Inference method and device based on state adaptive action space, equipment and medium

Abstract

The application provides a reasoning method and device based on a state adaptive action space, equipment and a medium. The method comprises: acquiring a visual observation image and a text instruction; inputting the visual observation image and the text instruction into a mask prediction model, extracting multi-modal features of the visual observation image and the text instruction, and predicting an action dimension mask of the visual observation image; wherein the mask prediction model is trained by using an optimal mask label screened based on dynamic consistency testing; inputting the action dimension mask into a visual language action generation model, decoding a to-be-decoded action dimension according to the action dimension mask, skipping an action dimension that does not need to be decoded, combining the decoded action dimension and the skipped action dimension, and generating a complete execution action. The application not only avoids the interference of irrelevant action dimensions on instruction execution logic and feature dilution, but also greatly reduces the autoregressive or generation calculation overhead of the visual language action generation model, thereby ensuring accurate output of actions.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to a reasoning method, apparatus, device, and medium based on state-adaptive action space. Background Technology

[0002] Visual Language Action Generation (VLA), a core architecture in the field of embodied intelligence, leverages large language models to generate executable output actions from multimodal inputs, achieving a mapping from high-level instruction understanding to low-level robot control. These models typically fall into two main paradigms: autoregressive and parallel decoding. Autoregressive methods mimic the step-by-step text generation process of large language models, generating actions sequentially and exhibiting strong contextual consistency and expressive power. Parallel decoding methods employ a generative structure, directly predicting the entire action sequence in a single forward propagation, significantly improving inference speed and further reducing query frequency through action segmentation for smoother execution. These two paradigms each present trade-offs in terms of latency, control accuracy, and computational requirements, highlighting the importance of exploring hybrid or more efficient optimization methods to better balance these factors.

[0003] To address the need for efficient inference using VLA models in practical robot control, existing methods are mainly divided into two categories: structural optimization methods and reuse strategy methods. Structural optimization methods aim to reduce computational overhead through actions, parallelization, or model pruning. For example, autoregressive models can be modified to parallel decoding to reduce the number of inference steps. Reuse strategies, on the other hand, attempt to avoid redundant computation through caching mechanisms. They retain static visual or linguistic features in continuous interaction time steps and only recalculate tokens that are dynamically related to the current task, thereby reducing latency.

[0004] However, existing architectures typically perform full computation on all action dimensions indiscriminately during the inference phase, ignoring the fact that not all action dimensions contribute substantially to task completion under specific task conditions. This redundant computation is particularly severe in parallel decoding mode, resulting in significant waste of hardware power consumption and time. In autoregressive mode, the redundancy of action dimensions directly leads to a linear increase in inference latency with the number of dimensions. Furthermore, traditional model compression and structure optimization techniques often require additional fine-tuning of the original model, consuming significant resources and being highly task-dependent. More critically, the accuracy loss introduced by these techniques is easily amplified in continuous control tasks with cumulative error effects, leading to trajectory deviations or even task failure. Similarly, while reuse strategies and spatial similarity-based structure optimization methods can reduce latency, they often sacrifice the finesse of actions during caching and approximate computation, easily losing control information of key dimensions during dynamic decoupling, ultimately creating an irreconcilable trade-off between performance and efficiency. Summary of the Invention

[0005] This invention provides a reasoning method, apparatus, device, and medium based on state-adaptive action space to address the shortcomings of existing technologies that require full calculation of all possible action dimensions, resulting in huge computational overhead and significant time delay. It avoids interference and feature dilution of irrelevant action dimensions on instruction execution logic, and significantly reduces the autoregressive or generation computational overhead of visual language action generation models, ensuring that intelligent agents can output highly consistent, accurate, and efficient physical execution actions when faced with complex cross-modal instructions.

[0006] This invention provides a reasoning method based on state-adaptive action space, comprising: acquiring a visual observation image and text instructions; inputting the visual observation image and text instructions into a mask prediction model to extract multimodal features of the visual observation image and text instructions, and predicting an action dimension mask of the visual observation image; wherein, the mask prediction model is trained using the optimal mask labels selected based on dynamic consistency checks; inputting the action dimension mask into a visual language action generation model, decoding the action dimensions to be decoded according to the action dimension mask, skipping action dimensions that do not need to be decoded, and combining the decoded action dimensions with the skipped action dimensions to generate a complete execution action.

[0007] According to the inference method based on state-adaptive action space provided by the present invention, before inputting visual observation images and text instructions into a mask prediction model, the method includes: acquiring visual observation samples, predicted actions generated by the original strategy under the visual observation samples, and the baseline observation state at the next moment after executing the predicted actions; generating corresponding masking actions using multiple pre-generated strategies with different masking patterns, and acquiring the masked observation state at the next moment after executing each masking action; comparing each masked observation state with the baseline observation state, and determining that the difference meets the preset dynamic consistency condition based on the comparison results, using the corresponding masking pattern as the optimal masking label for the visual observation sample, and training the mask prediction model to be trained.

[0008] According to the inference method based on state-adaptive action space provided by the present invention, for each masked observation state, a comparison is made with a reference observation state, including: for each masked observation state, combining each masked observation state with the reference observation state to determine the difference metric value of the corresponding masked observation state; when the difference metric value of the masked observation state is less than or equal to a preset threshold, the corresponding mask mode is determined to meet the preset dynamic consistency condition.

