An Embodied Intelligent Control Method Based on Frequency Domain Sensing and State-Adaptive Timing Integration

Abstract

This invention discloses an embodied intelligent control method based on frequency domain perception and state-adaptive temporal integration. The method includes acquiring an expert demonstration dataset and collecting state data of the robotic arm; segmenting the motion intensity sequence using a penalized exact linear time algorithm, extracting frequency domain features from each segment and mapping them to the target motion block size; assigning weights to historical predicted actions in a temporal integration module based on a dynamic exponential decay rate and fusing them to generate original motion commands; and outputting the original motion commands separately after decoupling them from the control domain. This invention achieves adaptive dynamic adjustment of the motion block size, allowing for switching of the field of view length according to the operation stage without retraining the model, and is applicable to the motion generation and control of embodied intelligent robots.

Description

Technical Field

[0001] This invention relates to the field of embodied intelligent robot control technology, and in particular to an embodied intelligent control method based on frequency domain perception and state adaptive timing integration. Background Technology

[0002] During complex robot operations, the robot needs to generate a continuous sequence of motion commands based on perception information and task planning to complete multi-stage operations such as object grasping and precision insertion. Different operation stages have significantly different requirements for the control precision and field of view of motion generation. When the robot is in a contact-type operation stage such as grasping or precision insertion, the trajectory signal variance is large and the frequency abrupt changes are obvious. At this time, precise local motion generation is required to reduce open-loop cumulative error. When the robot is in a rough operation stage such as translation towards a target point, where the precision requirements are not high, the trajectory signal is relatively smooth. At this time, a wider field of view is needed to ensure the correctness of the movement direction and the smoothness of the trajectory. Fixed-size motion blocks cannot simultaneously meet the differentiated requirements of these multi-stage operations. Therefore, how to adaptively adjust the size of the motion blocks according to the trajectory characteristics of the robot's current operation stage has become a pressing technical problem to be solved in this field.

[0003] In existing embodied intelligent action block generation algorithms, the action block size is typically set based entirely on engineering experience and remains fixed throughout the task cycle. A typical approach, exemplified by the action block transformer model, outputs a predicted action sequence with a fixed step size each control cycle. The historical prediction results are then fused using a weighted average with a fixed exponential decay rate through a temporal ensemble module to achieve smooth action. While this fixed exponential weighting works to some extent in basic smoothing scenarios, it cannot accommodate the varying requirements for historical forgetting rates at different operational stages. Some other solutions attempt to directly truncate the action block size at the neural network prediction stage, achieving viewpoint switching by rigidly shortening the length of the transformer output sequence. However, this approach directly disrupts the complete motion primitives learned by the model during pre-training, causing the generated actions to lose physical coherence, and in severe cases, even leading to control system collapse.

[0004] The existing technologies described above have the following significant shortcomings: the fixed empirical values ​​for motion block sizes cannot adapt to the differentiated needs of different stages during robot operation; in complex contact-type operation stages, the excessive field of view leads to a significant increase in open-loop cumulative error; and in stable movement stages, insufficient field of view leads to a decrease in trajectory smoothness. Furthermore, the method of rigidly truncating motion block sizes destroys the integrity of the motion primitives in the pre-trained model, preventing the underlying neural network from achieving dynamic field-of-view switching without retraining. Therefore, it is necessary to design an embodied intelligent control method based on frequency domain perception and state-adaptive temporal integration to solve these problems. Summary of the Invention

[0005] The technical problem to be solved by the present invention is to provide an embodied intelligent control method based on frequency domain perception and state adaptive timing integration. It aims to overcome the problems of fixed action block size in the prior art, which cannot adapt to the differences in multi-stage operation and hard truncation destroys the integrity of the model motion primitives. It has the characteristics of objectively quantifying the intensity of operation stages based on frequency domain features and using dynamic attenuation soft mask to achieve retraining-free field of view switching.

