

[0079] 3) Sim-to-Real Domain Adaptation and Closed-Loop Calibration

[0080] To address the discrepancy between simulation and reality, data collected from real robots can be used to automatically calibrate key physical parameters in the simulation model. For example, by allowing a real robot to move under specific excitation signals, joint angle and torque data can be collected. Then, optimization algorithms (such as gradient descent) can be used to minimize the error between real and simulation data, thereby identifying more accurate parameters such as inertia and friction. During training, the parameters of the simulation environment are randomized to enhance the robustness of the VLA model. The randomized parameter set includes:

[0081] Physical parameters: mass of the object, coefficient of friction, coefficient of restitution, and direction and intensity of illumination;

[0082] Visual parameters: texture, color, camera intrinsic and extrinsic parameters, and lens distortion;

[0083] Task parameters: initial position and orientation of the object and target position.

[0084] By training in a wide range of simulated environments, the model can learn to ignore irrelevant environmental features and focus on the task itself.

[0085] Furthermore, noise models that match the characteristics of real sensors can be injected into the simulation. For example, Gaussian noise and holes can be added to the depth values ​​of a simulated depth camera, white noise and bias drift can be added to the torque sensor signal, and response delay and control error can be simulated for the motor actuator. This allows policies trained in simulation to better adapt to real-world sensor uncertainties. Safety force control thresholds and contact force patterns extracted from real data acquisition can also be directly imported into the simulation environment. In the reinforcement learning reward function, behaviors exceeding these thresholds can be subject to large negative rewards, or directly used as constraints, thus "imprinting" safety boundaries into the policies generated by the VLA model during the training phase.

[0086] 4) Automated task and scene generation

[0087] A scripting interface can be provided to programmatically generate a large number of training and testing scenarios. For example, for a meta-skill like "tightening screws," screw holes in different positions and angles, as well as screws of different sizes, can be automatically generated. The simulation environment can automatically adjust the task difficulty according to the learning progress of the VLA model. Starting with simple tasks (such as grasping in a fixed position), it gradually transitions to complex tasks (such as coordinating two-handed assembly under disturbance), thereby accelerating model convergence.

[0088] 5) Simulation as a Service Interface Configuration

[0089] The digital twin simulation platform provides API interfaces conforming to the OpenAI Gym / Gymnasium standards to the upper layer (i.e., the meta-skill factory in the cloud). Reinforcement learning algorithms can initialize the environment using `env.reset()` and submit actions using `env.step(action)` to obtain the next moment's observations, rewards, and completion flags. The observation space of the digital twin simulation platform is multimodal, including: RGB images and depth maps generated by the simulation camera, robot body states (such as joint angles and joint velocities), end effector states (such as Cartesian pose, forces, and torques), and task-related information (such as target position and object state). The actions accepted by the digital twin simulation platform can be joint space commands or Cartesian space commands, consistent with the control interface of a real robot.

[0090] Specifically, VLA models obtained through supervised learning are often suboptimal. Reinforcement (RL) learning platforms can leverage digital twin simulation environments to fine-tune and optimize these models online. To deeply integrate security into the policy, a constrained RL framework can be employed. Its optimization objective is to maximize task rewards while ensuring that the expected value of security costs does not exceed a pre-defined minimum threshold. Malicious actors can use algorithms such as PPO-Lagrangian or CPO to solve this constrained optimization problem, thereby generating a policy that is both efficient and inherently secure. RL platforms can interact with digital twin simulation platforms at high speed through standardized APIs. Utilizing the parallelization advantages of GPUs, experience can be collected simultaneously in thousands of parallel simulation environments, significantly accelerating the training process.

[0091] S204. The model parameters of the target model, the input parameters of the target model, the threshold parameters of the meta-skills corresponding to the model parameters, the semantic description of the meta-skills corresponding to the model parameters, the training sample parameters of the meta-skills, and the version parameters of the meta-skills corresponding to the model parameters are packaged to obtain a meta-skill data package.

[0092] Specifically, each verified meta-skill is packaged into a standardized data package, which contains the following:

[0093] 1) Model parameters of the target model: parameter file of VLA policy network.

[0094] 2) Input parameters of the target model: Hyperparameters required for model inference, such as motion smoothing coefficient and sampling temperature.

