A Management Method for End-Side Robot Memory Systems that Combines Dynamic Compression and Efficient Recall

Abstract

This invention discloses a method for managing the memory system of an edge robot that combines dynamic compression and efficient recall, relating to the field of edge artificial intelligence technology. The method includes: mapping data signals to a feature space to generate a basic memory vector, inputting it into a value evaluator, and outputting a value score; compressing the basic memory vector based on the value score to obtain a compressed memory vector and a compression level; inserting the compressed memory vector into an incremental index structure according to the compression level, and dynamically adjusting the insertion method based on real-time resource status; inputting the query vector and index information from the incremental index structure into a recall control module to form a recall request to be executed; selecting the optimal recall path based on the compression level and real-time resource status, and outputting a candidate memory set; determining whether to perform local dequantization based on the compression level of the memories in the candidate memory set, and generating the final memory. This application solves the problem of inefficient memory management in edge robots through dynamic compression and efficient recall.

Description

Technical Field

[0001] This invention relates to the field of edge artificial intelligence technology, and more specifically, to an edge robot memory system management method that combines dynamic compression and efficient recall. Background Technology

[0002] In modern intelligent robot systems, edge memory management is a core component for achieving autonomous perception, decision-making, and long-term learning. However, with the increasing complexity of robot application scenarios and the massive generation of multimodal sensory data, traditional edge memory management technologies face numerous bottlenecks. Existing edge robot systems mostly employ fixed storage or simple indexing strategies, directly storing sensory data or managing it chronologically. While this method is simple to implement, it suffers from low storage resource utilization, poor retrieval efficiency, and difficulty in maintaining historical information over the long term. With the introduction of multimodal sensors (such as vision, hearing, and touch) and complex tasks, the amount of data is growing exponentially. Fixed storage methods can no longer meet the needs of real-time decision-making and long-term learning in edge robots, leading to response delays, information redundancy, and low task execution efficiency under resource-constrained conditions.

[0003] For example, the invention patent with publication number CN120524372A discloses a memory optimization system and method for an intelligent companion robot, belonging to the field of computer technology. It includes a multimodal memory encoding unit for processing raw interaction data to generate structured memory nodes containing emotional valence; a memory association initialization unit for calculating association weights based on the temporal proximity and semantic relevance between memory nodes to generate initial association weights; a memory node importance assessment unit for evaluating the value of memory nodes to generate node importance scores; and an association weight adaptive adjustment unit for iteratively updating the initial association weights by combining the node importance scores with the semantic and emotional associations between memory nodes to generate updated association weights. This system can retain key emotional cues, thereby exhibiting higher logical consistency and emotional continuity in long-term interactions with users.

[0004] For example, the invention patent with announcement number CN119283042B discloses an adaptive memory-enhanced imitation learning method for robot control tasks. This method includes standardizing and extracting temporal features from raw multimodal data to construct a hierarchical feature representation; establishing a hierarchical structure containing short-term memory, working memory, and long-term memory; evaluating memory importance based on usage frequency, task relevance, and novelty, and dynamically updating memory content; retrieving relevant memories based on the current state and performing multimodal feature fusion; calculating shaping rewards based on nearest neighbor sample search; and optimizing strategy network parameters through a combination of multiple loss functions. Through adaptive memory management and multimodal feature fusion mechanisms, the learning efficiency and execution performance of the robot control system are improved.

[0005] The above-disclosed technical solutions have at least the following technical problems:

[0006] Traditional edge robot systems often employ fixed storage or simple indexing, failing to dynamically select retention or compression strategies based on memory value. This leads to wasted storage resources and low memory retrieval efficiency. Furthermore, existing technologies separate compression operations from system resource status management, ignoring dynamic changes in storage, computing power, and write load, which can easily cause index write conflicts, local sorting disorder, or increased system latency. To address these issues, this invention proposes a solution. Summary of the Invention

[0007] To overcome the aforementioned deficiencies of the prior art, embodiments of the present invention provide a method for managing the memory system of an end-side robot that combines dynamic compression and efficient recall. By combining dynamic compression and efficient recall, the method solves the problem of inefficient memory management of end-side robots.

[0008] To achieve the above objectives, the present invention provides the following technical solution:

[0009] A method for managing the memory system of an edge robot that coordinates dynamic compression and efficient recall includes: mapping the collected data signals to a unified feature space to generate a basic memory vector, inputting it into a value evaluator, and outputting a value score; selecting a corresponding compression method based on the value score to compress the basic memory vector, obtaining a compressed memory vector and a compression level; inserting the compressed memory vector into an incremental index structure according to the compression level, and dynamically adjusting the insertion method of the compressed memory vector based on the real-time resource status; when the robot issues a query request, inputting the query vector and the index information in the incremental index structure into the recall control module to form a recall request to be executed; after receiving the recall request, the recall control module selects the best recall path based on the compression level and the real-time resource status, and outputs a candidate memory set; and determining whether to perform local dequantization based on the compression level of each memory in the candidate memory set to generate the final high-quality memory.