[0009] According to the inference method based on state-adaptive action space provided by the present invention, when the difference is determined to meet the preset dynamic consistency condition based on the comparison result, the corresponding mask pattern is used as the optimal mask label for the visual observation sample, and the mask prediction model to be trained is trained. The method includes: when the difference is determined to meet the preset dynamic consistency condition based on the comparison result, when at least two mask patterns meet the preset dynamic consistency condition, determining the number of masked dimensions in the mask patterns that meet the condition, selecting the mask pattern with the most masked dimensions as the optimal mask label for the corresponding visual observation sample, and supervising the training of the mask prediction model to be trained; or, taking a weighted average of all mask patterns that meet the condition to obtain a continuous soft label as the optimal mask label for the corresponding visual observation sample, and supervising the training of the mask prediction model to be trained; wherein the weight of each mask pattern that meets the condition is determined according to the difference metric between its corresponding masked observation state and the baseline observation state; or, using all mask patterns that meet the condition as positive samples and mask patterns that do not meet the condition as negative samples, constructing a contrastive loss function based on the positive and negative samples to train the mask prediction model to be trained.

[0010] According to the inference method based on state-adaptive action space provided by the present invention, before generating corresponding masking actions using multiple pre-generated strategies with different masking patterns, the method includes: generating a masking pattern with a preset structure based on the physical coupling relationship of each dimension in the action space; wherein, the physical coupling relationship is used to characterize the correlation between the physical effects generated by different action dimensions during execution, and the masking pattern with the preset structure includes a masking pattern that only masks action dimensions belonging to the same physical group, or a progressive masking pattern that gradually increases the number of masking dimensions according to a preset incrementing rule; and / or, generating multiple different masking patterns using a mask generator trained with historical interaction data; and / or, when it is determined that the action space dimension is lower than a preset dimension threshold, generating all non-empty subsets as corresponding masking patterns according to the preset action dimensions; wherein, the non-empty subsets are used to characterize that not all retained action dimensions are default masking values; when it is determined that the action space dimension is greater than or equal to the preset dimension threshold, generating corresponding masking patterns independently based on the preset retention probability of each action dimension, or, based on the randomly selected number of retained dimensions, randomly selecting a corresponding number of action dimensions from all action space dimensions for retention, and generating corresponding masking patterns.

[0011] According to the inference method based on state-adaptive action space provided by the present invention, before decoding the action dimension to be decoded and skipping the action dimensions that do not need to be decoded, the method includes: determining that the flag of the action dimension in the action dimension mask is a non-default masking value, then determining that the corresponding action dimension is a key dimension and needs to be decoded by the action decoder; determining that the flag of the dimension in the action dimension mask is a default masking value, then determining that the corresponding action dimension is a redundant dimension and does not need to be decoded.

[0012] According to the inference method based on state-adaptive action space provided by the present invention, after generating a complete execution action, the method includes: sending the execution action to the underlying execution of the agent.

[0013] This invention also provides a reasoning device based on a state-adaptive action space, comprising: a data acquisition module for acquiring visual observation images and text commands; a masking module for inputting the visual observation images and text commands into a masking prediction model, extracting multimodal features of the visual observation images and text commands, and predicting the action dimension mask of the visual observation images; wherein the masking prediction model is trained using the optimal mask labels selected based on dynamic consistency checks; and an action generation module for inputting the action dimension mask into a visual language action generation model, decoding the action dimensions to be decoded according to the action dimension mask, skipping action dimensions that do not need to be decoded, and combining the decoded action dimensions with the skipped action dimensions to generate a complete execution action.

[0014] The present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the reasoning method based on state adaptive action space as described above.

[0015] The present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the reasoning method based on state-adaptive action space as described above.

[0016] The present invention also provides a computer program product, including a computer program that, when executed by a processor, implements the reasoning method based on state adaptive action space as described above.

[0017] The present invention provides a reasoning method, apparatus, device, and medium based on state-adaptive action space. By jointly inputting visual observation images and text commands into a mask prediction model, it extracts multimodal features and predicts action dimension masks, achieving a direct mapping from a high-dimensional cross-modal semantic space to a structured action space. By training the mask prediction model with optimal mask labels selected based on dynamic consistency checks, the model can accurately identify and remove redundant action dimensions unrelated to the current text command and visual context, thus completing the physical-level filtering of non-control features during the multimodal fusion stage. Simultaneously, by predicting... Action dimension masking serves as a structured prior route to guide the decoding process of the visual language action generation model. This allows the decoder to perform fine-grained reasoning only on dimensions determined to be core dynamics, while bypassing redundant dimensions. Ultimately, the decoded values ​​of key dimensions are combined with the default values ​​of redundant dimensions to generate complete execution actions. In multimodal instruction-driven tasks, this avoids interference and feature dilution of irrelevant action dimensions on instruction execution logic, while significantly reducing the autoregressive or generative computational overhead of the visual language action generation model. This ensures that the agent can output highly consistent, accurate, and efficient physical execution actions when faced with complex cross-modal instructions. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0019] Figure 1 This is one of the flowcharts illustrating the reasoning method based on state-adaptive action space provided by the present invention; Figure 2 This is the second flowchart of the reasoning method based on state adaptive action space provided by the present invention; Figure 3 This is a schematic diagram of the structure of the inference device based on state adaptive action space provided by the present invention; Figure 4 This is a schematic diagram of the structure of the electronic device provided by the present invention. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0021] Figure 1 This is a flowchart illustrating the reasoning method based on state-adaptive action space provided by the present invention, as shown below. Figure 1 As shown, the method includes the following: S11, acquire visual observation images and text commands; S12, input the visual observation image and text command into the mask prediction model, extract the multimodal features of the visual observation image and text command, and predict the action dimension mask of the visual observation image; wherein, the mask prediction model is trained using the optimal mask label selected based on dynamic consistency test. S13. Input the action dimension mask into the visual language action generation model. Based on the action dimension mask, decode the action dimensions to be decoded, skip the action dimensions that do not need to be decoded, and combine the decoded action dimensions with the skipped action dimensions to generate a complete execution action.

