Causal anchor point-based vla model control optimization method and device, and electronic equipment

Abstract

The application discloses a VLA model control optimization method and device based on a causal anchor point, and electronic equipment, comprising: obtaining historical interaction data of a VLA model performing a target task, identifying a causal anchor point and an ideal state thereof which have a key influence on the success of the task from a task execution track; detecting the deviation between the current state and the ideal state of the target causal anchor point during the task execution process, and constructing the corresponding state difference; performing differential semantic coding and hierarchical vector quantization on the state difference to obtain a multi-layer discrete code word representing the recovery semantics required for the current state to recover to the target causal anchor point; generating a correction action and a local execution strategy based on the multi-layer discrete code word, and recovering to the original long-time sequence strategy after completing the anchor point correction, thereby reducing error accumulation in a long-time sequence task and improving the success rate of task execution.

Description

Technical Field

[0001] This application belongs to the field of artificial intelligence and natural language processing technology, and in particular relates to a VLA model control optimization method, device, and electronic device based on causal anchors. Background Technology

[0002] In recent years, with the development of embodied intelligence technology, the Vision-Language-Action Model (VLA) has gradually become the mainstream solution for robots to complete complex tasks. This type of model takes multimodal information such as visual observation, natural language commands, and proprioception as input, and predicts and outputs robot actions step by step in an autoregressive manner, thereby completing long-sequential tasks composed of multiple sequential sub-stages. In industrial scenarios, such as logistics sorting and packaging workstations, robotic arms need to sequentially complete steps such as identification, approach, grasping, handling, alignment, and placement according to packaging instructions, which places high demands on the stable execution of long-sequential tasks.

[0003] However, VLA models generally face the problem of error accumulation when performing long-term tasks. Because actions are generated progressively in an autoregressive manner, small deviations in each prediction step propagate as input to subsequent predictions and are gradually amplified, ultimately leading to inaccurate end-effector pose, grasping, or deployment failures. There is currently no fundamental solution to this problem within existing frameworks.

[0004] To suppress deviations during execution, existing methods combat deviations by increasing the replanning frequency or repeatedly re-predicting the entire trajectory during execution. However, such methods require repeatedly running large models for state prediction over long time series, significantly increasing computational overhead and real-time burden. Furthermore, the re-predicted action sequences are also generated in an autoregressive manner, which will still produce new deviations, making it difficult to fundamentally eliminate error accumulation. Moreover, this type of encoding targets absolute actions or absolute states, with each step independent of the others. Therefore, the resulting discrete representation is difficult to directly serve as a directional deviation correction constraint.

[0005] Therefore, how to perform directional and convergent local corrections of deviations around a few decisive key intermediate states without replanning the entire long-term time-series path, thereby suppressing error accumulation in the execution of long-term time-series tasks while ensuring real-time performance, has become a technical problem that urgently needs to be solved in this field. Summary of the Invention

[0006] To address the problems existing in the prior art, this application provides a VLA model control optimization method, apparatus, and electronic device based on causal anchors. By acquiring intermediate motion states and then performing masking intervention to obtain a set of causal anchors, deviation trigger correction is further calculated to obtain multimodal conditional features. Finally, the correction action is output through decoding, thereby achieving directional and convergent local correction around a few decisive key states without replanning the entire long time-series path. This ensures real-time performance, reduces state prediction overhead, and suppresses error accumulation in long-series task execution from the source.

[0007] According to a first aspect of the embodiments of this application, a VLA model control optimization based on causal anchor points is provided, including: Obtain historical interaction data of the VLA model executing the target task, extract the task execution trajectory from it, and use the intermediate states in the trajectory as candidate intermediate states; The candidate intermediate states are masked based on the masked autoencoder, and the decrease in the success rate of the target task before and after masking is used as the causal score. Candidate intermediate states with causal scores higher than a preset threshold are determined as causal anchors, and a set of causal anchors including causal anchors and their ideal states is generated. When the VLA model performs the target task, the target causal anchor point is determined based on the set of anchor points, and the state difference between the current actual state and the ideal state corresponding to the target causal anchor point is calculated. The state difference is used to obtain a continuous representation of the object to be encoded by an encoder, and the continuous representation is then subjected to hierarchical discretization encoding to obtain a multi-layer discrete codeword that characterizes the deviation direction and degree of the difference between the current actual state and the ideal state. The multilayer discrete codeword is input into the intervention decoder, decoded to obtain the correction action, and the original action output by the VLA model is compensated based on the correction action to obtain the optimized control action; When the correction action is triggered, the entire long-sequence path of the target task is not replanned. Instead, a local execution strategy is regenerated only for the optimized control action. After the correction action is completed, the original long-sequence strategy of the target task is resumed to suppress error accumulation.

