Amr controller model cross-fleet incremental learning method based on federated learning

Abstract

The application relates to the technical field of federated learning, and discloses an AMR controller model cross-vehicle fleet incremental learning method based on federated learning, which comprises the following steps: establishing and maintaining a global basic model library comprising a plurality of different AMR vehicle fleet controller models; when a new vehicle fleet joins, selecting a local controller model for the vehicle fleet from the global basic model library according to the running environment characteristics and task requirements; collecting running data during local operation of the vehicle fleet, training the local controller model based on the running data, extracting local adjustment parameters generated by the training, performing safety verification on the parameters, and updating the parameters to the global basic model library; and the vehicle fleet selects relevant local adjustment parameters from the global basic model library according to the current running environment and task requirements, and performs fusion training on the relevant local adjustment parameters and the local controller model to realize continuous individual optimization of the local controller model. The application effectively solves the problems of data islands and privacy security.

Description

Technical Field

[0001] This application relates to the field of federated learning technology, and more specifically, to a federated learning-based method for incremental learning of AMR controller models across fleets. Background Technology

[0002] With the increasing demand for automation and intelligence in modern industrial production, autonomous mobile robots (AMRs) are playing an increasingly important role in logistics and manufacturing. However, enterprises often face numerous challenges when deploying AMR fleets on a large scale. For example, different fleets generate large amounts of data in their own complex operating environments. This data is often scattered across different regions, forming data silos that are difficult to share, and directly centralizing the processing of this data may raise privacy and security concerns. To overcome these challenges, federated learning, as a distributed learning approach, offers a potential way to enable different fleets to learn collaboratively and improve model performance while protecting data privacy. However, directly applying traditional federated learning methods to the training of AMR controller models still has some significant shortcomings.

[0003] Specifically, in the scenario of incremental learning of AMR controller models across fleets, traditional federated learning methods struggle to effectively balance the general adaptability of the global model with the local personalized needs of each fleet, given the complex conditions of dynamically changing operating environments, highly inconsistent data distribution, and stringent requirements for privacy, security, and communication efficiency. Traditional methods typically pursue a single globally optimal model, making it difficult to adapt to the inconsistent data distribution among different fleets, resulting in a difficulty in finding a balance between model generality and meeting the specific needs of each fleet. Secondly, its static batch learning mode easily leads to the forgetting of previously acquired knowledge when learning new tasks or environments, and it cannot quickly and effectively absorb new information. Furthermore, traditional synchronous aggregation mechanisms suffer from communication efficiency bottlenecks, consuming significant computational resources, and lack effective response and security protection capabilities against performance differences between different devices within the system and potential malicious attacks.

[0004] To address the aforementioned issues, existing technologies urgently need improvement. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this application provides a federated learning-based method for cross-team incremental learning of AMR controller models. This method addresses the difficulty in effectively balancing the general adaptability of the global model with the local personalized needs of each team in cross-team incremental learning scenarios, as well as the deficiencies of traditional federated learning methods in adapting to dynamic and changing environments, inconsistent data distribution, privacy and security, and communication efficiency.

[0006] In response, this application discloses a method for incremental learning of AMR controller models across fleets based on federated learning, specifically including: Establish and maintain a global basic model library that includes multiple different AMR fleet controller models; When a new AMR fleet joins, a matching controller model is selected and distributed from the global basic model library as the local controller model, based on the operating environment characteristics and task requirements of the new AMR fleet. Each AMR fleet collects its own operational data during local operation, trains the local controller model based on its own operational data, extracts the local adjustment parameters generated by the training, performs safety verification on the local adjustment parameters, and updates the local adjustment parameters that have passed the safety verification to the global basic model library. Each AMR team selects relevant local adjustment parameters from the global basic model library based on its current operating environment and mission requirements, and uses these local adjustment parameters to perform fusion training with the local controller model in order to achieve continuous personalized optimization of the local controller model.

[0007] Through this technical solution, this application constructs a federated learning framework, which enables incremental learning and personalized optimization of the AMR controller model across different fleets, effectively solving the problems of data silos and privacy security, while taking into account both the model's universality and the specific needs of each fleet.