[0010] In a preferred embodiment, the step of mapping the collected data signals to a unified feature space, generating a basic memory vector, and inputting it into a value evaluator to output a value score is as follows: The output signal of the target to be tested is acquired to form a multimodal raw dataset, and preprocessing is performed on the raw data of each modality; the preprocessed data of each modality is input into a lightweight modal backbone network to obtain modal feature vectors of different modalities; linear mapping and regularization are performed on the modal feature vectors to transform them into projection vectors with consistent dimensions; multiple projection vectors are fused using a lightweight attention fusion method to obtain a single basic memory vector; a lightweight prediction head is constructed, and the basic memory vector is input into the Softmax distribution layer of the prediction head. Obtain the probability distribution and output the prediction entropy based on the probability distribution; input the base memory vector into the Dropout layer of the prediction head, perform several forward propagations on the same base memory vector to obtain several sets of outputs, and output the variance, i.e., the dropout forward variance; output the confidence score based on the prediction entropy and the dropout forward variance; maintain a lightweight index of recent vectors on the edge, calculate the distance distribution between the vector to be evaluated and its neighborhood to form the novelty; calculate the cosine similarity between the base memory vector and the task vector to obtain the task relevance score; obtain the time decay based on the difference between the vector timestamp and the current time; combine the task relevance score, novelty, confidence score and time decay according to the set weights to output the value score.

[0011] In a preferred embodiment, the step of maintaining a lightweight index of recent vectors on the endpoint, calculating the distance distribution between the vector to be evaluated and its neighborhood, and forming novelty is as follows: A lightweight index is established for the written compressed memory vectors, the index key is stored using locality-sensitive hashing, and a portion of the basic memory vectors are selected as the vectors to be evaluated; using the same hash function as the index structure, the index key of the vector to be evaluated is calculated, and the corresponding neighborhood bucket is located in the lightweight index to determine the candidate neighborhood range; the basic memory vectors in the neighborhood buckets are read, and redundant vectors are filtered according to timestamps and quantity limits to restrict the neighborhood size, resulting in a neighborhood vector set; the distance between the vector to be evaluated and each vector in the neighborhood vector set is calculated using a cosine distance function, and the distance distribution characteristics are statistically analyzed; based on the distance distribution characteristics, the novelty is output using a local outlier probability algorithm.

[0012] In a preferred embodiment, the step of inserting the compressed memory vector into the incremental index structure according to the compression level, and dynamically adjusting the insertion method of the compressed memory vector in conjunction with the real-time resource status, is as follows: The compressed memory vector is parsed to obtain the real-time resource status; the real-time resource status is input into the edge resource prediction model to predict resource changes within a short future time window, obtaining predicted storage changes, predicted write load, and predicted available computing power, and converting them into resource prediction vectors; the compressed memory vector is mapped to the target level in the multi-layer incremental index structure according to the compression level, and the disturbance impact characteristics of this insertion are output in conjunction with the resource prediction vectors; the optimal insertion strategy is selected from a preset set of insertion paths based on the compression level, resource prediction vectors, and disturbance impact characteristics; and the index write operation is performed according to the selected insertion strategy.

[0013] In a preferred embodiment, the step of mapping the compressed memory vector to the target level in the multi-layer incremental index structure according to the compression level, and outputting the perturbation impact features of this insertion in conjunction with the resource prediction vector, specifically includes: obtaining the compression level information of the compressed memory vector, and mapping different compression levels to the target level in the multi-layer incremental index structure according to the preset index level division rules; reading the corresponding structure data from the target level to obtain the target index partition and corresponding bucket structure to which the compressed memory vector will be written; inputting the resource prediction vector into the edge resource prediction model, and generating corresponding resource constraint data based on the output predicted storage occupancy, predicted write load, and predicted available computing power; determining the compressed memory based on the bucket density, key distribution, and sorting structure of the target level, combined with the semantic cluster number of the compressed memory vector. The expected insertion position range of the vector within the target level; based on the structural data of the target level, the node split probability and local key movement are calculated using the B-tree insertion cost model, and combined with the LSM-tree inter-layer compaction trigger probability, the index migration scale caused by the insertion operation is obtained, thus outputting the expected migration amount; the local region sort rank difference of the expected insertion position is obtained, and the number of sorted nodes adjusted after the new vector is inserted is obtained based on the incremental sorting model, thus obtaining the local sort change scale, i.e., the reordering cost; the predicted write load is input into the optimistic concurrency control conflict model, the competition probability within the operation window is output, and combined with the write event model based on Poisson arrival rate, the probability of multiple concurrent insertions colliding in the same index region is predicted, thus obtaining the write conflict probability.