[0008] According to a second aspect of the application, a VLA model control optimization device based on causal anchor points is provided, comprising: The extraction module is used to acquire historical interaction data of the VLA model executing the target task, extract the task execution trajectory from it, and use the intermediate states in the trajectory as candidate intermediate states. The causal anchor discovery module is used to perform masking intervention on the candidate intermediate states based on the mask autoencoder, and use the decrease in the success rate of the target task before and after masking as the causal score. Candidate intermediate states with causal scores higher than a preset threshold are identified as causal anchors, and a set of causal anchors including causal anchors and their ideal states is generated. The deviation detection module is used to determine the target causal anchor point currently aligned based on the anchor point set when the VLA model is performing the target task, and to calculate the state difference between the current actual state and the ideal state corresponding to the target causal anchor point. The differential coding module is used to obtain a continuous representation of the object to be encoded by the state difference through an encoder, and to perform hierarchical discretization coding on the continuous representation to obtain a multi-layer discrete codeword that represents the deviation direction and degree of deviation between the current actual state and the ideal state. The action decoding module is used to input the multi-layer discrete codeword into the intervention decoder, decode it to obtain the correction action, and compensate the original action output by the VLA model based on the correction action to obtain the optimized control action; The local replanning module is used to replan the entire long-sequence path of the target task when issuing the correction action, but only regenerates the local execution strategy for the optimized control action. After the correction action is completed, the original long-sequence strategy of the target task is restored to suppress error accumulation.

[0009] According to a third aspect of the embodiments of this application, a computer-readable storage medium is provided that stores computer instructions thereon, which, when executed by a processor, implement the steps of the method as described in the first aspect.

[0010] According to a fourth aspect of the embodiments of this application, a computer-readable storage medium is provided that stores computer instructions thereon, which, when executed by a processor, implement the steps of the method as described in the first aspect.

[0011] Compared with the prior art, the embodiments of the present invention have at least the following beneficial effects: As can be seen from the above technical solutions, this invention utilizes counterfactual masking intervention based on a masked autoencoder, using the decrease in task success rate before and after masking as the causal score. It identifies key intermediate states that play a decisive role in task success from historical interaction data as causal anchor points, compressing the lengthy execution trajectory into a sparse causal skeleton. Secondly, during execution, deviations are detected only around the causal anchor points. The action-induced state differences, rather than absolute states, are fused with multimodal conditions and encoded into multi-layer discrete codewords via hierarchical vector quantization. These codewords serve both as directional differential semantic constraints guiding local replanning only for the anchor point and as convergence criteria for whether the deviation returns to zero. Finally, this invention aligns the correction actions to the ideal actions under uninterrupted model conditions to suppress correction drift. This achieves directional and convergent local correction around a few decisive key states without replanning the entire long time-series path, thereby ensuring real-time performance, reducing state prediction overhead, and suppressing error accumulation in long-series task execution from the source. Attached Figure Description

[0012] Figure 1 This is a flowchart illustrating a VLA model control optimization method based on causal anchors, according to an exemplary embodiment.

[0013] Figure 2 This is a block diagram illustrating a VLA model control optimization device based on causal anchors, according to an exemplary embodiment.

[0014] Figure 3 This is a block diagram illustrating an electronic device according to an exemplary embodiment. Detailed Implementation

[0015] The specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0016] In this example, an n-DOF industrial robotic arm used in a logistics sorting and packaging workstation is taken as an example. The robotic arm is equipped with a two-finger adaptive gripper at its end, an RGB-D camera on its wrist, and a global camera above the workstation. A conveyor belt continuously feeds in different types of items to be packaged. It should be noted that this invention is not limited to the above example of an industrial robotic arm; any intelligent mobile robot or VLA logistics robot equipped with a VLA model can be processed and implemented using this invention.

[0017] Figure 1 This is a flowchart illustrating a VLA model control optimization method based on causal anchors, according to an exemplary embodiment. The VLA model control optimization method based on causal anchors provided by this invention may include the following sub-steps: S1: Obtain historical interaction data of the VLA model executing the target task, extract the task execution trajectory from it, and use the intermediate states in the trajectory as candidate intermediate states; this step includes the following sub-steps: S11: Extract task information from the historical interaction data of the VLA model executing the target task to obtain the task execution trajectory represented as an alternating sequence of states and actions; Specifically, during the deployment and commissioning phase, data is collected from industrial robotic arms using a policy network. The historical interactive dataset for the disassembly task T is executed on this workstation, and observations are recorded at each time step t. ,state Actions performed and the task-level result label for this assignment .

[0018] Organize the interaction sequence corresponding to a complete packaging operation into a task execution trajectory according to time sequence:

[0019] in This represents the initial state when the item is in place. Indicates the initial action to be performed. This represents the termination status upon completion of deployment or abnormal termination. A historical trajectory set is obtained from the M collected operations. : .

[0020] S12: Take each state in the task execution trajectory as a candidate intermediate state; Specifically, candidate intermediate states are found by using points of significant state change as criteria:

[0021] in As a criterion for judging significant changes, the value is set based on the specific robotic arm configuration; in this case, it is set to 0.1.

[0022] Obtain candidate intermediate states :

[0023] In this way, this embodiment can filter out candidate intermediate states related to the execution process of the target task from the historical interaction trajectory, providing a data foundation for subsequent causal anchor point identification and avoiding relying solely on single-frame states or human experience to determine key task nodes.