[0008] Furthermore, based on the operating environment characteristics and mission requirements of the new AMR fleet, the steps of selecting and distributing a matching controller model from the global basic model library as the local controller model for the new AMR fleet include: Obtain the warehouse type, channel features, lighting conditions, and main task types of the AMR fleet, and calculate the matching score between the warehouse type, channel features, lighting conditions, and main task types and the scene labels adapted to each controller model in the global basic model library. Select and distribute the controller model with the highest matching score as the local controller model for the corresponding AMR fleet.

[0009] This technical solution enables the intelligent selection of the most suitable base model based on the specific operating environment and mission requirements of the AMR fleet, improving the efficiency and accuracy of model initialization and laying a solid foundation for subsequent personalized learning.

[0010] Based on this, the steps of training the local controller model using their respective operational data and extracting the local adjustment parameters generated during training include: The local controller model is trained using efficient parameter fine-tuning techniques based on their respective operational data; Only the adapter parameters or low-rank matrix parameters that change during training and represent new knowledge are retained and extracted as local adjustment parameters.

[0011] This technical solution employs efficient parameter fine-tuning techniques, which significantly reduces communication overhead and computational resource consumption, while effectively avoiding catastrophic forgetting problems and ensuring that the model can efficiently absorb new knowledge during incremental learning.

[0012] In some preferred embodiments, security verification includes: Check whether the value range of the local adjustment parameters is within the preset reasonable range; And / or, analyze whether there are any anomalies in the gradient statistics of the locally adjusted parameters; And / or, compare local adjustment parameters with local adjustment parameters from historical global model updates to identify potentially malicious data.

[0013] This technical solution introduces multiple security verification mechanisms, effectively identifying and resisting potential malicious data and abnormal parameters, ensuring the security and stability of global model updates.

[0014] Furthermore, the steps for selecting relevant local adjustment parameters from the global basic model library based on the current operating environment and task requirements include: maintaining labels representing the current operating environment and task requirements of each AMR fleet; calculating the matching degree between the labels and the association descriptions of each local adjustment parameter in the global basic model library; and downloading and applying only one or more local adjustment parameters whose matching degree exceeds a preset threshold.

[0015] This technical solution enables AMR fleets to accurately select and apply relevant local adjustment parameters according to their actual needs, avoiding unnecessary parameter downloads and calculations, and further improving learning efficiency and personalization.

[0016] As a technical improvement, the steps of updating the locally adjusted parameters that have passed security verification to the global base model library also include: For each local adjustment parameter that passes security verification, perform a contextual attribute assessment to determine whether it is a general local adjustment parameter or a professional local adjustment parameter; If it is determined to be a general local adjustment parameter, it will be integrated into the core general capability module of the global basic model library; If it is determined to be a professional local adjustment parameter, it is encapsulated as an independent context-specific adaptation module and stored in the global basic model library.

[0017] This technical solution classifies local adjustment parameters by contextual attributes, effectively separating and managing general knowledge and professional knowledge, making the structure of the global model library more reasonable, and improving the model's generalization ability and adaptability in specific scenarios.

[0018] To improve the solution, the scenario attribute evaluation includes: applying the local adjustment parameters to a preset simulated global model copy, evaluating the performance improvement on a general task test set in the preset simulated global model copy, and evaluating the performance improvement on a preset specific extreme scenario task set; if the performance improvement on the preset specific extreme scenario task set is higher than the performance improvement on the preset general task test set, it is judged as a professional local adjustment parameter; otherwise, it is judged as a general local adjustment parameter.

[0019] In one implementation, multiple different AMR fleet controller models are pre-trained for different AMR operating scenarios.

[0020] Based on the above, different AMR operating scenarios include indoor flat ground navigation scenarios, outdoor complex terrain obstacle avoidance scenarios, and heavy-load transport path planning scenarios.

[0021] As a further improvement, the AMR fleet controller model includes a navigation policy network based on deep reinforcement learning, which is used to output motion control commands for the AMR fleet based on sensor data.