[0014] In a preferred embodiment, the step of inputting the query vector and index information in the incremental index structure into the recall control module to form a recall request to be executed is as follows: The input data corresponding to the query is extracted from the task scheduling module and input into the feature encoder to generate a query vector consistent with the memory vector space; the current index metadata is read from the multi-level incremental index structure; the query vector is normalized, and the similarity between the normalized query vector and each semantic cluster center is calculated using the semantic cluster center in the index metadata as a reference benchmark; based on the similarity ranking result, several semantic clusters with the highest similarity are selected as candidate clusters; according to the semantic cluster-level mapping relationship and level activity matrix recorded in the index metadata, the level number, level span, and cross-level activity corresponding to the candidate clusters are aggregated and analyzed to obtain the level number range covered by the candidate clusters in the multi-level incremental index structure, generating a candidate search range; the normalized query vector features, index structure data, and candidate search range are input into the recall control module to generate a recall request to be executed, and it is represented as a recall request object containing the search range, candidate levels, and estimated recall depth.

[0015] In a preferred embodiment, the step of combining compression level and real-time resource status to select the optimal recall path and output a candidate memory set is as follows: The compressed memory vectors within the candidate level range are statistically analyzed to form compression features; the real-time system status is obtained and organized into a resource constraint set as a condition for limiting the recall operation; the compression features of the candidate region are combined with resource constraints to analyze the retrieval overhead generated by different recall paths, and multiple candidate recall paths are constructed based on the analysis results; the reconstruction overhead, access latency, conflict risk, and comprehensive cost in terms of recall completeness of the candidate recall paths are compared to select the optimal recall path; the retrieval operation is performed according to the optimal recall path, and the candidate memory set is output.

[0016] In a preferred embodiment, the step of determining whether to perform local dequantization based on the compression level of each memory in the candidate memory set to generate the final high-quality memory is as follows: The compression level, quantization residual information, semantic cluster number, and historical access frequency of each candidate memory are read, and it is determined whether to perform local dequantization; the memories requiring dequantization are grouped by semantic cluster, and local dequantization is performed in batches within each cluster; the locally dequantized candidate memories and the candidate memories that do not require dequantization are uniformly mapped to the same feature space and normalized; all locally dequantized memories and the memories that do not require dequantization are integrated to form the final high-quality memory set.

[0017] The technical effects and advantages of the present invention, which combines dynamic compression and efficient recall in a collaborative end-side robot memory system management method, are as follows:

[0018] 1. This invention fuses query vectors with multi-layered index information from an incremental index structure to form precise recall requests to be executed, and intelligently selects the optimal recall path based on compression level and real-time resource status. This process fully considers reconstruction overhead, access latency, conflict risk, and recall completeness, achieving efficient and resource-sensitive memory retrieval. In particular, performing local dequantization operations on memories at different compression levels in the candidate memory set can reconstruct high-quality memory representations while ensuring computational efficiency, significantly improving the robot's accuracy in utilizing historical information and its task completion capabilities.

[0019] 2. This invention achieves efficient management of the end-user robot's memory system through value-driven memory selection, intelligent mapping at compression levels, multi-layer incremental index structure management, and dynamically adjusted recall strategies. This method not only significantly reduces storage and computational overhead and improves memory retrieval efficiency, but also ensures memory reconstruction quality and task relevance, enabling the end-user robot to achieve long-term learning and efficient execution under resource-constrained conditions. It provides a stable, reliable, and scalable memory management solution for intelligent robot systems. Attached Figure Description

[0020] Figure 1 This is a flowchart illustrating a method for managing an end-side robot memory system that combines dynamic compression and efficient recall, as described in this invention. Detailed Implementation

[0021] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0022] Example 1, Figure 1 This invention presents a method for managing the memory system of an end-side robot that combines dynamic compression and efficient recall, comprising:

[0023] S1 maps the collected data signals to a unified feature space, generates a basic memory vector, and inputs it into the value evaluator to output a value score;

[0024] In this embodiment, the collected data signals are mapped to a unified feature space to generate a basic memory vector, which is then input into a value evaluator to output a value score, as detailed below:

[0025] The robot sequentially acquires output signals from vision, speech, touch, and other sensors to form a multimodal raw dataset. It then performs preprocessing on the raw data of each modality, including image normalization and denoising, speech endpoint detection and filtering, and tactile signal smoothing, to obtain a data format adapted to a unified encoder.

[0026] The preprocessed modal data are input into the lightweight modal backbone network to obtain modal feature vectors for different modalities.

[0027] Perform linear mapping and regularization on the modal feature vectors to transform them into dimensionally consistent projection vectors;

[0028] By using a lightweight attention fusion method, multiple projection vectors are fused to obtain a single basic memory vector, which serves as a unified representation for each data point.

[0029] The underlying memory vector is scaled and normalized based on end-side sliding statistics, and then lightly calibrated based on the most recent high-value samples to mitigate long-term drift.

[0030] Construct a lightweight prediction head, which includes: a 1-layer linear transformation + Softmax (for generating category or semantic distributions), and a 1-layer or multi-layer feedforward network with Dropout (for Monte Carlo Dropout inference).