[0095] 3) Threshold parameters of meta-skills corresponding to model parameters: Defines the safety boundary for the operation of meta-skills, including force and torque thresholds for each joint, speed and acceleration limits, and virtual walls in the task space.

[0096] 4) Semantic description of meta-skills corresponding to model parameters: usually a YAML or JSON file, describing the meta-skill function, input (such as target object type), output (such as final state), and preconditions and postconditions for execution in natural language.

[0097] 5) Metadata: version number of meta-skill, generation date, meta-skill training sample parameters (such as training dataset hash), performance metrics (success rate and time consumption), etc.

[0098] The meta-skill library, which stores meta-skill data packages, provides a RESTful API, allowing clients on the local terminal to search, query, and download meta-skill capsules based on keywords such as name, function, and tags. It supports skill version control, allows rollback to stable versions, and provides a hot deployment interface for skill updates. Once a new version of a skill passes all verifications, it can be automatically pushed to the client's candidate update list.

[0099] S205. Receive user-inputted meta-skill editing information through a preset visual meta-skill editor.

[0100] The preset visual meta-skill editor can output graphical meta-skill nodes and meta-skill parameter configuration panels through the display interface. The meta-skill editing information includes editing the meta-skill parameters corresponding to the meta-skill data package and / or combining and sorting the meta-skills corresponding to multiple meta-skill data packages.

[0101] S206. Generate a user task based on the meta-skill editing information input by the user and the meta-skill data package, and run the user task.

[0102] The user task includes a subtask corresponding to the edited meta-skill data package.

[0103] Specifically, the goal of this method is to transform the abstract meta-skills generated by the cloud-based meta-skill factory into concrete task flows that robots can execute in real-world industrial scenarios in an intuitive, flexible, and secure manner. This lowers the barrier to entry for humanoid robots through low-code or no-code operations. The visual meta-skill editor is a web-based or native application-based visual programming environment that allows users to build complex task sequences through drag-and-drop and connection methods.

[0104] First, the node-based programming interface in the visual meta-skill editor includes:

[0105] 1) Skill Nodes: Each meta-skill data package synchronized from the cloud meta-skill library is represented as an independent graphical node in the editor, such as "Grab Object", "Tighten Screws", and "Visual Inspection". The color or icon of the node can indicate its function type (such as operation, movement, and perception).

[0106] 2) Logic nodes: Provide standard programming logic control, such as nodes for "start", "end", "condition judgment", "loop" and "parallel execution".

[0107] 3) Data Flow Connections: Users define the execution order and data flow of tasks by dragging and dropping lines between nodes. For example, the output (screw coordinates) of a "Visual Positioning Screw" node can be connected to the input of a "Move to Target" node.

[0108] 4) Parameter configuration panel:

[0109] Clicking on any skill node will bring up a parameter configuration panel. Users can set the specific execution parameters for that skill in this panel.

[0110] 5) Static parameters: such as moving speed, gripping force, and number of tightening turns.

[0111] 6) Dynamic parameter binding: Parameters can be bound to the output of the previous node or to a global variable. For example, the target object ID of the "Grab Object" node can be dynamically provided by the previous "Identify Object to Grab" node.

[0112] 7) Safety parameter fine-tuning: Allows authorized users to make minor adjustments to the force control threshold on site (e.g., reduce it by 5%) within the safety envelope set in the cloud to adapt to specific workpieces or environments, but cannot exceed the absolute upper limit set in the cloud.

[0113] Secondly, to further simplify the creation process of common tasks, a pre-built template library can be provided on the client side, such as:

[0114] Official and Community Templates: The library contains complete workflow templates for typical industrial applications such as circuit board assembly, logistics sorting, and quality inspection.

[0115] It also offers one-click application and customization services. After a user selects a template, the complete skill node diagram can be automatically loaded. Users only need to modify a few key parameters (such as workbench coordinates or camera calibration files) according to the site conditions to quickly deploy a complete application. Users can also modify the template to create processes that meet their specific needs and save them as new custom templates.

[0116] Furthermore, an automatic orchestration assistant driven by a large language model can be provided. This assistant can automatically convert natural language requirements into executable task flows. Leveraging the powerful logical reasoning capabilities of the large language model, it arranges meta-skill nodes matching the user's input in a logical order and automatically fills in the parameters extracted from the user input to generate the user task.