[0024] S2: Mask the candidate intermediate states based on a masked autoencoder, and use the decrease in the success rate of the target task before and after masking as the causal score. Determine the candidate intermediate states with causal scores higher than a preset threshold as causal anchor points, generating a set of causal anchor points including their ideal states. This step includes the following sub-steps: S21: Randomly mask and reconstruct the intermediate states in the historical interaction data to reconstruct the error-supervised scoring module based on the mask autoencoder for representation learning, and obtain the trained scoring module. Specifically, state samples from historical interaction data are input into the Masked AutoEncoder (MAE) for training.

[0025] For candidate intermediate states any candidate state within :

[0026] in, Represents visual images, Indicates the pose of the robotic arm's end effector. Indicates the gripper status. This indicates the cargo container's position or the cargo container's inspection result. Represents the characteristics of language task instructions. It represents a historical sequence of actions.

[0027] First, a portion of the visual regions and state features in the randomly occluded state are obtained to obtain the masked state features. :

[0028] in, ⊙ represents a random mask, and ⊙ represents an element-by-element masking operation.

[0029] The masked state features The latent representation is obtained after inputting encoder E. :

[0030] The original state is then reconstructed using decoder D. :

[0031] Using reconstruction error As a training objective:

[0032] After training, the scoring module F(⋅) is obtained.

[0033] S22: Use the trained scoring module to mask each candidate intermediate state, and input the masked candidate intermediate state into the policy network to predict subsequent actions. Use the normalized value of the decrease in the success rate of the target task before and after masking as the causal score of the candidate intermediate state. Specifically, for each candidate state in the set of candidate intermediate states A masking intervention experiment was conducted. First, the original trajectory was used to execute the target task, and the success rate of the original task was obtained. .

[0034] Then the candidate intermediate states Remove or replace the mask state from the trajectory to obtain the mask trajectory. :

[0035] The masked trajectory is input into the policy network corresponding to the VLA model. By performing replay reasoning and statistical analysis of task completion, the success rate of the masked trajectory after intervention masking is obtained. :

[0036] Define candidate intermediate states Causal score for:

[0037] Among them, to prevent An error occurred when the value was zero; add [something]. It is a very small constant.

[0038] When masking a certain state leads to a significant decrease in the task success rate, it indicates that the state has a high impact on the task success.

[0039] S23: Collect the successful trajectory and the failed trajectory of the target task, compare the state distribution differences of the two types of trajectories at each time step, and take the state inflection point in the successful trajectory that meets the preset condition as the candidate state of the causal anchor point. Specifically, further collection of a large number of successful trajectories and failure trajectory The state distribution at each time step is statistically analyzed.

[0040] For the t-th time step, calculate the success rate of the successful trajectory. With failure trajectory State distribution distance :

[0041] in Let KL divergence be denoted by . This is expressed as the state distribution distance threshold, when the following conditions are met. At this point, a critical turning point in the task is considered to have occurred. Specifically, significant differences in state distribution will be observed when the robotic arm successfully grabs a box, the box completely detaches from the shelf, or the box accurately enters the target pallet during the logistics handling process.

[0042] From candidate intermediate state Find candidate states that meet the requirements for causal anchor points. The corresponding state is denoted as the candidate state set of the causal anchor point. , represented as: .

[0043] S24: Combine the causal score with the candidate states of the causal anchor points, and determine the states where the causal score is higher than a preset threshold as causal anchor points, to obtain a set of causal anchor points that includes each causal anchor point and its ideal state.

[0044] Specifically, for the candidate state of the causal anchor point If the causal score in its state is satisfied Then it is determined to be a causal anchor point. :

[0045] in, The preset causal score threshold is used.

[0046] Furthermore, from the trajectory of success Extract the state sample set corresponding to the causal anchor point :

[0047] Calculate the sample mean as the ideal state of this causal anchor point. :

[0048] Finally, the set of causal anchor points is obtained. :

[0049] in Indicates causal anchor point, This represents the corresponding ideal state.

[0050] In this case, the industrial robotic arm ultimately obtains the following in the logistics cargo handling task: A = {(Complete gripping, ideal gripping action), (Remove from shelf, ideal removal from shelf action), (Reach the target location, ideal arrival at the target location action)} three causal anchor points and their corresponding ideal states. The set of causal anchor points will serve as the basis for generating the real-time deviation detection and local correction strategy in the subsequent step S3.

[0051] In this way, this embodiment can evaluate the causal impact of candidate intermediate states based on mask intervention, and identify the intermediate states that truly affect the success or failure of the task as causal anchor points, thereby improving the accuracy and interpretability of key state identification.

[0052] S3: When the VLA model performs the target task, the target causal anchor point currently aligned is determined based on the anchor point set, and the state difference between the current actual state and the ideal state corresponding to the target causal anchor point is calculated; this step includes the following sub-steps: S31: Based on the set of causal anchor points and the current execution progress of the target task, determine the target causal anchor point currently being aligned, and obtain the ideal state corresponding to the target causal anchor point; Specifically, during the execution of the target task by the industrial robotic arm, the current status is acquired in real time. :

[0053] in, Represents the current visual image. Indicates the current pose of the robotic arm's end effector. Indicates the current opening and closing state of the grippers. This indicates the current position or orientation of the cargo container or the cargo container inspection result. Indicates the current language task instruction. This indicates the current historical action sequence.

