A method and system for speculative execution acceleration of robot control based on physical drives

Abstract

The application provides a robot control speculation execution acceleration method and system based on physical driving, and relates to the technical fields of artificial intelligence and robot control. The method first performs data-driven task criticality labeling based on an offline data set to obtain task criticality labels; then a multi-task draft model with residual dynamics prediction capability is trained; the draft model is a multi-task lightweight network, and outputs action trajectory prediction, residual latent dynamics prediction and step-by-step criticality score prediction; in the reasoning phase of robot task execution, after initial reasoning is performed by a VLA model, the draft model generates draft actions and performs double-gated verification, and the draft model and the VLA model are alternately used according to the verification result. The method of the application does not need to perform invasive modification on the VLA model architecture, and is suitable for various VLA model architectures and robot control tasks with different complexities.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and robot control technology, and in particular to a method and system for accelerating the speculative execution of robot control based on physical drive. Background Technology

[0002] Vision-Language-Action Models (VLA) directly transfer the perception and reasoning capabilities of pre-trained multimodal large models to robot manipulation tasks, enabling robots to understand natural language commands and perform diverse manipulations without requiring task-specific engineering. However, large-scale models, represented by 7B-parameter VLA models, suffer from a core bottleneck in terms of inference latency: the inference frequency on standard GPUs is only about 5Hz, while real-time fine-grained robot manipulation typically requires control frequencies of 50Hz or higher.

[0003] In the inference process of large language models, the process from user input to generating the first word is called the prefill stage, and the subsequent autoregressive word-by-word generation process is called the decoding stage. Speculative decoding is an effective paradigm for accelerating the decoding stage, using a draft model with fewer parameters to quickly predict several subsequent words, and then the original large model performs parallel verification on the candidate words. However, VLA models differ fundamentally from standard large language models: VLA models need to reset the context each time an action block is generated, re-performing the complete prefill process for multimodal inputs that include the current visual observation, resulting in inference latency far exceeding that of pure text large language models.

[0004] Existing VLA inference acceleration methods mainly include the following categories: model architecture and fine-tuning optimization methods improve inference throughput by integrating parallel decoding, action segmentation and other techniques, but require special modifications to the model architecture; dynamic inference strategies achieve adaptive acceleration through multi-exit early termination or visual lexical pruning; prefill stage caching acceleration utilizes the temporal continuity of visual input in adjacent frames to reuse KV cache representations; and decode stage speculative decoding introduces the speculative decoding framework into the autoregressive action generation of VLA.

[0005] The core limitation of the above method is that each action generated by the VLA model must go through the entire prefill-decode chain, and local optimization at only a certain stage cannot fundamentally overcome the latency bottleneck. In addition, in robot control, since action execution depends on the real evolution of the physical world, it is impossible to predict future observations after the execution of the draft action. Therefore, unlike speculative text decoding, it is not possible to directly use the original large model to verify the draft output afterward. Summary of the Invention

[0006] The technical problem to be solved by the present invention is to address the shortcomings of the prior art by providing a physical-driven speculative execution vision-language-action model inference acceleration method and system. This method achieves end-to-end acceleration by reducing the number of VLA model calls at the task global level, calling the VLA model only at critical moments of the task for fine-grained decision-making, while using a lightweight draft model to quickly generate and execute actions during normal stages.

[0007] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows:

[0008] On one hand, the present invention provides a method for accelerating speculative execution of robot control based on physical drive, comprising:

[0009] Based on offline datasets, data-driven task key annotation is performed to obtain task key labels;

[0010] A multi-task draft model with residual dynamics prediction capability is trained; the draft model is a multi-task lightweight network that takes the current observation and language command as input and outputs motion trajectory prediction, residual potential dynamics prediction and stepwise criticality score prediction; the overall training loss of the draft model is the weighted sum of the three prediction training losses.

[0011] During the reasoning phase of the robot's task execution, after the VLA model performs initial reasoning, the draft model generates draft actions and performs dual gating verification. Based on the verification results, the draft model and the VLA model are used alternately.

