Robot action correction method and apparatus, electronic device, medium, and program product

Abstract

The present disclosure provides a robot action correction method and device, electronic equipment, medium and program product, by freezing the world action model backbone, only updating the lightweight adapter online, avoiding the high cost of full parameter fine tuning and the risk of overfitting; the reference action block is used as a coarse-grained prior, so that the adapter does not need to explore the action space from zero, and the fine-grained information in the latent variable feature is extracted to support precise correction, overcoming the problem of information loss in the prior art; through the execution of feedback driven closed loop iteration, the adapter continuously optimizes and converges to the optimal strategy of the current environment, effectively suppressing the influence of deployment error on precise operation; finally, the target adapter is output when the preset condition is met, balancing performance and efficiency.

Description

Technical Field

[0001] This disclosure relates to robot control technology, and in particular to a robot motion correction method, apparatus, electronic device, medium, and program product. Background Technology

[0002] World-Action Models (WAMs) are a class of fundamental models used for robot manipulation. They typically generate long-term action sequences or action blocks directly from multi-view images, the initial robot state, and the initial task instructions. Thanks to training on large-scale offline datasets, WAMs can acquire strong visual understanding, scene prediction, and long-range planning capabilities, exhibiting good generalization in common tasks such as grasping and moving.

[0003] However, when world motion models are deployed for contact tasks such as insertion, precision assembly, and high-precision alignment, their performance often degrades significantly. The fundamental reason is that inherent camera calibration errors, perception biases, actuator accuracy limitations, and subtle differences in contact dynamics in real-world deployment environments are amplified dramatically during the robot's end-effector execution phase. Although existing world motion models can guide the end-effector to the vicinity of the target, the lack of rapid adaptation to the current physical environment during the "last few millimeters" of insertion, alignment, or contact makes the task highly prone to failure.

[0004] To address the aforementioned deployment errors, existing countermeasures can be categorized into four main technical approaches: 1) online fine-tuning of all parameters of a large-scale base model; 2) freezing the base model and learning independent action residual modules for local correction; 3) optimizing small-scale policies from scratch or from the base policy using online reinforcement learning or human demonstration data; and 4) compressing the internal representation of the base model into compact tokens and then training a small adaptive network.

[0005] The four technical approaches mentioned above all demonstrate some effectiveness in conventional robot strategy adaptation scenarios. However, they primarily target general policy models or vision-language-action models and are not designed specifically for the structural characteristics of the world action model itself. The action generation mechanism of the world action model is unique; its task-related information is not only reflected in the final output action but is also widely distributed in intermediate representations such as visual encoding, world state prediction, and action generation latent variables. If only residual correction is applied to the final action, or if internal features are excessively compressed into a single compact representation, fine-grained information essential for precise operation, such as target position, object geometry, contact area, and action generation context, is easily lost, thus limiting the accuracy of corrections in the final stage. Summary of the Invention

[0006] To address the technical problems in related technologies, embodiments of this disclosure provide a robot motion correction method, apparatus, electronic device, medium, and program product.

[0007] According to a first aspect of the present disclosure, a robot motion correction method is provided, the method comprising: Acquire the robot's first multi-view visual observation, first robot body state, and first task instruction; The first multi-view visual observation, the first robot body state, and the first task instruction are input into the pre-trained world action model to obtain reference action blocks and latent variable features, wherein the network parameters of the world action model are in a frozen state. A first reinforcement learning state representation is constructed based on the first robot body state, the reference action block, and the latent variable features. The first reinforcement learning state representation is then input into a lightweight adapter, which generates a first absolute action block based on the attention mechanism and the first reinforcement learning state representation. The lightweight adapter is an action refinement network with fewer trainable parameters than the world action model. The first absolute action block is the action instruction to be verified output by the lightweight adapter during the calibration process. In response to receiving execution feedback after the robot executes the first absolute action block, the parameters of the lightweight adapter are updated online based on the execution feedback, which includes a second multi-view visual observation, a second robot body state, a second task instruction, a task completion status signal, and an intervention signal. The above steps are executed iteratively until a preset termination condition is met, and a target lightweight adapter is output. The target lightweight adapter is the iteratively updated lightweight adapter, which is used to output a second absolute action block to control the robot to complete the operation task. The second absolute action block is the correction action command output by the lightweight adapter after correction and convergence.

[0008] As an optional embodiment, the step of inputting the first multi-view visual observation, the first robot body state, and the first task instruction into a pre-trained world motion model to obtain reference action blocks and latent variable features includes: In response to the world action model with the first multi-view visual observation, the first robot body state, and the first task instruction input network parameters in a frozen state, the first multi-view visual observation is feature-encoded through the visual encoding branch of the world action model to obtain visual encoding features. The visual encoding features are fused with the first robot body state and the first task instruction through multimodal fusion and temporal modeling to obtain the fused world state features; The world state features are input into the action generation branch of the world action model. During the forward propagation of the action generation branch, action generation latent variable features are extracted from the preset intermediate layer. The action generation latent variable features are decoded by the action decoding branch of the world action model, and a reference action block of preset length is output.

[0009] As an optional embodiment, the step of constructing a first reinforcement learning state representation based on the first robot body state, the reference action block, and the latent variable features, and inputting the first reinforcement learning state representation into a lightweight adapter, so that the lightweight adapter generates a first absolute action block based on an attention mechanism and the first reinforcement learning state representation, includes: The first robot body state, the reference action block, and the latent variable features are encoded to obtain a first query-side token sequence and a first memory token sequence. The first query-side token sequence and the first memory token sequence are input into the lightweight adapter, so that the lightweight adapter executes the attention mechanism based on the first query-side token sequence and the first memory token sequence to generate the first absolute action block.

[0010] As an optional embodiment, the encoding process of the first robot body state, the reference action block, and the latent variable features to obtain the first query-side token sequence and the first memory token sequence includes: The state of the first robot body is encoded as a state token, and the first preset number of reference actions in the reference action block are encoded as reference action tokens, wherein the preset number is less than or equal to the length of the reference action block; The status token is concatenated with each of the reference action tokens to form a first query-side token sequence; The latent variable features are encoded into at least one memory token to obtain a first memory token sequence.

[0011] As an optional embodiment, the step of inputting the first query-side token sequence and the first memory token sequence into the lightweight adapter, so that the lightweight adapter executes the attention mechanism based on the first query-side token sequence and the first memory token sequence to generate the first absolute action block, includes: Input the first query-side token sequence and the first memory token sequence into the lightweight adapter; The first query-side token sequence is processed by the lightweight adapter to model the structural relationship between the first robot body state and the reference action block, and the first query-side token sequence after self-attention processing is obtained. Through the lightweight adapter, cross-attention processing is performed on the first query-side token sequence and the first memory token sequence after self-attention processing to obtain a query-side representation that integrates the latent variable features; Based on the query-side representation that integrates the latent variable features, the action head outputs the first absolute action block with a length of the preset number.

[0012] As an optional embodiment, the step of responding to the execution feedback received after the robot executes the first absolute action block, and updating the parameters of the lightweight adapter online based on the execution feedback, includes: In response to the execution feedback after the robot executes the first absolute action block, a second reinforcement learning state representation, a termination state, and a reward signal after the robot executes the first absolute action block are determined based on the execution feedback, wherein the termination state is used to indicate whether the robot has completed the task or whether the current round needs to be terminated; The first reinforcement learning state representation, the first absolute action block, the reward signal, the second reinforcement learning state representation, and the termination state are stored as a transition sample in the experience replay pool. In response to sampling transfer samples from the experience replay pool, a loss function of the lightweight adapter is calculated based on the sampled transfer samples, and the parameters of the lightweight adapter are updated according to the loss function.

[0013] As an optional embodiment, the step of determining a second reinforcement learning state representation, a termination state, and a reward signal after the robot executes the first absolute action block based on the execution feedback, in response to the execution feedback, includes: The second multi-view visual observation, the second robot body state, the second task instruction, the task completion status signal, and the intervention signal are obtained from the execution feedback after the robot executes the first absolute action block; Confirm the type of indication for the task completion status signal in the execution feedback; In response to the indication type indicating that the robot has successfully completed the task, a sparse reward signal of a first value is generated; and in response to the indication type indicating that the robot has not completed the task, a sparse reward signal of a second value is generated. The state type of the termination state is determined based on the task completion status signal and / or intervention signal in the execution feedback; The second multi-view visual observation, the second robot body state, and the second task instruction are encoded to obtain a second query-side token sequence and a second memory token sequence, which are then used to generate the second reinforcement learning state representation.

[0014] As an optional embodiment, the step of sampling transfer samples from the experience replay pool, calculating the loss function of the lightweight adapter based on the sampled transfer samples, and updating the parameters of the lightweight adapter according to the loss function includes: In response to sampling transfer samples from the experience replay pool, a temporal difference objective is determined based on the second reinforcement learning state, the reward signal, the second reinforcement learning state representation, and the termination state; The value loss of the lightweight adapter is determined based on the difference between the temporal difference objective and the value estimate of the first reinforcement learning state representation and the first absolute action block by the lightweight adapter. Based on the value estimate of the first reinforcement learning state representation by the lightweight adapter, the value-oriented term in the policy loss is determined, and the first absolute action block is added as a constraint term to the policy loss to obtain the policy loss of the lightweight adapter. The parameters of the lightweight adapter are updated based on the value loss and the policy loss.

[0015] According to a second aspect of the present disclosure, a robot motion correction device is provided, the device comprising: The acquisition module is used to acquire the robot's first multi-view visual observation, first robot body state, and first task instruction; The reference data extraction module is used to input the first multi-view visual observation, the first robot body state and the first task instruction into the pre-trained world action model to obtain reference action blocks and latent variable features, wherein the network parameters of the world action model are in a frozen state. An absolute action block generation module is used to construct a first reinforcement learning state representation based on the first robot body state, the reference action block, and the latent variable features, and input the first reinforcement learning state representation into a lightweight adapter so that the lightweight adapter generates a first absolute action block based on the attention mechanism and the first reinforcement learning state representation. The lightweight adapter is an action refinement network with fewer trainable parameters than the world action model, and the first absolute action block is the action instruction to be verified output by the lightweight adapter during the correction process. The parameter update module is used to update the parameters of the lightweight adapter online based on the execution feedback received after the robot executes the first absolute action block. The execution feedback includes a second multi-view visual observation, a second robot body state, a second task instruction, a task completion status signal, and an intervention signal. The iteration module is used to iteratively execute the above steps until a preset termination condition is met, and output a target lightweight adapter. The target lightweight adapter is the iteratively updated lightweight adapter, which is used to output a second absolute action block to control the robot to complete the operation task. The second absolute action block is the correction action instruction output by the lightweight adapter after correction and convergence.

[0016] As an optional embodiment, the reference data extraction module includes: The first encoding unit is configured to respond to the world action model, in which the first multi-view visual observation, the first robot body state, and the first task instruction are input into the network parameters in a frozen state, by performing feature encoding on the first multi-view visual observation through the visual encoding branch of the world action model to obtain visual encoding features. The modeling unit is used to perform multimodal fusion and temporal modeling of the visual encoded features with the first robot body state and the first task instruction to obtain the fused world state features; The feature extraction unit is used to input the world state features into the action generation branch of the world action model, and extract action generation latent variable features from a preset intermediate layer during the forward propagation of the action generation branch. The decoding unit is used to decode the latent variable features of the action generation through the action decoding branch of the world action model and output a reference action block of preset length.

[0017] As an optional embodiment, the absolute action block generation module includes: The second encoding unit is used to encode the first robot body state, the reference action block and the latent variable features to obtain the first query side token sequence and the first memory token sequence. An absolute action block generation unit is used to input the first query-side token sequence and the first memory token sequence into the lightweight adapter, so that the lightweight adapter executes the attention mechanism based on the first query-side token sequence and the first memory token sequence to generate the first absolute action block.

[0018] As an optional embodiment, the second encoding unit is specifically used for: The state of the first robot body is encoded as a state token, and the first preset number of reference actions in the reference action block are encoded as reference action tokens, wherein the preset number is less than or equal to the length of the reference action block; The status token is concatenated with each of the reference action tokens to form a first query-side token sequence; The latent variable features are encoded into at least one memory token to obtain a first memory token sequence.