[0054] The industrial robotic arm determines the target causal anchor point from the set of causal anchor points based on the current task progress. The task progress can be determined by the following information: the current action stage, whether the gripper is close to the box, whether the gripper is closed, whether the box has left the shelf, whether the box is close to the target sorting station, and the representational distance between the current state and the ideal state of each causal anchor point.

[0055] Input the current status into the task progress recognition module R to obtain the progress of the current task stage. :

[0056] In this case, when the end effector of the industrial robotic arm has entered the preset gripping area of ​​the cargo box and the gripper is before closing, the task progress recognition module outputs the current task stage progress. This is the "preparation phase for capture". At this point, the causal anchor point is... The target causal anchor point is determined as the current target.

[0057] If the grippers are closed and the cargo box detection result shows that the cargo box has been lifted, but the cargo box is still near the shelf, the task progress recognition module outputs the current task stage progress. This is the "out-of-shelf stage." At this point, the causal anchor point is... The target causal anchor point is determined as the current target.

[0058] When the container has been stably clamped and moved to the vicinity of the target sorting station, the task progress recognition module outputs the current task stage progress. This is the "placement alignment phase". At this point, the causal anchor point is... Set as the current target causal anchor point.

[0059] The causal anchor point set is evaluated using a state filtering function. and current task phase progress By filtering, the current target causal anchor point is obtained. :

[0060] in This represents the set of causal anchor points adapted to the current task phase. The state filtering function is used to filter out the causal anchor points of the current task stage.

[0061] S32: Real-time detection of the deviation between the current actual state and the ideal state, and determination of whether the deviation corresponds to the state offset of the target causal anchor point; Specifically, determine the target causal anchor point. Then, its corresponding ideal state can be obtained.

[0062] Current actual state and ideal state Input the shared state encoder to obtain the current actual state representation. and ideal state representation :

[0063]

[0064] in, This indicates a shared state encoder.

[0065] Further calculate the deviation between the current actual state and the ideal state. :

[0066] in, The function is represented as Euclidean distance.

[0067] In logistics cargo handling scenarios, deviation Further including visual deviation, end-effector pose deviation, cargo box pose deviation, and gripper status deviation:

[0068] in: This represents the distance between the current visual feature and the ideal visual feature, i.e., the visual deviation; This represents the distance between the current pose of the robotic arm's end effector and the ideal end effector pose, i.e., the end effector pose deviation. This indicates the distance between the current pose of the cargo box and the ideal pose of the cargo box, i.e., the pose deviation of the cargo box; This represents the distance between the current gripper state and the ideal gripper state, i.e., the gripper state deviation. These are preset weighting coefficients.

[0069] S33: If the current actual state is formed by the ideal state offset of the target causal anchor point, then calculate the state difference between the actual state induced by the current action and the ideal state; if the deviation does not correspond to any causal anchor point in the set of causal anchor points, continue to rely on the action generation capability of the VLA model itself to perform the task. Specifically, when the deviation satisfies When the current state deviates from the ideal state of the target causal anchor point, a correction is triggered. Specifically, Target causal anchor The corresponding anchor point deviation threshold.

[0070] Further calculation of state difference :

[0071] in, The function represents the state difference calculation function.

[0072] In this embodiment, the state difference can be expressed as:

[0073] in Semantic differences in the shared state representation space:

[0074] This represents the pose difference at the end effector of the robotic arm;

[0075] Indicates the cargo container position difference;

[0076] Indicates the difference in gripper state:

[0077] When state difference Once calculated, the system generates a correction trigger signal. :

[0078] and state difference Current target causal anchor Ideal state Current actual status Send to subsequent step S4.

[0079] If satisfied If the current state is still within the allowable deviation range of the target causal anchor point, no correction is triggered, and the system continues to execute the original time series strategy generated by the VLA model.

[0080] In this embodiment, not all state deviations trigger correction. Correction is only triggered when the current task stage matches a causal anchor in the causal anchor set, and the deviation between the current actual state and the ideal state of that causal anchor exceeds the corresponding threshold.

[0081] After the above processing, the following trigger information set U is output:

[0082] in, This indicates the causal anchor point of the currently aligning target. This represents the ideal state of the target causal anchor point. This indicates the current actual status of the industrial robotic arm. This represents the state difference between the current actual state and the ideal state. Indicates whether correction is triggered.

[0083] when Then, proceed to step S4 to encode the state difference and extract multimodal conditional features.

[0084] when If necessary, continue executing the original long-sequence strategy for the target task.

[0085] In this way, this embodiment can trigger corrections only for state offsets related to causal anchors during the execution of the target task by the VLA model, avoiding excessive intervention in ordinary state fluctuations, thereby improving the targeting and stability of error correction triggering.

[0086] S4: Using the state difference as the object to be encoded, a continuous representation is obtained through an encoder, and the continuous representation is then subjected to hierarchical discretization encoding to obtain a multi-layer discrete codeword representing the deviation direction and degree of the difference between the current actual state and the ideal state; this step includes the following sub-steps: S41: Using the state difference as the object to be encoded, a continuous representation of the state difference is obtained through an encoder; Specifically, after triggering, the state differential input is sent to the differential encoder. This yields a continuous representation corresponding to the state difference. : 。

[0087] S42: Extract multimodal conditional features that characterize the difference between the actual state and the ideal state; Specifically, firstly, a visual encoder is used. Extracting visual conditional features :

[0088] Language encoder Extracting linguistic conditional features :

[0089] Anchor encoder Extracting anchor point condition features :

[0090] Motion encoder Extracting action history features :

[0091] Finally passed The connection function yields multimodal conditional features. : .