[0031] The basic memory vector is input into the Softmax distribution layer of the prediction head to obtain the probability distribution, and the prediction entropy is output based on the probability distribution.

[0032] The basic memory vector is input into the Dropout layer of the prediction head. Several forward propagations are performed on the same basic memory vector to obtain several sets of outputs, and the variance is output, i.e., dropout forward variance.

[0033] Output the confidence score based on the prediction entropy and dropout forward variance;

[0034] Maintain a lightweight index of recent vectors at the end, calculate the distance distribution between the vector to be evaluated and its neighborhood, and form the novelty.

[0035] The cosine similarity between the base memory vector and the task vector is calculated to obtain the task relevance score.

[0036] The time decay is calculated based on the difference between the vector timestamp and the current time, which is used to adjust the value of outdated memories.

[0037] The task relevance score, novelty, confidence, and time decay are combined according to a set weight to output a value score;

[0038] Value level labels are determined based on value scores and system dynamic thresholds, including high value, medium value, and low value.

[0039] If a sample exhibits high novelty but insufficient confidence, it is placed in a short-term buffer and awaits further contextual confirmation.

[0040] The confidence score is output based on the prediction entropy and dropout forward variance, as detailed below:

[0041]

[0042]

[0043]

[0044]

[0045]

[0046]

[0047] In the formula, It is a probability distribution, where i is the index of the basic memory vector. These are the weights of the prediction head (classification layer), used to map the base memory vector to the K-class semantic space. It is the basic memory vector. It is a classification layer bias term. It is predicting entropy. It is the number of semantic categories. It is the probability corresponding to the k-th class in the Softmax probability distribution. The confidence level is derived from the prediction entropy. It is variance. It is the number of forward propagations. It is the predicted value of the t-th forward output. It is the average of the outputs from T forward propagations. The confidence level is derived from the variance. It is the attenuation coefficient that normalizes the result to [0,1]. The confidence score is output based on the prediction entropy and dropout forward variance. It is the weighting coefficient.

[0048] In this embodiment, a lightweight index of recent vectors is maintained at the edge, and the distance distribution between the vector to be evaluated and its neighborhood is calculated to form the novelty, as follows:

[0049] A lightweight index is created for recently written compressed memory vectors, using Locality Sensitive Hash (LSH) to store the index key, and recording the vector digest and timestamp;

[0050] Based on the system strategy (such as newly written, potentially high-value or high-novelty vectors), a portion of the basic memory vectors are selected as vectors to be evaluated for further precise novelty calculation;

[0051] Using the same hash function as the index structure, calculate the index key (including hash signature, quantization center number or subspace codeword) of the vector to be evaluated, and locate the corresponding neighborhood bucket in the lightweight index to determine the candidate neighborhood range;

[0052] Read the basic memory vector in the neighborhood bucket, and filter redundant vectors according to the timestamp and the upper limit of the number to limit the neighborhood size and obtain the neighborhood vector set;

[0053] The distance between the vector to be evaluated and each vector in the neighborhood vector set is calculated using the cosine distance function, and the distance distribution characteristics such as the mean distance, distance variance, Top-K minimum distance, and local density value are statistically analyzed.

[0054] Based on the distance distribution characteristics, the novelty is output using a local outlier probability algorithm to measure the degree of difference between the vector to be evaluated and the recent memory space, and serves as the input to the value evaluation module.

[0055] The value level label is determined based on the value score and the system's dynamic threshold, including high value, medium value, and low value, as detailed below:

[0056] The system dynamically generates two levels of dynamic thresholds, high and low, based on the mean and standard deviation of the current value score set, combined with real-time storage usage and available computing power, using the score threshold method.

[0057] When the value score is greater than the high threshold, it is marked as high value;

[0058] Values ​​between the high and low thresholds are marked as medium value.

[0059] Values ​​below a low threshold are marked as low.

[0060] S2, Select the corresponding compression method based on the value score to compress the basic memory vector, and obtain the compressed memory vector and compression level;

[0061] In this embodiment, the corresponding compression method is selected based on the value score to compress the basic memory vector, resulting in a compressed memory vector and a compression level, as detailed below:

[0062] The underlying memory vector to be processed and its corresponding value score are obtained from the value assessment module and used as input to the compression module;

[0063] Value scores are mapped to a predefined set of compression levels, such as high value, medium value, and low value, to determine the priority of information retention for vectors.

[0064] Prepare compression methods of varying complexity for each compression level, including but not limited to:

[0065] High value: A compression method that uses slight dimensionality reduction while preserving the main semantic structure;

[0066] Medium value: Medium-intensity quantization or vector block compression;

[0067] Low value: Highly quantized or hashed processing to significantly reduce storage footprint;

[0068] Obtain real-time resource status information of the end-side device, including storage usage, write load and available computing power, and determine the final compression path based on the resource status and compression level.