[0019] As an optional embodiment, the absolute action block generation unit is specifically used for: Input the first query-side token sequence and the first memory token sequence into the lightweight adapter; The first query-side token sequence is processed by the lightweight adapter to model the structural relationship between the first robot body state and the reference action block, and the first query-side token sequence after self-attention processing is obtained. Through the lightweight adapter, cross-attention processing is performed on the first query-side token sequence and the first memory token sequence after self-attention processing to obtain a query-side representation that integrates the latent variable features; Based on the query-side representation that integrates the latent variable features, the action head outputs the first absolute action block with a length of the preset number.

[0020] As an optional embodiment, the parameter update module includes: A determining unit is configured to, in response to the execution feedback after the robot executes the first absolute action block, determine, based on the execution feedback, a second reinforcement learning state representation, a termination state, and a reward signal after the robot executes the first absolute action block, wherein the termination state is used to indicate whether the robot has completed the task or whether it needs to terminate the current round; The sample storage unit is used to store the first reinforcement learning state representation, the first absolute action block, the reward signal, the second reinforcement learning state representation, and the termination state as a transition sample into the experience replay pool. The parameter update unit is configured to, in response to sampling transfer samples from the experience replay pool, calculate the loss function of the lightweight adapter based on the sampled transfer samples, and update the parameters of the lightweight adapter according to the loss function.

[0021] As an optional embodiment, the determining unit is specifically used for: The second multi-view visual observation, the second robot body state, the second task instruction, the task completion status signal, and the intervention signal are obtained from the execution feedback after the robot executes the first absolute action block; Confirm the type of indication for the task completion status signal in the execution feedback; In response to the indication type indicating that the robot has successfully completed the task, a sparse reward signal of a first value is generated; and in response to the indication type indicating that the robot has not completed the task, a sparse reward signal of a second value is generated. The state type of the termination state is determined based on the task completion status signal and / or intervention signal in the execution feedback; The second multi-view visual observation, the second robot body state, and the second task instruction are encoded to obtain a second query-side token sequence and a second memory token sequence, which are then used to generate the second reinforcement learning state representation.

[0022] As an optional embodiment, the parameter update unit is specifically used for: In response to sampling transfer samples from the experience replay pool, a temporal difference objective is determined based on the second reinforcement learning state, the reward signal, the second reinforcement learning state representation, and the termination state; The value loss of the lightweight adapter is determined based on the difference between the temporal difference objective and the value estimate of the first reinforcement learning state representation and the first absolute action block by the lightweight adapter. Based on the value estimate of the first reinforcement learning state representation by the lightweight adapter, the value-oriented term in the policy loss is determined, and the first absolute action block is added as a constraint term to the policy loss to obtain the policy loss of the lightweight adapter. The parameters of the lightweight adapter are updated based on the value loss and the policy loss.

[0023] According to a third aspect of the present disclosure, an electronic device is provided, comprising: Memory, used to store computer program products; A processor is configured to execute a computer program product stored in the memory, and when the computer program product is executed, to implement the method described in the first aspect above.

[0024] According to a fourth aspect of the present disclosure, a computer-readable storage medium is provided that stores computer program instructions thereon, which, when executed by a processor, implement the method described in the first aspect.

[0025] According to a fifth aspect of the present disclosure, a computer program product is provided, including computer program instructions that, when executed by a processor, implement the method described in the first aspect.

[0026] The technical solutions of this disclosure will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0027] The accompanying drawings, which form part of this specification, illustrate embodiments of this disclosure and, together with the description, serve to explain the principles of this disclosure.

[0028] This disclosure will become clearer with reference to the accompanying drawings and the following detailed description, wherein: Figure 1 This is a schematic diagram of the system structure according to an embodiment of the present disclosure.

[0029] Figure 2 This is a flowchart illustrating a robot motion correction method according to an embodiment of the present disclosure.

[0030] Figure 3 This is a flowchart illustrating a robot motion correction method according to another embodiment of the present disclosure.

[0031] Figure 4 This is a flowchart illustrating a robot motion correction method according to yet another embodiment of the present disclosure.

[0032] Figure 5 This is a flowchart illustrating a robot motion correction method according to another embodiment of the present disclosure.

[0033] Figure 6 This is a schematic diagram of the structure of a robot motion correction device according to an embodiment of the present disclosure.

[0034] Figure 7 This is a schematic diagram of the structure of an electronic device according to an embodiment of the present disclosure. Detailed Implementation

[0035] World motion models (WMAs) are a class of fundamental models that generate long-sequence action blocks based on multi-view images, robot body states, and task instructions, exhibiting good generalization ability in routine tasks such as grasping and movement. However, when WMAs are deployed in contact-based tasks such as socketing, precision assembly, and high-precision alignment, inherent camera calibration errors, perception biases, actuator accuracy limitations, and contact dynamics differences in the deployment environment are amplified dramatically during the end-effector stage, leading to frequent failures in the high-precision adjustment phase of the model.

[0036] To address the aforementioned deployment errors, existing technologies mainly include online fine-tuning of all parameters, motion residual correction, training small-scale strategies from scratch, and training adaptive networks after compressing internal representations. However, none of these solutions are designed for the structural characteristics of the world action model. Task-related information from the world action model is not only distributed in the final motion output but also widely distributed in intermediate latent variables such as visual encoding, world state prediction, and motion generation. If only the final motion is corrected or internal features are excessively compressed, fine-grained information essential for precise operation, such as target position, object geometry, and contact area, is easily lost, making effective end-effector accuracy correction difficult.

[0037] To address this, the present invention provides a robot motion correction method. This method trains a lightweight adapter on a frozen world motion model and utilizes the reference motion blocks output by the world motion model and their internal latent variable features to achieve rapid online correction of deployment errors. The invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0038] Figure 1 This is a schematic diagram of the system structure according to an embodiment of this disclosure. Figure 1 As shown, the online adaptive correction system of this disclosure adopts a cloud-robot collaborative architecture, including a frozen world action model as the backbone model, a lightweight adapter (actor-evaluator network), an experience replay buffer (experience replay pool), and a real robot execution environment.

[0039] It includes a frozen world action model as the backbone model, a lightweight adapter (actor-evaluator network), an experience replay buffer, and a realistic robot execution environment.

[0040] At the data flow level, multi-view visual observations (images) are first acquired using cameras deployed on the robot, simultaneously obtaining the robot's state and task instructions. This information is uploaded to the cloud and then used as the backbone model, with its network parameters frozen. The backbone model performs one forward inference step, outputting a reference action block of a preset length from its action decoding end as a coarse-grained long-term action prior. Simultaneously, it extracts latent variable features (as shown in the VAE / diffusion backbone in the figure) rich in target position, object geometry, and contact area information from the preset intermediate layers of its action generation branches. The reference action block and latent variable features, along with the robot's state, are input into a lightweight adapter. This lightweight adapter is an action refinement network with a far fewer trainable parameters than the backbone model (the world action model). It first models the structural relationship between the state and the reference action block through self-attention, then uses cross-attention to read information related to the current correction from the latent variable features as needed, generating the corrected action block (the first absolute action block). The action block is sent from the cloud to the robot for execution.

[0041] At the execution and feedback level, after the robot executes an action block, a state transition occurs in the real environment. Multi-view visual observations, robot state, and task completion status signals are collected after execution, while simultaneously monitoring for human intervention signals. This information serves as execution feedback. This feedback is used to calculate sparse reward signals (positive rewards for task success, negative rewards for failure) and determine the termination state. Furthermore, it, together with the reinforcement learning state representation and action block before execution, constitutes a complete transition sample, which is stored in the experience replay buffer. The system periodically samples batch transition samples from the experience replay buffer, uses an actor-evaluator reinforcement learning algorithm to calculate the value loss Q(s,a) and policy loss, and updates the network parameters of the lightweight adapter accordingly. Meanwhile, the parameters of the world action model, which serves as the backbone model, remain frozen.

[0042] At the task level, the system supports a variety of precision operation tasks that require end alignment, contact, or insertion, such as inserting sticks, inserting network cables, assembling blocks, inserting plugs, ringing bells, and hammering blocks. It can be trained online on a real robot system or pre-learned in a simulation environment.

[0043] The aforementioned system architecture forms a continuous optimization closed loop: perception → coarse-grained planning of the world action model → fine-grained correction of the lightweight adapter → execution → feedback → online update. This is achieved by freezing the world action model as the backbone model, performing lightweight adapter inference and online gradient updates in the cloud, executing actions on the robot and sending back feedback. This enables rapid online adaptive correction of deployment environment errors.

[0044] The technical solutions of the embodiments of this disclosure will now be described in detail with reference to the accompanying drawings.

[0045] Figure 2 This is a schematic flowchart illustrating a robot motion correction method according to an embodiment of this disclosure. Figure 2 As shown, the robot motion correction method of this disclosure embodiment may include the following steps: Step 210: Obtain the robot's first multi-view visual observation, first robot body state, and first task instruction.

[0046] In this embodiment of the disclosure, step 210 is used as the data acquisition entry point for the entire online adaptive correction method. Multiple types of sensors deployed on the robot itself and in its environment can be used to collect, at time t, at least three types of modal data, including but not limited to visual observations, body state, and task commands, providing raw input for subsequent inference of the world action model (WAM) and correction of the lightweight adapter.

[0047] For example, this can be achieved through the following hardware devices and software interfaces: Visual sensors: The robot can be equipped with, for example, a global-view RGB camera and / or a wrist-mounted RGB camera. The global camera is fixed above the workspace, providing scene-level contextual information; the wrist camera is mounted at the end of the robotic arm, providing a localized, fine-grained view. Both cameras are RGB cameras, but can be replaced with RGB-D cameras to acquire depth information. This enables the acquisition of first-level multi-view visual observations.

[0048] Body state sensors: The robot body integrates joint encoders, end effector pose sensors, gripper state sensors, and velocity / torque sensors. The joint encoders measure the angles of each joint; the end effector pose sensors measure the three-dimensional position and orientation of the actuator in Cartesian space; the gripper state sensors measure the gripper opening degree; the velocity sensors measure the movement speed of each joint; and the torque sensors measure the forces and torques acting on the joints or end effector. These sensors are used to acquire the state of the first robot body.

[0049] Task instruction input device: Receives task instructions through a human-computer interaction interface, voice recognition module, or host task planning system to acquire the first task instruction.

[0050] Data acquisition and transmission software: Data from each sensor is synchronously collected by the robot's acquisition module and uploaded to a cloud server via wired or wireless network for use by the world motion model inference. The cloud server is equipped with a GPU (e.g., NVIDIA A800, A100, H100 or equivalent GPU) to perform world motion model inference, lightweight adapter inference, and online gradient updates.

[0051] In this embodiment, the first multi-view visual observation can be RGB image data, for example, synchronously acquired by a global camera and a wrist camera, with a resolution of, for example, 224×224×3 and a frame rate of, for example, 30fps, providing scene-level context and end-effector local detail information respectively; the first robot body state is numerical vector data, acquired by a joint encoder, an end-effector pose sensor, and a torque sensor, including parameters such as joint angle vectors, the three-dimensional position and orientation of the end effector, gripper opening degree, speed, and torque; the first task instruction is text or encoded data, obtained through a human-machine interface or a host task planning system, using natural language strings or discrete task identifier encoding, used to specify the robot operation task to be executed.

[0052] For example, multi-view visual observation can provide visual context information about the robot's environment—the global view captures scene-level layout (such as workpiece position and target area), while the wrist view provides local details near the end effector (such as hole positions and contact surfaces). This multi-view information is input into the world motion model and processed by the visual encoding branch to generate visually encoded features for scene understanding and target recognition. The robot's body state provides its own kinematic and dynamic information, including current joint configuration and end effector pose. This data is used for multimodal fusion and temporal modeling of the world motion model, and also serves as an important component of the reinforcement learning state representation for motion correction in lightweight adapters. Task instructions specify the robot's current task to be performed (such as "inserting a stick" or "inserting a network cable"), guiding the world motion model to generate action priors aligned with the task semantics, ensuring that the reference action blocks conform to the task objectives.

[0053] This step-by-step embodiment provides complete and complementary multimodal input for subsequent world action model inference through multi-view visual observation and simultaneous acquisition of ontology state and task instructions. The multi-view design overcomes the visual blind spots of a single viewpoint, enabling the world action model to simultaneously acquire scene-level context and end-effector local details; ontology state data aligns action generation with the robot's current configuration; and task instructions ensure that action priors have task semantic consistency, laying the foundation for the entire online adaptive correction method.

[0054] Step 220: Input the first multi-view visual observation, the first robot body state and the first task instruction into the pre-trained world action model to obtain reference action blocks and latent variable features, wherein the network parameters of the world action model are in a frozen state.