[0092] S43: The multimodal conditional features are fused with the continuous representation, and then hierarchical vector quantization is performed. Each quantization layer sequentially quantizes the quantization residual of the previous quantization layer, and the discrete codewords output by each quantization layer together constitute a multi-layer discrete codeword. Specifically, continuous difference representation Multimodal conditional features The fusion representation is obtained by using the fusion function Fuse. :

[0093] Fusion representation This is represented as the recovery semantics required to recover from the current state to the target causal anchor point under the current task objective. To obtain a reusable recovery pattern, the conditional difference representation is... Perform hierarchical residual vector quantization. Let the first... The layer codebook is as follows:

[0094] in, Indicates the first The first in the layer codebook Each code character Indicates the first The number of codewords in the layer codebook.

[0095] First, search for the corresponding code in the first-level codebook. The nearest codeword index :

[0096] And obtain the first layer of discrete codewords :

[0097] First-level codeword representation This represents the large-granularity category of the recovery task. In this case, it is represented as: adjusting the gripping position, restabilizing the clamp, correcting the transport attitude, and correcting the placement position.

[0098] Calculate the first-level quantization residual :

[0099] Continue with the second layer of quantization to obtain the second layer of codewords. :

[0100]

[0101] Second layer of code This is used to describe a finer-grained recovery strategy. In this case, it is represented as: leftward movement, rightward movement, clockwise rotation, and increasing clamping force.

[0102] Continue calculating the second-level quantization residual :

[0103] And so on, for the first Layer quantization, satisfying:

[0104] Obtain the i-th layer codeword and the quantization residual of the i-th layer :

[0105]

[0106] Ultimately, the discrete codewords output from each quantization layer together constitute a multi-layer discrete codeword. :

[0107] in, This indicates the number of layers in the layer vector quantization.

[0108] In this way, the continuous state difference can be converted into a discrete recovery codeword with a hierarchical semantic structure, so that the system can not only represent the degree of deviation of the current actual state from the ideal state, but also represent the recovery direction and recovery strategy in a reusable discrete intervention mode, thereby providing a structured input for the generation of subsequent correction actions.

[0109] S5: Input the multilayer discrete codeword into the intervention decoder, decode to obtain the correction action, and compensate the original action output by the VLA model based on the correction action to obtain the optimized control action; this step includes the following sub-steps: S51: Input the multilayer discrete codeword into the decoder, decode to obtain the features representing the current global state after the intervention is applied, and decode accordingly to obtain the correction action; Specifically, multi-layer discrete codewords Current actual status Ideal state of target causal anchor point and the language task instruction L input decoder The next global state features predicted after the corrective intervention are obtained. :

[0110] The predicted global state features are used to represent the state that the industrial robotic arm may reach after performing a correction action. In this case, they include the cargo box pose, gripper state, robotic arm end-effector pose, and the spatial relationship between the cargo box and the target area.

[0111] Predict global state features Input correction action generation module , to obtain the correction action : .

[0112] S52: The anchor point alignment error is corrected between the correction action and the ideal action output by the VLA model for the ideal state without intervention. The correction action is corrected with the goal of making the norm of the difference between the correction action and the ideal action approach zero, so as to obtain the optimized control action and drive the industrial robotic arm to execute. Specifically, in order to avoid the correction action merely approximating a certain historical action in the action space without truly eliminating the state deviation, this embodiment uses state alignment as the correction action correction target.

[0113] Using state prediction models Predict the state after the corrective action is performed. :

[0114] Predicted state Ideal state of causal anchor point with target Input shared state encoder The predicted state characterization is obtained. Ideal state representation of the causal anchor point of the target :

[0115]

[0116] Calculate state alignment error :

[0117] The correction action was then modified to reduce the state alignment error, resulting in an optimized control action. :

[0118] In this way, this embodiment can generate correction actions oriented towards target causal anchor points based on multi-layer discrete codewords, so that the current state of the industrial robot arm can be restored to the ideal state, thereby reducing the action execution error caused by the deviation of key intermediate states.

[0119] S6: When the correction action is triggered, the entire long-sequence path of the target task is not replanned. Instead, a local execution strategy is regenerated only for the optimized control action. After the correction is completed, the original long-sequence strategy of the target task is resumed to suppress error accumulation. This step includes the following sub-steps: S61: When the correction is triggered, the entire long time-series path of the target task is not replanned. Instead, the multi-layer discrete codeword is reused as a differential semantic constraint from the actual state to the target causal anchor point to regenerate the local execution strategy and drive the industrial robotic arm to execute. Specifically, when an industrial robotic arm performs a target task, the VLA model has already generated or implicitly maintained the original long-term time-series strategy based on the task instructions. :

[0120] in, This represents the j-th policy fragment in the target task. This represents the total number of strategy fragments.