[0117] Next, after the user task is generated, the client can provide real-time monitoring and debugging tools to ensure the task runs smoothly, such as providing a real-time status visualization interface, which includes:

[0118] 1) 3D Digital Twin View: A 3D model synchronized with the real robot, displaying the robot's posture, joint status, and interaction with the environment in real time;

[0119] 2) Data Dashboard: Displays key data streams in real time, such as joint torques, end-effector forces, camera views, etc. in chart form; currently executing skill nodes are highlighted in the graphical editor;

[0120] 3) Online parameter adjustment panel: When a task is paused, users can modify the parameters of the node to be executed online without redeploying the entire process.

[0121] The client automatically records detailed operation logs, including the execution time of each skill, success / failure status, triggered exceptions, and peak values ​​of force control data. When a task fails, it automatically captures multimodal data (video and sensor readings, etc.) from the period preceding the failure and packages it into a "failure case package." Users can anonymize these logs and failure case packages and upload them to the cloud-based meta-skill factory with a single click. This valuable real-world data will become a key input for the next round of model iteration and optimization, forming a complete data-driven closed-loop optimization.

[0122] Optionally, the architecture of this method involves multiple distributed components, including local terminals and the cloud. The data flows between these components exhibit diverse characteristics, including low-latency real-time control, high-bandwidth data acquisition, and non-real-time model updates. A robust, efficient, and layered communication and synchronization mechanism is the cornerstone of ensuring the stable operation of the entire method. Data interaction between the cloud and local terminals can be achieved through multi-layered data transmission strategies. The advantage of this setup is that it effectively addresses the challenges posed by complex network environments and massive data volumes in industrial scenarios, ensuring efficient and reliable collaboration among all components of the system.

[0123] For example, for real-time control flow content, such as VR teleoperation commands, robot body state feedback (joint angles, end effector posture), and low-frequency force feedback signals, a state synchronization rather than event-driven transmission mechanism is used for data transmission. For instance, instead of sending a "move forward 1cm" command, the target pose is sent directly; even if packets are lost in the middle, subsequent data packets can correct the robot's state. For high-frequency sensor data stream content, such as high-resolution images, depth maps, full-body torque data, and haptic sensor array data, data transmission is achieved through local caching and asynchronous upload. For data streams related to model and meta-skill deployment, such as meta-skill data packages, data transmission is achieved through version control and incremental updates. For log and telemetry data streams, such as robot operation logs, performance metrics (e.g., CPU and GPU utilization), task success / failure statistics, and triggered safety events, a lightweight log collection agent (such as Fluentd) is used to batch package and compress the log data before periodically sending it to the cloud.

[0124] The robot meta-skill management method provided in this invention constructs a dual-mode data acquisition system. By integrating VR teleoperation and UMI devices, it achieves multimodal teaching both remotely and locally. Simultaneously, it utilizes a video generation model to translate human-view data into robot-view samples, significantly reducing the acquisition threshold. Semantic data governance and meta-skill slicing are established in the cloud, unifying the modeling of motion trajectories, visual information, language commands, and full-body force control thresholds. Using a VLA model as the core generation engine, and combined with a digital twin simulation platform for constraint reinforcement learning, meta-skills incorporating safety thresholds are formed. Meta-skill encapsulation is achieved by packaging model weights and semantic descriptions, enabling visual combination and online fine-tuning. This method provides convenient data acquisition and augmentation for remote clients or factories where data cannot be exported. It offers a standardized encapsulation and combination method for meta-skills, greatly simplifying the application development process of humanoid robots and establishing a continuous learning channel from deployment to feedback, enabling meta-skills to continuously evolve and improve in real industrial environments.

[0125] Example 3

[0126] Figure 3 This is a structural schematic diagram of a robot meta-skill management device provided in Embodiment 3 of the present invention. Figure 3 As shown, the device includes: a training sample determination module 301, a meta-skill data package generation module 302, and a task execution module 303, wherein:

[0127] The training sample determination module is used to acquire the robot's task execution data and slice the task execution data to obtain meta-skill training samples.

[0128] The meta-skill data package generation module is used to train an initial model using the meta-skill training samples to obtain a target model, and to generate a meta-skill data package based on the model parameters of the meta-skill training samples and the target model.