[0012] Furthermore, the specific method for obtaining task key labels by performing data-driven task key annotation based on offline datasets is as follows:

[0013] Based on an offline dataset, an image encoder is used to map observed images to a latent space, and image feature similarity is used to retrieve data at each time step. The nearest neighbor samples are used to calculate the action discreteness signal and motion precision signal, respectively. After normalization, they are weighted and fused to obtain the task key label.

[0014] Furthermore, the offline dataset is configured to contain... Each action trajectory, each time step Corresponding observations Relative motion to the end effector ,in, For current observations, For language instructions; a lightweight image encoder is used to map observations to a latent space, and retrieval is performed at each time step based on image feature similarity. The nearest neighbor samples are used as the neighborhood action distribution of the end effector;

[0015] Two complementary task-critical signals are extracted from the neighborhood motion distribution of the end effector: motion dispersion and motion precision. Motion dispersion measures the degree of dispersion of the motion distribution of the end effector in the local neighborhood. Motion precision captures the low-amplitude, high-sensitivity operation phase of the end effector based on the translation and rotation of the motion in each neighborhood.

[0016] The two types of task-critical signals are weighted and fused after robust Sigmoid normalization to obtain the final task-criticality score, which serves as the task-criticality label.

[0017] Furthermore, the predicted motion trajectory is for the future. The key action sequence of the step, the training loss is the masked L1 loss over the effective time step;

[0018] The residual latent dynamics prediction is the future state residual relative to the current observation, and the training loss combines cosine similarity and L2 norm;

[0019] Stepwise criticality score prediction uses task criticality labels as monitoring signals to predict future... The key scores at each time step are used, and the training loss is the mean squared error.

[0020] Furthermore, the dual-gated verification includes critical gating verification and dynamic gating verification. The critical gating verification actively identifies the fine operation stage based on the stepwise critical prediction to make the draft model yield to the VLA model in advance. The dynamic gating verification passively triggers a safety interruption and calls the VLA model to take over based on the dynamic deviation between the actual observation and the predicted state in the potential space.

[0021] Furthermore, the critical gating verification is an active prediction mechanism, with the signal originating from the progressive criticality prediction of the draft model itself. When the predicted value of the progressive criticality prediction exceeds a set threshold, it indicates that the task is about to enter the fine operation stage, and the robot control system actively transfers control to the VLA model. Simultaneously, before the robot executes the draft actions frame by frame, the robot control system pre-scans the sequence of critical actions for the next H steps to determine the maximum safe execution length. ,when At that time, in the Step call VLA model, if all If all draft actions pass the dynamic gating verification, then continue generating draft actions.

[0022] Furthermore, the dynamic gating verification is a passive detection mechanism. After each draft action is executed, the robot control system obtains the potential representation of the current real observation through a lightweight image encoder and calculates the dynamic deviation between the real observation and the predicted observation in the potential space. When the deviation exceeds a set threshold, a safety interruption is triggered and the VLA model is called to take over the inference task.

[0023] On the other hand, the present invention also provides a physics-driven robot control speculative execution acceleration system, including a task criticality annotation module, a multi-task draft model training module, a dual gating verification module, and a control switching module.

[0024] The task key labeling module automatically extracts action discreteness and motion precision signals based on offline data, and generates task key labels through weighted fusion.

[0025] The multi-task draft model training module is used to train a multi-task draft model with residual dynamics prediction capability. The draft model is a multi-task lightweight network that takes the current observation and language command as input and outputs motion trajectory prediction, residual potential dynamics prediction and stepwise criticality score prediction.

[0026] The dual-gated verification module includes a critical gating unit and a dynamic gating unit. The critical gating unit actively identifies the fine operation stage based on progressive critical predictions to relinquish power to the VLA model. The dynamic gating unit passively triggers a safety interruption based on the potential spatial dynamic deviation between the actual observation and the predicted state.

[0027] The control switching module is used to dynamically switch between VLA model inference mode and draft model generation mode.

[0028] Thirdly, this application proposes a computer-readable storage medium storing executable instructions that, when executed, cause a processor to perform the described physical-driven robot control speculative execution acceleration method.

[0029] Fourthly, this application proposes a computer program product, including a computer program or instructions that, when executed by a processor, implement the described physical-driven robot control speculative execution acceleration method.