[0121] When step S3 triggers the correction, the system does not replan all the above strategy segments, but instead determines the current target causal anchor point. The corresponding local policy fragment, while keeping other policies unchanged.

[0122] The multi-layer discrete codewords obtained in step S4:

[0123] Differential semantic constraints serve as the causal anchor point between the current actual state and the ideal state.

[0124] The local execution strategy generation module generates strategies based on the current actual state. Target causal anchor Ideal state Language task instructions and multi-layer discrete codewords Generate local execution strategy :

[0125] The local execution strategy is used to guide the industrial robotic arm from its current actual state to the ideal state of the target causal anchor point, rather than regenerating the complete task path.

[0126] S62: Taking the state after executing the local execution strategy as input, re-execute the encoding and hierarchical vector quantization in step S4 on the state to obtain the re-encoded multi-layer discrete codeword; Specifically, the industrial robotic arm executes the local execution strategy to obtain the state for the next time period. :

[0127] When implementing the local execution strategy, a rolling time-domain control method is adopted. That is, after each one or more correction actions are executed, the current visual image, the state of the industrial robot body, and the state of the cargo box are reacquired, and the deviation between the current state and the ideal state of the target causal anchor point is recalculated.

[0128] Specifically, in the first In the next local correction cycle, the industrial robotic arm performs... To obtain a new actual state Then the alignment error was recalculated. :

[0129] After executing the partial execution strategy, the system recalculates the state difference between the current actual state and the ideal state of the target causal anchor point:

[0130] Then, the differential coding and hierarchical vector quantization in step S4 are re-executed to obtain the re-coded multi-layer discrete codewords. :

[0131] in, Indicates the first Layer code words, This indicates the number of layers in the layer vector quantization.

[0132] S63: When the multi-layer discrete codeword obtained in step S62 converges to an empty codeword representing zero deviation, it is determined that the correction is complete, the regeneration of the local execution strategy is stopped, and the original long-sequence strategy for executing the target task is restored.

[0133] Using similarity function A preset threshold for the similarity of empty codewords can be set. The result is 0.8, and the newly obtained multi-layer discrete codeword satisfies:

[0134] When the current deviation has converged to zero or near-zero deviation, it is determined that the current deviation has converged to zero or near-zero deviation.

[0135] Once the correction is determined to be complete, the regeneration of the local execution strategy is stopped, and the original long-sequence strategy for executing the target task is restored. Specifically, in a task with an anchor length of M, the target causal anchor is determined. Location index in the original long-series strategy And continue execution from the policy segment following that target causal anchor point:

[0136] In this way, this embodiment can regenerate the local execution strategy only around the target causal anchor point without replanning the entire long-term task path, and restore the original long-term strategy after correction, thereby reducing computational overhead, suppressing the accumulation of long-term errors, and improving the continuity of task execution.

[0137] As can be seen from the above technical solutions, this invention utilizes counterfactual masking intervention based on a masked autoencoder, using the decrease in task success rate before and after masking as the causal score. It identifies key intermediate states that play a decisive role in task success from historical interaction data as causal anchor points, compressing the lengthy execution trajectory into a sparse causal skeleton. Secondly, during execution, deviations are detected only around the causal anchor points. The action-induced state differences, rather than absolute states, are fused with multimodal conditions and encoded into multi-layer discrete codewords via hierarchical vector quantization. These codewords serve both as directional differential semantic constraints guiding local replanning only for the anchor point and as convergence criteria for whether the deviation returns to zero. Finally, this invention aligns the correction actions to the ideal actions under uninterrupted model conditions to suppress correction drift. This achieves directional and convergent local correction around a few decisive key states without replanning the entire long time-series path, thereby ensuring real-time performance, reducing state prediction overhead, and suppressing error accumulation in long-series task execution from the source.

[0138] Corresponding to the aforementioned embodiment of a VLA model control optimization method based on causal anchors, this application also provides an embodiment of a VLA model control optimization device based on causal anchors.

[0139] Figure 2 This is a block diagram illustrating a VLA model control optimization device based on causal anchors according to an exemplary embodiment. The device may include: Extraction module 1 is used to obtain historical interaction data of the VLA model executing the target task, extract the task execution trajectory from it, and take the intermediate state in the trajectory as the candidate intermediate state; The causal anchor discovery module 2 is used to perform masking intervention on the candidate intermediate states based on the mask autoencoder, and use the decrease in the success rate of the target task before and after masking as the causal score. Candidate intermediate states with causal scores higher than a preset threshold are identified as causal anchors, and a set of causal anchors including causal anchors and their ideal states is generated. The deviation detection module 3 is used to determine the target causal anchor point currently aligned based on the anchor point set when the VLA model performs the target task, and to calculate the state difference between the current actual state and the ideal state corresponding to the target causal anchor point. The differential coding module 4 is used to obtain a continuous representation of the object to be encoded by the state difference through an encoder, and to perform hierarchical discretization coding on the continuous representation to obtain a multi-layer discrete codeword that represents the deviation direction and degree of deviation between the current actual state and the ideal state. The action decoding module 5 is used to input the multilayer discrete codeword into the intervention decoder, decode it to obtain the correction action, and compensate the original action output by the VLA model based on the correction action to obtain the optimized control action. The local replanning module 6 is used to replan the entire long-sequence path of the target task when issuing the correction action, but only regenerate the local execution strategy for the optimized control action. After the correction is completed, the original long-sequence strategy of the target task is restored to suppress error accumulation.