[0055] In this embodiment, the reference action block can serve as a coarse-grained long-term action prior, while latent variable features serve as a fine-grained representation including target position, object geometry, and contact area information. These two types of outputs, along with the robot's body state, constitute the input to the subsequent lightweight adapter. As an embodiment of this disclosure, all network parameters of the world action model remain frozen throughout this step and all subsequent steps, and do not participate in any form of parameter update.

[0056] As one implementation method, this step can be performed using a GPU server deployed in the cloud. The first multi-view visual observation, the first robot body state, and the first task command are collected by the robot and uploaded to the cloud via wired or wireless network. The cloud server is equipped with a high-performance GPU (e.g., NVIDIA A800, A100, H100, or equivalent GPU) to load the pre-trained world action model and perform forward inference. The world action model is a pre-trained world action model on a large-scale offline dataset, and its network parameters are fixed at deployment time. After inference is completed, the cloud outputs reference action blocks and latent variable features to a lightweight adapter module for use in subsequent steps.

[0057] The world action model, based on multi-view visual observations, robot state, and task instructions, can generate reference action blocks and latent variable features through the coordinated processing of its internal functional branches. In specific implementations, the acquisition of these two types of outputs can be achieved through methods including, but not limited to, the following: (1) The visual encoding branch of the world action model can be used to first encode the features of multi-view visual observations to obtain visual encoding features; then, the visual encoding features are fused with the robot's body state and task instructions in a multimodal manner and temporal modeling to obtain the fused world state features; the world state features are then input into the action generation branch. During the forward propagation of the action generation branch, on the one hand, the action generation latent variable features are extracted from the preset intermediate layer, and on the other hand, the action decoding branch decodes the latent variable features and outputs a reference action block of preset length.

[0058] (2) After the visual encoding branch of the world action model completes the multi-view image feature encoding, visual latent variable features can be extracted from the intermediate or output layer of the visual encoder, rather than from the action generation branch. The latent variable features obtained in this way focus on the visual semantic understanding of the scene and are suitable for operation tasks that rely on strong visual priors (such as object recognition and localization).

[0059] (3) After performing multimodal fusion and temporal modeling on the world action model, latent variable features are extracted from the output of the world state prediction module. These features focus on the temporal evolution prediction of the environmental state and are suitable for tasks that require predicting the trajectory of an object or changes in its contact state.

[0060] (4) If the world action model uses a variational autoencoder (VAE) or diffusion model architecture for action generation, features are extracted from the latent space of the VAE or the intermediate latent variables of the diffusion process. These features have probability distribution characteristics, can reflect the uncertainty of action generation, and are suitable for tasks that require multimodal action selection.

[0061] (5) Features are extracted from multiple intermediate layers of the world action model, such as the visual encoding layer, action generation layer, and world state prediction layer, and multi-scale fusion is performed to obtain a comprehensive latent variable feature representation. This approach can take into account semantic information at different levels and is suitable for complex tasks that require the simultaneous use of visual semantics, temporal dynamics, and action generation context.

[0062] All of the above implementation methods can complete the acquisition of reference action blocks and latent variable features. The specific method used can be selected according to the architectural characteristics of the world action model and the accuracy requirements of the target task. This disclosure does not limit this.

[0063] In this embodiment, the network parameters of the world action model are kept frozen, thus fully preserving its visual understanding, scene prediction, and long-term planning capabilities learned on large-scale offline data. This avoids the prior waste caused by training small strategies from scratch and also avoids the computational overhead and overfitting risk caused by online fine-tuning of all parameters. Simultaneously, reference action blocks, as coarse-grained long-term action priors, provide global action directions for the subsequent lightweight adapter, while latent variable features, as fine-grained intermediate representations, carry information such as target position, object geometry, contact area, and action generation context. These two types of outputs provide complementary multi-level information from the action output layer and the intermediate representation layer, respectively, laying a data foundation for subsequent precise calibration. Furthermore, by extracting latent variable features from a pre-defined intermediate layer instead of simply splicing or over-compressing all outputs of the world action model, the loss of fine-grained spatial information is avoided. At the same time, the extraction location of latent variable features can be flexibly selected according to the world action model architecture and task accuracy requirements, enabling this method to adapt to various types of world action model architectures. In addition, the world action model only performs forward inference once, without involving backpropagation and gradient calculation, and all network parameters are always frozen. This makes the computational burden of the entire online adaptive correction method mainly concentrated on the lightweight adapter, reducing the consumption of computing resources and inference latency of cloud GPUs, and meeting the real-time requirements of real robot deployment scenarios.

[0064] Step 230: Construct a first reinforcement learning state representation based on the first robot body state, reference action block, and latent variable features, and input the first reinforcement learning state representation into the lightweight adapter so that the lightweight adapter generates a first absolute action block based on the attention mechanism and the first reinforcement learning state representation.

[0065] The lightweight adapter is an action refinement network with fewer trainable parameters than the world action model, and the first absolute action block is the action instruction to be verified output by the lightweight adapter during the calibration process.

[0066] In this step, reference action blocks and latent variable features are combined with the first robot body state to construct a first reinforcement learning state representation, which is then input into a lightweight adapter. The lightweight adapter can fuse the above information based on an attention mechanism, ultimately outputting the first absolute action block. The lightweight adapter is an action refinement network with a far fewer trainable parameters than the world action model as the backbone model. The first absolute action block it outputs during the calibration process is the action instruction to be verified, used to drive the real robot to execute, and to collect feedback in subsequent steps to drive its own iterative updates.

[0067] As one implementation, this step can be executed using a GPU server deployed in the cloud. The reference action block and latent variable features, along with the first robot's body state, are input into the lightweight adapter. In this embodiment, the lightweight adapter can also be deployed in the cloud, with its network parameters stored in GPU memory, enabling forward inference and backward propagation. The lightweight adapter has a much smaller number of parameters than the world action model used as the backbone model (e.g., only 1% to 5% of the world action model's parameters), resulting in extremely low computational overhead and enabling frequent online gradient updates while maintaining real-time inference speed.

[0068] In this embodiment, the construction of the first reinforcement learning state representation can uniformly encode three types of heterogeneous information—robot ontology state (numerical vector), reference action block (action sequence), and latent variable features (spatiotemporal feature map)—into a token sequence format that can be processed by a lightweight adapter. In specific implementations, the construction methods include, but are not limited to: Optionally, these can be encoded as query-side representations and memory-side representations, respectively. Specifically, the first robot body state and the reference action block are encoded as query-side representations, and the latent variable features are encoded as memory-side representations; together, they constitute the first reinforcement learning state representation. This approach provides a clear input paradigm for subsequent attention mechanisms through the query / memory separation design.

[0069] Optionally, all information can be encoded into a single joint representation. Specifically, the first robot body state, complete reference action block, and latent variable features can be directly fused into a joint representation after projection, without distinguishing between the query side and the memory side. The lightweight adapter performs global information interaction within this joint representation. This approach is simple to implement and suitable for scenarios with low-dimensional latent variable features.

[0070] Optionally, cross-attention can be used to construct the state representation. Specifically, the robot's own state is used as the query, and the reference action block and latent variable features are used as both keys and values. The state representation is obtained by directly fusing them through cross-attention. This approach is suitable for scenarios that require retrieving information on demand, with the robot's current state as the core.

[0071] Alternatively, iterative attention can be used to construct the state representation. Specifically, a fixed set of potential tokens can be set as queries, and information can be extracted from the multimodal input through multiple iterations of cross-attention to finally obtain a compact reinforcement learning state representation. This approach is suitable for scenarios with extremely high input dimensionality and can effectively control computational complexity.

[0072] All of the above implementation methods can complete the construction of reinforcement learning state representations. The specific method used can be selected based on the architectural characteristics of the world action model, the computational resource constraints of the target task, and the accuracy requirements.

[0073] Furthermore, based on the constructed first reinforcement learning state representation, a lightweight adapter is used to generate the first absolute action block through an attention mechanism. In specific implementations, this can include, but is not limited to, the following methods: Optionally, a lightweight adapter is used to first perform self-attention processing on the query-side representation to model the structural relationship between the robot's body state and the reference action block, maintaining the temporal consistency of the action block; then, the processed query-side representation is used as the query and the memory-side representation is used as the key and value to perform cross-attention processing, obtaining information such as target position, object geometry, end-effector alignment, and potential contact area related to the action correction to be generated from the latent variable features; finally, the first absolute action block is output through the action head.

[0074] Optionally, a lightweight adapter employs a multi-layered alternating self-attention and cross-attention structure. In each layer, self-attention is first performed on the query side, followed by cross-attention with the memory side. This multi-layered stacking achieves more comprehensive multimodal information fusion. This approach is suitable for complex tasks requiring iterative refinement and multiple retrievals of information at different levels from latent variables.

[0075] Optionally, a gating mechanism is introduced after cross-attention to allow the lightweight adapter to adaptively determine how much information to read from the memory-side representation for the current action correction. When the latent variable features are weakly correlated with the current correction requirement, the gating value approaches zero, and the lightweight adapter mainly relies on the reference action block to generate the output, thereby avoiding interference from irrelevant features.

[0076] It is understood that all of the above implementation methods can generate the first absolute action block. The specific method used can be selected according to the architectural characteristics of the world action model and the needs of the target task. This disclosure does not impose any restrictions on this.

[0077] In this step, the reference action block, as a coarse-grained prior, provides the lightweight adapter with long-term action directions and overall trends, reducing learning costs while preserving the behavioral priors of the world action model. Meanwhile, latent variable features, as fine-grained information, provide spatial details such as target location, object geometry, and contact areas. Through a cross-attention mechanism, these features are read on demand rather than simply concatenated or over-compressed, avoiding the loss of crucial spatial information and enabling the lightweight adapter to obtain sufficient contextual support during the final precision operation stage. The lightweight adapter's structural design of performing self-attention followed by cross-attention on the query-side representation maintains the continuity and stability of the output action block in the temporal dimension through the structural relationship between the self-attention modeling state and the reference action block, reducing jitter that may be caused by action-by-action correction. Finally, the lightweight adapter has far fewer parameters than the world action model as the backbone model, resulting in extremely low computational overhead and enabling real-time inference on cloud GPUs.

[0078] Step 240: In response to receiving the execution feedback after the robot executes the first absolute action block, update the parameters of the lightweight adapter online based on the execution feedback. The execution feedback includes the second multi-view visual observation, the second robot body state, the second task instruction, the task completion status signal, and the intervention signal.

[0079] In this embodiment, after the first absolute motion block is generated by the lightweight adapter, it can be sent to the robot end controller via the cloud-robot communication link. The robot end controller parses it into specific joint torque or end-effector pose control commands and drives the robotic arm to execute them. After the robot executes the first absolute motion block, a state transition occurs in the real environment. The system collects multi-dimensional feedback information after execution, which is the execution feedback. This execution feedback includes at least a second multi-view visual observation, a second robot body state, a second task command, a task completion status signal, and an intervention signal. In this step, based on the above execution feedback, on the one hand, the second reinforcement learning state representation, the termination state, and the reward signal are determined; on the other hand, the above information is stored as transition samples in the experience replay pool, and batch data is sampled from it to update the network parameters of the lightweight adapter online. Through this step, the lightweight adapter can continuously optimize its own parameters according to the execution results in the real environment, gradually converging to the optimal correction strategy in the current deployment environment.

[0080] As one implementation approach, this step can rely on a cloud-robot collaborative architecture. After receiving the first absolute action block, the robot executes the action. Upon completion, it acquires a second multi-view visual observation and a second robot body state, while simultaneously recording task completion status signals and intervention signals. The execution feedback is transmitted back to the cloud via wired or wireless network. A lightweight adapter on the cloud server performs forward inference to generate actions, calculates the loss based on the feedback, and updates the parameters. The updated parameters are stored in the GPU memory for use in the next round of inference. An experience replay pool can also be deployed in the cloud to store transferred samples and provide batch sampling. Data transmission between the cloud and the robot uses a real-time communication protocol to ensure a low-latency closed loop for action delivery and feedback transmission.

[0081] For example, in this step, the execution feedback generated by the robot after executing the first absolute action block can be obtained first. Among them, the first three items of the execution feedback are perception layer data, such as the second multi-view visual observation collected by the global camera and wrist camera after execution, the second robot body state collected by the joint encoder, end-effector pose sensor, etc., and the second task instruction is the semantic identifier of the current task; the task completion status signal is decision layer data, indicating whether the current task has been successfully completed or failed; the intervention signal is human input data, triggered by the operator through physical buttons or teleoperation devices, indicating whether human intervention has occurred.