[0069] The basic memory vector is reduced in dimension, quantized, or encoded according to the selected compression path to obtain the compressed memory vector.

[0070] The process involves acquiring real-time resource status information of the end-side device, including storage usage, write load, and available computing power. Based on this resource status and the compression level, the final compression path is determined, as follows:

[0071] The storage management module reads the currently used storage space, the remaining available space, and the historical growth rate to determine whether it is necessary to increase the compression ratio or delay some write tasks.

[0072] Monitor the queue length of current vector write operations, the arrival rate of write requests, and the index update frequency to determine if the system is under high write load.

[0073] Real-time computing power data, including processor utilization, acceleration unit load, power consumption level, and allocable computing time slices, is obtained from the edge scheduler to determine whether the conditions are met for performing high-complexity compression operations.

[0074] The detection results of storage usage, write load, and available computing power are combined into a resource status description to reflect the overall operating pressure of the edge device at the current moment;

[0075] The final compression path is selected based on the resource status description and the compression level of the underlying memory vector:

[0076] When storage space is tight, prioritize methods with higher compression ratios (sparse thresholding, binary hashing).

[0077] When the write load is high, the fast compression method with fewer processing steps (fixed matrix projection, simplified bucketing) should be given priority.

[0078] When computing power is sufficient, high-precision compression methods with higher reconstruction quality (low-rank approximation, vector piecewise residual quantization) should be given priority.

[0079] When the compression level is high, even with limited resources, necessary information is retained to a certain extent to avoid loss of semantic importance (preservation of local subspace).

[0080] S3 inserts the compressed memory vector into the incremental index structure according to the compression level, and dynamically adjusts the insertion method of the compressed memory vector in combination with the real-time resource status.

[0081] In this embodiment, the compressed memory vector is inserted into the incremental index structure according to the compression level, and the insertion method of the compressed memory vector is dynamically adjusted in combination with the real-time resource status, as follows:

[0082] It receives the compressed memory vector output by the upstream compression module, parses its compression level, semantic cluster number, multi-level quantization parameters and other structural tags, and obtains the real-time resource status from the end-side resource monitoring module, including storage occupancy, write load, available computing power and energy consumption limits.

[0083] Input the real-time resource status into the edge resource prediction model (lightweight LSTM or edge Kalman predictor) to predict resource changes in the short time window, including predicting storage changes, predicting write load and predicting available computing power.

[0084] The prediction results are transformed into resource prediction vectors, which serve as pre-constraints for insertion disturbance assessment and strategy selection.

[0085] Based on the compression level, the compressed memory vector is mapped to the target level in the multi-level incremental index structure, and the perturbation impact features of this insertion are output in combination with the resource prediction vector. The perturbation impact features include the expected migration amount, reordering cost, and write conflict probability.

[0086] Based on the compression level, resource prediction vector, and perturbation impact characteristics, the optimal insertion strategy is selected from the preset insertion path set, and combined with the future recall cost output by the retrieval cost prediction model, to ensure that the selected strategy achieves a balance between write stability and recall performance.

[0087] The index write operation is performed according to the selected insertion strategy. The structural consistency during the write process is ensured through the corresponding stability control mechanism. When the disturbance score is high, the execution timing of the write task is adjusted according to the resource prediction results.

[0088] The process involves mapping the compressed memory vector to the target level in the multi-layer incremental index structure based on the compression level, and combining this with the resource prediction vector to output the perturbation impact features of this insertion, as detailed below:

[0089] Obtain the compression level information of the compressed memory vector, and according to the preset index level division rules, map different compression levels to the target level in the multi-level incremental index structure, so that high compression level vectors are written to the deep structure and low compression level vectors are written to the high-activity level.

[0090] Read the corresponding structural data from the target level to obtain the target index partition and corresponding bucket structure into which the compressed memory vector will be written. The structural data includes bucket density, key distribution, local sorting method and level depth.

[0091] Input the resource prediction vector into the edge resource prediction model, and generate corresponding resource constraint data based on the output predicted storage usage, predicted write load and predicted available computing power;

[0092] Based on the bucket density, key distribution, and sorting structure of the target level, combined with the semantic cluster number of the compressed memory vector, the expected insertion position range of the compressed memory vector within the target level is determined.

[0093] Based on the target level structured data, the node splitting probability and local key movement amount are calculated through the B-tree insertion cost model, and combined with the LSM-tree inter-layer compaction triggering probability, the index migration scale caused by the insertion operation is obtained, and the expected migration amount is output.

[0094] Obtain the local rank difference of the expected insertion position, and based on the incremental ranking update model, by comparing the rank difference of the newly inserted vector with the surrounding vectors, obtain the number of sorting nodes adjusted after the new vector is inserted, and obtain the scale of local sorting change, i.e., the reordering cost.

