A Deep Learning-Based Method for Robot Action Intent Recognition

Abstract

This invention discloses a deep learning-based method for recognizing robot action intentions, comprising the following steps: collecting multimodal perception data of the robot during task execution; performing semantic analysis on task context information and voice command data; constructing a semantically guided differentiable neural computer model, inputting the multimodal input feature vector and the memory vector read in the previous round into the controller submodule; updating the short-term memory matrix and long-term memory matrix in the write controller submodule; reading content from the short-term memory slot and long-term memory slot in the read controller submodule; inputting the control signal vector, read vector, and task semantic label vector into the decoding network; dynamically adjusting the semantic priority threshold in the write controller submodule and updating the utilization of the memory slots. This invention employs a semantically guided differentiable neural computer to achieve high-precision recognition of robot action intentions.

Description

Technical Field

[0001] This invention relates to the field of robot intelligent control and deep learning technology, and in particular to a method for recognizing robot action intentions based on deep learning. Background Technology

[0002] In existing technologies, robot action intent recognition methods primarily rely on predefined rules, traditional machine learning models, or single deep learning architectures to process robot sensor data or task-related input information. These methods typically process acceleration, gyroscope, and visual sensor data through fixed rules or feature engineering extraction techniques, and then input this data into a classification model for intent recognition. However, these methods have weak adaptability to dynamic environments and struggle to effectively handle changes in task semantics or long-term dependencies on historical operations. Furthermore, while existing deep learning-based robot control systems can improve feature representation capabilities to some extent, they still have limitations in handling multimodal inputs, long short-term memory coupling, cross-task semantic representation, and dynamic decision-making reasoning.

[0003] Existing memory-based neural networks (such as LSTM or GRU) struggle to achieve task-aware control of explicit memory writing strategies. In complex task scenarios, the lack of semantically driven memory update mechanisms leads to decreased intent recognition accuracy and delayed or ineffective control strategies.

[0004] At the same time, some systems only extract features from the input information of the current round, without combining the robot's historical actions, environmental background and task labels for correlation reasoning, which limits the model's ability to understand complex long-term behavior.

[0005] Therefore, how to provide a deep learning-based method for recognizing robot action intentions is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0006] One objective of this invention is to propose a deep learning-based method for robot action intent recognition. This invention utilizes a semantically guided differentiable neural computer as its core support, based on a deep learning model, and integrates multimodal sensor data, task semantic labels, and historical behavioral states of the robot. It achieves efficient action intent reasoning through a semantically guided dynamic memory writing and reading strategy. This invention details the key steps of multimodal input feature fusion, controller signal generation, a semantically priority-driven memory writing control mechanism, a memory reading method based on content and semantic joint addressing, and intent and confidence decoding. It possesses advantages such as high action recognition accuracy, good task transfer adaptability, and high interpretability of intent expression, effectively enhancing the robot's interactive intelligence and autonomous decision-making capabilities in dynamic environments.

[0007] A robot action intent recognition method based on deep learning according to an embodiment of the present invention includes the following steps:

[0008] Collect multimodal perception data of the robot during task execution, and extract features from each modal data to generate multimodal input feature vectors;

[0009] Semantic analysis is performed on task context information and voice command data to obtain task semantic label vectors;

[0010] A semantically guided differentiable neural computer model is constructed. The multimodal input feature vector and the memory vector read in the previous round are input into the controller submodule to generate the control signal vector and interface vector set for the current round.

[0011] In the write controller submodule, the semantic priority of each memory slot is calculated based on the control signal vector and the task semantic tag vector, and the short-term memory matrix and long-term memory matrix are updated.

[0012] In the read controller submodule, the addressing weight is calculated based on content similarity and semantic matching weight, and the content is read from the short-term memory slot and the long-term memory slot to obtain a unified read vector;

[0013] The control signal vector, reading vector, and task semantic label vector are concatenated and input into the decoding network, which outputs the current action intent recognition result and intent confidence.

[0014] Based on the action intent recognition results and intent confidence, the semantic priority threshold written to the controller submodule is dynamically adjusted, and the memory slot usage is updated.

[0015] Optionally, the process of collecting multimodal perception data of the robot during task execution and extracting features from each modality of data to generate a multimodal input feature vector specifically includes:

[0016] Collect multimodal perception data of the robot during task execution, including motion sensor data, environmental image data, voice command data, and task context information;

[0017] The motion sensor data is processed, including the three-axis acceleration vector collected by the accelerometer, the three-axis angular velocity vector collected by the gyroscope, and the angles and angular velocities of each joint. The robot's posture change information is calculated by low-pass filtering and numerical integration, and then combined with the current motion parameters of the robotic arm to construct the motion state vector.

[0018] The environmental image data is processed to obtain the image frame sequence in the robot's working environment. A convolutional neural network is used to extract features from the images and output the environmental image feature vector.

[0019] The voice command data is processed by using a speech recognition module to convert the voice signal into a text command, and then encoding the text command through a language model to generate a voice command vector.

[0020] The task context information, including the current task objective, task type, and historical task tags, is processed. Semantic information is extracted through word vector embedding and sequence modeling to generate a task context vector.

[0021] The action state vector, environmental image feature vector, voice command vector, and task context vector are concatenated to form a multimodal input feature vector.

[0022] Optionally, the step of performing semantic analysis on the task context information and voice command data to obtain the task semantic label vector specifically includes:

[0023] The task context information is processed by text processing. The task context information includes the task target name, task type description, historical task tags and task execution status. A word segmentation algorithm based on dictionary matching and statistical rules is used to segment the task context information to obtain a context word segmentation sequence.

[0024] Entity recognition is performed on the context segmented word sequence. The named entity recognition model is used to classify each word in the segmented word sequence by label, and the task type entity, target object entity, and historical state entity are identified. The recognition result is then represented as a structured context sequence.

[0025] The voice command data is processed to convert speech into text. The speech recognition engine is used to convert the input speech signal into a text command string. The same entity recognition process is used to parse the speech text, extract the task semantic fields in the speech command, and generate a speech command semantic sequence.

[0026] The structured context sequence is merged with the semantic sequence of voice commands to form a semantic input sequence, which is then passed through a semantic coding network to output a context-aware semantic embedding vector.

[0027] Attention-weighted pooling is applied to the semantic embedding vector sequence to filter out vector fragments with high semantic relevance in the current task, and task semantic label vectors are constructed.