[0030] The beneficial effects of adopting the above technical solutions are as follows: The physical-driven robot control speculative execution acceleration method and system provided by the present invention (1) reduces the number of VLA model calls from the global task level, breaks through the limitation of existing methods that only perform local optimization in a certain stage of prefill or decode, achieves end-to-end acceleration, significantly improves inference throughput while maintaining task success rate, and is applicable to various VLA model architectures and robot control tasks of different complexities; (2) introduces a physical-driven dual-gating verification mechanism, through active prediction of key gating verification and passive detection of dynamic gating verification, achieves reliable physical verification without relying on forward inference of VLA model, and ensures that the execution of draft model is always within a safe and controllable range; (3) does not require intrusive modification of VLA model architecture, has good model independence and universality, and is applicable to VLA models with different parameter scales and architectures; (4) the task key annotation method does not require additional manual annotation, can be automatically extracted from offline data, and has good scalability; (5) in long-term continuous execution scenarios, the dual-gating verification mechanism can effectively suppress error accumulation and maintain the stability of task success rate. Attached Figure Description

[0031] Figure 1 The flowchart illustrates the physics-driven robot control speculative execution acceleration method provided in Embodiment 1 of the present invention. The figure shows the working principle of the lightweight draft model responsible for rapidly generating action sequences and simultaneously predicting residual potential dynamics and progressive criticality scores, as well as the two independent gating mechanisms of criticality gating and dynamics gating.

[0032] Figure 2 This is a schematic diagram of the data-driven task-critical annotation process provided in Embodiment 1 of the present invention; the diagram shows the process of extracting two types of critical signals, motion discreteness and motion fineness, from offline teaching data and performing robust normalization fusion;

[0033] Figure 3 This is a schematic diagram of the network structure of the multi-task draft model provided in Embodiment 1 of the present invention; the figure shows the structure of the draft model that simultaneously outputs three branches: motion trajectory prediction, residual potential dynamics prediction, and stepwise criticality score prediction;

[0034] Figure 4 This is a schematic diagram of the dual-gated verification reasoning process provided in Embodiment 1 of the present invention; the diagram shows the dynamic switching logic between the VLA model and the draft model, as well as the triggering mechanism of critical gating and dynamic gating. Detailed Implementation

[0035] The specific embodiments of the present invention will be described in further detail below with reference to the accompanying drawings and examples. The following examples are for illustrative purposes only and are not intended to limit the scope of the invention.

[0036] Example 1:

[0037] This embodiment provides a method for accelerating speculative execution in robot control based on physical drive. See [link to relevant documentation]. Figures 1 to 4 The specific steps are as follows:

[0038] Step 1: Data-driven task key annotation; Based on the offline dataset, an image encoder is used to map the observed images to the latent space, and image feature similarity is used to retrieve key information at each time step. The nearest neighbor samples are used to calculate the action discreteness signal and motion precision signal respectively, and after normalization, they are weighted and fused to obtain the task key label;

[0039] In physical manipulation tasks, the decision-making difficulty and control sensitivity vary significantly across different time steps. Certain stages, such as free-space motion, have lower requirements for strategy precision, while fine-grained operation stages such as insertion, alignment, and contact transition are extremely sensitive to errors. This embodiment automatically extracts key task signals from offline data to obtain key task labels, eliminating the need for additional manual annotation.

[0040] Suppose the offline dataset contains Each action trajectory, each time step Corresponding observations Relative motion to the end effector ,in, For current observations, For language instructions. A lightweight image encoder maps observations to a latent space, retrieving them at each time step based on image feature similarity. The nearest neighbor samples are used as the neighborhood action distribution of the end effector.

[0041] Two complementary task-critical signals, motion discreteness and motion fineness, are extracted from the neighborhood motion distribution of the end effector:

[0042] Motion dispersion measures the degree of dispersion of motion distribution of an end effector within a local neighborhood:

[0043] ;

[0044] in, For action discreteness, The average of neighborhood actions. For the first Each neighborhood action.

[0045] The higher the action dispersion, the more likely there are multiple feasible control strategies with significant differences in that state. The higher the decision ambiguity, the more likely the lightweight draft model is to produce prediction bias.