[0129] The task execution module is used to generate user tasks based on the meta-skill editing information input by the user and the meta-skill data package, and to run the user tasks, wherein the user tasks include sub-tasks corresponding to the edited meta-skill data package.

[0130] The robot meta-skill management device provided in this invention realizes a complete system for generating robot training samples, training models, generating robot meta-skills, editing robot meta-skills, and customizing user tasks. By providing a complete closed-loop system that integrates multi-source data acquisition, model training, meta-skill encapsulation, and task customization, this device improves the generation efficiency, reliability, and reusability of industrial robot meta-skills, enables the rapid construction of industrial robot meta-skills, reduces reliance on professional personnel, and significantly improves the deployment efficiency and reuse value of industrial robots in complex tasks.

[0131] Optionally, the training sample determination module includes:

[0132] The data acquisition unit is used to acquire multimodal task execution data of the robot through VR teleoperation technology and / or UMI robot data acquisition kit, wherein the task execution data includes at least visual data, motion data, force control data and task semantic data;

[0133] The sample determination unit is used to remove abnormal trajectories in the task execution data using an anomaly detector based on self-supervised contrastive learning, to obtain data to be processed, and to perform semantic slicing on the data to be processed to obtain meta-skill training samples.

[0134] Optionally, the training sample determination module also includes:

[0135] The visual supplementation unit is used to supplement the visual images of the robot from the non-eye perspective in the meta-skill training samples by using a video generation model after slicing the task execution data to obtain meta-skill training samples from the robot's eye perspective.

[0136] The sample update unit is used to evaluate the smoothness of robot movements, stability of robot movements, safety of robot torque, and success rate of task in the meta-skill training samples under the robot's eye view, so as to obtain a trajectory quality score, and determine the updated meta-skill training samples based on the trajectory quality score.

[0137] Optionally, the meta-skill data package generation module includes:

[0138] The model training unit is used to train an initial VLA model using the meta-skill training samples to obtain the model to be determined.

[0139] The model optimization unit is used to fine-tune and optimize the model to be determined using a digital twin simulation platform and a reinforcement learning platform to obtain the target model after training.

[0140] Optionally, the meta-skill data package generation module includes:

[0141] The meta-skill data package generation unit is used to package the model parameters of the target model, the input parameters of the target model, the threshold parameters of the meta-skills corresponding to the model parameters, the semantic description of the meta-skills corresponding to the model parameters, the training sample parameters of the meta-skills, and the version parameters of the meta-skills corresponding to the model parameters to obtain a meta-skill data package.

[0142] Optionally, the device may also include:

[0143] The editing information receiving module is used to receive the user-inputted meta-skill editing information through a preset visual meta-skill editor before generating the user task based on the user-inputted meta-skill editing information and the meta-skill data package. The preset visual meta-skill editor can output graphical meta-skill nodes and meta-skill parameter configuration panels through a display interface. The meta-skill editing information includes editing the meta-skill parameters corresponding to the meta-skill data package and / or combining and sorting the meta-skills corresponding to multiple meta-skill data packages.

[0144] Optionally, the method is applied in the cloud.

[0145] The robot meta-skill management device provided in this embodiment of the invention can execute the robot meta-skill management method provided in any embodiment of the invention, and has the corresponding functional modules and beneficial effects of the execution method.

[0146] Example 4

[0147] Figure 4 A schematic diagram of an electronic device 40 that can be used to implement embodiments of the present invention is shown. The electronic device is intended to represent various forms of digital computers, such as laptop computers, desktop computers, workstations, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. The electronic device can also represent various forms of mobile devices, such as personal digital processors, cellular phones, smartphones, wearable devices (e.g., helmets, glasses, watches, etc.), and other similar computing devices. The components shown herein, their connections and relationships, and their functions are merely illustrative and are not intended to limit the implementation of the invention described and / or claimed herein.

[0148] like Figure 4As shown, the electronic device 40 includes at least one processor 41 and a memory, such as a read-only memory (ROM) 42 or a random access memory (RAM) 43, communicatively connected to the at least one processor 41. The memory stores computer programs executable by the at least one processor. The processor 41 can perform various appropriate actions and processes based on the computer program stored in the ROM 42 or loaded from storage unit 48 into the RAM 43. The RAM 43 may also store various programs and data required for the operation of the electronic device 40. The processor 41, ROM 42, and RAM 43 are interconnected via a bus 44. An input / output (I / O) interface 45 is also connected to the bus 44.