[0082] After obtaining execution feedback, the second multi-view visual observation, the second robot body state, and the second task instruction can be encoded into a second query-side representation and a second memory-side representation in the same way as in step 230 above, together forming the second reinforcement learning state representation. Based on the task completion state signal, it is determined whether the current task was successful or failed, and a sparse reward signal is generated accordingly. Based on the task completion state signal and / or intervention signal, it is determined whether the current round should be terminated. After the above processing is completed, the first reinforcement learning state representation, the first absolute action block, the reward signal, the second reinforcement learning state representation, and the termination state are combined into a complete transition sample.

[0083] Furthermore, the aforementioned transferred samples can be stored in an experience replay pool. In this embodiment, the capacity of the experience replay pool can be preset according to actual needs, for example, storing the latest 5000 transferred samples by default, and discarding old samples in a first-in-first-out manner when the pool is full. After storage, the system periodically randomly samples a batch of transferred samples from the experience replay pool (the number of samples can be preset according to computing resources, for example, 256 samples each time), calculates the value loss and policy loss of the lightweight adapter based on the sampled transferred samples, and performs gradient updates on the lightweight adapter network parameters according to the above losses. The design of the experience replay pool breaks the temporal correlation between samples, improving training stability and sample utilization efficiency; at the same time, periodic batch sampling updates have higher computational efficiency than single-step updates.

[0084] As an implementation of this disclosure, the online update of the lightweight adapter is not limited to a single reinforcement learning algorithm; various actor-critic variants can be used in specific implementations. For example, the TD3 (Twin Delayed Deep Deterministic Policy Gradient) algorithm or its variants can be used for updating. This algorithm utilizes a target network to generate stable temporal difference targets, employs a dual-Q network to reduce overestimation bias in value estimation, and improves training stability through delayed policy updates and policy noise regularization, making it a preferred solution for continuous action control tasks. Alternatively, the SAC (Soft Actor-Critic) algorithm can be used, introducing an entropy regularization term to encourage the policy to remain exploratory during training, suitable for tasks with continuous action spaces and requiring high exploration efficiency. Or, the DDPG (Deep Deterministic Policy Gradient) algorithm can be used, as deterministic policy gradient methods are suitable for scenarios with relatively low action space dimensions and limited computational resources. For tasks requiring high training stability, PPO (Proximal Policy Gradient) can be used. The Optimization algorithm achieves stable policy updates through importance sampling and trust domain constraints. However, as an on-policy algorithm, its compatibility with offline sampled experience replay pools requires specific adaptation. In scenarios with limited online data, an offline-to-online reinforcement learning algorithm can be used. This involves pre-training the lightweight adapter on an offline dataset and then fine-tuning it using online interactive data during actual deployment to obtain a stable initial policy more quickly. Alternatively, Advantage-Weighted Regression (AWR) or its variants can be used for policy updates based on weighted regression. This approach is simple to implement and provides stable training, making it suitable for scenarios requiring avoidance of complex value function estimation. All of these implementation methods can achieve online updates of the lightweight adapter. The specific algorithm chosen depends on the computational resource constraints, training stability requirements, and exploration efficiency requirements of the target task.

[0085] In this embodiment, the lightweight adapter update relies solely on task completion status and intervention signals from the execution feedback, without depending on precise contact dynamics models or environmental calibration parameters. This allows the method to achieve adaptive correction without requiring precise environmental modeling, lowering the deployment threshold. Secondly, by storing interaction data in an experience replay pool and updating it in batches, the temporal correlation between continuous samples is broken, improving training stability and sample utilization efficiency. It also supports stable policy learning even with limited real robot interaction samples. Thirdly, the online update process only adjusts the lightweight adapter parameters, while the world motion model, as the backbone model, remains frozen. This avoids the computational burden and overfitting risk associated with full parameter fine-tuning, while preserving the universal prior learning capabilities of the world motion model on large-scale offline data. Finally, through continuous iteration, the lightweight adapter gradually converges to the optimal correction strategy in the current robot deployment environment, achieving comprehensive suppression of camera calibration errors, perception biases, actuator accuracy limitations, and contact dynamics differences. The resulting second absolute motion block can complete precision operation tasks with higher accuracy and success rate.

[0086] Step 250: Iterate through steps 210 to 240 until the preset termination condition is met, and output the target lightweight adapter.

[0087] This step is the convergence judgment and output stage of the entire online adaptive calibration method. Specifically, after completing a single parameter update in step 240, the method flow returns to step 210, using the multi-view visual observation, robot body state, and task instructions at the second moment as input, and re-executes the complete loop from step 210 to step 240. The above process is repeated until the preset termination condition is met. At this point, iteration stops and the lightweight adapter updated in the current iteration is output as the target lightweight adapter. The target lightweight adapter is the lightweight adapter after online adaptive calibration convergence. Its network parameters have been optimized for the current robot deployment environment and can be directly used to generate the second absolute action block to control the robot to complete the operation task without further online updates. The second absolute action block is the calibration action instruction output after the target lightweight adapter calibration convergence. It has higher accuracy and execution success rate than the first absolute action block generated in the initial stage of online calibration.

[0088] This iterative process operates continuously using a cloud-robot collaborative architecture. In each loop, the robot collects multi-view visual observations, its own state, and task instructions at the current moment and uploads them to the cloud. The cloud then sequentially executes world action model inference (e.g., step 220 above), lightweight adapter inference to generate actions (e.g., step 230 above), action delivery and execution, and feedback collection (e.g., step 240 above), before entering the next loop. During each iteration of the loop, the lightweight adapter parameters are continuously optimized based on the previous update. The monitoring of termination conditions is performed by the cloud-based training management module, which continuously calculates metrics such as the success rate, intervention rate, and cumulative training time for the most recent N rounds. When the preset termination condition is met, the training management module sends a termination signal to the lightweight adapter training process, stopping the loop and fixing the current lightweight adapter parameters into the target lightweight adapter, storing it in the cloud model library for subsequent deployment. After training terminates, the robot receives the target lightweight adapter and runs in inference mode (i.e., parameters are frozen, only forward inference to generate actions), without performing gradient updates.

[0089] In this embodiment, a preset termination condition is used to determine when the online adaptive correction reaches a convergence state. In specific implementations, one or more of the following conditions can be combined: For example, the task success rate in the most recently executed preset number of rounds can be directly calculated. When the moving average reaches a preset success rate threshold, the iteration terminates. This method directly reflects the actual execution effect of the strategy and is the most crucial basis for termination judgment. Alternatively, the proportion of rounds in the most recently executed preset number of rounds where human intervention occurs can be calculated. When the intervention rate is lower than a preset intervention rate threshold, the iteration terminates. A low intervention rate indicates that the strategy requires less human intervention during execution, which is an important indicator of strategy maturity. Furthermore, when the total online training time reaches a preset time... The upper limit of the training interval serves as a termination condition, acting as a safety net to prevent unlimited training due to environmental anomalies or strategies failing to converge. It can also monitor the fluctuation of success rate or cumulative reward over a preset number of consecutive rounds, terminating the iteration when performance metrics stabilize (i.e., fluctuations fall below a preset threshold) to avoid premature termination due to random fluctuations from single successes or failures. Alternatively, it can monitor the changes in the value loss and policy loss of the lightweight adapter over a preset number of updates, terminating the iteration when the change in loss falls below a preset threshold, directly determining model convergence from the perspective of training loss. It can also terminate iteration when the total number of iterations reaches a preset upper limit, suitable for deployment scenarios with clear limitations on the total number of training steps. These various implementation methods can be used individually or in combination (e.g., terminating when the success rate reaches a threshold or the training duration exceeds a limit). For example, a termination condition combining success rate, intervention rate, and training duration can be used: terminating iteration when the task success rate exceeds 90% and the intervention rate is below 10% in the last 20 rounds, or forcibly terminating and saving the current model when the training duration reaches 2 hours.

[0090] Through continuous iterative optimization, the lightweight adapter gradually converges to the optimal correction strategy in the current robot deployment environment. This effectively suppresses the impact of camera calibration errors, perception biases, actuator accuracy limitations, and contact dynamics differences on precision operations, resulting in high accuracy and execution success rate for the final output second absolute motion block. Secondly, the success rate threshold and intervention rate threshold in the preset termination conditions directly reflect the execution effect and reliability of the strategy in actual tasks. Training automatically stops when the strategy stably meets the task requirements and requires minimal human intervention, balancing model performance and training efficiency while avoiding unnecessary computational resource consumption. Thirdly, the training duration upper limit serves as a safety net mechanism, ensuring that even if the strategy fails to reach the ideal convergence state within the expected time, it can terminate in time and output the current optimal model for subsequent analysis, guaranteeing the engineering usability of the entire online adaptive correction method. Finally, the target lightweight adapter can be directly used for deployment inference after output. Its network parameters remain fixed, requiring no online updates, and the inference computation is minimal, enabling efficient operation on robot terminals or edge devices, reducing the computational requirements during the production deployment phase.

[0091] Figure 3 This is a flowchart illustrating a robot motion correction method according to another embodiment of the present disclosure. The robot motion correction method according to this embodiment may further include the following steps: Step 221: In response to inputting the first multi-view visual observation, the first robot body state, and the first task instruction into the world action model with the network parameters frozen, the first multi-view visual observation is feature-encoded through the visual encoding branch of the world action model to obtain visual encoding features.

[0092] In this step, the acquired multi-view visual observations are converted into visually encoded features. The visual encoding branch of the world action model receives multi-view RGB images from the global camera and the wrist camera. A visual encoder (e.g., ResNet, ViT, or a multi-view fusion network composed of them) extracts features and spatially encodes each frame, mapping the high-dimensional pixel space into compact, semantically rich visual feature vectors or feature maps to capture the position, geometry, spatial relationships, and task-related visual cues of objects in the scene. The visually encoded features differ from the original images; for example, redundant texture and lighting information may be discarded, retaining only the semantic and spatial information most relevant to the task. This step is executed first in the forward inference process of the world action model, and its output serves as the input for subsequent multimodal fusion. Furthermore, depending on the architecture of the world action model and the task requirements, the visually encoded features from this step can also serve as one of the sources for extracting latent variable features.

[0093] For example, assuming a 224×224×3 resolution three-view RGB image (one from the global camera and two from the wrist camera), the visual encoding branch can use a pre-trained ResNet-34 or ViT as the backbone network to encode each image into a 512-dimensional feature vector. Then, a multi-view attention fusion module aggregates the features from the three views into a unified visual encoding feature tensor. This tensor has dimensions such as 512 or a 7×7×512 spatial feature map structure, used to represent the overall visual context of the current scene. In specific implementations, if ResNet is used as the encoder, the output of the last convolutional layer is taken as the visual encoding feature; if ViT is used, the [CLS] token at the output or the average pooling result of all tokens is taken.

[0094] In this way, by encoding high-dimensional image data into compact visual coding features, the computational burden of subsequent multimodal fusion and temporal modeling is effectively reduced. At the same time, the visual coding features retain information such as the position, geometry and spatial relationships of objects in the scene, providing sufficient visual semantic input for scene understanding and action planning of subsequent world action models.

[0095] Step 222: Perform multimodal fusion and temporal modeling of the visual encoding features with the first robot body state and the first task instruction to obtain the fused world state features.

[0096] The extracted visually encoded features are deeply fused and temporally modeled with the first robot body state and the first task instruction. Specifically, the world action model, as a multimodal foundation model, needs to simultaneously process three types of heterogeneous data: visual perception information, robot kinematic state, and task semantic constraints. This step aligns these three types of information into a unified feature space through a multimodal fusion module, and then integrates historical information temporally through a temporal modeling module (such as Transformer, LSTM, or state-space model), outputting fused world state features to provide sufficient contextual information for action generation. Visually encoded features provide a visual semantic understanding of the current environment, including the position, geometry, and spatial relationships of target objects; the robot body state provides the robot's own kinematic and dynamic information, including joint angles, end effector pose, and velocity; and the task instruction provides task semantic constraints and desired objectives. After fusing and temporally modeling these three types of information, a comprehensive representation of the current world state can be obtained, serving as the conditional input for subsequent action generation.

[0097] For example, a Transformer decoder structure can be used as a multimodal fusion and temporal modeling module. Visual encoded features serve as keys and values, while the encoded sequences of robot body state and task instructions serve as queries. Multimodal information fusion is achieved through cross-attention. Then, the fused multimodal features are input into a temporal Transformer layer to perform temporal modeling on the features of the current moment and several preceding moments. The output is a fused world state feature with a dimension of, for example, 1024. The features of preceding moments can come from the preceding step data in the historical observation window. The length of this window can be set according to task requirements, for example, covering the observation and state information of the most recent 3 to 5 moments.