[0046] Motion precision capture end effector in low-amplitude, high-sensitivity operation phase:

[0047] ;

[0048] in, For motion precision, Let the translation and rotation components of the j-th domain action be... To prevent division by zero of extremely small constants;

[0049] When the end effector is in the fine adjustment stage, the movement amplitude is extremely small, and the motion precision is high.

[0050] The two types of signals are complementary: motion dispersion captures highly ambiguous decision points, while motion precision captures low-amplitude, highly sensitive operational phases; after robust sigmoid normalization, the two types of signals are weighted and fused to obtain the final task criticality score, which serves as the task criticality label.

[0051] In this embodiment, both types of key signals undergo robust Sigmoid normalization, specifically as follows:

[0052] ;

[0053] in, Indicates a key signal, The key signal after normalization, g represents disp or fine; The Sigmoid normalization function, For the clipping function, and These are the upper and lower quantile cutoff values ​​for the key signal, respectively. The median of the sequence after the key signal has been truncated to the percentile. Interquartile range;

[0054] The final task criticality score is used as the task criticality label, as shown in the following formula:

[0055] ;

[0056] in, As the score for the task's criticality, and These are the weighting coefficients for motion dispersion and motion precision, respectively.

[0057] Step 2: Train a multi-task draft model with residual dynamics prediction capabilities;

[0058] Draft model This is a lightweight, multi-task network; in this embodiment, the draft model uses an Action Chunking Transformer (ACT) model based on Transformer. ACT will display the current observations... With language instructions Joint coding is performed, where the current observation This can include one or more of the following: image observation, robot body state, end effector state, historical actions, or environmental state; and language commands. Used to describe the current task objective or operational constraints. ACT extracts cross-modal correlations between observation information and linguistic semantics through a Transformer encoder, and generates motion trajectory predictions, residual latent dynamics predictions, and stepwise criticality score predictions in parallel using a Transformer decoder or action query vectors for a future period.

[0059] It should be noted that ACT is only a draft model in this embodiment. This is one exemplary implementation; the actual draft model is not limited to the ACT structure. Anything that can be observed at the current time... and language instructions Any model that takes motion trajectory prediction, residual latent dynamics prediction, and stepwise criticality score prediction as input can be used as a draft model in this application. For example, the draft model can also use a lightweight Transformer, a diffusion policy network, a recurrent neural network, a temporal convolutional network, a hybrid expert network, a multilayer perceptron policy network, or any combination or variation of the above networks.

[0060] Among them, motion trajectory prediction is for the future. Key movement sequence of steps The training loss is the mask L1 loss over the effective time steps;

[0061] The residual latent dynamics prediction is the future state residual relative to the current observation. The training loss combines cosine similarity and L2 norm. The residual dynamics prediction can compare the draft model's expectation of the future physical state with the actual observation after execution, providing a physical verification signal for subsequent dynamic gating.

[0062] Stepwise criticality score prediction uses task criticality labels as monitoring signals to predict future... Key scores for each time step The training loss is the mean squared error.

[0063] The overall training loss for training the draft model is the weighted sum of the training losses of the three predictions.

[0064] In this embodiment, the training loss for motion trajectory prediction is:

[0065] ;

[0066] in, The training loss is for motion trajectory prediction. To predict actions, For actual actions, For the effective time step set;

[0067] The training loss for residual latent dynamics prediction is:

[0068] ;

[0069] in, The training loss is used to predict the residual latent dynamics. and All of these are hyperparameters. and These are the potential observations of the predicted residuals and the potential observations of the actual residuals, respectively.

[0070] The training loss for stepwise critical score prediction is:

[0071] ;

[0072] in, The training loss is used for progressive key score prediction. To predict the criticality score of the task, The score is based on the key performance indicator of the actual task.

[0073] The overall training loss of the draft model is:

[0074] ;

[0075] in, , , These are the weight coefficients of the training loss for the three predictions.