[0022] It should be noted that the following will be combined with specific details. Figure 2 The present invention describes a reasoning method based on state-adaptive action space.

[0023] Step S11: Obtain visual observation images and text instructions; Step S12: Input the visual observation image and text command into the mask prediction model, extract the multimodal features of the visual observation image and text command, and predict the action dimension mask of the visual observation image; wherein, the mask prediction model is trained using the optimal mask label selected based on dynamic consistency test.

[0024] It's worth noting that by introducing a lightweight mask prediction model, before the visual language action generation model executes each step of action inference, the mask prediction model dynamically evaluates the importance of each action dimension in the current task stage based on the current visual observation and language instructions, and generates a binary or continuous action mask. This breaks the rigid mode of traditional visual language action models (VLA) that calculates all dimensions equally. For example, when the robot approaches a target over a long distance, the calculation of rotation and gripper dimensions is redundant; while in fine grasping, the change in translation dimension may be very small. Through dynamic masking, the visual language action generation model can calculate on demand, accurately eliminating redundant action dimensions that do not need to be changed in the current frame.

[0025] In one alternative embodiment, reference Figure 2 Before inputting visually observed images and text instructions into the mask prediction model, the process includes: acquiring visually observed samples, and developing the original strategy (…). Predicted actions generated from visual observation samples ) and the baseline observation state at the next moment after performing the prediction action ( ); Utilizing multiple pre-generated strategies with different mask patterns ( ), respectively generate corresponding masking actions ( ), and obtain the masking observation state at the next moment after each masking action is performed ( For each masked observation state, it is compared with the baseline observation state. Based on the comparison results, when the difference meets the preset dynamic consistency condition, the corresponding mask pattern is used as the optimal mask label for the visual observation sample, and the mask prediction model to be trained is trained.

[0026] It should be noted that by acquiring visual observation samples, the predicted actions generated by the original strategy under the visual observation samples, and executing the complete actions of the original strategy in the environment, the resulting state at the next moment is recorded to obtain an objective state transition benchmark that is not affected by masking and conforms to the underlying physical laws. Furthermore, by executing masking actions of different masking patterns in the same physical environment and observing their impact on the evolution of the environmental state, it is equivalent to conducting a series of controlled variable interference experiments on the action space. Then, by comparing the state transitions caused by each masking action with the objective state transition benchmark, the masking pattern that meets the dynamic consistency condition is selected as the optimal mask label using the comparison results. This enables the mask prediction model to use the real response of the physical environment as a supervision signal, automatically mine and remove non-controlled redundant dimensions in the action space, and force the model to focus only on the core action semantics that drive environmental changes. This endows the model with underlying dynamic perception and feature anti-interference ability, fundamentally avoiding the risk of representation pollution and semantic collapse caused by random or heuristic masks.

[0027] Specifically, for each masked observation state, a comparison is made with the baseline observation state, including: for each masked observation state, each masked observation state is combined with the baseline observation state to determine the difference metric value of the corresponding masked observation state; when the difference metric value of the masked observation state is less than or equal to a preset threshold, the corresponding masking mode is determined to meet the preset dynamic consistency condition.

[0028] It should be added that the difference metric can be determined based on L2 squared distance, cosine similarity, or KL divergence, etc., and the specific choice can be made according to the actual design requirements. No further limitations are made here. For example, when determining the difference metric based on L2 squared distance, the difference metric should satisfy... When the corresponding mask pattern is determined to meet the preset dynamic consistency condition, it indicates that the masked action dimension is redundant; among which, Indicates the baseline observation status; Indicates the first individual mask patterns The corresponding masked observation state; This indicates a preset threshold.

[0029] In addition, when there is at least one mask pattern that meets the preset dynamic consistency condition, training samples can be constructed based on the corresponding mask pattern that meets the preset dynamic consistency condition and the corresponding visual observation sample to train the mask prediction model to be trained. The network weights of the mask model are updated through backpropagation, so that it learns to accurately predict which action dimensions are critical (need to be calculated) and which are redundant (can be skipped) according to the current visual scene.