[0098] Thus, by fusing and temporally modeling three types of heterogeneous information—visual encoding features, robot body state, and task instructions—the world action model can simultaneously utilize visual semantics, kinematic state, and task semantics for comprehensive understanding. This avoids scene understanding bias or inconsistency between action planning and task objectives caused by insufficient information in a single modality. At the same time, temporal modeling enables the model to remember historical states and predict the future, laying the foundation for the subsequent generation of long-term temporal action blocks with temporal consistency.

[0099] Step 223: Input the world state features into the action generation branch of the world action model. During the forward propagation of the action generation branch, extract the action generation latent variable features from the preset intermediate layer.

[0100] In this step, latent variable features are extracted during the action generation process of the world action model. After obtaining the fused world state features and inputting them into the action generation branch of the world action model, the action generation branch performs a complete forward propagation, progressively converting the world state features into latent representations that can be decoded into specific actions. During this forward propagation, latent variable features can be extracted from preset intermediate layers of the action generation branch (e.g., an intermediate layer of the Transformer decoder, a latent space layer of a VAE, or a time step of a diffusion model denoising network). These latent variable features differ from the reference action block (which is the final action output after decoding). They contain richer information about the target location, object geometry, contact area, and action generation context, serving as fine-grained features to provide the spatial detail information required for precise operation in the subsequent lightweight adapter. Extracting latent variable features and decoding the reference action block are two events that occur in parallel during the same forward inference process. For example, the latent variable features come from an intermediate layer of the action generation branch, while the reference action block comes from the output of the action decoding branch.

[0101] The specific intermediate layer outputs extracted can be determined based on the architectural characteristics of the World Action Model (VAE) and the accuracy requirements of the target task. For example, the latent space vectors of the VAE can be extracted as latent variable features to obtain the probability distribution information of action generation; or the output of the penultimate layer of the Transformer decoder can be extracted to preserve the contextual semantic information of action generation. In a preferred embodiment, the output of a single intermediate layer directly related to action generation in the action generation branch is extracted as a latent variable feature.

[0102] For example, the world action model adopts an architecture based on Transformer and VAE, with the action generation branch containing 6 Transformer decoders. The fused world state features are sequentially propagated forward through the 6 decoders. When the computation reaches the 4th decoder layer, the feature tensor is extracted from the output of this layer as the action generation latent variable features. These features are a sequence of latent variables that preserves temporal resolution, with one latent variable token corresponding to each time step. After extraction, the forward propagation continues, passing through the 5th and 6th decoders, and finally the action decoding branch outputs the reference action block. Alternatively, if the world action model adopts a diffusion model architecture, latent variable features can be extracted from the intermediate time steps of the inverse denoising process of the diffusion process.

[0103] In this way, by extracting latent variable features from the preset intermediate layer of the action generation branch, rather than relying solely on the final output reference action block, the subsequent lightweight adapter can obtain intermediate representations rich in fine-grained information such as target position, object geometry, contact area, and action generation context. At the same time, the extraction of latent variable features does not change the forward propagation computation flow of the world action model, nor does it introduce additional computational overhead or network parameter updates, ensuring that the world action model is completely frozen as the backbone model. In addition, the extraction location can be flexibly selected according to task requirements, taking into account the adaptability flexibility of different world action model architectures.

[0104] Step 224: Decode the latent variable features of action generation through the action decoding branch of the world action model, and output a reference action block of preset length.

[0105] In this step, the motion decoding branch receives the motion generation latent variable features finally output by the motion generation branch (or continues to complete the calculation of subsequent decoding layers), and can decode and map them from the latent representation space back to the robot motion space, outputting a reference motion block of a preset length. This reference motion block is the original motion prior generated by the world motion model based on the current multi-view visual observations, robot body state, and task instructions. It contains a long-term motion sequence, with each time step corresponding to a robot motion instruction (e.g., a joint angle target value or an end effector pose increment). This reference motion block, as a coarse-grained long-term motion prior, provides the overall motion direction and trend for the subsequent lightweight adapter, eliminating the need for the lightweight adapter to plan motions from scratch. It should be noted that in this embodiment, the reference motion block is a motion prior learned based on offline training data. It is not corrected for specific errors in the current deployment environment (camera calibration errors, contact dynamics differences, etc.), and therefore may not be accurate enough in the final stage of high-precision operation, requiring refinement by the subsequent lightweight adapter.

[0106] For example, the motion decoding branch receives the motion generation latent variable features (e.g., a latent variable sequence with a dimension of 48×512, where 48 is the number of time steps) finally output in step 223, and maps them to joint angle increments through a linear layer or a small MLP to obtain a reference motion block with a preset length H=48. Each time step outputs a 7-dimensional motion instruction (e.g., a 6-dimensional end pose increment plus a 1-dimensional gripper opening degree), for a total of 48 time steps of robot motion sequence, forming a complete motion block.

[0107] In this way, by outputting a reference action block of preset length through the action decoding branch, a high-quality long-term action prior is provided for the subsequent lightweight adapter, avoiding the need for the lightweight adapter to explore and plan actions from scratch, and significantly improving the efficiency of online learning. At the same time, the reference action block, as a coarse-grained prior, complements the extracted latent variable features, which serve as fine-grained information, providing multi-level correction basis from the "action output layer" and the "intermediate representation layer" respectively, providing sufficient prior support for end-point correction in subsequent precise operations. In addition, the world action model only performs forward inference once, without involving backpropagation and gradient calculation. The online update of the lightweight adapter does not affect the inference results of the world action model as the backbone model, ensuring the complete preservation of the world action model's prior capabilities.

[0108] Figure 4 This is a flowchart illustrating yet another embodiment of a robot motion correction method according to the present disclosure. Figure 4 As shown, the robot motion correction method of this disclosure embodiment may further include the following steps: Step 231: Encode the first robot body state, reference action block and latent variable features to obtain the first query side token sequence and the first memory token sequence.

[0109] This step encodes three types of heterogeneous data—the first robot body state, the generated reference action block, and latent variable features—into a unified token sequence format that can be processed by the lightweight adapter, in order to construct the first reinforcement learning state representation. The reinforcement learning state representation needs to map information of different types and dimensions to a unified feature space, enabling the lightweight adapter to jointly process this information and generate corrective actions.

[0110] Specifically, this step may include two encoding paths: The first encoding path (query side): The first robot body state is encoded into a state token, which represents the robot's current kinematic and dynamic state, including joint angles, end effector pose, velocity, and other information. Simultaneously, a preset number of reference actions are extracted from the reference action block. This preset number is less than or equal to the total length of the reference action block, used to control the length of the action block refined by the lightweight adapter each time, ensuring that this length is suitable for the action block-level update granularity of subsequent reinforcement learning training. Each extracted reference action is then encoded as a separate reference action token, rather than encoding all reference actions as a single token, to preserve the independent temporal information of each action at each time step. The state token and each reference action token are concatenated in chronological order to obtain the first query side token sequence. The state token and the reference action tokens together constitute the query side, indicating what the action sequence given by the reference action prior is in the robot's current state, and that the structural relationship between the two will be explicitly modeled in subsequent self-attention processing.

[0111] The second encoding path (memory side): The extracted latent variable features are encoded into at least one memory token, resulting in the first memory token sequence. The latent variable features carry fine-grained information such as target location, object geometry, contact area, and action generation context. After being encoded into a memory token sequence, they serve as key-value pairs for the query side to retrieve on demand during subsequent cross-attention processing. When the latent variable features are spatiotemporal feature maps with spatial resolution, they can be flattened into multiple memory tokens to preserve spatial details; when the latent variable features are globally compact vectors, they can be encoded into a single memory token. This encoding process transfers the fine-grained spatial information inherent in the world action model to the lightweight adapter in a structured manner, avoiding the loss of key information due to over-compression.

[0112] The first query-side token sequence and the first memory token sequence together constitute the first reinforcement learning state representation. This structured representation fully preserves the three elements of the current state, action prior, and visual geometric context, providing a clear input paradigm for the subsequent hybrid attention processing of the lightweight adapter.

[0113] For example, suppose the total length of the reference action block is H=48, and the preset number is C=12. The first robot body state (e.g., including 7-dimensional joint angles and 6-dimensional end-effector pose, totaling 13 dimensions) is encoded into a state token with a dimension of 512 through a linear projection layer. The first 12 reference actions are extracted from the 48-step reference action block, and each reference action (e.g., a 7-dimensional action command) is encoded into a reference action token with a dimension of 512 through a linear projection layer, resulting in 12 reference action tokens. The state tokens are concatenated with the 12 reference action tokens to obtain the first query-side token sequence with a length of 13 and a dimension of 512. Simultaneously, the latent variable features are spatiotemporal feature maps (e.g., feature tensors with dimensions C×T×H×W=48×12×7×7), which are flattened and encoded into multiple memory tokens (e.g., 1024), each with a dimension of 512, resulting in the first memory token sequence. At this point, the first reinforcement learning state representation is constructed.

[0114] In this embodiment, by encoding the first robot body state and reference action blocks as query-side token sequences and the latent variable features as memory token sequences, a structured input is provided for the subsequent hybrid attention processing of the lightweight adapter. First, the separate encoding of state tokens and each reference action token preserves the independent semantic information of the robot state and action priors, while the concatenation into a query-side sequence enables subsequent self-attention to explicitly model the structural relationship between the two. Second, by extracting a preset number of actions from the reference action blocks instead of using complete action blocks, the length of the action blocks is aligned with the granularity of reinforcement learning updates, facilitating updates in subsequent online reinforcement learning training. Third, by encoding the latent variable features as multiple memory tokens instead of compressing them into a single token, the spatial distribution information and multi-granular semantic information in the latent variable features are preserved, avoiding the loss of fine-grained information caused by over-compression, and providing sufficient visual geometric context for end-point correction in subsequent precision operations. Finally, by designing a preset number less than or equal to the length of the reference action blocks, the lightweight adapter can flexibly control the length of the action blocks for each refinement, balancing the consistency of long-term actions with the fineness of local corrections.

[0115] Step 232: Input the first query-side token sequence and the first memory token sequence into the lightweight adapter so that the lightweight adapter performs an attention mechanism based on the first query-side token sequence and the first memory token sequence to generate the first absolute action block.

[0116] In this step, based on the constructed first reinforcement learning state representation, the core inference process of the corrected first absolute action block is generated through a hybrid attention mechanism. This step inputs the first query-side token sequence and the first memory token sequence into a lightweight adapter. The lightweight adapter internally performs self-attention processing and cross-attention processing sequentially, and finally outputs the first absolute action block via the action head.

[0117] Specifically, the lightweight adapter can first perform self-attention processing on the first query-side token sequence. The self-attention mechanism enables sufficient information exchange between the state tokens in the query-side sequence and each reference action token. That is, each token pays attention to other tokens in the sequence through attention weights, thereby modeling the structural relationship between the first robot body state and the actions at each time step in the reference action block, as well as the temporal dependencies between actions within the reference action block. Through self-attention processing, the overall trend information of the reference action sequence is incorporated into the state tokens, and the constraint information of the current robot state is incorporated into each action token. Simultaneously, the continuity of actions between adjacent time steps within the reference action block is enhanced. This processing ensures the continuity and stability of the subsequently output corrected action blocks in the time dimension, avoiding the jitter caused by independent correction of each action.

[0118] Then, the lightweight adapter uses each token in the first query-side token sequence after self-attention processing as a query, and the first memory token sequence as a key and value, to perform cross-attention processing. During this process, each query-side token can retrieve, on demand, the latent variable information most relevant to the current correction requirement from the memory token sequence through an attention mechanism. For example, query tokens related to the end position will pay more attention to tokens in the memory token sequence that encode the target position and contact area, while query tokens related to the action direction will pay more attention to tokens in the memory token sequence that encode the object geometry and task context. Through cross-attention, fine-grained information such as target position, object geometry, contact area, and action generation context contained in the latent variable features is dynamically read and fused into the query-side representation, resulting in a query-side representation that integrates latent variable features.

[0119] Finally, based on the query-side representation of the fused latent variable features, the action head outputs a first absolute motion block of a preset length. The action head can consist of one or more linear layers or a small MLP, and its function is to map the fused query-side representation to the robot's motion space and output specific motion commands (such as joint angle target values ​​or end effector pose increments). The first absolute motion block is the motion command to be verified output by the lightweight adapter during the calibration process. It sends this command to the robot for execution and drives the online update of the lightweight adapter based on the execution feedback.