[0006] To achieve the above technical solution, the technical solution adopted by the present invention is as follows: A method for generating dynamic action blocks based on frequency domain analysis includes the following steps: Step S1: Obtain the expert demonstration dataset and collect data on the robotic arm at time steps. ontology state data Ontology state data This includes the angles and angular velocities of the continuous spatial joints, the opening and closing parameters of the end effector in discrete states, and the total step size of the training data extracted from the expert demonstration dataset. ; Step S2, based on frequency domain prior dynamic vision generation, transforms the continuous joint action sequence into a one-dimensional discrete time sequence. The one-dimensional discrete time sequence is then segmented using a penalized exact linear time algorithm. Frequency domain features of each segment are extracted and mapped to the target action block size. The penalized exact linear-time algorithm iterates through all data in the dataset to obtain the average split point and average variance of each action segment. Step S3, based on the state-adaptive dynamic decay mechanism, in the timing integration module, according to the target action block size... Dynamically calculate the exponential decay rate at the current time step. Assign time-series index weights to historical predicted actions, and generate original action instructions through weighted average fusion; Step S4: Decouple the original motion command from the control domain, output the continuous spatial joint motion component after low-pass filtering, and output the discrete end effector motion component directly.

[0007] Preferably, in step S2, converting the continuous joint movement sequence into a one-dimensional discrete time sequence specifically includes: calculating the exercise intensity index. ; ; in, For time steps The joint position vector, This represents the total dimension of the joint after the end clamps have been removed.

[0008] Preferably, in step S2, the one-dimensional discrete-time series is segmented using a penalized exact linear-time algorithm, by solving: ; Obtain the optimal number of blocks and the segmentation position, where To the time The optimal total cost of partitioning so far. The cost function within a single slice is based on radial basis functions. This is a penalty item.

[0009] Preferably, in step S2, extracting the frequency domain features of each segment specifically includes: performing time-frequency domain transformation on each segmented action sequence, and extracting the peak frequency from the power spectrum. and variance Calculate the first The total kinetic energy of each action segment And the combined peak frequency: ; in For the number of joints, For joint indexing.

[0010] Preferably, in step S2, the mapping to the target action block size specifically includes: performing global extremum normalization on the integrated peak frequency and total motion energy respectively to obtain a normalized frequency index. and normalized energy index Calculate the overall intensity score: ; Will Mapped to target action block size: ; in and These are the frequency weighting coefficient and the energy weighting coefficient, respectively. This is the total weighting coefficient. and These are the preset minimum and maximum target action block lengths, respectively. This is the amplitude limiting function.

[0011] Preferably, the formula for calculating the normalized frequency index is: ; The formula for calculating the normalized energy index is as follows: ; in , These are the maximum and minimum composite peak frequencies across all segments, respectively. , These represent the maximum and minimum total kinetic energy among all segments.

[0012] Preferably, in step S3, the exponential decay rate of the current time step is dynamically calculated. The specific formula is as follows: ; in The preset base smooth decay rate, The total step size of the output sequence is fixed for the neural network.

[0013] Preferably, in step S3, assigning time-series index weights to historical predicted actions specifically includes: assigning weights to candidates in the candidate set based on their distance from the current time. for Every historical prediction action, according to Calculate its time-series exponential weights and normalize the exponential weights of all valid historical predicted actions. ; in The number of actions that can be effectively predicted from historical data.

[0014] Preferably, in step S4, the control domain decoupling specifically includes: separating the original motion command into continuous spatial joint motion components and discrete end effector motion components according to the robot arm's own decoupling algorithm; inputting the continuous spatial joint motion components into a preset low-pass filter for smoothing; and outputting the filtered spatial control command to the robot arm joint; the discrete end effector motion components are output directly without filtering; and the low-pass filter is a Kalman filter.

[0015] Preferably, an embodied intelligent control system based on frequency domain sensing and state adaptive timing integration is provided for executing the embodied intelligent control method based on frequency domain sensing and state adaptive timing integration. The system includes: The data acquisition module is used to acquire the expert demonstration dataset and collect data on the robotic arm at time steps. ontology state data The body state data This includes the angles and angular velocities of continuous spatial joints, as well as the opening and closing parameters of the end effector in discrete states. The dynamic field-of-view generation module, connected to the data acquisition module, is used to receive the body state data. The continuous joint motion sequence is transformed into a one-dimensional discrete-time sequence. The one-dimensional discrete-time sequence is then segmented using a penalized exact linear-time algorithm. The frequency domain features of each segment are extracted and mapped to the target motion block size. ; The adaptive attenuation module, connected to the dynamic field-of-view generation module, is used to receive the target action block size. In the timing integration module, according to Dynamically calculate the exponential decay rate at the current time step. Assign time-series index weights to historical predicted actions, and generate original action instructions through weighted average fusion; The control decoupling output module, connected to the adaptive attenuation module, is used to receive the original motion command, decouple the original motion command in the control domain, output the continuous spatial joint motion component to the robotic arm joint after low-pass filtering, and output the discrete end effector motion component directly to the end effector.