[0028] Optionally, the construction of the semantically guided differentiable neural computer model, which involves inputting the multimodal input feature vector and the memory vector read in the previous round into the controller submodule to generate the control signal vector and interface vector set for the current round, specifically includes:

[0029] The multimodal input feature vector of the current round, the memory vector read from the hierarchical memory matrix of the previous round, and the semantic label vector corresponding to the current task are obtained. The above three vectors are concatenated in a fixed order to form the controller input vector. The memory vector is a unified reading vector.

[0030] A semantically guided differentiable neural computer model is constructed. The semantically guided differentiable neural computer model includes a controller submodule, a write controller submodule, a read controller submodule, and a hierarchical memory matrix. The hierarchical memory matrix includes a short-term memory matrix and a long-term memory matrix, both of which are composed of several memory slots.

[0031] The controller input vector is input to the controller submodule for processing. The controller submodule is a feedforward neural network with a gating mechanism, consisting of two linear mapping layers and one nonlinear activation layer. The first linear mapping layer transforms the dimension of the input vector to adapt to the neural network structure. Then, the nonlinear transformation of the ReLU activation function is integrated to enhance the feature representation capability. The second linear mapping layer then maps the intermediate result after nonlinear transformation into a control signal vector.

[0032] Based on the control signal vector output by the controller submodule, an interface vector set is generated through multiple sub-mapping networks with independent parameter structures.

[0033] Optionally, the step of generating an interface vector set by sequentially passing the control signal vector output by the controller submodule through multiple sub-mapping networks with independent parameter structures specifically includes:

[0034] The process involves generating read and write key vectors for controlling read and write operations of the memory matrix based on the control signal vector. Specifically, the control signal vector is input into a linear transformation layer for key vector generation, and the read key vector is obtained by weighted summation and bias correction.

[0035] The same control signal vector is input into another linear transformation layer, and combined with the processing of a nonlinear activation function, the write key vector is obtained.

[0036] Based on the control signal vector, read strength scalars and write strength scalars for adjusting the intensity of memory operations are generated. Specifically, the control signal vector is input to the intensity coefficient generation unit, and through a combination of linear transformation and nonlinear activation function, read strength scalars for reading memory content and write strength scalars for writing memory content are obtained respectively.

[0037] The erase vector and add vector for updating memory content are generated based on the control signal vector. Specifically, the control signal vector is input to the erase vector generation unit, and after being processed by linear mapping and normalization function, the erase vector with a value range between 0 and 1 is obtained.

[0038] The control signal vector is input to the vector generation unit, and after being processed by linear transformation and nonlinear activation function, the vector with a value range between -1 and 1 is obtained.

[0039] The allocation gating coefficient vector is generated based on the control signal vector to guide the memory write allocation strategy. Specifically, the control signal vector is input to the allocation coefficient generation unit, and after processing by linear mapping and restricted activation function, the allocation gating coefficient vector is obtained. Each component of the allocation gating coefficient vector takes a value between 0 and 1.

[0040] The usage vector for recording the historical usage frequency of the memory slot is generated based on the control signal vector. Specifically, the control signal vector is input to the usage generation unit, processed by linear transformation and nonlinear activation function to obtain the usage vector, and the values ​​of each component are constrained to be non-negative real numbers.

[0041] The generated read key vector, read strength scalar, write key vector, write strength scalar, erase vector, add vector, allocation gating coefficient vector, and usage vector are combined sequentially to form a complete set of interface vectors.

[0042] Optionally, in the write controller submodule, calculating the semantic priority of each memory slot based on the control signal vector and the task semantic tag vector, and updating the short-term memory matrix and the long-term memory matrix specifically includes:

[0043] Calculate the semantic priority score for each memory slot based on the control signal vector and the task semantic label vector;

[0044] The semantic priority scores of all memory slots are normalized to generate a normalized priority weight vector;

[0045] The normalization process includes performing element-wise division on the semantic priority score of each memory slot generated in the previous step, so that the sum equals 1, thereby generating a normalized priority weight vector. The value of each component in the priority weight vector is between 0 and 1, and the sum of the components is 1.

[0046] The normalized priority weight vector is combined with the allocation gating coefficient vector to generate write allocation weights for short-term memory slots and long-term memory slots.

[0047] Specifically, this involves allocating the overall write ratio to short-term memory slots and long-term memory slots based on the allocation gating coefficient, multiplying the portion of the normalized priority weight vector corresponding to the short-term memory slot with the short-term write gating coefficient, and obtaining the write allocation weight for each short-term memory slot.

[0048] Multiply the portion of the normalized priority weight vector corresponding to the long memory slot by the long write gating coefficient to obtain the write allocation weight for each long memory slot, where the sum of the short write gating coefficient and the long write gating coefficient is 1.

[0049] Based on the write key vector and write strength scalar in the interface vector set, calculate the write addressing weight vectors of short-term memory slots and long-term memory slots;

[0050] Specifically, this involves extracting write key vectors and write strength scalars from the interface vector set, and calculating the similarity between the write key vectors and each memory slot in the short-term memory matrix and long-term memory matrix to obtain the similarity vector between the short-term and long-term memory slots.

[0051] Each similarity vector is multiplied by a write strength scalar to adjust its influence on the write operation. The above write influence values ​​are normalized to obtain the write addressing weight vectors for short-term memory and long-term memory.

[0052] Based on the erase vector and write allocation weight generated in the interface vector set, differentiable erase operations are performed on the short-term memory matrix and the long-term memory matrix respectively.

[0053] The differentiable erasure operation includes performing a slot-level combination operation on the erasure vector and the corresponding write allocation weight to obtain a slot-by-slot erasure coefficient matrix.

[0054] Element-level scaling is performed on each slot in the short-term memory matrix and long-term memory matrix according to the corresponding erasure coefficient, thereby achieving partial erasure of existing memory content;

[0055] After the erasure operation is completed, incremental write operations are performed on the short-term memory matrix and long-term memory matrix according to the write addressing weight vector and the add vector;

[0056] The incremental write operation includes combining the corresponding write addressing weight vector and the add vector at the slot level to generate new content to be written to each memory slot, and adding the new content element by element to the corresponding slots in the short-term memory matrix and long-term memory matrix to supplement and update the erased memory content.