[0030] Furthermore, when a certain action dimension is forcibly masked, the policy distribution output by the masked prediction model is evaluated to ensure high consistency with the output during full-dimensional inference. If, after masking a dimension, the prediction results of the masked prediction model for other key dimensions do not show significant shifts (i.e., the dynamic consistency condition is met, and the difference metric is less than or equal to a preset threshold), then that dimension is automatically marked as the default mask value, such as "redundant (Mask=0)"; otherwise, it is marked as "critical (Mask=1)". Compared to relying on manually set static rules or simple action change rate thresholds, the "dynamic consistency" method is model-aware, considering not only changes in physical trajectories but also deeply reflecting the dependencies of feature representations within the VLA model. The labels generated by this method ensure that while removing redundant dimensions, the control accuracy of the original policy is preserved to the greatest extent, fundamentally preventing covariate shifts caused by acceleration.

[0031] It should be noted that by comparing the features of each masked observation state with the baseline observation state to calculate the difference metric representing the state offset between the two, the degree of environmental dynamic offset under different masking patterns can be accurately quantified. Furthermore, by determining that the masked action dimension has no substantial impact on the underlying dynamic evolution when the difference metric is less than or equal to a preset difference threshold, the state tolerance of the physical environment is transformed into a strict mathematical screening boundary. This allows the model to automatically and objectively identify and filter out non-control redundant dimensions in the action space, forcing the masking prediction model to reconstruct and learn only the core action semantics that truly drive the environmental state transition. This fundamentally avoids the risk of the feature space being contaminated by non-dynamic information due to masked irrelevant dimensions, ensuring that the extracted potential action representation is strictly aligned with the real physical evolution law.

[0032] In an optional embodiment, when the difference is determined to meet the preset dynamic consistency condition based on the comparison result, the corresponding mask pattern is used as the optimal mask label for the visual observation sample, and the mask prediction model to be trained is trained, including: when the difference is determined to meet the preset dynamic consistency condition based on the comparison result, when it is determined that at least two mask patterns meet the preset dynamic consistency condition, the number of masked dimensions in the mask patterns that meet the condition is determined, the mask pattern with the most masked dimensions is selected as the optimal mask label for the corresponding visual observation sample, and the mask prediction model to be trained is trained under supervision.

[0033] It should be noted that, when at least two masking patterns meet the preset dynamic consistency condition, the number of masked dimensions in each candidate masking pattern is further compared, and the masking pattern with the most masked dimensions is forcibly selected as the optimal mask label, so as to maximize the removal of redundant degrees of freedom in the action space. Furthermore, by using this maximized mask label as a supervision signal to train the mask prediction model, the model is forced to reason and reconstruct the complete action from the minimal unmasked dimensions, while ensuring that the evolution of environmental dynamics is not substantially affected. This greatly improves the model's ability to compress and extract the core action semantics, effectively suppresses the model's dependence on redundant uncontrolled features and overfitting, and ensures that the learned potential action representation has the highest information density and the strongest dynamic transferability.

[0034] In an optional embodiment, when the difference is determined to meet the preset dynamic consistency condition based on the comparison result, the corresponding mask pattern is used as the optimal mask label for the visual observation sample, and the mask prediction model to be trained is trained. The method further includes: when at least two mask patterns meet the preset dynamic consistency condition, a weighted average of all mask patterns that meet the condition is performed to obtain a continuous soft label as the optimal mask label for the corresponding visual observation sample, and the mask prediction model to be trained is trained under supervision. The weight of each mask pattern that meets the condition is determined based on the difference metric between its corresponding masked observation state and the baseline observation state; the smaller the difference metric, the larger the weight.

[0035] In an optional embodiment, when the difference is determined to meet the preset dynamic consistency condition based on the comparison result, the corresponding mask pattern is used as the optimal mask label of the visual observation sample to train the mask prediction model to be trained. The method further includes: when it is determined that at least two mask patterns meet the preset dynamic consistency condition, all mask patterns that meet the condition are used as positive samples, and mask patterns that do not meet the condition are used as negative samples. A contrastive loss function is constructed based on the positive and negative samples to train the mask prediction model to be trained.

[0036] It should be noted that, by abandoning the single hard label selection mechanism when at least two mask patterns meet the preset dynamic consistency condition, all candidate mask patterns that meet the condition are weighted and fused according to their dynamic offset degree to construct a continuous soft label as the optimal mask label. This transforms the discrete mask selection problem into a probability distribution model based on physical consistency. Furthermore, by dynamically assigning weights based on the difference metric between each mask observation state and the baseline observation state, mask patterns with a weaker impact on environmental state offset (i.e., smaller difference metric) receive higher confidence. This accurately maps the dynamic tolerance of the physical environment to the continuous probability expression of the soft label, enabling the mask prediction model to absorb multiple equivalent non-core dimensions of mask priors during supervised training. This effectively alleviates the risk of overfitting and feature oscillation caused by a single hard label, and gives the model a smoother generalization perception of action space redundancy and a more robust semantic representation ability.

[0037] It should be noted that, by abandoning the limitation of using only a single positive sample for regression reconstruction when at least two mask patterns meet the preset dynamic consistency condition, all mask patterns that meet the condition are constructed as a positive sample set, and mask patterns that do not meet the dynamic consistency condition are constructed as a negative sample set, thus forming a contrastive learning space based on physical dynamic response. Furthermore, by constructing a contrastive loss function based on the positive and negative sample sets to train the mask prediction model, the model is forced to actively narrow the distance with mask representations that conform to the laws of physical evolution in the feature space, while distancing itself from mask representations that violate dynamic logic. This transforms the dynamic consistency constraint of the environment into a rigid feature geometric boundary, enabling the model not only to identify core action semantics but also to have the ability to explicitly exclude interference from non-control factors at the feature level. This greatly enhances the discriminativeness and physical consistency of potential action representations, fundamentally eliminating the risk of the model confusing non-dynamic factors such as background or lighting as action representations.