[0022] In summary, the federated learning-based cross-team incremental learning method for AMR controller models provided in this application achieves personalized model initialization by establishing a global basic model library and distributing matching local controller models according to the operating environment and task requirements of the new AMR fleet. After local training, each AMR fleet extracts local adjustment parameters and performs security verification, then updates the verified parameters to the global model library. Finally, each fleet selects relevant parameters from the global model library for fusion training according to its own needs, achieving continuous optimization of the local model. This method effectively solves the problems of data silos, privacy and security, and the difficulty in balancing model generality and personalized needs in cross-team incremental learning scenarios for AMR controller models in existing technologies, as well as the shortcomings of traditional federated learning methods in adapting to dynamic and changing environments, inconsistent data distribution, communication efficiency, and security protection. Through the distributed collaboration mechanism of federated learning, this application achieves collaborative learning and performance improvement of controller models from different AMR fleets while protecting data privacy, taking into account both the general adaptability of the global model and the personalized needs of each fleet, overcoming the limitations of traditional methods in handling dynamic environments and data heterogeneity, and improving communication efficiency and security. Attached Figure Description

[0023] Figure 1 This is a flowchart illustrating the cross-vehicle incremental learning method for AMR controller models based on federated learning provided in this application embodiment. Detailed Implementation

[0024] The technical solutions of this application will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are merely some embodiments of this application, and not all embodiments. The components of this application described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of this application provided in the accompanying drawings is not intended to limit the scope of the claimed application, but merely to illustrate selected embodiments of this application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without inventive effort are within the scope of protection of this application.

[0025] It should be noted that similar reference numerals and letters in the following figures indicate similar items; therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures. Furthermore, in the description of this application, terms such as "first," "second," etc., are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0026] Traditional AMR controller models often present numerous challenges for enterprises when deploying AMR fleets on a large scale. For example, different fleets generate vast amounts of data in their complex operating environments, often scattered across different regions, creating "data silos" that are difficult to share. Furthermore, directly centralizing this data can raise privacy and security concerns. Applying traditional federated learning methods directly to AMR controller model training still has significant shortcomings. It struggles to effectively balance the general adaptability of the global model with the specific needs of each fleet, and its static batch learning model can lead to the forgetting of previously acquired knowledge when learning new tasks or environments, hindering the rapid and effective absorption of new information. Moreover, traditional synchronous aggregation mechanisms suffer from communication efficiency bottlenecks, consuming significant computational resources, and lack effective countermeasures and security protection against performance differences between internal devices and potential malicious attacks.

[0027] In this regard, refer to Figure 1 This application proposes a cross-vehicle incremental learning method for AMR controller models based on federated learning, including: Establish and maintain a global basic model library that includes multiple different AMR fleet controller models; When a new AMR fleet joins, a matching controller model is selected and distributed from the global basic model library as the local controller model, based on the operating environment characteristics and task requirements of the new AMR fleet. Each AMR fleet collects its own operational data during local operation, trains the local controller model based on its own operational data, extracts the local adjustment parameters generated by the training, performs safety verification on the local adjustment parameters, and updates the local adjustment parameters that have passed the safety verification to the global basic model library. Each AMR team selects relevant local adjustment parameters from the global basic model library based on its current operating environment and mission requirements, and uses these local adjustment parameters to perform fusion training with the local controller model in order to achieve continuous personalized optimization of the local controller model.

[0028] This application introduces a federated learning mechanism to achieve collaborative incremental learning of the AMR controller model across different fleets while protecting data privacy. This effectively solves problems such as data silos, privacy and security, balancing model universality and personalized needs, and knowledge forgetting, and significantly improves the adaptability, learning efficiency, and security of the AMR fleet controller model.