[0057] Optionally, in the read controller submodule, calculating the addressing weight based on content similarity and semantic matching weight, and reading content from short-term memory and long-term memory to obtain a unified read vector specifically includes:

[0058] The read key vector and read intensity scalar are extracted from the interface vector set, and the similarity between the read key vector and the content of each memory slot in the short-term memory matrix and the long-term memory matrix is ​​calculated to obtain the short-term memory slot content similarity vector and the long-term memory slot content similarity vector; the similarity calculation adopts the normalized cosine similarity method.

[0059] The content similarity vector is fused with the semantic task matching weight to generate the read addressing weight vector for short-term memory slots and long-term memory slots.

[0060] Specifically, this involves applying a read intensity scalar to the short-term memory slot content similarity vector and the long-term memory slot content similarity vector respectively to adjust the influence of content relevance on the read operation; converting the task semantic tag vector into a semantic weight vector that matches the memory slot structure through a preset semantic projection matrix; and weighting it together with the aforementioned adjusted content similarity vector to introduce task semantic guidance into the read decision.

[0061] The above weighted results are normalized to obtain the read addressing weight vectors for short-term memory slots and long-term memory slots.

[0062] Based on the read addressing weight vectors of the short-term memory slots and long-term memory slots, extract the content from the corresponding memory matrix and generate the read vector;

[0063] Specifically, this involves multiplying the read addressing weights of the short-term memory slots one by one with the contents of each slot in the short-term memory matrix, and then summing all the weighted contents to obtain the short-term read vector.

[0064] The long-term memory slot read addressing weights are weighted and summed with the contents of each slot in the long-term memory matrix in the same way to obtain the long-term read vector;

[0065] The two read vectors obtained represent the valid information extracted from the short-term memory slot and the long-term memory slot in the current round, respectively;

[0066] The short-term read vector and the long-term read vector are weighted and synthesized according to a preset fusion ratio to generate a unified read vector for the current round.

[0067] Specifically, this involves determining the contribution weights of the short-term and long-term read vectors in the final output based on the read fusion coefficient, scaling the short-term read vector according to the fusion coefficient, scaling the long-term read vector according to its complementary coefficient, and adding the two element-wise to obtain a unified read vector.

[0068] Optionally, the step of concatenating the control signal vector, the reading vector, and the task semantic label vector, inputting them into the decoding network, and outputting the current action intent recognition result and intent confidence specifically includes:

[0069] The control signal vector, unified reading vector, and task semantic label vector are obtained from the previous round of memory processing results, and the three are concatenated in sequence to form the decoding input vector;

[0070] The decoding input vector is input to the decoding network, which is a multilayer perceptron structure, including an input layer, at least one hidden layer and an output layer. The hidden layer uses the ReLU function as the activation function, and the output layer uses the softmax function and the sigmoid function for normalization according to the task configuration.

[0071] In the output layer of the decoding network, the action intent label vector for the current round is generated by normalizing the forward propagation results.

[0072] Specifically, this involves taking the decoded input vector and performing a linear transformation on the output layer weight matrix and bias vector to obtain an unnormalized classification score. The classification score is then input into the softmax function for normalization, so that each component corresponds to the prediction probability of different action intent categories, thereby forming an action intent label vector.

[0073] The decoded input vector is fed into the confidence generation unit to generate the intent confidence of the current action intent prediction result. The confidence generation unit converts the vector data into a real number between 0 and 1 through a combination of linear operation and sigmoid activation function, which is then used as the intent confidence output.

[0074] The action intent label vector is classified and the category corresponding to the highest probability component is selected as the action intent recognition result of the current round. The intent confidence is then output together.

[0075] Optionally, the step of dynamically adjusting the semantic priority threshold written to the controller submodule and updating the memory slot usage based on the action intent recognition result and intent confidence specifically includes:

[0076] Obtain the action intent recognition result of the current round and its corresponding intent confidence, and dynamically update the semantic priority threshold based on the intent confidence, and calculate the updated semantic priority threshold;

[0077] Specifically, this includes reading the semantic priority threshold that was last written into the controller as the initial baseline for the current adjustment, calculating the correction value using the difference between the intent confidence and the initial threshold, and updating the threshold linearly according to the preset learning rate.

[0078] The updated semantic priority threshold is input into the write controller submodule to adjust the write priority allocation relationship between the short-term memory slot and the long-term memory slot in the next round of write operation, so that action association information with high intent confidence is written to the long-term memory slot first, and vice versa.

[0079] Based on the action intent recognition results and intent confidence of the current round, the memory slots of the previous round are updated using the degree vector to obtain the updated degree vector.

[0080] Specifically, this includes scaling each element in the degree vector used in the previous round according to a preset attenuation coefficient to obtain the basic attenuation part of the degree vector used; and scaling the read weight vector generated in this round of memory read operation according to a ratio complementary to the attenuation coefficient to obtain the new part reflecting the current memory slot access status.

[0081] The two parts above are added element by element to obtain the updated usage vector, which is used to reflect the dynamic usage frequency of memory slots in the write and read control strategy, thereby optimizing the cyclic allocation of memory resources.

[0082] The updated semantic priority threshold and memory slots are returned to the controller input process using degree vectors to optimize the next round of control signal vector generation, interface vector construction, and semantic memory read / write strategy.

[0083] The beneficial effects of this invention are:

[0084] This invention introduces a semantically guided differentiable neural computer architecture to achieve joint encoding of multimodal sensor information, task semantic labels, and robot historical behavior. It can dynamically distinguish and allocate short-term and long-term memory resources according to different task requirements, improving the adaptability and robustness of action intent recognition. During memory writing, this invention uses a semantic priority control mechanism and allocation gating strategy to explicitly control different types of information, enabling on-demand erasure and incremental addition of information in memory slots, making the memory update process task-guided and interpretable. Furthermore, this invention performs memory retrieval operations based on a combination of content similarity and task semantic weights. By weighted fusion of short-term and long-term information, a unified memory vector for intent reasoning is generated, providing a reliable basis for action determination.