[0095] The predicted write load is input into the Optimistic Concurrency Control (OCC) model, which outputs the contention probability within the operation window. Combined with the Poisson arrival rate-based write event model, the probability of multiple concurrent inserts colliding in the same index region is predicted, thus obtaining the write conflict risk, i.e., the write conflict probability. Specifically, the OCC calculates the write event arrival rate of the target index region within a given time window based on the predicted write load, and obtains the write set crossover probability by combining bucket density and local access frequency. The Poisson arrival rate-based write event model calculates the probability of concurrent write overlap (the probability of multiple concurrent inserts colliding in the same index region) based on the write set crossover probability, thereby obtaining the write conflict risk of this insert operation.

[0096] S4, When the robot issues a query request, it inputs the query vector and the index information in the incremental index structure into the recall control module to form a recall request to be executed.

[0097] In this embodiment, the query vector and the index information in the incremental index structure are input into the recall control module to form a recall request to be executed, as follows:

[0098] When the robot generates a query request, it extracts the input data corresponding to the query from the task scheduling module and inputs it into the feature encoder to generate a query vector consistent with the memory vector space. The input data includes text-based instructions, visual perception results, multimodal sensing data, and task context.

[0099] Read the current index metadata from the multi-level incremental index structure. The index metadata includes index information such as bucket structure data of each index level, semantic cluster center and hierarchical mapping relationship;

[0100] The query vector is normalized to make its feature scale consistent with the compressed memory vector in the index structure, so as to ensure spatial alignment and dimensional consistency in subsequent recall calculations.

[0101] Using the semantic cluster centers in the index metadata as a reference, the similarity between the normalized query vector and each semantic cluster center is calculated, and the semantic clusters with the highest similarity are selected as candidate clusters based on the similarity ranking results.

[0102] Based on the semantic cluster-hierarchy mapping relationship and hierarchical activity matrix recorded in the index metadata, the hierarchical number, hierarchical span and cross-hierarchical activity corresponding to the candidate cluster are aggregated and analyzed to obtain the hierarchical number range covered by the candidate cluster in the multi-level incremental index structure, thereby generating the candidate search range.

[0103] The normalized query vector features, index structure data, and candidate search range are input into the recall control module to generate a recall request to be executed, which is represented as a recall request object containing the search range and the estimated recall depth.

[0104] S5. After receiving the recall request, the recall control module combines the compression level and the real-time resource status to select the best recall path and output the candidate memory set.

[0105] In this embodiment, the optimal recall path is selected by combining the compression level and the real-time resource status, and a candidate memory set is output, as follows:

[0106] Statistical analysis is performed on the compressed memory vectors within the candidate level range to form compression features, which include compression level, quantization residual size, local information preservation degree, and cross-level reconstruction pressure.

[0107] The system obtains real-time resource status from the edge resource monitoring module, including available computing power, storage access latency, read / write load, and energy consumption limits, and organizes them into a set of resource constraints as conditions for restricting recall operations.

[0108] By combining the compression features of candidate regions with resource constraints, the retrieval overhead generated by different recall paths is analyzed. The retrieval overhead includes vector reconstruction overhead (obtained through a hierarchical vector indexing method), search expansion overhead (based on a graph search method, obtained according to the number of graph jump levels and the number of node visits), access latency, and the risk of concurrent access conflicts.

[0109] Based on the analysis results, multiple candidate recall paths were constructed, including: low reconstruction overhead path (prioritizing the retrieval of low compression level vectors), hierarchical expansion path (gradually expanding from shallow to deep layers), resource adaptive path (adjusting the search scope and reconstruction quantity according to current resource constraints), and high similarity cluster priority path (prioritizing access to semantic clusters that are closest to the query vector).

[0110] The recall control module compares the overall costs of rebuilding, access latency, conflict risk, and recall completeness of candidate solutions to select the best recall path.

[0111] Perform the retrieval operation based on the optimal recall path and output the candidate memory set.

[0112] S6, determine whether to perform local dequantization based on the compression level of each memory in the candidate memory set, and generate the final high-quality memory;

[0113] In this embodiment, the decision to perform local dequantization is made based on the compression level of each memory in the candidate memory set to generate the final high-quality memory, as follows:

[0114] The compression level, quantization residual information, semantic cluster number, and historical access frequency of each candidate memory are read to provide a basis for local dequantization judgment;

[0115] Based on the compression level, residual information, and historical access frequency of each memory, determine whether to perform local dequantization;

[0116] The memories that need to be dequantized are grouped into semantic clusters, and local dequantization is performed in batches within the clusters. By utilizing the common features within the clusters, repetitive operations are reduced, the computational efficiency on the edge is improved, and the relative semantic relationships between memories within the clusters are maintained.

[0117] The candidate memories after local dequantization and the candidate memories that do not need to be dequantized are uniformly mapped to the same feature space and normalized to ensure that memories of different compression levels have consistent scale in subsequent similarity calculation, fusion and ranking.