[0120] For example, the lightweight adapter uses a Transformer decoder architecture. Let the length of the first query-side token sequence be 13 (1 state token + 12 reference action tokens), the length of the first memory token sequence be 1024, and all tokens have a dimension of 512. The lightweight adapter first performs multi-head self-attention (8 heads) on the 13 query-side tokens, obtaining a query-side token sequence after self-attention processing. At this point, each token has incorporated information from other tokens in the sequence, the structural relationship between state tokens and action tokens is explicitly modeled, and temporal dependencies are established between action tokens. Subsequently, the lightweight adapter uses the 13 self-attention-processed tokens as queries and the 1024 memory tokens as keys and values, performing multi-head cross-attention (8 heads) to obtain 13 query-side representation tokens that incorporate latent variable features. Finally, the above 13 tokens are input into the action head (e.g., a two-layer MLP with a hidden layer dimension of 256 and an output dimension of action dimension × preset number, i.e., 7 × 12 = 84 dimensions), and the first absolute action block with a length of 12 is output. Each time step corresponds to a 7-dimensional action instruction (6-dimensional end pose increment + 1-dimensional gripper opening degree).

[0121] Thus, by performing self-attention processing on the query-side token sequence, the structural relationship between the robot's body state and the reference action block, as well as the temporal dependencies between actions within the reference action block, are explicitly modeled. This ensures that the output corrected action block remains continuous and stable in the time dimension, avoiding motion jitter that may result from gradual correction, which is crucial for precision operation tasks. Secondly, by using a cross-attention mechanism with the query-side token as the query and the memory token sequence as the key and value, the most relevant information to the current correction requirement is dynamically read from the latent variable features on demand, rather than simply splicing or over-compressing all latent variable information. This preserves the fine-grained spatial information of the latent variable features while avoiding irrelevant information. Interference allows the lightweight adapter to obtain sufficient contextual support, such as target position, object geometry, and contact area, during the precision operation phase at the end effector, effectively improving correction accuracy. Secondly, the lightweight adapter generates absolute action blocks rather than residuals relative to reference action blocks, allowing the output of the lightweight adapter to be directly used as robot execution instructions, simplifying the control chain and reducing the risk of motion drift caused by residual accumulation. Finally, the number of parameters in the lightweight adapter is much smaller than that of the world action model as the backbone model, and the aforementioned self-attention and cross-attention calculations are performed only on low-dimensional token sequences, resulting in extremely low computational overhead. This enables real-time inference on cloud GPUs, meeting the low-latency control requirements of robot deployment scenarios.

[0122] Figure 5 This is a flowchart illustrating a robot motion correction method according to another embodiment of the present disclosure. Figure 5 As shown, the robot motion correction method of this disclosure embodiment may include the following steps: Step 241: In response to the execution feedback after the robot executes the first absolute action block, determine the second reinforcement learning state representation, termination state and reward signal after the robot executes the first absolute action block based on the execution feedback, wherein the termination state is used to indicate whether the robot has completed the task or whether the current round needs to be terminated.

[0123] This step is used to transform the raw execution feedback after the robot executes the first absolute action block into the three key elements required for reinforcement learning training: the second reinforcement learning state representation, the termination state, and the reward signal.

[0124] Specifically, in this embodiment of the disclosure, the execution feedback may include at least five types of information: second multi-view visual observation, second robot body state, second task instruction, task completion status signal, and intervention signal. Among them, the first three are perception layer data, used to construct the second reinforcement learning state representation; the task completion status signal is decision layer data, used to generate reward signals and assist in determining the termination state; and the intervention signal is manually input data, used to determine the termination state.

[0125] When determining the second reinforcement learning state representation, the second multi-view visual observation, the second robot body state, and the second task instruction are encoded to obtain the second query-side token sequence and the second memory token sequence, which together constitute the second reinforcement learning state representation. This state representation characterizes the new state the robot is in after executing the first absolute action block and will serve as the input for lightweight adapter inference in the next loop. The encoding process and method can be found in the aforementioned encoding process identical to step 231, and will not be repeated here for simplicity.

[0126] When generating reward signals, this can be achieved by confirming the indication type of the task completion status signal: for example, if the indication type is that the robot has successfully completed the task, a sparse reward signal with a first value (e.g., +1) is generated; if the indication type is that the robot has not completed the task (including task failure or incomplete completion), a sparse reward signal with a second value (e.g., -1 or 0) is generated. Reward signals are a core metric in reinforcement learning for measuring the quality of actions, driving the lightweight adapter to optimize towards obtaining higher rewards.

[0127] When determining the termination state, the current round can be terminated based on the task completion status signal and / or intervention signal. If the task completion status signal indicates that the task has been completed or confirmed as failed, the termination state is determined to be terminated; if the intervention signal indicates that human intervention has occurred, the termination state is also determined to be terminated. If none of the above conditions are met, the termination state is determined to be not terminated, and the process flow will continue to the next loop. The termination state is used to indicate whether the agent's current round has ended and is the basis for determining whether to reset the environment and start a new round in reinforcement learning.

[0128] For example, after the robot executes the first absolute action block, the task completion status signal in the execution feedback is "success," and the intervention signal is "no intervention occurred." The system confirms that the indication type of the task completion status signal is "success" and generates a +1 sparse reward signal; simultaneously, based on the task completion status signal (success), the termination state is determined to be "termination." Furthermore, the system encodes the second multi-view visual observation, the second robot body state, and the second task instruction after execution to obtain a second query-side token sequence and a second memory token sequence, which serve as the second reinforcement learning state representation. Thus, the system determines the second reinforcement learning state representation, the termination state (termination), and the reward signal (+1).

[0129] For example, if the task completion status signal in the execution feedback is "failure" and the intervention signal is "no intervention occurred," the system confirms the indication type as "task not completed" and generates a sparse reward signal of -1; based on the task completion status signal (failure), the termination status is determined to be "termination." Alternatively, if the task completion status signal in the execution feedback is "not completed," but the intervention signal is "human intervention occurred," the system determines the termination status to be "termination" based on the intervention signal, while the reward signal remains -1 (task not completed).

[0130] Thus, by processing the perception layer data and decision layer data in the execution feedback separately, the determination of the reinforcement learning state representation, reward signal, and termination state each perform their respective functions without interference. The reward signal comes directly from the task completion state signal, eliminating the need for manually designing dense reward functions, thus lowering the threshold for reward engineering. At the same time, sparse rewards avoid policy bias caused by improper reward function design. The determination of the termination state considers both the task completion state signal and the intervention signal. The round termination can be triggered when the task succeeds, fails, or human intervention occurs, accommodating both the need for automated termination judgment and safe human intervention. Finally, by simultaneously determining the second reinforcement learning state representation, the termination state, and the reward signal, data preparation is completed for constructing a complete transfer sample.

[0131] Step 242: Store the first reinforcement learning state representation, the first absolute action block, the reward signal, the second reinforcement learning state representation, and the termination state as a transition sample into the experience replay pool.

[0132] In this step, the determined second reinforcement learning state representation, termination state, and reward signal are combined with the existing first reinforcement learning state representation and first absolute action block to form a complete transition sample, which is then stored in the experience replay pool. A transition sample is a quintuple in reinforcement learning used to represent (current state, action, reward, next state, termination state). It records the complete process of the environment at a given moment: transitioning from state s, executing action a, transitioning to state s', receiving reward r, and determining whether to terminate. A complete transition sample can include the entire chain of information from decision-making to execution to feedback in the online adaptive correction method, and serves as the basic data unit for subsequent lightweight adapter parameter updates.

[0133] In this embodiment, the experience replay pool can be a fixed-capacity data storage structure used to cache historical transfer samples. The default capacity is 5000 samples, and old samples are discarded in a first-in, first-out manner when the pool is full. This mechanism breaks the temporal correlation between continuous interaction data by caching historical experience and supporting random sampling, so that the parameter updates of the lightweight adapter are no longer limited by the latest interaction data, thereby improving training stability and sample utilization efficiency.

[0134] Let the transition sample for the current round be (s, a, r, s', done), where s is the first reinforcement learning state representation, a is the first absolute action block, r is the reward signal generated in step 241 (e.g., +1), s' is the second reinforcement learning state representation, and done is the termination state (e.g., "terminated"). The system stores this quintuple as a complete transition sample in the experience replay pool. The current capacity of the experience replay pool is 5000, and 3820 samples have been stored. New samples are stored at index 3821 in the pool. When the number of samples in the pool reaches 5000, subsequent new samples will overwrite the oldest samples.

[0135] Through this step, the transition samples in the experience replay pool can completely record the causal chain of state transitions, including the current state, action, reward, next state, and termination state, ensuring the integrity and consistency of the training data. Secondly, by caching a large number of historical samples in the experience replay pool, the lightweight adapter's updates are no longer limited to the latest interaction data, but can fully utilize historical experience for batch learning, significantly improving sample utilization efficiency. Thirdly, the random sampling mechanism breaks the temporal correlation between consecutive samples, effectively reducing variance during training and improving the stability and convergence speed of policy learning. Finally, the termination state in the transition samples provides a basis for accurately determining the round boundaries, enabling correct state resetting and value backtracking in multi-round reinforcement learning training.

[0136] Step 243: In response to sampling transfer samples from the experience replay pool, calculate the loss function of the lightweight adapter based on the sampled transfer samples, and update the parameters of the lightweight adapter according to the loss function.

[0137] In this step, a batch of transfer samples can be periodically sampled from the experience replay pool. The value loss and policy loss of the lightweight adapter are calculated based on the sampled transfer samples, and the network parameters of the lightweight adapter are updated using the above losses.

[0138] Specifically, this step can be achieved through the following steps: (1) Determine the temporal difference objective. After sampling the transfer samples from the experience replay pool, the temporal difference objective is determined based on the second reinforcement learning state representation (s'), reward signal (r), and termination state (done), for example, by accumulating future rewards in a discounted manner. The temporal difference objective can be an estimate representing the expected cumulative reward that can be obtained in subsequent time periods starting from the current state and action. Accumulating future rewards in a discounted manner multiplies distant rewards by a power of the discount factor, and the further away the reward is from the current state, the smaller its impact on the current decision. When the termination state is "termination", the future reward is set to zero, and the temporal difference objective is the current reward; when the termination state is "non-termination", the temporal difference objective is the current reward plus the discounted future reward estimate. In the specific calculation, it is necessary to combine the next state action block output by the target network for value estimation. The temporal difference objective serves as a supervisory objective for updating the value network parameters, guiding the estimated value of the value network to gradually approach the true expected cumulative reward, thereby providing an accurate regression objective for the optimization of the value network, enabling the estimated value of the value network to gradually approach the true expected cumulative reward.

[0139] (2) Determine the value loss. The value loss is determined by calculating the difference between the temporal difference objective and the lightweight adapter's value estimate for the first reinforcement learning state representation and the first absolute action block. The value loss measures the gap between the lightweight adapter's value estimation network's value estimate for the current state-action pair and the temporal difference objective. The value loss is typically expressed as mean squared error, calculated as the squared difference between the predicted value and the target value. By minimizing the value loss, the value estimation network progressively learns to accurately evaluate the value of different state-action pairs, providing a reliable value signal for policy optimization.

[0140] (3) Determine the policy loss. Based on the value estimate of the first reinforcement learning state representation by the lightweight adapter, determine the value-oriented term in the policy loss, and add the first absolute action block as a constraint term to the policy loss to obtain the policy loss. The policy loss consists of two parts: the value-oriented term and the constraint term. The value-oriented term encourages the policy to choose actions that can obtain high value estimates, that is, the actions generated by the lightweight adapter can maximize the output of the value network; the constraint term constrains the actions generated by the lightweight adapter to not deviate from the first absolute action block (i.e., the actions that are actually executed and stored in the experience pool). Its role is to prevent the policy from over-exploring and deviating from the known reliable action distribution in the early stage of training, thereby improving the safety and stability of training. Combining the above two points, the policy loss, while pursuing high-value actions, avoids deviating from the verified reliable actions through the constraint term.

[0141] (4) Update the lightweight adapter parameters. Based on the value loss and policy loss, update the network parameters of the lightweight adapter using backpropagation and gradient descent algorithms. Specifically, the value loss is used to update the parameters of the value estimation network (Critic network) to make its value estimation more accurate; the policy loss is used to update the parameters of the policy network (Actor network) to make its generated actions of higher quality.