[0085] The decoding module of this invention, through joint processing of control signals, memory output vectors, and task semantics, can output action intent labels and corresponding confidence levels, achieving interpretable decision-making based on deep feature semantic association. Furthermore, this invention dynamically adjusts the semantic priority and memory slot utilization in the next round of memory management based on the intent recognition results and confidence levels, forming an adaptively optimized memory closed-loop structure, effectively improving the model's long-term learning ability in continuous interaction scenarios. Combining the above mechanisms, the method of this invention exhibits higher accuracy, flexibility, and scalability in action intent recognition in complex environments, while simultaneously improving the depth of intelligent behavior reasoning and memory management efficiency of the robot system. Attached Figure Description

[0086] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0087] Figure 1 This is a flowchart of a robot action intent recognition method based on deep learning proposed in this invention;

[0088] Figure 2 This is a schematic diagram of the semantically guided differentiable neural computer (SG-DNC) in a deep learning-based robot action intent recognition method proposed in this invention.

[0089] Figure 3 This is a flowchart illustrating the memory management process of the controller submodule in a deep learning-based robot action intent recognition method proposed in this invention. Detailed Implementation

[0090] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0091] refer to Figure 1-3 A deep learning-based method for robot action intent recognition includes the following steps:

[0092] Collect multimodal perception data of the robot during task execution, and extract features from each modal data to generate multimodal input feature vectors;

[0093] Semantic analysis is performed on task context information and voice command data to obtain task semantic label vectors;

[0094] A semantically guided differentiable neural computer model is constructed. The multimodal input feature vector and the memory vector read in the previous round are input into the controller submodule to generate the control signal vector and interface vector set for the current round.

[0095] In the write controller submodule, the semantic priority of each memory slot is calculated based on the control signal vector and the task semantic tag vector, and the short-term memory matrix and long-term memory matrix are updated.

[0096] In the read controller submodule, the addressing weight is calculated based on content similarity and semantic matching weight, and the content is read from the short-term memory slot and the long-term memory slot to obtain a unified read vector;

[0097] The control signal vector, reading vector, and task semantic label vector are concatenated and input into the decoding network, which outputs the current action intent recognition result and intent confidence.

[0098] Based on the action intent recognition results and intent confidence, the semantic priority threshold written to the controller submodule is dynamically adjusted, and the memory slot usage is updated.

[0099] In this embodiment, the process of collecting multimodal perception data of the robot during task execution and extracting features from each modal data to generate a multimodal input feature vector specifically includes:

[0100] Collect multimodal perception data of the robot during task execution, including motion sensor data, environmental image data, voice command data, and task context information;

[0101] The motion sensor data is processed, including the three-axis acceleration vector collected by the accelerometer, the three-axis angular velocity vector collected by the gyroscope, and the angles and angular velocities of each joint. The robot's posture change information is calculated by low-pass filtering and numerical integration, and then combined with the current motion parameters of the robotic arm to construct the motion state vector.

[0102] The environmental image data is processed to obtain the image frame sequence in the robot's working environment. A convolutional neural network is used to extract features from the images and output the environmental image feature vector.

[0103] The voice command data is processed by using a speech recognition module to convert the voice signal into a text command, and then encoding the text command through a language model to generate a voice command vector.

[0104] The task context information, including the current task objective, task type, and historical task tags, is processed. Semantic information is extracted through word vector embedding and sequence modeling to generate a task context vector.

[0105] The action state vector, environmental image feature vector, voice command vector, and task context vector are concatenated to form a multimodal input feature vector.

[0106] In this embodiment, the step of performing semantic analysis on task context information and voice command data to obtain task semantic tag vectors specifically includes:

[0107] The task context information is processed by text processing. The task context information includes the task target name, task type description, historical task tags and task execution status. A word segmentation algorithm based on dictionary matching and statistical rules is used to segment the task context information to obtain a context word segmentation sequence.

[0108] Entity recognition is performed on the context segmented word sequence. The named entity recognition model is used to classify each word in the segmented word sequence by label, and the task type entity, target object entity, and historical state entity are identified. The recognition result is then represented as a structured context sequence.

[0109] The voice command data is processed to convert speech into text. The speech recognition engine is used to convert the input speech signal into a text command string. The same entity recognition process is used to parse the speech text, extract the task semantic fields in the speech command, and generate a speech command semantic sequence.

[0110] The structured context sequence is merged with the semantic sequence of voice commands to form a semantic input sequence, which is then passed through a semantic coding network to output a context-aware semantic embedding vector.

[0111] The semantic encoding network adopts a deep neural network structure based on a pre-trained language model. It performs word embedding encoding on each semantic segment in the input sequence to obtain an initial embedding representation vector sequence. The initial embedding representation vector sequence is input into a context modeling network containing a multi-layer self-attention mechanism to capture the semantic dependencies and context interaction features between semantic segments. Through position encoding and normalization processing, a set of context-aware semantic embedding vectors is output.

[0112] Attention-weighted pooling is applied to the semantic embedding vector sequence to filter out vector fragments with high semantic relevance in the current task and construct task semantic label vectors.

[0113] The attention-weighted pooling method highlights vector segments in the semantic embedding vector sequence that have high semantic relevance to the current task target and execution state through weighting. It extracts task type information, target object information and historical task state information according to different semantic fields. The above three types of semantic information are constructed into task type field vector, target object field vector and historical task state field vector, respectively. Then, these three field vectors are concatenated in order to obtain the task semantic label vector.

[0114] In this embodiment, the construction of the semantically guided differentiable neural computer model, which involves inputting the multimodal input feature vector and the memory vector read in the previous round into the controller submodule to generate the control signal vector and interface vector set for the current round, specifically includes:

[0115] The multimodal input feature vector of the current round, the memory vector read from the hierarchical memory matrix of the previous round, and the semantic label vector corresponding to the current task are obtained. The above three vectors are concatenated in a fixed order to form the controller input vector. The memory vector is a unified reading vector.

[0116] A semantically guided differentiable neural computer model is constructed. The semantically guided differentiable neural computer model includes a controller submodule, a write controller submodule, a read controller submodule, and a hierarchical memory matrix. The hierarchical memory matrix includes a short-term memory matrix and a long-term memory matrix, both of which are composed of several memory slots.

[0117] The controller input vector is input to the controller submodule for processing. The controller submodule is a feedforward neural network with a gating mechanism, consisting of two linear mapping layers and one nonlinear activation layer. The first linear mapping layer transforms the dimension of the input vector to adapt to the neural network structure. Then, the nonlinear transformation of the ReLU activation function is integrated to enhance the feature representation capability. The second linear mapping layer then maps the intermediate result after nonlinear transformation into a control signal vector.