[0038] In an optional embodiment, before generating corresponding masking actions using multiple pre-generated strategies with different masking patterns, the method includes: generating a masking pattern with a preset structure based on the physical coupling relationship of each dimension in the action space; wherein, the physical coupling relationship is used to characterize the correlation between the physical effects generated by different action dimensions during execution, and the masking pattern with the preset structure includes a masking pattern that only masks action dimensions belonging to the same physical group, or a progressive masking pattern that gradually increases the number of masking dimensions according to a preset incremental rule.

[0039] It should be noted that by generating a mask pattern with a preset structure based on the physical coupling relationship of each dimension in the action space, the real mechanical linkage and dynamic dependence of robot joints are injected into the mask strategy generation process, thereby avoiding random and independent masking that destroys the physical integrity of the action dimensions. Furthermore, by adopting a mask pattern that only masks action dimensions belonging to the same physical group, or a progressive mask pattern that gradually increases the number of masking dimensions according to a preset incremental rule, the mask interference experiment strictly follows the physical kinematics of the robot. The former ensures that the linkage dimensions within the same functional unit are evaluated as a whole to prevent the fragmentation of key physical cooperative features, while the latter realizes a progressive probe scan of the redundancy of the action space from coarse to fine granular, effectively avoiding the chaotic interference caused by invalid or destructive mask combinations to the environmental dynamics evaluation, thereby ensuring the physical interpretability of the subsequent dynamic consistency comparison results and the efficiency of label mining, forcing the mask prediction model to learn a coherent action representation that conforms to the real physical topology.

[0040] In an alternative embodiment, before generating corresponding masking actions using multiple pre-generated strategies with different masking patterns, the method further includes: generating a mask generator trained using historical interaction data to generate multiple different masking patterns that tend to reveal redundant dimensions.

[0041] It should be noted that by training the mask generator using historical interaction data, redundant dimensional priors in the action space can be adaptively mined in a data-driven manner, thus overcoming the blindness and limitations of relying on manual rules or random combinations to generate mask patterns. Furthermore, by generating multiple different mask patterns that tend to reveal redundant dimensions through this mask generator, subsequent mask interference experiments can directly focus on dimensional intervals with a high probability of non-core action semantics for targeted detection. This greatly reduces the search space for invalid mask combinations and improves the efficiency of mining optimal mask labels. At the same time, it forces the mask prediction model to more keenly identify and extract the core dynamic features that truly dominate the environmental state transition in the adversarial game against high-density redundant interference, giving the model the adaptive evolutionary ability to efficiently strip away non-controlling factors in complex high-dimensional action spaces.

[0042] In an optional embodiment, before generating corresponding masking actions using multiple pre-generated strategies with different masking patterns, the method further includes: when the action space dimension is determined to be lower than a preset dimension threshold, generating all non-empty subsets as corresponding masking patterns based on the preset action dimension; wherein, the non-empty subsets are used to characterize that not all retained action dimensions are default masking values; or, when the action space dimension is determined to be greater than or equal to the preset dimension threshold, independently generating corresponding masking patterns based on the preset retention probability of each action dimension; or, based on the randomly selected number of retained dimensions, randomly selecting a corresponding number of action dimensions from all action space dimensions for retention, and generating corresponding masking patterns.

[0043] It should be noted that by adaptively switching the mask pattern generation strategy based on the comparison results between the action space dimension and the preset dimension threshold, the optimal balance between mask search complexity and action space scale is achieved. When the action space dimension is low, all non-empty subsets are traversed and generated as mask patterns, ensuring exhaustive dynamic interference testing of the low-dimensional action space under controllable computational overhead, eliminating the risk of missing potential optimal mask combinations, and thus obtaining absolutely accurate dynamic consistency evaluation. At the same time, when the action space dimension is high, mask patterns are generated independently based on preset retention probabilities or randomly sampled based on random dimension numbers. An efficient random sampling strategy breaks the combinatorial explosion dilemma brought about by high-dimensional space. While ensuring the diversity and coverage of mask pattern distribution, meaningless computation is greatly reduced. Thus, the training of the mask prediction model can take into account both rigorous exhaustive search in low-dimensional scenarios and sparse exploration in high-dimensional scenarios, ensuring the universality and computational robustness of the optimal mask label mining mechanism under action spaces of different sizes.

[0044] Step S13: Input the action dimension mask into the visual language action generation model. Based on the action dimension mask, decode the action dimensions to be decoded, skip the action dimensions that do not need to be decoded, and combine the decoded action dimensions with the skipped action dimensions to generate a complete execution action.

[0045] In this embodiment, before decoding the action dimension to be decoded and skipping the action dimensions that do not need to be decoded, the process includes: if the flag indicating the presence of an action dimension in the action dimension mask is a non-default masking value, then the corresponding action dimension is determined to be a critical dimension and needs to be decoded by the action decoder; if the flag indicating the presence of a dimension in the action dimension mask is a default masking value, then the corresponding action dimension is determined to be a redundant dimension and does not need to be decoded.