[0149] Multiple components in electronic device 40 are connected to I / O interface 45, including: input unit 46, such as keyboard, mouse, etc.; output unit 47, such as various types of monitors, speakers, etc.; storage unit 48, such as disk, optical disk, etc.; and communication unit 49, such as network card, modem, wireless transceiver, etc. Communication unit 49 allows electronic device 40 to exchange information / data with other devices through computer networks such as the Internet and / or various telecommunications networks.

[0150] Processor 41 can be a variety of general-purpose and / or special-purpose processing components with processing and computing capabilities. Some examples of processor 41 include, but are not limited to, central processing unit (CPU), graphics processing unit (GPU), various special-purpose artificial intelligence (AI) computing chips, various processors running machine learning model algorithms, digital signal processors (DSPs), and any suitable processor, controller, microcontroller, etc. Processor 41 performs the various methods and processes described above, such as robot meta-skill management methods.

[0151] In some embodiments, the robot meta-skill management method may be implemented as a computer program tangibly contained in a computer-readable storage medium, such as storage unit 48. In some embodiments, part or all of the computer program may be loaded and / or installed on electronic device 40 via ROM 42 and / or communication unit 49. When the computer program is loaded into RAM 43 and executed by processor 41, one or more steps of the robot meta-skill management method described above may be performed. Alternatively, in other embodiments, processor 41 may be configured to execute the robot meta-skill management method by any other suitable means (e.g., by means of firmware).

[0152] Various embodiments of the systems and techniques described above herein can be implemented in digital electronic circuit systems, integrated circuit systems, field-programmable gate arrays (FPGAs), application-specific integrated circuits (ASICs), application-specific standard products (ASSPs), system-on-a-chip (SoC) systems, complex programmable logic devices (CPLDs), computer hardware, firmware, software, and / or combinations thereof. These various embodiments may include implementations in one or more computer programs that can be executed and / or interpreted on a programmable system including at least one programmable processor, which may be a dedicated or general-purpose programmable processor, capable of receiving data and instructions from a storage system, at least one input device, and at least one output device, and transmitting data and instructions to the storage system, the at least one input device, and the at least one output device.

[0153] Computer programs used to implement the methods of the present invention may be written in any combination of one or more programming languages. These computer programs may be provided to a processor of a general-purpose computer, a special-purpose computer, or other programmable data processing device, such that when executed by the processor, the computer programs cause the functions / operations specified in the flowcharts and / or block diagrams to be performed. The computer programs may be executed entirely on a machine, partially on a machine, or as a standalone software package, partially on a machine and partially on a remote machine, or entirely on a remote machine or server.

[0154] The computer equipment provided above can be used to execute the robot meta-skill management method provided in any of the above embodiments, and has corresponding functions and beneficial effects.

[0155] Example 5

[0156] In the context of this invention, the computer-readable storage medium may be a tangible medium, and the computer-executable instructions, when executed by a computer processor, are used to perform a robot meta-skill management method, the method comprising:

[0157] Obtain the robot's task execution data and slice the task execution data to obtain meta-skill training samples;

[0158] An initial model is trained using the meta-skill training samples to obtain a target model, and a meta-skill data package is generated based on the model parameters of the meta-skill training samples and the target model.

[0159] A user task is generated based on the user-input meta-skill editing information and the meta-skill data package, and the user task is run. The user task includes a sub-task corresponding to the edited meta-skill data package.

[0160] In the context of this invention, a computer-readable storage medium can be a tangible medium that may contain or store a computer program for use by, or in conjunction with, an instruction execution system, apparatus, or device. A computer-readable storage medium may include, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatus, or devices, or any suitable combination thereof. Alternatively, a computer-readable storage medium may be a machine-readable signal medium. More specific examples of machine-readable storage media include electrical connections based on one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fibers, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof.

[0161] The computer equipment provided above can be used to execute the robot meta-skill management method provided in any of the above embodiments, and has corresponding functions and beneficial effects.

[0162] It is worth noting that in the embodiments of the robot meta-skill management device described above, the various units and modules included are only divided according to functional logic, but are not limited to the above division, as long as the corresponding functions can be achieved; in addition, the specific names of each functional unit are only for easy differentiation and are not used to limit the scope of protection of the present invention.