[0118] The control signal vector serves as the intermediate instruction result in the current round of robot action intent recognition task, and is used to drive the read and write control operations of the subsequent memory interface. All parameters of the controller submodule remain consistent and participate in joint training in multiple rounds of inference throughout the entire action recognition process, thereby ensuring the continuity and stability of the recognition behavior.

[0119] Based on the control signal vector output by the controller submodule, an interface vector set is generated through multiple sub-mapping networks with independent parameter structures.

[0120] In this embodiment, the step of generating an interface vector set by sequentially passing the control signal vector output by the controller submodule through multiple sub-mapping networks with independent parameter structures specifically includes:

[0121] The process involves generating read and write key vectors for controlling read and write operations of the memory matrix based on the control signal vector. Specifically, the control signal vector is input into a linear transformation layer for key vector generation, and the read key vector is obtained by weighted summation and bias correction.

[0122] The same control signal vector is input to another linear transformation layer and processed by a nonlinear activation function to obtain a write key vector; the dimensions of the read key vector and the write key vector are consistent with the word width of the memory matrix, and are used to perform content similarity matching on the short-term memory matrix and the long-term memory matrix respectively;

[0123] Based on the control signal vector, read intensity scalars and write intensity scalars are generated to adjust the intensity of memory operations. Specifically, the control signal vector is input to the intensity coefficient generation unit, and through a combination of linear transformation and nonlinear activation function, read intensity scalars for reading memory content and write intensity scalars for writing memory content are obtained respectively. The read intensity scalars and write intensity scalars are both non-negative real values, used to control the influence range and intensity of read and write operations respectively.

[0124] The erase vector and add vector for updating memory content are generated based on the control signal vector. Specifically, the control signal vector is input to the erase vector generation unit, and after being processed by linear mapping and normalization function, the erase vector with a value range between 0 and 1 is obtained, which is used to determine the position and proportion of elements in the memory matrix that need to be cleared or partially covered.

[0125] The control signal vector is input to the add vector generation unit. After processing by linear transformation and nonlinear activation function, an add vector with a value range between -1 and 1 is obtained, which is used to represent the new information content to be written to the memory matrix. The erase vector and the add vector are both consistent with the memory word width in terms of dimension.

[0126] The allocation gating coefficient vector is generated based on the control signal vector to guide the memory write allocation strategy. Specifically, the control signal vector is input to the allocation coefficient generation unit, and after processing by linear mapping and restricted activation function, the allocation gating coefficient vector is obtained. Each component of the allocation gating coefficient vector takes a value between 0 and 1, which is used to indicate the degree of participation of different memory slots in the write operation. The allocation gating coefficient vector dynamically balances and adjusts the write ratio of the short-term memory matrix and the long-term memory matrix, so that the write control strategy can be flexibly switched between different types of memory slots, thereby realizing the collaborative encoding and management of short-term sensitive information and long-term stable information.

[0127] A usage degree vector is generated based on the control signal vector to record the historical usage frequency of memory slots. Specifically, the control signal vector is input to the usage degree generation unit, processed by linear transformation and nonlinear activation function to obtain the usage degree vector, and the values ​​of each component are constrained to be non-negative real numbers. Each component of the usage degree vector corresponds one-to-one with different memory slots in the hierarchical memory matrix, which is used to provide a dynamic constraint mechanism based on usage frequency in memory read and write operations to indicate the location of memory slots to be replaced or retained first, thereby improving the rationality of memory allocation and long-term usage efficiency.

[0128] The generated read key vector, read strength scalar, write key vector, write strength scalar, erase vector, add vector, allocation gating coefficient vector, and usage vector are sequentially combined to form a complete interface vector set. The interface vector set is used to guide subsequent memory read and write operations and is an important basis for performing memory addressing, content updating, and slot allocation in semantically guided differentiable neural computers.

[0129] In this embodiment, the step of calculating the semantic priority of each memory slot and updating the short-term memory matrix and long-term memory matrix based on the control signal vector and task semantic tag vector in the write controller submodule specifically includes:

[0130] Calculate the semantic priority score for each memory slot based on the control signal vector and the task semantic label vector:

[0131] ;

[0132] in, For the first The semantic priority score of each memory slot. It is the Sigmoid activation function. These are trainable weight vectors used to project the control signal vector and the task semantic label vector onto the priority score space. This represents the transpose operation of a vector or matrix. To concatenate the control signal vector and the task semantic label vector to obtain a combined vector, For the control signal vector, For task semantic label vectors, This is a bias term used to adjust the intercept in the semantic priority score calculation. This represents the number of memory slots in the short-term memory matrix. This represents the number of memory slots in the long-term memory matrix. For memory slot index, This is a vector concatenation symbol;

[0133] The semantic priority scores of all memory slots are normalized to generate a normalized priority weight vector;

[0134] The normalization process includes performing element-wise division on the semantic priority score of each memory slot generated in the previous step, so that the sum equals 1, thereby generating a normalized priority weight vector. The value of each component in the priority weight vector is between 0 and 1, and the sum of the components is 1. The priority weight vector is used to reasonably allocate writing resources between short-term memory slots and long-term memory slots, so that information with higher semantic relevance can obtain a larger proportion of memory writing weight.

[0135] The normalized priority weight vector is combined with the allocation gating coefficient vector to generate write allocation weights for short-term memory slots and long-term memory slots.

[0136] Specifically, this involves allocating the overall write ratio to short-term memory slots and long-term memory slots based on the allocation gating coefficient, multiplying the portion of the normalized priority weight vector corresponding to the short-term memory slot with the short-term write gating coefficient, and obtaining the write allocation weight for each short-term memory slot.

[0137] Multiply the portion of the normalized priority weight vector corresponding to the long-term memory slot by the long-term write gating coefficient to obtain the write allocation weight for each long-term memory slot. The sum of the short-term write gating coefficient and the long-term write gating coefficient is 1, which is used to ensure the total amount of write resources is conserved between the two types of memory slots.

[0138] Based on the write key vector and write strength scalar in the interface vector set, calculate the write addressing weight vectors of short-term memory slots and long-term memory slots;

[0139] Specifically, this involves extracting write key vectors and write strength scalars from the interface vector set, and calculating the similarity between the write key vectors and each memory slot in the short-term memory matrix and long-term memory matrix to obtain similarity vectors between the short-term and long-term memory slots; multiplying each similarity vector by the write strength scalar to adjust its influence on the write operation; and normalizing the above write influence values ​​to obtain write addressing weight vectors for the short-term and long-term memory slots, which are used to indicate the weight allocation of each memory slot in this round of write operation. The higher the value, the more preferentially the slot is written.