[0046] It should be noted that by parsing the flag bits of each dimension in the action dimension mask, the action dimensions corresponding to non-default mask values ​​are identified as key dimensions and routed to the action decoder for decoding, while the action dimensions corresponding to default mask values ​​are identified as redundant dimensions and directly skipped for decoding. This constructs a sparse decoupling decoding mechanism based on mask priors. By strictly dividing the action space into key dynamic dimensions that require fine reasoning and redundant dimensions that can be directly bypassed, the computational resources of the action decoder are fully concentrated on the core action semantic reconstruction that has a substantial driving effect on the transition of environmental states. This not only blocks the pollution and interference of redundant features on the potential representation space from the bottom layer of the network architecture, but also significantly reduces the computational overhead of invalid dimensions in the reasoning stage. Thus, while ensuring the physical consistency and high fidelity of the generated actions, the reasoning efficiency and real-time response capability of the policy model are significantly improved.

[0047] Furthermore, in the autoregressive decoding stage of the visual language action generation model, by directly applying the action mask generated in the above manner to the underlying computation graph, for action tokens marked as redundant, the decoder will skip the dense computation of their corresponding attention mechanisms and feedforward neural networks, directly reusing historical states or outputting default values. This allows for the allocation of computational resources only for action tokens marked as critical, achieving a cost-effective and efficient execution phase. By directly reducing the amount of floating-point operations in redundant dimensions at the mathematical and system levels, the single-step inference latency of the visual language action generation model is significantly reduced. Combined with extremely high mask accuracy, this successfully breaks through the bottleneck of the traditional acceleration method where speed and accuracy cannot be simultaneously achieved, enabling robots to achieve high-frequency, high-precision closed-loop control on resource-constrained edge devices.

[0048] In addition, the action dimensions obtained from the combined decoding and the action dimensions that are skipped from the decoding are combined to generate a complete action, including: keeping the decoded values ​​of the key dimensions unchanged, filling the default values ​​of redundant dimensions into the corresponding dimension positions, and splicing them together to form a complete action vector that is consistent with the original action space dimensions, which is then used as the action.

[0049] It should be noted that by keeping the decoded values ​​of key dimensions unchanged and filling the corresponding dimension positions with the preset default values ​​of redundant dimensions, a complete action vector consistent with the original action space dimensions is formed. This achieves perfect alignment between high-dimensional sparse semantics and low-dimensional dense execution at the feature representation level. Furthermore, by filling the default values ​​of non-core dimensions (such as zero values ​​or neutral physical poses) with constant values, it ensures that the output action vector structurally meets the dimensional rigidity constraints of the underlying environment execution interface, and completely cuts off the random interference of uncontrolled redundant dimensions on the dynamic evolution of the environment in terms of physical semantics. This forces the state transition of the physical environment to be driven only by the key dimensions accurately inferred by the decoder. Thus, while ensuring the compliance and security of policy execution, it achieves lossless mapping and dynamic purification of the action space from sparse semantic representation to dense physical execution.

[0050] In one optional embodiment, after generating the complete execution action, the process includes: sending the execution action to the lower-level execution of the agent, thereby based on the visual observation image at the next moment after the action is completed.

[0051] It should be noted that by directly sending the complete execution action generated by splicing to the bottom layer of the agent, a closed-loop connection is achieved between the sparse core semantic decisions learned by the mask prediction model in the feature space and the real motion control link in the physical world. Furthermore, by applying the action instructions containing only the key dimension decoded values ​​and the default fill values ​​of redundant dimensions to the bottom layer controller, it is ensured that the state transition of the agent in the real physical or high-fidelity simulation environment is strictly controlled by the extracted core dynamic features. This completely shields the random jitter or undesirable dynamic interference that may be caused by redundant degrees of freedom at the execution end, which not only ensures the safety and physical consistency of policy execution, but also enables the agent to maintain high-precision task completion performance and robust generalization ability under lightweight inference with low latency and low computing power consumption.

[0052] It is worth noting that by introducing the lightweight dynamic masking mechanism described above, the visual language action generation model is endowed with context awareness, enabling it to accurately identify and eliminate redundant action dimensions that do not need to be changed in the current frame based on the current stage of the task (such as approaching the target or fine grasping). This strategy of allocating computing power on demand significantly optimizes the utilization efficiency of computing resources. As shown in Table 1, throughout the entire task cycle, the above method successfully masks an average of 26.79% of action dimensions compared to the baseline VLA model.

[0053] Table 1 Action Dimension Mask Rate Table

[0054] The mask label generation and training framework based on dynamic consistency eliminates the reliance on manually set static rules or simple rate-of-change thresholds. By automatically evaluating the consistency of the model's output policy distribution, it generates high-quality mask labels at low cost and high efficiency. As shown in Table 2, the average success rate of this method reached 82.4% across various tasks, a significant improvement compared to the 76.8% success rate of the original VLA model. In contrast, the success rate of the method based on simple thresholds dropped to 64.4% due to covariate shift.