[0142] For example, a batch of transition samples (e.g., 256) is sampled from the experience replay pool, and one transition sample (s, a, r, s', done) is selected from it. The system inputs s' into the target network of the lightweight adapter to obtain the target action block a'. Based on the termination state done, the system accumulates future rewards with a discount and determines the temporal difference objective y = r + γ^c·Q_target(s', a') (when done is the termination state, y = r, where γ is the discount factor and C is the action block length). At the same time, the system inputs (s, a) into the value estimation network of the lightweight adapter to obtain Q(s, a), and calculates the value loss as (Q(s, a) - y)². During policy updates, the system inputs s into the policy network to obtain π(s), calculates the value-oriented term as -Q(s, π(s)), the constraint term as ||π(s) - a||², and the policy loss as -Q(s, π(s)) + λ·||π(s) - a||², where λ is the constraint term coefficient (initially 50, decaying exponentially with training). Finally, the lightweight adapter parameters are updated based on the value loss and policy loss. This process is computed in parallel on a batch of samples and accumulated, followed by a gradient update.

[0143] In this embodiment, the lightweight adapter parameters are updated online by sampling transferred samples from the experience replay pool and calculating the loss function. First, batch sampling updates offer higher computational efficiency and more stable convergence compared to single-step updates. The sample diversity provided by the experience replay pool further reduces training variance, making policy learning smoother. Second, the temporal difference objective combines current and future reward estimates, enabling the value network to accurately assess the long-term impact of actions. This provides a reliable value orientation for the action correction of the lightweight adapter, allowing it to correctly assess the impact of the current action on future gains in long-term action block scenarios. Third, the value orientation term in the policy loss drives the lightweight adapter to generate actions that yield higher value, i.e., continuously optimizing the correction strategy to improve task success rate. The constraint term, on the other hand, constrains the lightweight adapter from deviating from the validated reliable action distribution. In the early stages of training when experience data is scarce, this prevents the lightweight adapter from generating unsafe or invalid actions due to overexploration, ensuring training safety and stability in real robot deployment environments. Finally, the entire update process only adjusts the parameters of the lightweight adapter, while the world action model, as the backbone model, remains frozen. This achieves continuous adaptive correction of deployment environment errors while avoiding the high computational overhead and overfitting risk associated with full parameter fine-tuning.

[0144] In summary, by freezing the pre-trained world action model as the backbone model to fully preserve its visual understanding, scene prediction, and long-term planning capabilities learned on large-scale offline data, and by only updating a lightweight adapter with far fewer parameters than the backbone model online, the high computational cost and overfitting risk associated with full-parameter fine-tuning are avoided, significantly reducing the online adaptation cost in real-world robot deployment scenarios. Simultaneously, by using the reference action blocks output by the world action model as coarse-grained action priors, the adapter can directly obtain high-quality action directions and trends without exploring the action space from scratch. Furthermore, latent variable features within the world action model are extracted as components of the reinforcement learning state representation, enabling the adapter to utilize target position, object geometry, and contact area... Fine-grained information such as domains is used for action refinement, overcoming the information loss problem caused by only correcting the final action or over-compressing internal representations in existing technologies. Through a closed-loop mechanism of "execution → feedback acquisition → online update → repeated iteration", the adapter can continuously optimize its own parameters based on execution feedback in the real deployment environment, gradually converging to the optimal correction strategy for the current environment. This effectively suppresses the impact of camera calibration errors, perception biases, actuator precision limitations, and contact dynamic differences on precision operations. Finally, by setting a termination condition, training is automatically stopped and the target adapter is output when the strategy performance meets the target. This balances model performance and training efficiency, and the inference computation of the converged adapter is minimal, so it can be directly used for deployment, reducing the computing power requirements in the production deployment stage.

[0145] This disclosure also provides a robot motion correction device. Figure 6 This is a schematic diagram of the structure of a robot motion correction device according to an embodiment of this disclosure. Figure 6 As shown in the figure, the robot motion correction device of this disclosure embodiment may include: The acquisition module 610 is used to acquire the robot's first multi-view visual observation, first robot body state, and first task instruction; The reference data extraction module 620 is used to input the first multi-view visual observation, the first robot body state and the first task instruction into the pre-trained world action model to obtain reference action blocks and latent variable features, wherein the network parameters of the world action model are in a frozen state. The absolute action block generation module 630 is used to construct a first reinforcement learning state representation based on the first robot body state, the reference action block and the latent variable features, and input the first reinforcement learning state representation into a lightweight adapter so that the lightweight adapter generates a first absolute action block based on the attention mechanism and the first reinforcement learning state representation. The lightweight adapter is an action refinement network with fewer trainable parameters than the world action model, and the first absolute action block is the action instruction to be verified output by the lightweight adapter during the correction process. The parameter update module 640 is used to update the parameters of the lightweight adapter online based on the execution feedback received after the robot executes the first absolute action block, in response to receiving the execution feedback. The execution feedback includes a second multi-view visual observation, a second robot body state, a second task instruction, a task completion status signal, and an intervention signal. The iteration module 650 is used to iteratively execute the above steps until a preset termination condition is met, and output a target lightweight adapter. The target lightweight adapter is the iteratively updated lightweight adapter, which is used to output a second absolute action block to control the robot to complete the operation task. The second absolute action block is the correction action instruction output by the lightweight adapter after correction and convergence.

[0146] As an optional embodiment, the reference data extraction module 620 includes: The first encoding unit is configured to respond to the world action model, in which the first multi-view visual observation, the first robot body state, and the first task instruction are input into the network parameters in a frozen state, by performing feature encoding on the first multi-view visual observation through the visual encoding branch of the world action model to obtain visual encoding features. The modeling unit is used to perform multimodal fusion and temporal modeling of the visual encoded features with the first robot body state and the first task instruction to obtain the fused world state features; The feature extraction unit is used to input the world state features into the action generation branch of the world action model, and extract action generation latent variable features from a preset intermediate layer during the forward propagation of the action generation branch. The decoding unit is used to decode the latent variable features of the action generation through the action decoding branch of the world action model and output a reference action block of preset length.

[0147] As an optional embodiment, the absolute action block generation module 630 includes: The second encoding unit is used to encode the first robot body state, the reference action block and the latent variable features to obtain the first query side token sequence and the first memory token sequence. An absolute action block generation unit is used to input the first query-side token sequence and the first memory token sequence into the lightweight adapter, so that the lightweight adapter executes the attention mechanism based on the first query-side token sequence and the first memory token sequence to generate the first absolute action block.

[0148] As an optional embodiment, the second encoding unit is specifically used for: The state of the first robot body is encoded as a state token, and the first preset number of reference actions in the reference action block are encoded as reference action tokens, wherein the preset number is less than or equal to the length of the reference action block; The status token is concatenated with each of the reference action tokens to form a first query-side token sequence; The latent variable features are encoded into at least one memory token to obtain a first memory token sequence.

[0149] As an optional embodiment, the absolute action block generation unit is specifically used for: Input the first query-side token sequence and the first memory token sequence into the lightweight adapter; The first query-side token sequence is processed by the lightweight adapter to model the structural relationship between the first robot body state and the reference action block, and the first query-side token sequence after self-attention processing is obtained. Through the lightweight adapter, cross-attention processing is performed on the first query-side token sequence and the first memory token sequence after self-attention processing to obtain a query-side representation that integrates the latent variable features; Based on the query-side representation that integrates the latent variable features, the action head outputs the first absolute action block with a length of the preset number.

[0150] As an optional embodiment, the parameter update module 640 includes: A determining unit is configured to, in response to the execution feedback after the robot executes the first absolute action block, determine, based on the execution feedback, a second reinforcement learning state representation, a termination state, and a reward signal after the robot executes the first absolute action block, wherein the termination state is used to indicate whether the robot has completed the task or whether it needs to terminate the current round; The sample storage unit is used to store the first reinforcement learning state representation, the first absolute action block, the reward signal, the second reinforcement learning state representation, and the termination state as a transition sample into the experience replay pool. The parameter update unit is configured to, in response to sampling transfer samples from the experience replay pool, calculate the loss function of the lightweight adapter based on the sampled transfer samples, and update the parameters of the lightweight adapter according to the loss function.

[0151] As an optional embodiment, the determining unit is specifically used for: The second multi-view visual observation, the second robot body state, the second task instruction, the task completion status signal, and the intervention signal are obtained from the execution feedback after the robot executes the first absolute action block; Confirm the type of indication for the task completion status signal in the execution feedback; In response to the indication type indicating that the robot has successfully completed the task, a sparse reward signal of a first value is generated; and in response to the indication type indicating that the robot has not completed the task, a sparse reward signal of a second value is generated. The state type of the termination state is determined based on the task completion status signal and / or intervention signal in the execution feedback; The second multi-view visual observation, the second robot body state, and the second task instruction are encoded to obtain a second query-side token sequence and a second memory token sequence, which are then used to generate the second reinforcement learning state representation.

[0152] As an optional embodiment, the parameter update unit is specifically used for: In response to sampling transfer samples from the experience replay pool, a temporal difference objective is determined based on the second reinforcement learning state, the reward signal, the second reinforcement learning state representation, and the termination state; The value loss of the lightweight adapter is determined based on the difference between the temporal difference objective and the value estimate of the first reinforcement learning state representation and the first absolute action block by the lightweight adapter. Based on the value estimate of the first reinforcement learning state representation by the lightweight adapter, the value-oriented term in the policy loss is determined, and the first absolute action block is added as a constraint term to the policy loss to obtain the policy loss of the lightweight adapter. The parameters of the lightweight adapter are updated based on the value loss and the policy loss.

[0153] In summary, by freezing the pre-trained world action model as the backbone model to fully preserve its visual understanding, scene prediction, and long-term planning capabilities learned on large-scale offline data, and by only updating a lightweight adapter with far fewer parameters than the backbone model online, the high computational cost and overfitting risk associated with full-parameter fine-tuning are avoided, significantly reducing the online adaptation cost in real-world robot deployment scenarios. Simultaneously, by using the reference action blocks output by the world action model as coarse-grained action priors, the adapter can directly obtain high-quality action directions and trends without exploring the action space from scratch. Furthermore, latent variable features within the world action model are extracted as components of the reinforcement learning state representation, enabling the adapter to utilize target position, object geometry, and contact area... Fine-grained information such as domains is used for action refinement, overcoming the information loss problem caused by only correcting the final action or over-compressing internal representations in existing technologies. Through a closed-loop mechanism of "execution → feedback acquisition → online update → repeated iteration", the adapter can continuously optimize its own parameters based on execution feedback in the real deployment environment, gradually converging to the optimal correction strategy for the current environment. This effectively suppresses the impact of camera calibration errors, perception biases, actuator precision limitations, and contact dynamic differences on precision operations. Finally, by setting a termination condition, training is automatically stopped and the target adapter is output when the strategy performance meets the target. This balances model performance and training efficiency, and the inference computation of the converged adapter is minimal, so it can be directly used for deployment, reducing the computing power requirements in the production deployment stage.

[0154] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional units and modules is merely an example. In practical applications, the above functions can be assigned to different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above. The functional units and modules in the embodiments can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit. Furthermore, the specific names of the functional units and modules are only for easy differentiation and are not intended to limit the scope of protection of this disclosure. The specific working process of the units and modules in the above system can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0155] Below, for reference Figure 7 This describes an electronic device according to embodiments of the present disclosure. The electronic device may be either or both of a first device and a second device, or a standalone device independent of them, which may communicate with the first device and the second device to receive acquired input signals from them.

[0156] Figure 7 A block diagram of an electronic device according to an embodiment of the present disclosure is shown.

[0157] like Figure 7 As shown, the electronic device includes one or more processors and memory.

[0158] A processor can be a central processing unit (CPU) or other form of processing unit with data processing and / or instruction execution capabilities, and can control other components in an electronic device to perform desired functions.

[0159] The memory can store one or more computer program products, and the memory can include various forms of computer-readable storage media, such as volatile memory and / or non-volatile memory. The volatile memory may include, for example, random access memory (RAM) and / or cache memory. The non-volatile memory may include, for example, read-only memory (ROM), hard disk, flash memory, etc. One or more computer program products can be stored on the computer-readable storage medium, and the processor can run the computer program products to implement the robot motion correction methods of the various embodiments of this disclosure described above and / or other desired functions.

[0160] In one example, the electronic device may also include input devices and output devices, which are interconnected via a bus system and / or other forms of connection mechanism (not shown).

[0161] In addition, the input device may also include, for example, a keyboard, a mouse, etc.