[0140] Based on the erase vector and write allocation weight generated in the interface vector set, differentiable erase operations are performed on the short-term memory matrix and the long-term memory matrix respectively.

[0141] The differentiable erasure operation includes performing a slot-level combination operation on the erasure vector and the corresponding write allocation weight to obtain a slot-by-slot erasure coefficient matrix; performing element-level scaling operations on each slot in the short-term memory matrix and long-term memory matrix according to the corresponding erasure coefficient, thereby realizing partial erasure of existing memory content; the above erasure operation does not completely clear the content of each memory slot, but gradually decays element by element, so that each memory slot can be superimposed with new content, realizing a differentiable memory update mechanism;

[0142] After the erasure operation is completed, incremental write operations are performed on the short-term memory matrix and long-term memory matrix according to the write addressing weight vector and the add vector;

[0143] The incremental write operation includes combining the corresponding write addressing weight vector and the add vector at the slot level to generate new content to be written to each memory slot. The new content is then added element-wise to the corresponding slots in the short-term memory matrix and long-term memory matrix to supplement and update the erased memory content. The dimension of the add vector is consistent with the memory word width to ensure that the write operation can perform complete and consistent incremental updates on each memory slot.

[0144] In this embodiment, the step of calculating the addressing weight based on content similarity and semantic matching weight in the read controller submodule, and reading content from the short-term memory slot and long-term memory slot to obtain a unified read vector specifically includes:

[0145] The key vector and read intensity scalar are extracted from the interface vector set, and the similarity between the key vector and the content of each memory slot in the short-term memory matrix and the long-term memory matrix is ​​calculated to obtain the short-term memory slot content similarity vector and the long-term memory slot content similarity vector. The similarity calculation adopts the normalized cosine similarity method to measure the semantic matching degree between the key vector and the content of each memory slot.

[0146] The content similarity vector is fused with the semantic task matching weight to generate the read addressing weight vector for short-term memory slots and long-term memory slots.

[0147] Specifically, this involves applying a read intensity scalar to the short-term memory slot content similarity vector and the long-term memory slot content similarity vector respectively to adjust the influence of content relevance on the read operation; converting the task semantic tag vector into a semantic weight vector that matches the memory slot structure through a preset semantic projection matrix; and weighting it together with the aforementioned adjusted content similarity vector to introduce task semantic guidance into the read decision.

[0148] The above weighted results are normalized to obtain read addressing weight vectors for short-term memory and long-term memory. Each component in the weight vector is used to indicate the importance of the corresponding memory slot in the current read operation.

[0149] Based on the read addressing weight vectors of the short-term memory slots and long-term memory slots, extract the content from the corresponding memory matrix and generate the read vector;

[0150] Specifically, this involves multiplying the read addressing weights of the short-term memory slots one by one with the contents of each slot in the short-term memory matrix, and then summing all the weighted contents to obtain the short-term read vector.

[0151] The long-term memory slot read addressing weights are weighted and summed with the contents of each slot in the long-term memory matrix in the same way to obtain the long-term read vector;

[0152] The two read vectors obtained represent the effective information extracted from the short-term memory slot and the long-term memory slot in the current round, respectively, which are used to support semantic judgment and task decision-making in the subsequent intent recognition process;

[0153] The short-term read vector and the long-term read vector are weighted and synthesized according to a preset fusion ratio to generate a unified read vector for the current round.

[0154] Specifically, this involves determining the contribution weights of short-term and long-term read vectors in the final output based on the read fusion coefficient, scaling the short-term read vector according to the fusion coefficient, scaling the long-term read vector according to its complementary coefficient, and adding the two element-wise to obtain a unified read vector, which is used to guide the final judgment and response generation of the robot's action intent.

[0155] In this embodiment, the step of concatenating the control signal vector, the reading vector, and the task semantic label vector, inputting them into the decoding network, and outputting the current action intent recognition result and intent confidence specifically includes:

[0156] The control signal vector, unified reading vector, and task semantic label vector are obtained from the previous round of memory processing results, and the three are concatenated in sequence to form the decoding input vector;

[0157] The decoding input vector is input to the decoding network, which is a multilayer perceptron structure, including an input layer, at least one hidden layer and an output layer. The hidden layer uses the ReLU function as the activation function, and the output layer uses the softmax function and the sigmoid function for normalization according to the task configuration.

[0158] In the output layer of the decoding network, the action intent label vector for the current round is generated by normalizing the forward propagation results.

[0159] Specifically, the decoding input vector is linearly transformed by the output layer weight matrix and bias vector to obtain an unnormalized classification score. The classification score is then input into a softmax function for normalization, so that each component corresponds to the predicted probability of different action intention categories, thereby forming an action intention label vector. Each component of the action intention label vector takes a value between 0 and 1, and the sum of all components is 1, which is used to represent the probability distribution of the current robot action intention on multiple candidate categories.

[0160] The decoded input vector is fed into the confidence generation unit to generate the intent confidence of the current round of action intent prediction results. The confidence generation unit converts the vector data into a real number between 0 and 1 through a combination of linear operation and sigmoid activation function as the intent confidence output. The confidence scalar is used to indicate the credibility or decision reliability of the decoding network in the current round of recognition results.

[0161] The action intent label vector is classified and the category corresponding to the highest probability component is selected as the action intent recognition result of the current round. The intent confidence is then output together.

[0162] In this embodiment, the step of dynamically adjusting the semantic priority threshold written to the controller submodule and updating the memory slot usage based on the action intent recognition result and intent confidence specifically includes:

[0163] Obtain the action intent recognition result and corresponding intent confidence score for the current round, and dynamically update the semantic priority threshold based on the intent confidence score, and calculate the updated semantic priority threshold.

[0164] Specifically, this includes reading the semantic priority threshold from the previous memory writing controller as the initial benchmark for the current adjustment, calculating the correction value using the difference between the intent confidence and the initial threshold, and linearly updating the threshold according to a preset learning rate. The updated semantic priority threshold is used to reflect the reliability of the current recognition result in the semantic dimension, thereby increasing the writing priority of high-confidence action intents in the next round of memory writing.