[0055] Table 2 Task Success Rate Table

[0056] Based on mask-based selective / sparse autoregressive decoding, this method substantially reduces floating-point computation at both the mathematical and system levels. Combined with the extremely high accuracy guaranteed by key point two, this mechanism successfully breaks the deadlock of "speed and accuracy cannot be achieved simultaneously" in traditional model acceleration, significantly reducing single-step inference latency. This allows the originally large VLA model to be successfully deployed on computationally limited robot edge devices and perfectly supports high-frequency, high-precision real-time closed-loop control. As shown in Table 3, compared with full-dimensional autoregressive decoding, the selective decoding mechanism of this method significantly reduces the inference latency of single-step inference from 270.10 ms to 212.10 ms.

[0057] Table 3 Inference Delay Table

[0058] In summary, this invention, by jointly inputting visual observation images and text commands into a mask prediction model, extracts multimodal features and predicts action dimension masks, achieving a direct mapping from a high-dimensional cross-modal semantic space to a structured action space. By training this mask prediction model with optimal mask labels selected based on dynamic consistency checks, the model can accurately identify and remove redundant action dimensions unrelated to the current text command and visual context, thus achieving physical-level filtering of non-control features during the multimodal fusion stage. Simultaneously, by using the predicted action dimension masks as structured prior routes to guide the decoding process of the visual language action generation model, the decoder performs fine-grained reasoning only on dimensions determined to be core dynamics, bypassing redundant dimensions. Finally, the decoded values ​​of key dimensions are combined with the default values ​​of redundant dimensions to generate complete execution actions. Therefore, in multimodal command-driven tasks, this avoids interference and feature dilution of irrelevant action dimensions on command execution logic, significantly reduces the autoregressive or generative computational overhead of the visual language action generation model, and ensures that the agent can output highly consistent, accurate, and efficient physical execution actions when faced with complex cross-modal commands.

[0059] The inference device based on state-adaptive action space provided by the present invention will be described below. The inference device based on state-adaptive action space described below and the inference method based on state-adaptive action space described above can be referred to in correspondence.

[0060] Figure 3 A schematic diagram of a reasoning device based on a state-adaptive action space is shown. The device includes: Data acquisition module 31 acquires visual observation images and text commands; The masking module 32 inputs the visual observation image and text command into the masking prediction model, extracts the multimodal features of the visual observation image and text command, and predicts the action dimension mask of the visual observation image; wherein, the masking prediction model is trained using the optimal mask labels selected based on dynamic consistency test; The action generation module 33 inputs the action dimension mask into the visual language action generation model, decodes the action dimensions to be decoded based on the action dimension mask, skips the action dimensions that do not need to be decoded, and combines the decoded action dimensions with the skipped action dimensions to generate a complete execution action.

[0061] It should be noted that the specific principles of the embodiments of the present invention are the same as those of the above method embodiments. For details, please refer to the method embodiments above. More detailed explanations will not be repeated here.

[0062] Figure 4 An example is a schematic diagram of the physical structure of an electronic device, such as... Figure 4 As shown, the electronic device may include a processor 410, a communications interface 420, a memory 430, and a communication bus 440. The processor 410, communications interface 420, and memory 430 communicate with each other via the communication bus 440. The processor 410 can call logical instructions in the memory 430 to execute a state-adaptive action space-based reasoning method. This method includes: acquiring a visual observation image and text instructions; inputting the visual observation image and text instructions into a mask prediction model to extract multimodal features from the visual observation image and text instructions, and predicting an action dimension mask for the visual observation image; wherein the mask prediction model is trained using the optimal mask labels selected based on a dynamic consistency check; inputting the action dimension mask into a visual language action generation model; decoding the action dimensions to be decoded according to the action dimension mask, skipping action dimensions that do not need to be decoded, and combining the decoded action dimensions with the skipped action dimensions to generate a complete execution action.

[0063] Furthermore, the logical instructions in the aforementioned memory 430 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0064] On the other hand, the present invention also provides a computer program product, which includes a computer program that can be stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer can execute the reasoning method based on state-adaptive action space provided by the above methods. The method includes: acquiring a visual observation image and a text instruction; inputting the visual observation image and the text instruction into a mask prediction model, extracting multimodal features of the visual observation image and the text instruction, and predicting an action dimension mask of the visual observation image; wherein the mask prediction model is trained using the optimal mask label selected based on dynamic consistency test; inputting the action dimension mask into a visual language action generation model, decoding the action dimension to be decoded according to the action dimension mask, skipping action dimensions that do not need to be decoded, and combining the decoded action dimension with the skipped action dimension to generate a complete execution action.

[0065] In another aspect, the present invention also provides a non-transitory computer-readable storage medium storing a computer program thereon. When executed by a processor, the computer program implements the reasoning method based on state-adaptive action space provided by the above methods. The method includes: acquiring a visual observation image and a text instruction; inputting the visual observation image and the text instruction into a mask prediction model, extracting multimodal features of the visual observation image and the text instruction, and predicting an action dimension mask of the visual observation image; wherein the mask prediction model is trained using the optimal mask label selected based on dynamic consistency check; inputting the action dimension mask into a visual language action generation model, decoding the action dimension to be decoded according to the action dimension mask, skipping action dimensions that do not need to be decoded, and combining the decoded action dimension with the skipped action dimension to generate a complete execution action.

[0066] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. 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 embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0067] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0068] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A reasoning method based on state-adaptive action space, characterized in that, include: Acquire visual observation images and text commands; The visual observation image and text command are input into the mask prediction model to extract the multimodal features of the visual observation image and text command, and to predict the action dimension mask of the visual observation image; wherein, the mask prediction model is trained using the optimal mask labels selected based on dynamic consistency test; The action dimension mask is input into the visual language action generation model. Based on the action dimension mask, the action dimensions to be decoded are decoded, and the action dimensions that do not need to be decoded are skipped. The decoded action dimensions are combined with the skipped action dimensions to generate a complete execution action.