[0016] Furthermore, in step S3, the timing integration module performs the following steps at each control time step: The policy network outputs a full-length predicted action sequence, which is then written to the global historical action buffer. Simultaneously, the dynamic field-of-view length configuration table generated in step S2 is queried to obtain the current target action block size. .

[0017] Furthermore, the comprehensive intensity score The mapping principle is: The higher the value, the smaller the size of the output target action block; The lower the value, the larger the size of the target action block output.

[0018] The beneficial effects of this invention are as follows: 1. This invention transforms continuous joint motion sequences into one-dimensional discrete time sequences and uses a penalized exact linear time algorithm to objectively segment the operation trajectory. Combined with power spectrum estimation, it extracts the peak frequency and variance features of each segment from the frequency domain, realizing dynamic vision generation based on frequency domain priors. This solves the defect in the prior art where the motion block size is set entirely by experience and remains unchanged, enabling the motion block size to be adaptively adjusted according to the trajectory frequency domain characteristics of the robot's current operation stage.

[0019] 2. This invention introduces a dynamic decay mechanism based on exponential decay rate. Without changing the output structure of the policy network, it uses a smooth scaling mathematical model to calculate the target decay rate at the current time step. Through normalization processing, it forms a soft mask effect on the historical prediction weights, realizing dynamic view switching without retraining a large model. This solves the defect in the prior art where hard truncation of action block size will destroy the integrity of motion primitives.

[0020] 3. This invention decouples and separates the control domain of the original motion command, outputs the continuous spatial joint motion component after low-pass filtering to eliminate mechanical jitter caused by high-frequency switching, and outputs the discrete end effector motion component directly without filtering to ensure response speed. This solves the defect of the prior art that directly filters the complete motion command, resulting in end effector response delay. Attached Figure Description

[0021] Figure 1 This is a schematic diagram of the model framework of the present invention; Figure 2 This is a schematic diagram of the PELT segmentation result in Embodiment 2 of the present invention; Figure 3 This is a schematic diagram of the Kalman filtering effect in Example 2 of the present invention. Detailed Implementation

[0022] Example 1: like Figure 1 As shown, a method for generating dynamic action blocks based on frequency domain analysis includes the following steps: Step S1: Obtain the expert demonstration dataset and collect data on the robotic arm at time steps. ontology state data Ontology state data This includes the angles and angular velocities of the continuous spatial joints, the opening and closing parameters of the end effector in discrete states, and the total step size of the training data extracted from the expert demonstration dataset. ; Step S2, based on frequency domain prior dynamic vision generation, transforms the continuous joint action sequence into a one-dimensional discrete time sequence. The one-dimensional discrete time sequence is then segmented using a penalized exact linear time algorithm. Frequency domain features of each segment are extracted and mapped to the target action block size. The penalized exact linear-time algorithm iterates through all data in the dataset to obtain the average split point and average variance of each action segment. Step S3, based on the state-adaptive dynamic decay mechanism, in the timing integration module, according to the target action block size... Dynamically calculate the exponential decay rate at the current time step. Assign time-series index weights to historical predicted actions, and generate original action instructions through weighted average fusion; Step S4: Decouple the original motion command from the control domain, output the continuous spatial joint motion component after low-pass filtering, and output the discrete end effector motion component directly.

[0023] Preferably, in step S2, converting the continuous joint movement sequence into a one-dimensional discrete time sequence specifically includes: calculating the exercise intensity index. ; ; in, For time steps The joint position vector, This represents the total dimension of the joint after the end clamps have been removed.