[0163] Note that the above description is merely a preferred embodiment of the present invention and the technical principles employed. Those skilled in the art will understand that the present invention is not limited to the specific embodiments described herein, and various obvious changes, readjustments, and substitutions can be made without departing from the scope of protection of the present invention. Therefore, although the present invention has been described in detail through the above embodiments, the present invention is not limited to the above embodiments, and may include many other equivalent embodiments without departing from the concept of the present invention, the scope of which is determined by the scope of the appended claims.

Claims

1. A method for managing robot meta-skills, characterized in that, include: Obtain the robot's task execution data and slice the task execution data to obtain meta-skill training samples; An initial model is trained using the meta-skill training samples to obtain a target model, and a meta-skill data package is generated based on the model parameters of the meta-skill training samples and the target model. A user task is generated based on the user-input meta-skill editing information and the meta-skill data package, and the user task is run. The user task includes a sub-task corresponding to the edited meta-skill data package.

2. The method according to claim 1, characterized in that, The process of acquiring the robot's task execution data and slicing the task execution data to obtain meta-skill training samples includes: Multimodal task execution data of the robot is acquired through VR teleoperation technology and / or UMI robot data acquisition kit, wherein the task execution data includes at least visual data, motion data, force control data and task semantic data; An anomaly detector based on self-supervised contrastive learning is used to remove abnormal trajectories from the task execution data to obtain data to be processed. Then, semantic slicing is performed on the data to be processed to obtain meta-skill training samples.

3. The method according to claim 2, characterized in that, After slicing the task execution data to obtain meta-skill training samples, the method further includes: The video generation model is used to supplement the visual images of the robot from the non-eye perspective in the meta-skill training samples to obtain the meta-skill training samples from the robot's eye perspective. The smoothness of robot movements, stability of robot movements, safety of robot torque, and success rate of task in the meta-skill training samples under the robot's eye view are evaluated to obtain a trajectory quality score, and the updated meta-skill training samples are determined based on the trajectory quality score.

4. The method according to any one of claims 1-3, characterized in that, The step of training an initial model using the meta-skill training samples to obtain a target model includes: An initial VLA model is trained using the aforementioned meta-skill training samples to obtain the model to be determined. The model to be determined is fine-tuned and optimized using a digital twin simulation platform and a reinforcement learning platform to obtain the target model after training.

5. The method according to claim 1, characterized in that, The step of generating a meta-skill data package based on the meta-skill training samples and the model parameters of the target model includes: The model parameters of the target model, the input parameters of the target model, the threshold parameters of the meta-skills corresponding to the model parameters, the semantic description of the meta-skills corresponding to the model parameters, the training sample parameters of the meta-skills, and the version parameters of the meta-skills corresponding to the model parameters are packaged to obtain a meta-skill data package.

6. The method according to claim 1, characterized in that, Before generating the user task based on the user-input meta-skill editing information and the meta-skill data package, the method further includes: The system receives meta-skill editing information input by the user through a preset visual meta-skill editor. The preset visual meta-skill editor can output graphical meta-skill nodes and meta-skill parameter configuration panels through a display interface. The meta-skill editing information includes editing the meta-skill parameters corresponding to the meta-skill data package and / or combining and sorting the meta-skills corresponding to multiple meta-skill data packages.

7. The method according to claim 1, characterized in that, The method is applied in the cloud.

8. A robot meta-skill management device, characterized in that, include: The training sample determination module is used to acquire the robot's task execution data and slice the task execution data to obtain meta-skill training samples. The meta-skill data package generation module is used to train an initial model using the meta-skill training samples to obtain a target model, and to generate a meta-skill data package based on the model parameters of the meta-skill training samples and the target model. The task execution module is used to generate user tasks based on the meta-skill editing information input by the user and the meta-skill data package, and to run the user tasks, wherein the user tasks include sub-tasks corresponding to the edited meta-skill data package.

9. An electronic device, characterized in that, The electronic device includes: At least one processor; and A memory communicatively connected to the at least one processor; wherein, The memory stores a computer program that can be executed by the at least one processor, the computer program being executed by the at least one processor to enable the at least one processor to perform the robot meta-skill management method according to any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer instructions that cause a processor to execute the robot meta-skill management method according to any one of claims 1-7.

Images