[0118] All locally dequantized memories and memories that do not require dequantization are integrated to form a final high-quality memory set. At the same time, it can be prioritized according to compression level, historical access frequency and cluster activity to provide high-value and usable memory information for downstream decision-making or cross-modal fusion modules.

[0119] The determination of whether to perform local dequantization based on the compression level, residual information, and historical access frequency of each memory is as follows:

[0120] The characteristics of each memory are analyzed to determine the dequantization depth and range, including compression level, quantization residual information, semantic cluster number and historical access frequency;

[0121] Compression level: The higher the compression level, the greater the information loss, and the deeper the dequantization is required;

[0122] Residual information content: The larger the residual, the more feature dimensions need to be recovered;

[0123] Historical access frequency: Prioritize the memory of high-frequency accesses to ensure the recovery of core features;

[0124] Semantic cluster distribution: Other memory features within a cluster can be used to determine the feature dimensions of the shared reconstruction;

[0125] A comprehensive dequantification strategy for each memory based on its characteristics:

[0126] The memory with high compression level and a lot of residual information is subjected to complete local dequantization to recover all key feature dimensions;

[0127] Local dequantization of the memory execution portion at medium compression levels or historically frequently accessed levels only restores the core feature dimensions.

[0128] Memory with low compression levels or high information retention does not require dequantization and can be directly retained.

[0129] The above formulas are all dimensionless calculations. The formulas are derived from software simulations based on a large amount of collected data to obtain the most recent real-world results. The preset parameters in the formulas are set by those skilled in the art according to the actual situation.

[0130] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented, in whole or in part, in the form of a computer program product.

[0131] Those skilled in the art will recognize that the modules and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0132] In addition, the functional modules in the various embodiments of this application can be integrated into one processing module, or each module can exist physically separately, or two or more modules can be integrated into one module.

[0133] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0134] In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for managing the memory system of an end-side robot that combines dynamic compression and efficient recall, characterized in that, include: The collected data signals are mapped to a unified feature space to generate a basic memory vector, which is then input into a value evaluator to output a value score. Based on the value score, the corresponding compression method is selected to compress the basic memory vector, resulting in a compressed memory vector and a compression level. The compressed memory vector is inserted into the incremental index structure according to the compression level, and the insertion method of the compressed memory vector is dynamically adjusted in combination with the real-time resource status. When the robot issues a query request, it inputs the query vector and the index information in the incremental index structure into the recall control module to form a recall request to be executed. After receiving a recall request, the recall control module combines the compression level and real-time resource status to select the best recall path and output a candidate memory set. Based on the compression level of each memory in the candidate memory set, determine whether to perform local dequantization to generate the final high-quality memory; The process involves mapping the collected data signals to a unified feature space, generating a basic memory vector, inputting it into a value evaluator, and outputting a value score, as detailed below: The output signal of the target under test is acquired to form a multimodal raw dataset, and preprocessing is performed on the raw data of each modality. The preprocessed modal data are input into the lightweight modal backbone network to obtain modal feature vectors for different modalities. Perform linear mapping and regularization on the modal feature vectors to transform them into projection vectors with consistent dimensions; By using a lightweight attention fusion method, multiple projection vectors are fused together to obtain a single basic memory vector; A lightweight prediction head is constructed. The basic memory vector is input into the Softmax distribution layer of the prediction head to obtain the probability distribution, and the prediction entropy is output based on the probability distribution. The basic memory vector is input into the Dropout layer of the prediction head. Several forward propagations are performed on the same basic memory vector to obtain several sets of outputs, and the variance is output, i.e., dropout forward variance. Output the confidence score based on the prediction entropy and dropout forward variance; Maintain a lightweight index of recent vectors at the end, calculate the distance distribution between the vector to be evaluated and its neighborhood, and form the novelty. The cosine similarity between the base memory vector and the task vector is calculated to obtain the task relevance score. The time decay is obtained by calculating the difference between the vector timestamp and the current time; The task relevance score, novelty, confidence, and time decay are combined according to a set weight to output a value score; The process of maintaining a lightweight index of recent vectors at the edge and calculating the distance distribution between the vector to be evaluated and its neighborhood to form novelty is as follows: A lightweight index is created for the written compressed memory vector, the index key is stored using locality-sensitive hashing, and a portion of the basic memory vector is selected as the vector to be evaluated. Using the same hash function as the index structure, calculate the index key of the vector to be evaluated, locate the corresponding neighborhood bucket in the lightweight index, and determine the candidate neighborhood range; Read the basic memory vector in the neighborhood bucket, and filter redundant vectors according to the timestamp and the upper limit of the number to limit the neighborhood size and obtain the neighborhood vector set; The distance between the vector to be evaluated and each vector in the set of neighboring vectors is calculated using the cosine distance function, and the distance distribution characteristics are statistically analyzed. Based on the distance distribution characteristics, the novelty is output using a local outlier probability algorithm. The process of inserting the compressed memory vector into the incremental index structure according to the compression level, and dynamically adjusting the insertion method of the compressed memory vector in conjunction with the real-time resource status, is as follows: Parse the compressed memory vector and obtain the real-time resource status; The real-time resource status is input into the edge resource prediction model to predict resource changes within a short future time window, resulting in predicted storage changes, predicted write load, and predicted available computing power, which are then converted into resource prediction vectors. Based on the compression level, the compressed memory vector is mapped to the target level in the multi-layer incremental index structure, and the perturbation impact features of this insertion are output in combination with the resource prediction vector. Based on the compression level, resource prediction vector, and disturbance impact characteristics, the optimal insertion strategy is selected from the preset insertion path set. Perform index write operations according to the selected insert strategy.