[0165] The updated semantic priority threshold is input into the write controller submodule to adjust the write priority allocation relationship between the short-term memory slot and the long-term memory slot in the next round of write operation, so that action association information with high intent confidence is written to the long-term memory slot first, and vice versa.

[0166] Based on the action intent recognition results and intent confidence of the current round, the memory slots of the previous round are updated using the degree vector to obtain the updated degree vector.

[0167] Specifically, this includes scaling each element in the degree vector used in the previous round according to a preset attenuation coefficient to obtain the basic attenuation part of the degree vector used; and scaling the read weight vector generated in this round of memory read operation according to a ratio complementary to the attenuation coefficient to obtain the new part reflecting the current memory slot access status.

[0168] The two parts above are added element by element to obtain the updated usage vector, which is used to reflect the dynamic usage frequency of memory slots in the write and read control strategy, thereby optimizing the cyclic allocation of memory resources.

[0169] The updated semantic priority threshold and memory slots are returned to the controller input flow using degree vectors to optimize the next round of control signal vector generation, interface vector construction, and semantic memory read / write strategy, thereby realizing a system-level adaptive loop optimization mechanism.

[0170] Example 1:

[0171] To verify the feasibility of this invention in practice, it was applied to an industrial-grade multi-functional humanoid service robot platform for multi-task collaborative control under complex voice commands. The robot is equipped with vision, voice, and tactile sensors and is suitable for indoor intelligent patrol tasks involving path planning, obstacle avoidance, object recognition, and voice interaction.

[0172] Traditional humanoid robots suffer from weak command ambiguity resolution, severe task memory loss, and inconsistent action responses in multi-tasking environments. Particularly under continuous dialogue control, robots struggle to understand the relationship between current actions and historical semantics, leading to erroneous actions, semantic drift, or operational interruptions. This invention constructs a semantically guided differentiable neural computer (SG-DNC) structure and introduces a dual-channel memory mechanism to achieve joint modeling and dynamic control of short-term action information and long-term task context.

[0173] During the deployment phase, 100 multi-turn voice control tasks were selected, covering navigation commands, target recognition, semantic question answering, and environmental interaction. Experiments compared the task completion rate, response latency, semantic consistency, and error recovery capability of the method described in this invention with traditional Transformer-based dialogue control methods and behavior generation methods based on Long Short-Term Memory (LSTM) networks.

[0174] Experimental results show that the present invention achieves a task completion rate of 92.4% in complex task sequences, an average improvement of 16.7% compared to traditional methods; under instructions with high semantic ambiguity, the intent recognition accuracy reaches 95.6%, higher than the 78.2% in the comparison model; at the same time, the present method supports a fast error recovery mechanism, shortening the response recovery time to less than 1.8 seconds, and the task semantic preservation rate remains stable at over 90%, fully demonstrating the superior performance of the present method in information preservation, semantic alignment and system adaptation.

[0175] Table 1. Performance evaluation results of the present invention and the comparative method in multi-turn speech tasks.

[0176]

[0177] As can be seen from the table above, the present invention exhibits significant advantages over traditional Transformer and LSTM control methods in several key performance indicators for multi-turn speech task control scenarios. Firstly, in terms of task completion rate, the present invention achieves 92.4%, representing improvements of 17.3 and 23.7 percentage points respectively compared to the Transformer method's 75.1% and the LSTM method's 68.7%. This demonstrates that the present invention possesses stronger execution stability and control robustness in complex semantic task environments. Secondly, in terms of intent recognition accuracy, the present invention achieves 95.6%, representing improvements of 14.3 and 17.4 percentage points respectively compared to the Transformer and LSTM methods. This indicates that the present invention effectively enhances multimodal semantic aggregation and contextual understanding capabilities by introducing a semantically guided differentiable neural computation mechanism.

[0178] In terms of response performance, the average response latency of this invention is 1.28 seconds, which is significantly better than the 1.96 seconds of Transformer and the 2.11 seconds of LSTM, effectively alleviating the response bottleneck of traditional methods in instruction redundancy and multi-round information processing. At the same time, the error recovery time is controlled within 1.8 seconds, which is better than the 3.2 seconds of Transformer and the 4.5 seconds of LSTM, indicating that this invention has a faster system adaptability when facing instruction conflicts or execution anomalies.

[0179] Furthermore, in terms of semantic retention rate, this invention maintained 91.3%, a significant improvement compared to the traditional methods' 73.4% and 69.8%, further validating the invention's collaborative memory retention effect on historical task information, action sequences, and the current semantic context. In multimodal integration scoring and action continuity scoring, this invention achieved scores of 94.7 and 92.8 respectively, both higher than the comparative methods, demonstrating the overall structural optimization advantages of this invention in information fusion and control logic continuity. Combined with user satisfaction scores, this invention also achieved a significant improvement in user experience, comprehensively illustrating that this semantically guided control method possesses higher practical value and system reliability.

[0180] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for recognizing a robot motion intention based on deep learning, characterized by, Includes the following steps: Collect multimodal perception data of the robot during task execution, and extract features from each modal data to generate multimodal input feature vectors; Semantic analysis is performed on task context information and voice command data to obtain task semantic label vectors; A semantically guided differentiable neural computer model is constructed. The multimodal input feature vector and the memory vector read in the previous round are input into the controller submodule to generate the control signal vector and interface vector set for the current round. In the write controller submodule, the semantic priority of each memory slot is calculated based on the control signal vector and the task semantic tag vector, and the short-term memory matrix and long-term memory matrix are updated. In the read controller submodule, the addressing weight is calculated based on content similarity and semantic matching weight, and the content is read from the short-term memory slot and the long-term memory slot to obtain a unified read vector; The control signal vector, reading vector, and task semantic label vector are concatenated and input into the decoding network, which outputs the current action intent recognition result and intent confidence. Based on the action intent recognition results and intent confidence, dynamically adjust the semantic priority threshold written to the controller submodule and update the memory slot usage. The construction of the semantically guided differentiable neural computer model, which involves inputting the multimodal input feature vector and the memory vector read in the previous round into the controller submodule to generate the control signal vector and interface vector set for the current round, specifically includes: The multimodal input feature vector of the current round, the memory vector read from the hierarchical memory matrix of the previous round, and the semantic label vector corresponding to the current task are obtained. The above three vectors are concatenated in a fixed order to form the controller input vector. A semantically guided differentiable neural computer model is constructed, which includes a controller submodule, a write controller submodule, a read controller submodule, and a hierarchical memory matrix, wherein the hierarchical memory matrix includes a short-term memory matrix and a long-term memory matrix. The controller input vector is input to the controller submodule for processing to obtain the control signal vector; Based on the control signal vector output by the controller submodule, an interface vector set is generated through multiple sub-mapping networks with independent parameter structures.