[0024] Preferably, in step S2, the one-dimensional discrete-time series is segmented using a penalized exact linear-time algorithm, by solving: ; Obtain the optimal number of blocks and the segmentation position, where To the time The optimal total cost of partitioning so far. The cost function within a single slice is based on radial basis functions. This is a penalty item.

[0025] Preferably, in step S2, extracting the frequency domain features of each segment specifically includes: performing time-frequency domain transformation on each segmented action sequence, and extracting the peak frequency from the power spectrum. and variance Calculate the first The total kinetic energy of each action segment And the combined peak frequency: ; in For the number of joints, For joint indexing.

[0026] Preferably, in step S2, the mapping to the target action block size specifically includes: performing global extremum normalization on the integrated peak frequency and total motion energy respectively to obtain a normalized frequency index. and normalized energy index Calculate the overall intensity score: ; Will Mapped to target action block size: ; in and These are the frequency weighting coefficient and the energy weighting coefficient, respectively. This is the total weighting coefficient. and These are the preset minimum and maximum target action block lengths, respectively. This is the amplitude limiting function.

[0027] Preferably, the formula for calculating the normalized frequency index is: ; The formula for calculating the normalized energy index is as follows: ; in , These are the maximum and minimum composite peak frequencies across all segments, respectively. , These represent the maximum and minimum total kinetic energy among all segments.

[0028] Preferably, in step S3, the exponential decay rate of the current time step is dynamically calculated. The specific formula is as follows: ; in The preset base smooth decay rate, The total step size of the output sequence is fixed for the neural network.

[0029] Preferably, in step S3, assigning time-series index weights to historical predicted actions specifically includes: assigning weights to candidates in the candidate set based on their distance from the current time. for Every historical prediction action, according to Calculate its time-series exponential weights and normalize the exponential weights of all valid historical predicted actions. ; in The number of actions that can be effectively predicted from historical data.

[0030] Preferably, in step S4, the control domain decoupling specifically includes: separating the original motion command into continuous spatial joint motion components and discrete end effector motion components according to the robot arm's own decoupling algorithm; inputting the continuous spatial joint motion components into a preset low-pass filter for smoothing; and outputting the filtered spatial control command to the robot arm joint; the discrete end effector motion components are output directly without filtering; and the low-pass filter is a Kalman filter.

[0031] Preferably, an embodied intelligent control system based on frequency domain sensing and state adaptive timing integration is provided for executing the embodied intelligent control method based on frequency domain sensing and state adaptive timing integration. The system includes: The data acquisition module is used to acquire the expert demonstration dataset and collect data on the robotic arm at time steps. ontology state data The body state data This includes the angles and angular velocities of continuous spatial joints, as well as the opening and closing parameters of the end effector in discrete states. The dynamic field-of-view generation module, connected to the data acquisition module, is used to receive the body state data. The continuous joint motion sequence is transformed into a one-dimensional discrete-time sequence. The one-dimensional discrete-time sequence is then segmented using a penalized exact linear-time algorithm. The frequency domain features of each segment are extracted and mapped to the target motion block size. ; The adaptive attenuation module, connected to the dynamic field-of-view generation module, is used to receive the target action block size. In the timing integration module, according to Dynamically calculate the exponential decay rate at the current time step. Assign time-series index weights to historical predicted actions, and generate original action instructions through weighted average fusion; The control decoupling output module, connected to the adaptive attenuation module, is used to receive the original motion command, decouple the original motion command in the control domain, output the continuous spatial joint motion component to the robotic arm joint after low-pass filtering, and output the discrete end effector motion component directly to the end effector.

[0032] Furthermore, in step S3, the timing integration module performs the following steps at each control time step: The policy network outputs a full-length predicted action sequence, which is then written to the global historical action buffer. Simultaneously, the dynamic field-of-view length configuration table generated in step S2 is queried to obtain the current target action block size. .

[0033] Furthermore, the comprehensive intensity score The mapping principle is: The higher the value, the smaller the size of the output target action block; The lower the value, the larger the size of the target action block output.