2. The end-side robot memory system management method with dynamic compression and efficient recall coordination according to claim 1, characterized in that, The process involves mapping the compressed memory vector to the target level in the multi-layer incremental index structure based on the compression level, and combining this with the resource prediction vector to output the perturbation impact features of this insertion, as detailed below: Obtain the compression level information of the compressed memory vector, and according to the preset index level division rules, map different compression levels to the target level in the multi-level incremental index structure; Read the corresponding structure data from the target level to obtain the target index partition and corresponding bucket structure into which the compressed memory vector will be written; Input the resource prediction vector into the edge resource prediction model, and generate corresponding resource constraint data based on the output predicted storage occupancy, predicted write load and predicted available computing power; Based on the bucket density, key distribution, and sorting structure of the target level, combined with the semantic cluster number of the compressed memory vector, the expected insertion position range of the compressed memory vector within the target level is determined. Based on the target level structured data, the node splitting probability and local key movement amount are calculated through the B-tree insertion cost model, and combined with the LSM-tree inter-layer compaction triggering probability, the index migration scale caused by the insertion operation is obtained, and the expected migration amount is output. Obtain the local region sort rank difference at the expected insertion position, and based on the incremental sorting model, obtain the number of sorted nodes adjusted after the new vector is inserted, thus obtaining the scale of local sorting change, i.e., the reordering cost. The predicted write load is input into an optimistic concurrency control conflict model, the output is the contention probability within the operation window, and combined with a write event model based on Poisson arrival rate, the probability of multiple concurrent inserts colliding in the same index region is predicted to obtain the write conflict probability.

3. The end-side robot memory system management method based on dynamic compression and efficient recall coordination according to claim 2, characterized in that, The process of inputting the query vector and the index information in the incremental index structure into the recall control module to form a recall request to be executed is as follows: The input data corresponding to the query is extracted from the task scheduling module and input into the feature encoder to generate a query vector consistent with the memory vector space. Read the current index metadata from the multi-level incremental index structure; The query vector is normalized, and the semantic cluster centers in the index metadata are used as reference benchmarks. The similarity between the normalized query vector and each semantic cluster center is calculated, and the semantic clusters with the highest similarity are selected as candidate clusters based on the similarity ranking results. Based on the semantic cluster-hierarchy mapping relationship and hierarchical activity matrix recorded in the index metadata, the hierarchical number, hierarchical span and cross-hierarchical activity corresponding to the candidate cluster are aggregated and analyzed to obtain the hierarchical number range covered by the candidate cluster in the multi-level incremental index structure, and to generate the candidate search range. The normalized query vector features, index structure data, and candidate search range are input into the recall control module to generate a recall request to be executed, which is represented as a recall request object containing the search range, candidate level, and estimated recall depth.

4. The end-side robot memory system management method with dynamic compression and efficient recall coordination according to claim 3, characterized in that, The optimal recall path is selected by combining the compression level and the real-time resource status, and a candidate memory set is output, as follows: Statistical analysis is performed on the compressed memory vectors within the candidate level range to form compressed features; Obtain the real-time resource status and organize it into a resource constraint set, which serves as the condition for restricting the recall operation; By combining the compression features of candidate regions with resource constraints, the retrieval overhead generated by different recall paths is analyzed, and multiple candidate recall paths are constructed based on the analysis results. Compare the overall costs of rebuilding, access latency, conflict risk, and recall completeness of candidate recall paths, and select the best recall path; Perform the retrieval operation based on the optimal recall path and output the candidate memory set.

5. The end-side robot memory system management method with dynamic compression and efficient recall coordination according to claim 4, characterized in that, The process of determining whether to perform local dequantization based on the compression level of each memory in the candidate memory set to generate the final high-quality memory is as follows: Read the compression level, quantization residual information, semantic cluster number and historical access frequency of each candidate memory, and determine whether to perform local dequantization; Group the memories that need to be dequantized into semantic clusters, and perform local dequantization in batches within each cluster; The candidate memories after local dequantization and the candidate memories that do not need to be dequantized are uniformly mapped to the same feature space and normalized. All locally dequantized memories and memories that do not require dequantization are integrated to form a final high-quality memory set.

Images