[0076] Step 3: Physically driven dual-gated verification pipeline; During the reasoning phase of the robot's task execution, after the VLA model performs initial reasoning, the draft model quickly generates draft actions and performs dual-gated verification. Based on the verification results, the draft model and VLA model are used alternately. Dual-gated verification includes critical gating verification and dynamic gating verification. Critical gating verification actively identifies the fine operation stage based on progressive critical predictions to allow the draft model to yield to the VLA model in advance. Dynamic gating verification passively triggers a safety interruption and calls the VLA model to take over based on the dynamic deviation between the actual observation and predicted states in the potential space.

[0077] During the reasoning phase of the robot's task execution, the robot control system dynamically switches between draft mode and VLA takeover mode based on dual gating verification. In draft mode, the draft model generates draft actions with extremely low latency, and the robot executes the draft actions frame by frame. In VLA takeover mode, the VLA model is awakened to perform complete visual-language-action reasoning for fine-grained decision-making.

[0078] The critical gating verification is an active prediction mechanism, with signals originating from the progressive criticality predictions of the draft model itself. When the predicted value of the progressive criticality prediction exceeds a threshold, it indicates that the task is about to enter the fine-tuning stage, and the robot control system actively transfers control to the VLA model. Before the robot executes the draft actions frame by frame, the robot control system pre-scans the sequence of critical actions for the next H steps to determine the maximum safe execution length.

[0079] In this embodiment, the trigger condition for criticality gating is that the criticality score of the predicted task exceeds a set threshold. The maximum safe execution length is determined as follows:

[0080] ;

[0081] in, Maximum safe execution length;

[0082] judge Is it less than If not, in The step actively calls the VLA model to take over, if all If all draft actions pass the dynamic gating verification, then continue generating draft actions;

[0083] The dynamic gating verification is a passive detection mechanism. After executing each draft action, the robot control system acquires the latent representation of the current real observation through a lightweight image encoder and calculates the dynamic deviation between the real and predicted observations in the latent space. When the deviation exceeds a set threshold... When this occurs, a safety interrupt is triggered, and the VLA model is invoked to take over the inference task.

[0084] In this embodiment, the dynamic deviation between actual observations and predicted observations in the potential space is considered during the dynamic gating verification. The calculation is as follows:

[0085] ;

[0086] in, This is the latent representation of the current true observations obtained through the image encoder. The potential states predicted by the draft model;

[0087] when When this occurs, a safety interrupt is triggered, and the VLA model is invoked to take over the inference task.

[0088] The two gating verifications correspond to two complementary safety mechanisms: "anticipating risks and proactively avoiding them" and "perceiving deviations and correcting them in a timely manner," which together ensure that the execution of the draft model is always within a controllable range.

[0089] In this embodiment, the overall reasoning process of the robot performing the task Figure 1 as follows:

[0090] Step S1: The robot control system uses the VLA model to perform a complete reasoning of the robot's initial state and outputs the action. The robot executes the initial action, and after the initial action is completed, proceed to step S2.

[0091] Step S2: Generate the future using the draft model Step-by-step draft motion sequence, residual potential dynamics prediction, and step-by-step criticality score prediction; the robot control system pre-scans the critical motion sequence to determine the maximum safe motion execution length. ;

[0092] Step S3: The robot executes draft actions frame by frame; after each draft action, the real observation is encoded into a latent state, and the dynamic deviation between the real observation and the predicted observation in the latent space is calculated; if the deviation exceeds the threshold, dynamic gating verification is triggered, and the VLA model is immediately invoked to take over the task inference.

[0093] Step S4: Determine Is it less than If not, in The step actively calls the VLA model to take over, if all If all draft actions pass the dynamic gating verification, return to step 2 to continue generating draft actions.

[0094] This reasoning process ensures that the VLA model is invoked only at critical moments of the mission or when dynamic deviations exceed limits, fundamentally reducing the number of times the VLA model is invoked and elevating the acceleration effect from single-stage local optimization to the global mission level.

[0095] Example 2:

[0096] This embodiment provides a physics-driven robot control speculative execution acceleration system, including a task criticality annotation module, a multi-task draft model training module, a dual-gating verification module, and a control switching module;

[0097] The task key labeling module automatically extracts action discreteness and motion precision signals based on offline data, and generates task key labels through weighted fusion.