[0034] Example 2: To objectively verify the authenticity and effectiveness of this invention, a comparative experiment was conducted with the classic imitation learning algorithm ACT in this embodiment to verify the effectiveness of the plug-and-play approach and the resulting performance improvement without increasing computational power requirements. Simultaneously, rigorous ablation experiments were performed on the method in this embodiment to ensure the effectiveness of each module.

[0035] This invention selects the classic block transfer task as the control task and does not modify the original ACT algorithm. Steps one and two are written in separate files and used through the main function of ACT. Step three only adds the state-adaptive dynamic timing integration module of this invention to the temporalagg judgment function. Step four adds a low-pass filter switch and corresponding function to the end of the original ACT code.

[0036] In this case, a Kalman filter was chosen as the low-pass filter. The power spectrum estimation method in step two uses the ARMA model, with an ACT fixed field-of-view planning value of 100 and an attenuation coefficient of 0.01. To avoid disrupting the underlying ACT network, the system presets the maximum and minimum target action block sizes (Cmin and Cmax) to 10 and 200, respectively. The task's random seed is set to 0, the training epochs are 1000, the batch size is 48, and the learning rate is 2e-5. The Kalman filter parameters are set to process noise Q = 1e-4 and measurement noise R = 1e-2.

[0037] In the block transfer task, the segmentation result of the dual-arm robotic arm using the PELT algorithm is as follows: Figure 2 As shown in the diagram. This illustrates the difference between using Kalman filtering and not using Kalman filtering in the cube transfer task. Figure 3 As shown.

[0038] Table 1: Comparative experimental results;

[0039] Table 2: Ablation Experiment Results;

[0040] The comparative experimental results for the block transfer task are shown in Table 1, and the ablation experimental results are shown in Table 2. As shown in Table 1, the adaptive ACT model, due to the absence of temporal, and the fact that the modified action chunksize in this invention is all within the temporal function, is essentially still an ACT model, with the success rate consistent with ACT. However, after enabling temporal, the success rate of this invention is significantly improved. This is because when transferring blocks between the two grippers, this invention benefits from the fine-tuning brought by the small action chunksize, greatly reducing collisions caused by insufficient accuracy. This comparative experiment also illustrates this well. This study demonstrates the benefits of using small action chunks for fine-grained tasks and large action chunks for coarse-grained tasks.

[0041] As shown in Table 2, all model components in this study are effective and indispensable. Enabling the Kalman filter alone provides limited benefits; while enabling the temporal filter alone leads to significant forgetting of past predictions when switching action block sizes. This can cause the robotic arm to replan, resulting in twitching and deviations from the target position, making the task more prone to failure. Therefore, both temporal and Kalman filters need to be enabled simultaneously to eliminate twitching while switching action block sizes.

Claims

1. A method for generating dynamic action blocks based on frequency domain analysis, characterized in that, Includes the following steps: Step S1: Obtain the expert demonstration dataset and collect data on the robotic arm at time steps. ontology state data Ontology state data This includes the angles and angular velocities of the continuous spatial joints, the opening and closing parameters of the end effector in discrete states, and the total step size of the training data extracted from the expert demonstration dataset. ; Step S2, based on frequency domain prior dynamic vision generation, transforms the continuous joint action sequence into a one-dimensional discrete time sequence. The one-dimensional discrete time sequence is then segmented using a penalized exact linear time algorithm. Frequency domain features of each segment are extracted and mapped to the target action block size. The penalized exact linear-time algorithm iterates through all data in the dataset to obtain the average split point and average variance of each action segment. Step S3, based on the state-adaptive dynamic decay mechanism, in the timing integration module, according to the target action block size... Dynamically calculate the exponential decay rate at the current time step. Assign time-series index weights to historical predicted actions, and generate original action instructions through weighted average fusion; Step S4: Decouple the original motion command from the control domain, output the continuous spatial joint motion component after low-pass filtering, and output the discrete end effector motion component directly.

2. The embodied intelligent control method based on frequency domain sensing and state adaptive timing integration according to claim 1, characterized in that, In step S2, converting the continuous joint movement sequence into a one-dimensional discrete time sequence specifically includes: calculating the exercise intensity index. ； ； in, For time steps The joint position vector, This represents the total dimension of the joint after the end clamps have been removed.