2. The robot action intent recognition method based on deep learning according to claim 1, characterized in that, The multimodal perception data includes motion sensor data, environmental image data, voice command data, and task context information.

3. The robot action intent recognition method based on deep learning according to claim 1, characterized in that, The step of performing semantic analysis on task context information and voice command data to obtain task semantic label vectors specifically includes: The task context information is processed by using a word segmentation algorithm based on dictionary matching and statistical rules to segment the task context information into words, resulting in a context word segmentation sequence. Entity recognition is performed on the context segmentation sequence, and the recognition result is represented as a structured context sequence; The voice command data is processed to convert speech to text, the task semantic fields in the voice command are extracted, and a voice command semantic sequence is generated. The structured context sequence is merged with the semantic sequence of voice commands to form a semantic input sequence, which is then passed through a semantic coding network to output a context-aware semantic embedding vector. Attention-weighted pooling is applied to the semantic embedding vector sequence to filter out vector fragments with high semantic relevance in the current task, and task semantic label vectors are constructed.

4. The robot action intent recognition method based on deep learning according to claim 1, characterized in that, The step of generating an interface vector set by sequentially passing the control signal vector output by the controller submodule through multiple sub-mapping networks with independent parameter structures specifically includes: The control signal vector is input into the linear transformation layer, and the read key vector is obtained by weighted summation and bias term correction. The same control signal vector is input into another linear transformation layer, and combined with the processing of a nonlinear activation function, the write key vector is obtained. The control signal vector is input to the intensity coefficient generation unit, and the read intensity scalar and write intensity scalar are obtained by combining linear transformation and nonlinear activation function, respectively. The control signal vector is input to the erase vector generation unit, and after processing by linear mapping and normalization function, the erase vector is obtained. The control signal vector is input to the vector generation unit, and after processing by a linear transformation and a nonlinear activation function, the added vector is obtained. The control signal vector is input to the allocation coefficient generation unit, and after processing by linear mapping and restricted activation function, the allocation gate coefficient vector is obtained. The control signal vector is input to the usage degree generation unit, and after being processed by linear transformation and nonlinear activation function, the usage degree vector is obtained. The generated read key vector, read strength scalar, write key vector, write strength scalar, erase vector, add vector, allocation gating coefficient vector, and usage vector are combined sequentially to form a complete set of interface vectors.

5. The robot action intent recognition method based on deep learning according to claim 1, characterized in that, In the write controller submodule, the process of calculating the semantic priority of each memory slot based on the control signal vector and the task semantic tag vector, and updating the short-term memory matrix and long-term memory matrix specifically includes: Calculate the semantic priority score for each memory slot based on the control signal vector and the task semantic label vector; The semantic priority scores of all memory slots are normalized to generate a normalized priority weight vector; The normalized priority weight vector is combined with the allocation gating coefficient vector to generate write allocation weights for short-term memory slots and long-term memory slots. Based on the write key vector and write strength scalar in the interface vector set, calculate the write addressing weight vectors of short-term memory slots and long-term memory slots; Based on the erase vector and write allocation weight generated in the interface vector set, differentiable erase operations are performed on the short-term memory matrix and the long-term memory matrix respectively. After the erasure operation is completed, incremental write operations are performed on the short-term memory matrix and long-term memory matrix according to the write addressing weight vector and the add vector.

6. The robot action intent recognition method based on deep learning according to claim 1, characterized in that, In the read controller submodule, the addressing weight is calculated based on content similarity and semantic matching weight, and content is read from short-term memory and long-term memory to obtain a unified read vector. Specifically, this includes: Extract the read key vector and read intensity scalar from the interface vector set, and calculate the similarity between the read key vector and the content of each memory slot in the short-term memory matrix and the long-term memory matrix respectively to obtain the short-term memory slot content similarity vector and the long-term memory slot content similarity vector. The content similarity vector is fused with the semantic task matching weight to generate the read addressing weight vector for short-term memory slots and long-term memory slots. Based on the read addressing weight vectors of the short-term memory slots and long-term memory slots, extract the content from the corresponding memory matrix and generate short-term read vectors and long-term read vectors; The short-term read vector and the long-term read vector are weighted and synthesized according to a preset fusion ratio to generate a unified read vector for the current round.

7. The robot action intent recognition method based on deep learning according to claim 1, characterized in that, The process of concatenating the control signal vector, the read vector, and the task semantic label vector, inputting them into the decoding network, and outputting the current action intent recognition result and intent confidence level specifically includes: The control signal vector, unified reading vector, and task semantic label vector are obtained from the previous round of memory processing results, and the three are concatenated in sequence to form the decoding input vector; The decoding input vector is input to the decoding network, which is a multilayer perceptron structure, including an input layer, at least one hidden layer and an output layer; In the output layer of the decoding network, the action intent label vector for the current round is generated by normalizing the forward propagation results. The decoded input vector is fed into the confidence generation unit to generate the intent confidence of the current round's action intent prediction result; The action intent label vector is classified and the category corresponding to the highest probability component is selected as the action intent recognition result of the current round. The intent confidence is then output together.

8. The robot action intent recognition method based on deep learning according to claim 1, characterized in that, The step of dynamically adjusting the semantic priority threshold written to the controller submodule based on the action intent recognition result and intent confidence, and updating the memory slot usage specifically includes: Obtain the action intent recognition result and corresponding intent confidence score for the current round, and dynamically update the semantic priority threshold based on the intent confidence score, and calculate the updated semantic priority threshold. The updated semantic priority threshold is input into the write controller submodule to adjust the write priority allocation relationship between short-term memory slots and long-term memory slots in the next round of write operations. Based on the action intent recognition results and intent confidence of the current round, the memory slots of the previous round are updated using the degree vector to obtain the updated degree vector. The updated semantic priority threshold and memory slots are returned to the controller input process using degree vectors to optimize the next round of control signal vector generation, interface vector construction, and semantic memory read / write strategy.

Images