2. The reasoning method based on state-adaptive action space according to claim 1, characterized in that, Before inputting the visually observed image and text instructions into the mask prediction model, the following steps are included: Acquire visual observation samples, the predicted action generated by the original strategy under the visual observation samples, and the baseline observation state at the next moment after executing the predicted action; Using multiple pre-generated strategies with different masking patterns, corresponding masking actions are generated respectively, and the masking observation state at the next moment is obtained after executing each masking action; For each of the masked observation states, a comparison is made with the baseline observation state. Based on the comparison results, when the difference meets the preset dynamic consistency condition, the corresponding mask pattern is used as the optimal mask label for the visual observation sample, and the mask prediction model to be trained is trained.

3. The reasoning method based on state-adaptive action space according to claim 2, characterized in that, For each of the aforementioned masked observation states, a comparison is made with the aforementioned baseline observation state, including: For each of the aforementioned masked observation states, the difference metric value for each masked observation state is determined by combining it with the baseline observation state. When the difference metric value of the masked observation state is determined to be less than or equal to a preset threshold, the corresponding mask mode is determined to meet the preset dynamic consistency condition.

4. The reasoning method based on state-adaptive action space according to claim 2, characterized in that, When the difference is determined to meet the preset dynamic consistency condition based on the comparison results, the corresponding mask pattern is used as the optimal mask label for the visual observation sample, and the mask prediction model to be trained is trained, including: When the difference is determined to meet the preset dynamic consistency condition based on the comparison results, and when at least two mask patterns meet the preset dynamic consistency condition, the number of masked dimensions in the mask patterns that meet the condition is determined, and the mask pattern with the most masked dimensions is selected as the optimal mask label for the corresponding visual observation sample, and supervised training is performed on the mask prediction model to be trained; or... All qualifying mask patterns are weighted and averaged to obtain a continuous soft label, which serves as the optimal mask label for the corresponding visual observation sample. This label is then used for supervised training of the mask prediction model to be trained. The weight of each qualifying mask pattern is determined based on a metric measuring the difference between its corresponding masked observation state and the baseline observation state. Alternatively, All mask patterns that meet the conditions are treated as positive samples, and mask patterns that do not meet the conditions are treated as negative samples. A contrastive loss function is constructed based on the positive and negative samples to train the mask prediction model to be trained.

5. The reasoning method based on state-adaptive action space according to claim 2, characterized in that, Before generating corresponding masking actions using multiple pre-generated strategies with different masking patterns, the process includes: Based on the physical coupling relationships of each dimension in the action space, a mask pattern with a preset structure is generated; wherein, the physical coupling relationship is used to characterize the correlation between the physical effects generated by different action dimensions during execution, and the mask pattern with the preset structure includes a mask pattern that only masks action dimensions belonging to the same physical group, or a progressive mask pattern that gradually increases the number of masked dimensions according to a preset incrementing rule; and / or, A mask generator trained using historical interaction data can generate multiple different mask patterns; and / or, When the action space dimension is determined to be lower than a preset dimension threshold, all non-empty subsets are generated as corresponding mask patterns according to the preset action dimension; wherein, the non-empty subsets are used to represent that not all of the retained action dimensions are the default mask values; When the action space dimension is determined to be greater than or equal to the preset dimension threshold, a corresponding mask pattern is independently generated based on the preset retention probability of each action dimension, or, based on the randomly selected number of retention dimensions, a corresponding number of action dimensions are randomly selected from all action space dimensions for retention, and a corresponding mask pattern is generated.

6. The reasoning method based on state-adaptive action space according to claim 1, characterized in that, Before decoding the action dimensions to be decoded and skipping action dimensions that do not need to be decoded, the process includes: If the flag indicating the presence of an action dimension in the action dimension mask is a non-default mask value, then the corresponding action dimension is determined to be a critical dimension and needs to be decoded by the action decoder. If the flag indicating the presence of a dimension in the action dimension mask is determined to be the default mask value, then the corresponding action dimension is determined to be a redundant dimension and does not need to be decoded.

7. The reasoning method based on state-adaptive action space according to claim 1, characterized in that, After generating the complete execution action, it includes: The execution action is sent down to the underlying layer of the intelligent agent for execution.

8. A reasoning device based on a state-adaptive action space, characterized in that, include: The data acquisition module acquires visually observed images and text commands; The masking module inputs the visual observation image and text command into the masking prediction model, extracts the multimodal features of the visual observation image and text command, and predicts the action dimension mask of the visual observation image; wherein, the masking prediction model is trained using the optimal mask labels selected based on dynamic consistency test; The action generation module inputs the action dimension mask into the visual language action generation model, decodes the action dimensions to be decoded based on the action dimension mask, skips action dimensions that do not need to be decoded, and combines the decoded action dimensions with the skipped action dimensions to generate a complete execution action.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the reasoning method based on state-adaptive action space as described in any one of claims 1 to 7.

10. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the reasoning method based on state-adaptive action space as described in any one of claims 1 to 7.

Images