3. The embodied intelligent control method based on frequency domain sensing and state adaptive timing integration according to claim 2, characterized in that, In step S2, the one-dimensional discrete-time series is segmented using a penalized exact linear-time algorithm, by solving: ； Obtain the optimal number of blocks and the segmentation position, where To the time The optimal total cost of partitioning so far, The cost function within a single slice is based on radial basis functions. This is a penalty item.

4. The embodied intelligent control method based on frequency domain sensing and state adaptive timing integration according to claim 3, characterized in that, In step S2, extracting the frequency domain features of each segment specifically includes: performing time-frequency domain transformation on each segmented action sequence, and extracting the peak frequency from the power spectrum. and variance Calculate the first The total kinetic energy of each action segment And the combined peak frequency: ； in For the number of joints, For joint indexing.

5. The embodied intelligent control method based on frequency domain sensing and state adaptive timing integration according to claim 4, characterized in that, In step S2, the mapping to the target action block size specifically includes: performing global extremum normalization on the integrated peak frequency and total motion energy respectively to obtain a normalized frequency index. and normalized energy index Calculate the overall intensity score: ； Will Mapped to target action block size: ； in and These are the frequency weighting coefficient and the energy weighting coefficient, respectively. This is the total weighting coefficient. and These are the preset minimum and maximum target action block lengths, respectively. This is the amplitude limiting function.

6. The embodied intelligent control method based on frequency domain sensing and state adaptive timing integration according to claim 5, characterized in that, The formula for calculating the normalized frequency index is as follows: ； The formula for calculating the normalized energy index is as follows: ； in , These are the maximum and minimum composite peak frequencies across all segments, respectively. , These represent the maximum and minimum total kinetic energy among all segments.

7. The embodied intelligent control method based on frequency domain sensing and state adaptive timing integration according to claim 1, characterized in that, In step S3, the exponential decay rate of the current time step is dynamically calculated. The specific formula is as follows: ； in The preset base smooth decay rate, The total step size of the output sequence is fixed for the neural network.

8. The embodied intelligent control method based on frequency domain sensing and state adaptive timing integration according to claim 1, characterized in that, In step S3, assigning time-series index weights to historical predicted actions specifically includes: assigning weights to candidates based on their distance from the current time. for Every historical prediction action, according to Calculate its time-series exponential weights and normalize the exponential weights of all valid historical predicted actions. ； in The number of actions that can be effectively predicted from historical data.

9. The embodied intelligent control method based on frequency domain sensing and state adaptive timing integration according to claim 1, characterized in that, In step S4, the control domain decoupling specifically includes: separating the original motion command into continuous spatial joint motion components and discrete end effector motion components according to the robot arm's own decoupling algorithm; inputting the continuous spatial joint motion components into a preset low-pass filter for smoothing; and outputting the filtered spatial control command to the robot arm joint; the discrete end effector motion components are output directly without filtering; the low-pass filter is a Kalman filter.

10. An embodied intelligent control system based on frequency domain sensing and state adaptive timing integration, used to execute the embodied intelligent control method based on frequency domain sensing and state adaptive timing integration as described in any one of claims 1-9, characterized in that the system include: The data acquisition module is used to acquire the expert demonstration dataset and collect data on the robotic arm at time steps. ontology state data The body state data This includes the angles and angular velocities of continuous spatial joints, as well as the opening and closing parameters of the end effector in discrete states. The dynamic field-of-view generation module, connected to the data acquisition module, is used to receive the body state data. The continuous joint motion sequence is transformed into a one-dimensional discrete-time sequence. The one-dimensional discrete-time sequence is then segmented using a penalized exact linear-time algorithm. The frequency domain features of each segment are extracted and mapped to the target motion block size. ; The adaptive attenuation module, connected to the dynamic field-of-view generation module, is used to receive the target action block size. In the timing integration module, according to Dynamically calculate the exponential decay rate at the current time step. Assign time-series index weights to historical predicted actions, and generate original action instructions through weighted average fusion; The control decoupling output module, connected to the adaptive attenuation module, is used to receive the original motion command, decouple the original motion command in the control domain, output the continuous spatial joint motion component to the robotic arm joint after low-pass filtering, and output the discrete end effector motion component directly to the end effector.

Images