[0140] Regarding the apparatus in the above embodiments, the specific manner in which each module performs its operation has been described in detail in the embodiments related to the method, and will not be elaborated upon here.

[0141] For the device embodiments, since they basically correspond to the method embodiments, the relevant parts can be referred to in the description of the method embodiments. The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this application according to actual needs. Those skilled in the art can understand and implement this without creative effort.

[0142] Accordingly, this application also provides an electronic device, comprising: one or more processors; a memory for storing one or more programs; and, when the one or more programs are executed by the one or more processors, causing the one or more processors to implement a VLA model control optimization method based on causal anchors as described above. Figure 3 The diagram shown is a hardware structure diagram of any data processing-capable device in which a memory-based, non-perceptive Prompt automatic enhancement and optimization device provided by an embodiment of the present invention is located. Except for... Figure 3 In addition to the processor and memory shown, any data processing device in the embodiment may also include other hardware depending on the actual function of the data processing device, which will not be described in detail here.

[0143] Accordingly, this application also provides a computer-readable storage medium storing computer instructions that, when executed by a processor, implement the VLA model control optimization method based on causal anchors as described above. The computer-readable storage medium can be an internal storage unit of any data-processing device as described in any of the foregoing embodiments, such as a hard disk or memory. The computer-readable storage medium can also be an external storage device, such as a plug-in hard disk, smart media card (SMC), SD card, flash card, etc., equipped on the device. Furthermore, the computer-readable storage medium can include both internal storage units of any data-processing device and external storage devices. The computer-readable storage medium is used to store the computer program and other programs and data required by the data-processing device, and can also be used to temporarily store data that has been output or will be output.

[0144] Other embodiments of this application will readily occur to those skilled in the art upon consideration of the specification and practice of the disclosure herein. This application is intended to cover any variations, uses, or adaptations of this application that follow the general principles of this application and include common knowledge or customary techniques in the art not disclosed herein. The specification and embodiments are to be considered exemplary only, and the true scope and spirit of this application are indicated by the claims.

[0145] It should be understood that this application is not limited to the precise structure described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope. The scope of this application is limited only by the appended claims.

Claims

1. A VLA model control optimization method based on causal anchor points, characterized in that, include: S1: Obtain historical interaction data of the VLA model executing the target task, extract the task execution trajectory from it, and use the intermediate state in the trajectory as a candidate intermediate state; S2: Mask the candidate intermediate states based on the mask autoencoder, and use the decrease in the success rate of the target task before and after masking as the causal score. Determine the candidate intermediate states with causal scores higher than a preset threshold as causal anchors, and generate a set of causal anchors including causal anchors and their ideal states. S3: When the VLA model performs the target task, determine the target causal anchor point currently aligned based on the anchor point set, and calculate the state difference between the current actual state and the ideal state corresponding to the target causal anchor point. S4: The state difference is used as the object to be encoded to obtain a continuous representation through an encoder, and the continuous representation is subjected to hierarchical discretization encoding to obtain a multi-layer discrete codeword that characterizes the deviation direction and degree of deviation between the current actual state and the ideal state. S5: Input the multilayer discrete codeword into the intervention decoder, decode to obtain the correction action, and compensate the original action output by the VLA model based on the correction action to obtain the optimized control action; S6: When the correction action is triggered, the entire long time-series path of the target task is not replanned. Instead, a local execution strategy is regenerated only for the optimized control action. After the correction action is completed, the original long time-series strategy of the target task is resumed to suppress error accumulation.

2. The VLA model control optimization method based on causal anchor points according to claim 1, characterized in that, Obtain historical interaction data of the VLA model executing the target task, extract the task execution trajectory from it, and use the intermediate states in the trajectory as candidate intermediate states, including: Extract task information from the historical interaction data of the VLA model executing the target task to obtain the task execution trajectory represented as an alternating sequence of states and actions; Each state in the task execution trajectory is taken as a candidate intermediate state.

3. The VLA model control optimization method based on causal anchor points according to claim 1, characterized in that, The candidate intermediate states are masked using a masked autoencoder, and the decrease in the target task success rate before and after masking is used as a causal score. Candidate intermediate states with causal scores higher than a preset threshold are identified as causal anchor points, generating a set of causal anchor points including causal anchor points and their ideal states, including: The intermediate states in the historical interaction data are randomly masked and reconstructed to reconstruct the error-supervised scoring module based on the mask autoencoder for representation learning, thereby obtaining the trained scoring module. The trained scoring module is used to mask each candidate intermediate state. The candidate intermediate state is then masked in the task execution trajectory and input into the policy network to predict subsequent actions. The normalized value of the decrease in the success rate of the target task before and after masking is used as the causal score of the candidate intermediate state. Collect the successful and failed trajectories of the target task, compare the state distribution differences of the two types of trajectories at each time step, and take the state inflection points in the successful trajectory that meet the preset conditions as candidate states of causal anchor points. The causal score is combined with the candidate state, and the state with the causal score higher than a preset threshold is determined as the causal anchor point, thus obtaining a causal anchor point set containing each causal anchor point and its ideal state.