[0098] The multi-task draft model training module is used to train a multi-task draft model with residual dynamics prediction capability. The draft model is a multi-task lightweight network that takes the current observation and language command as input and outputs motion trajectory prediction, residual potential dynamics prediction and stepwise criticality score prediction.

[0099] The dual-gated verification module includes a critical gating unit and a dynamic gating unit. The critical gating unit actively identifies the fine operation stage based on progressive critical predictions to relinquish power to the VLA model. The dynamic gating unit passively triggers a safety interruption based on the potential spatial dynamic deviation between the actual observation and the predicted state.

[0100] The control switching module is used to dynamically switch between VLA model inference mode and draft model rapid generation mode, so that the VLA model is only invoked at critical moments of the task or when the dynamic deviation exceeds the limit.

[0101] Example 3:

[0102] This embodiment proposes a computer-readable storage medium that stores executable instructions. When these instructions are executed, if they are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium.

[0103] The computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the physical-driven robot control speculative execution acceleration method described in the various embodiments of this application.

[0104] The aforementioned storage media include: flash memory, hard disks, multimedia cards, card-type memory (e.g., SD (Secure Digital Memory Card) or DX (Memory Data Register, MDR) memory), random access memory (RAM), static random-access memory (SRAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), programmable read-only memory (PROM), magnetic storage, disks, optical discs, servers, APP (Application) application stores, and other media capable of storing program verification codes. These media store computer programs, which, when executed by a processor, can implement the various steps of the aforementioned physical-driven robot control speculative execution acceleration method.

[0105] Example 4:

[0106] This embodiment proposes a computer program product, including a computer program or instructions, which, when executed by a processor, implements the aforementioned physical-driven robot control speculative execution acceleration method.

[0107] Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or part of the technical solution, can be embodied in the form of a computer program product.

[0108] The various embodiments in this application are described in a progressive manner. The same or similar parts between the various embodiments can be referred to each other. Each embodiment focuses on describing the differences from other embodiments.

[0109] 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 or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope defined by the present invention.

Claims

1. A method for accelerating speculative execution in robot control based on physical drive, characterized in that, include: Based on offline datasets, data-driven task key annotation is performed to obtain task key labels; Train a multi-task draft model with residual dynamics prediction capabilities; The draft model is a multi-task lightweight network that takes the current observation and language commands as input and outputs motion trajectory prediction, residual potential dynamics prediction and stepwise criticality score prediction. The overall training loss for training the draft model is a weighted sum of the three predicted training losses; During the reasoning phase of the robot's task execution, after the VLA model performs the initial reasoning, the draft model generates draft actions and performs dual gating verification. Based on the verification results, the draft model and the VLA model are used alternately. The dual-gated verification includes critical gating verification and dynamic gating verification. Critical gating verification actively identifies the fine operation stage based on progressive critical predictions to allow the draft model to yield power to the VLA model in advance. Dynamic gating verification passively triggers a safety interruption and calls the VLA model to take over based on the dynamic deviation between the actual observation and the predicted state in the potential space. The critical gating verification is an active prediction mechanism, with signals originating from the progressive criticality predictions of the draft model itself. When the predicted value of the progressive criticality prediction exceeds a set threshold, it indicates that the task is about to enter the fine-tuning stage, and the robot control system actively transfers control to the VLA model. Simultaneously, before the robot executes the draft actions frame by frame, the robot control system pre-scans the sequence of critical actions for the next H steps to determine the maximum safe execution length. ,when At that time, in the Step call VLA model, if all If all draft actions pass the dynamic gating verification, then continue generating draft actions; The dynamic gating verification is a passive detection mechanism. After each draft action is executed, the robot control system obtains the potential representation of the current real observation through a lightweight image encoder and calculates the dynamic deviation between the real observation and the predicted observation in the potential space. When the deviation exceeds a set threshold, a safety interruption is triggered and the VLA model is called to take over the inference task.

2. The method for accelerating speculative execution of robot control based on physical drive according to claim 1, characterized in that, The specific method for data-driven task key annotation based on offline datasets to obtain task key labels is as follows: Based on an offline dataset, an image encoder is used to map observed images to a latent space, and image feature similarity is used to retrieve data at each time step. The nearest neighbor samples are used to calculate the action discreteness signal and motion precision signal, respectively. After normalization, they are weighted and fused to obtain the task key label.