[0162] This output device can output various information to the outside, including determined distance information, direction information, etc. The output device may include, for example, a display, a speaker, a printer, and a communication network and its connected remote output devices, etc.

[0163] Of course, for the sake of simplicity, Figure 7 Only some of the components of the electronic device relevant to this disclosure are shown, omitting components such as buses, input / output interfaces, etc. In addition, the electronic device may include any other suitable components depending on the specific application.

[0164] In addition to the methods and apparatus described above, embodiments of this disclosure may also be computer program products, including computer program instructions that, when executed by a processor, cause the processor to perform the steps in the robot motion correction methods according to various embodiments of this disclosure as described in the foregoing portion of this specification.

[0165] The computer program product can be written in any combination of one or more programming languages ​​to perform the operations of the embodiments of this disclosure. The programming languages ​​include object-oriented programming languages ​​such as Java and C++, as well as conventional procedural programming languages ​​such as C or similar languages. The program code can be executed entirely on a user's computing device, partially on a user's computing device, as a standalone software package, partially on a user's computing device and partially on a remote computing device, or entirely on a remote computing device or server.

[0166] Furthermore, embodiments of this disclosure may also be computer-readable storage media storing computer program instructions that, when executed by a processor, cause the processor to perform the steps in the robot motion correction methods according to various embodiments of this disclosure as described in the foregoing portion of this specification.

[0167] The computer-readable storage medium may be any combination of one or more readable media. A readable medium may be a readable signal medium or a readable storage medium. A readable storage medium may be, for example, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of readable storage media (a non-exhaustive list) include: an electrical connection having one or more wires, a portable disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination thereof.

[0168] The basic principles of this disclosure have been described above with reference to specific embodiments. However, it should be noted that the advantages, benefits, and effects mentioned in this disclosure are merely examples and not limitations, and should not be considered as essential features of each embodiment of this disclosure. Furthermore, the specific details disclosed above are for illustrative and facilitative purposes only, and are not limitations. These details do not limit the scope of this disclosure to the necessity of employing the aforementioned specific details for implementation.

[0169] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For system embodiments, since they largely correspond to method embodiments, the description is relatively simple; relevant parts can be referred to the descriptions in the method embodiments.

[0170] The block diagrams of devices, apparatuses, devices, and systems disclosed herein are merely illustrative examples and are not intended to require or imply that they must be connected, arranged, or configured in the manner shown in the block diagrams. As those skilled in the art will recognize, these devices, apparatuses, devices, and systems can be connected, arranged, and configured in any manner. Words such as “comprising,” “including,” “having,” etc., are open-ended terms meaning “including but not limited to,” and are used interchangeably with them. The terms “or” and “and” as used herein refer to the terms “and / or,” and are used interchangeably with them unless the context clearly indicates otherwise. The term “such as” as used herein refers to the phrase “such as but not limited to,” and is used interchangeably with it.

[0171] The methods and apparatus of this disclosure may be implemented in many ways. For example, they may be implemented by software, hardware, firmware, or any combination of software, hardware, and firmware. The above-described order of steps for the methods is for illustrative purposes only, and the steps of the methods of this disclosure are not limited to the order specifically described above unless otherwise specifically stated. Furthermore, in some embodiments, this disclosure may also be implemented as a program recorded on a recording medium, the program including machine-readable instructions for implementing the methods according to this disclosure. Thus, this disclosure also covers recording media storing programs for performing the methods according to this disclosure.

[0172] It should also be noted that in the apparatus, devices, and methods of this disclosure, the components or steps can be disassembled and / or recombined. These disassemblies and / or recombinations should be considered as equivalent solutions to this disclosure.

[0173] The above description of the disclosed aspects is provided to enable any person skilled in the art to make or use this disclosure. Various modifications to these aspects will be readily apparent to those skilled in the art, and the general principles defined herein may be applied to other aspects without departing from the scope of this disclosure. Therefore, this disclosure is not intended to be limited to the aspects shown herein, but rather to be carried out within the widest scope consistent with the principles and novel features disclosed herein.

[0174] The above description has been given for purposes of illustration and description. Furthermore, this description is not intended to limit the embodiments of this disclosure to the forms disclosed herein. Although numerous exemplary aspects and embodiments have been discussed above, those skilled in the art will recognize certain variations, modifications, alterations, additions, and sub-combinations therein.

Claims

1. A method for correcting robot motion, characterized in that, The method includes: Acquire the robot's first multi-view visual observation, first robot body state, and first task command; The first multi-view visual observation, the first robot body state, and the first task instruction are input into the pre-trained world action model to obtain reference action blocks and latent variable features, wherein the network parameters of the world action model are in a frozen state. A first reinforcement learning state representation is constructed based on the first robot body state, the reference action block, and the latent variable features. The first reinforcement learning state representation is then input into a lightweight adapter, which generates a first absolute action block based on the attention mechanism and the first reinforcement learning state representation. The lightweight adapter is an action refinement network with fewer trainable parameters than the world action model. The first absolute action block is the action instruction to be verified output by the lightweight adapter during the calibration process. In response to receiving execution feedback after the robot executes the first absolute action block, the parameters of the lightweight adapter are updated online based on the execution feedback, which includes a second multi-view visual observation, a second robot body state, a second task instruction, a task completion status signal, and an intervention signal. The above steps are executed iteratively until a preset termination condition is met, and a target lightweight adapter is output. The target lightweight adapter is the iteratively updated lightweight adapter, which is used to output a second absolute action block to control the robot to complete the operation task. The second absolute action block is the correction action command output by the lightweight adapter after correction and convergence.

2. The method according to claim 1, characterized in that, The first multi-view visual observation, the first robot body state, and the first task instruction are input into the pre-trained world motion model to obtain reference action blocks and latent variable features, including: In response to the world action model with the first multi-view visual observation, the first robot body state, and the first task instruction input network parameters in a frozen state, the first multi-view visual observation is feature-encoded through the visual encoding branch of the world action model to obtain visual encoding features. The visual encoding features are fused with the first robot body state and the first task instruction through multimodal fusion and temporal modeling to obtain the fused world state features; The world state features are input into the action generation branch of the world action model. During the forward propagation of the action generation branch, action generation latent variable features are extracted from the preset intermediate layer. The action generation latent variable features are decoded by the action decoding branch of the world action model, and a reference action block of preset length is output.

3. The method according to claim 1 or 2, characterized in that, The step of constructing a first reinforcement learning state representation based on the first robot body state, the reference action block, and the latent variable features, and inputting the first reinforcement learning state representation into a lightweight adapter, so that the lightweight adapter generates a first absolute action block based on the attention mechanism and the first reinforcement learning state representation, includes: The first robot body state, the reference action block, and the latent variable features are encoded to obtain a first query-side token sequence and a first memory token sequence. The first query-side token sequence and the first memory token sequence are input into the lightweight adapter, so that the lightweight adapter executes the attention mechanism based on the first query-side token sequence and the first memory token sequence to generate the first absolute action block.

4. The method according to claim 3, characterized in that, The encoding process of the first robot body state, the reference action block, and the latent variable features to obtain the first query-side token sequence and the first memory token sequence includes: The state of the first robot body is encoded as a state token, and the first preset number of reference actions in the reference action block are encoded as reference action tokens, wherein the preset number is less than or equal to the length of the reference action block; The status token is concatenated with each of the reference action tokens to form a first query-side token sequence; The latent variable features are encoded into at least one memory token to obtain a first memory token sequence.

5. The method according to claim 3, characterized in that, The step of inputting the first query-side token sequence and the first memory token sequence into the lightweight adapter, so that the lightweight adapter executes the attention mechanism based on the first query-side token sequence and the first memory token sequence to generate the first absolute action block, includes: Input the first query-side token sequence and the first memory token sequence into the lightweight adapter; The first query-side token sequence is processed by the lightweight adapter to model the structural relationship between the first robot body state and the reference action block, and the first query-side token sequence after self-attention processing is obtained. Through the lightweight adapter, cross-attention processing is performed on the first query-side token sequence and the first memory token sequence after self-attention processing to obtain a query-side representation that integrates the latent variable features; Based on the query-side representation that integrates the latent variable features, the action head outputs the first absolute action block with a length of the preset number.

6. The method according to claim 1 or 2, characterized in that, The step of responding to the execution feedback received after the robot executes the first absolute action block, and updating the parameters of the lightweight adapter online based on the execution feedback, includes: In response to the execution feedback after the robot executes the first absolute action block, a second reinforcement learning state representation, a termination state, and a reward signal after the robot executes the first absolute action block are determined based on the execution feedback, wherein the termination state is used to indicate whether the robot has completed the task or whether the current round needs to be terminated; The first reinforcement learning state representation, the first absolute action block, the reward signal, the second reinforcement learning state representation, and the termination state are stored as a transition sample in the experience replay pool. In response to sampling transfer samples from the experience replay pool, a loss function of the lightweight adapter is calculated based on the sampled transfer samples, and the parameters of the lightweight adapter are updated according to the loss function.

7. The method according to claim 6, characterized in that, The step of responding to the execution feedback after the robot executes the first absolute action block, and determining the second reinforcement learning state representation, termination state, and reward signal after the robot executes the first absolute action block based on the execution feedback, includes: The second multi-view visual observation, the second robot body state, the second task instruction, the task completion status signal, and the intervention signal are obtained from the execution feedback after the robot executes the first absolute action block; Confirm the type of indication for the task completion status signal in the execution feedback; In response to the indication type indicating that the robot has successfully completed the task, a sparse reward signal of a first value is generated; and in response to the indication type indicating that the robot has not completed the task, a sparse reward signal of a second value is generated. The state type of the termination state is determined based on the task completion status signal and / or intervention signal in the execution feedback; The second multi-view visual observation, the second robot body state, and the second task instruction are encoded to obtain a second query-side token sequence and a second memory token sequence, which are then used to generate the second reinforcement learning state representation.

8. The method according to claim 6, characterized in that, The response to sampling transfer samples from the experience replay pool, calculating the loss function of the lightweight adapter based on the sampled transfer samples, and updating the parameters of the lightweight adapter according to the loss function includes: In response to sampling transfer samples from the experience replay pool, a temporal difference objective is determined based on the second reinforcement learning state, the reward signal, the second reinforcement learning state representation, and the termination state; The value loss of the lightweight adapter is determined based on the difference between the temporal difference objective and the value estimate of the first reinforcement learning state representation and the first absolute action block by the lightweight adapter. Based on the value estimate of the first reinforcement learning state representation by the lightweight adapter, the value-oriented term in the policy loss is determined, and the first absolute action block is added as a constraint term to the policy loss to obtain the policy loss of the lightweight adapter. The parameters of the lightweight adapter are updated based on the value loss and the policy loss.

9. A robot motion correction device, characterized in that, The device includes: The acquisition module is used to acquire the robot's first multi-view visual observation, first robot body state, and first task instruction; The reference data extraction module is used to input the first multi-view visual observation, the first robot body state and the first task instruction into the pre-trained world action model to obtain reference action blocks and latent variable features, wherein the network parameters of the world action model are in a frozen state. An absolute action block generation module is used to construct a first reinforcement learning state representation based on the first robot body state, the reference action block, and the latent variable features, and input the first reinforcement learning state representation into a lightweight adapter so that the lightweight adapter generates a first absolute action block based on the attention mechanism and the first reinforcement learning state representation. The lightweight adapter is an action refinement network with fewer trainable parameters than the world action model, and the first absolute action block is the action instruction to be verified output by the lightweight adapter during the correction process. The parameter update module is used to update the parameters of the lightweight adapter online based on the execution feedback received after the robot executes the first absolute action block. The execution feedback includes a second multi-view visual observation, a second robot body state, a second task instruction, a task completion status signal, and an intervention signal. The iteration module is used to iteratively execute the above steps until a preset termination condition is met, and output a target lightweight adapter. The target lightweight adapter is the iteratively updated lightweight adapter, which is used to output a second absolute action block to control the robot to complete the operation task. The second absolute action block is the correction action instruction output by the lightweight adapter after correction and convergence.

10. An electronic device, characterized in that, include: Memory, used to store computer program products; A processor for executing a computer program product stored in the memory, wherein when the computer program product is executed, it implements the method described in any one of claims 1-8.

11. A computer-readable storage medium storing computer program instructions thereon, characterized in that, When the computer program instructions are executed by the processor, they implement the method described in any one of claims 1-8.

12. A computer program product comprising computer program instructions, characterized in that, When the computer program instructions are executed by the processor, they implement the method described in any one of claims 1-8.

Images