4. The VLA model control optimization method based on causal anchor points according to claim 1, characterized in that, When the VLA model performs the target task, the target causal anchor point is determined based on the anchor point set, and the state difference between the current actual state and the ideal state corresponding to the target causal anchor point is calculated, including: Based on the set of causal anchor points and the current execution progress of the target task, determine the target causal anchor point to be aligned, and obtain the ideal state corresponding to the target causal anchor point; Real-time detection of the deviation between the current actual state and the ideal state, and determination of whether the deviation corresponds to the state offset of the target causal anchor point; If the current actual state is formed by the deviation of the ideal state of the target causal anchor point, then calculate the state difference between the actual state induced by the current action and the ideal state; if the deviation does not correspond to any causal anchor point in the set of causal anchor points, continue to rely on the action generation capability of the VLA model itself to perform the task.

5. The VLA model control optimization method based on causal anchor points according to claim 1, characterized in that, The state difference is used to obtain a continuous representation of the object to be encoded by an encoder. This continuous representation is then subjected to hierarchical discretization encoding to obtain a multi-layered discrete codeword representing the deviation direction and degree of the difference between the current actual state and the ideal state. This codeword includes: The state difference is the object to be encoded, and a continuous representation of the state difference is obtained by the encoder; Based on the current state, the ideal state, and the semantic representation of the current task, multimodal conditional features representing the difference between the actual state and the ideal state are extracted. The multimodal conditional features are fused with the continuous representation, and then hierarchical vector quantization is performed. Each quantization layer sequentially quantizes the quantization residual of the previous quantization layer, and the discrete codewords output by each quantization layer together constitute a multi-layer discrete codeword.

6. The VLA model control optimization method based on causal anchor points according to claim 1, characterized in that, The multilayer discrete codeword is input into the intervention decoder, decoded to obtain the correction action, and the original action output by the VLA model is compensated based on the correction action to obtain the optimized control action, including: The multilayer discrete codeword is input into the decoder to decode the features representing the current global state after the intervention is applied, and the correction action is obtained accordingly. The calibration action is compared with the ideal action output by the VLA model for the ideal state without intervention. The calibration action is then corrected with the goal of making the norm of the difference between the calibration action and the ideal action approach zero, resulting in an optimized control action to drive the smart terminal to execute.

7. The VLA model control optimization method based on causal anchor points according to claim 1, characterized in that, When the correction action is triggered, the entire long-sequence path of the target task is not replanned; instead, a local execution strategy is regenerated only for the optimized control action. After the correction action is completed, the original long-sequence strategy of the target task is resumed to suppress error accumulation, including: When the correction action is triggered, the entire long time-series path of the target task is not replanned. Instead, the multi-layer discrete codeword is reused as a differential semantic constraint from the actual state to the target causal anchor point to regenerate the local execution strategy and drive the smart terminal to execute it. Using the state after executing the local execution strategy as input, step S4 is re-executed on the state to obtain the re-encoded multi-layer discrete codeword; When the multilayer discrete codeword converges to an empty codeword representing zero bias, the correction action is determined to be complete, the regeneration of the local execution strategy is stopped, and the original long-sequence strategy for executing the target task is restored.

8. A VLA model control optimization device based on causal anchor points, characterized in that, include: The extraction module is used to obtain historical interaction data of the VLA model executing the target task, extract the task execution trajectory from it, and use the intermediate states in the trajectory as candidate intermediate states. The causal anchor discovery module is used to perform masking intervention on the candidate intermediate states based on the mask autoencoder, and use the decrease in the success rate of the target task before and after masking as the causal score. Candidate intermediate states with causal scores higher than a preset threshold are identified as causal anchors, and a set of causal anchors including causal anchors and their ideal states is generated. The deviation detection module is used to determine the target causal anchor point currently aligned based on the anchor point set when the VLA model is performing the target task, and to calculate the state difference between the current actual state and the ideal state corresponding to the target causal anchor point. The differential coding module is used to obtain a continuous representation of the object to be encoded by the state difference through an encoder, and to perform hierarchical discretization coding on the continuous representation to obtain a multi-layer discrete codeword that represents the deviation direction and degree of deviation between the current actual state and the ideal state. The action decoding module is used to input the multi-layer discrete codeword into the intervention decoder, decode it to obtain the correction action, and compensate the original action output by the VLA model based on the correction action to obtain the optimized control action; The local replanning module is used to replan the entire long-sequence path of the target task when issuing the correction action, but only regenerates the local execution strategy for the optimized control action. After the correction action is completed, the original long-sequence strategy of the target task is restored to suppress error accumulation.

9. An electronic device, characterized in that, include: One or more processors; Memory, used to store one or more programs; When the one or more programs are executed by the one or more processors, the one or more processors implement the method as described in any one of claims 1-7.

10. A computer-readable storage medium storing computer instructions thereon, characterized in that, When executed by the processor, this instruction implements the steps of the method as described in any one of claims 1-7.

Images