3. The method for accelerating speculative execution of robot control based on physical drive according to claim 2, characterized in that, The method sets the offline dataset to include Each action trajectory, each time step Corresponding observations Relative motion to the end effector ,in, For current observations, For language instructions; a lightweight image encoder is used to map observations to a latent space, and retrieval is performed at each time step based on image feature similarity. The nearest neighbor samples are used as the neighborhood action distribution of the end effector; Two complementary task-critical signals are extracted from the neighborhood motion distribution of the end effector: motion dispersion and motion precision. Motion dispersion measures the degree of dispersion of the motion distribution of the end effector in the local neighborhood. Motion precision captures the low-amplitude, high-sensitivity operation phase of the end effector based on the translation and rotation of the motion in each neighborhood. The two types of task-critical signals are weighted and fused after robust Sigmoid normalization to obtain the final task-criticality score, which serves as the task-criticality label.

4. The method for accelerating speculative execution of robot control based on physical drive according to claim 3, characterized in that, The predicted motion trajectory is for the future. The key action sequence of the step, the training loss is the masked L1 loss over the effective time step; The residual latent dynamics prediction is the future state residual relative to the current observation, and the training loss combines cosine similarity and L2 norm; Stepwise criticality score prediction uses task criticality labels as monitoring signals to predict future... The key scores at each time step are used, and the training loss is the mean squared error.

5. A physics-driven robot control speculative execution acceleration system, based on the physics-driven robot control speculative execution acceleration method of claim 1, characterized in that, It includes a task-critical annotation module, a multi-task draft model training module, a dual-gated verification module, and a control switching module; The task key labeling module automatically extracts action discreteness and motion precision signals based on offline data, and generates task key labels through weighted fusion. The multi-task draft model training module is used to train a multi-task draft model with residual dynamics prediction capability. The draft model is a multi-task lightweight network that takes the current observation and language command as input and outputs motion trajectory prediction, residual potential dynamics prediction and stepwise criticality score prediction. The dual-gated verification module includes a critical gating unit and a dynamic gating unit. The critical gating unit actively identifies the fine operation stage based on progressive critical predictions to relinquish power to the VLA model. The dynamic gating unit passively triggers a safety interruption based on the potential spatial dynamic deviation between the actual observation and the predicted state. The control switching module is used to dynamically switch between VLA model inference mode and draft model generation mode. The dual-gated verification includes critical gating verification and dynamic gating verification. Critical gating verification actively identifies the fine operation stage based on progressive critical predictions to allow the draft model to yield power to the VLA model in advance. Dynamic gating verification passively triggers a safety interruption and calls the VLA model to take over based on the dynamic deviation between the actual observation and the predicted state in the potential space. The critical gating verification is an active prediction mechanism, with signals originating from the progressive criticality predictions of the draft model itself. When the predicted value of the progressive criticality prediction exceeds a set threshold, it indicates that the task is about to enter the fine-tuning stage, and the robot control system actively transfers control to the VLA model. Simultaneously, before the robot executes the draft actions frame by frame, the robot control system pre-scans the sequence of critical actions for the next H steps to determine the maximum safe execution length. ,when At that time, in the Step call VLA model, if all If all draft actions pass the dynamic gating verification, then continue generating draft actions; The dynamic gating verification is a passive detection mechanism. After each draft action is executed, the robot control system obtains the potential representation of the current real observation through a lightweight image encoder and calculates the dynamic deviation between the real observation and the predicted observation in the potential space. When the deviation exceeds a set threshold, a safety interruption is triggered and the VLA model is called to take over the inference task.

6. A computer-readable storage medium storing executable instructions for performing the physical-driven robot control speculative execution acceleration method of claim 1, characterized in that, When the instruction is executed, it causes the processor to perform the described physics-driven robot control speculative execution acceleration method.

7. A computer program product for executing the physical-driven robot control speculative execution acceleration method of claim 1, characterized in that, This includes a computer program or instructions that, when executed by a processor, implement the described physical-driven robot control speculative execution acceleration method.

Images