[0029] AMR (Autonomous Mobile Robot) refers to a robot capable of autonomously perceiving its environment, planning paths, and performing tasks without human intervention. The AMR controller model is the core component of an AMR, responsible for receiving sensor data and outputting motion control commands to achieve functions such as navigation, obstacle avoidance, and material handling. Federated learning is a distributed machine learning paradigm that allows multiple participants to collaboratively train a shared model without sharing raw data. The global base model library is a knowledge base that centrally stores and manages controller models from multiple AMR fleets. It includes controller models pre-trained for different operating scenarios, as well as locally adjusted parameters contributed by each fleet. Local controller models are controller models distributed to a specific AMR fleet and trained and optimized locally within that fleet. Locally adjusted parameters are parameters generated by the AMR fleet during local training, fine-tuning its local controller model based on its own operating data. These parameters represent the fleet's unique knowledge and experience.

[0030] The federated learning-based cross-vehicle incremental learning method for AMR controller models proposed in this application aims to address challenges such as data silos, privacy protection, model adaptability, and efficiency in large-scale AMR deployments.

[0031] Specifically, this method first establishes and maintains a global basic model library. This library stores multiple controller models pre-trained for different AMR operating scenarios. For example, these models may include models for indoor flat ground navigation scenarios, models for outdoor complex terrain obstacle avoidance scenarios, and models for heavy-load transport path planning scenarios. These pre-trained models provide newly added AMR fleets with initial, somewhat general-purpose controller capabilities.

[0032] When a new AMR fleet joins, a suitable local controller model needs to be selected for it. One approach is for a manual selection of the best-matching controller model from a global base model library, based on the new fleet's operating environment characteristics and task requirements. For example, if the new fleet primarily performs light-load handling in warehouses, a model optimized for indoor navigation and light-load handling can be selected. Another approach is for the system to calculate a matching score between the new AMR fleet's operating environment characteristics and task requirements, such as warehouse type, passageway characteristics, lighting conditions, and primary task type, and the scene labels applicable to each controller model in the global base model library. The system then selects the controller model with the highest matching score as the local controller model for the corresponding AMR fleet.

[0033] Each AMR fleet continuously collects its own operational data during local operation. Based on this data, each fleet trains its local controller model. For example, a traditional full-parameter fine-tuning method can be used, which updates all parameters of the local controller model. After training, the locally adjusted parameters generated during training are extracted. These locally adjusted parameters can be complete copies of all model parameters or differences between model parameters.

[0034] After extracting the local adjustment parameters, these parameters need to be security verified. One verification method is for human experts to review the local adjustment parameters and determine whether they meet expectations. Another verification method is for the system to check whether the numerical range of the local adjustment parameters is within a preset reasonable range; for example, parameter values ​​should not be too large or too small to avoid model crashes. Furthermore, the gradient statistics of the local adjustment parameters can be analyzed for anomalies; for example, a sudden increase in gradient values ​​may indicate malicious attacks or data anomalies. Moreover, the local adjustment parameters can be compared with the local adjustment parameters updated in the historical global model to identify potentially malicious data; for example, parameters that deviate significantly from historical data may originate from contaminated data. Local adjustment parameters that pass security verification will be updated in the global base model library.

[0035] Each AMR fleet selects relevant local adjustment parameters from the global basic model library based on its current operating environment and mission requirements. For example, a fleet can periodically download all local adjustment parameters from the global basic model library and filter them locally. Another implementation method is for the system to maintain tags representing each AMR fleet's current operating environment and mission requirements, such as high-density shelving areas or nighttime operations. Then, the system calculates the matching degree between these tags and the association descriptions of each local adjustment parameter in the global basic model library. For example, a local adjustment parameter might be labeled as suitable for obstacle avoidance in narrow passages; if the fleet's tag includes high-density shelving areas, the parameter's matching degree will be high. Finally, only one or more local adjustment parameters with a matching degree exceeding a preset threshold are downloaded and applied.

[0036] Finally, each AMR team uses selected relevant local adjustment parameters to fuse and train its local controller model, enabling continuous personalized optimization of the local controller model. For example, downloaded local adjustment parameters can directly replace the corresponding parameters in the local controller model. Another fusion method is to use the downloaded local adjustment parameters as regularization terms, guiding the local controller model to align with the knowledge represented by these parameters during training. In this way, the local controller model can continuously absorb general knowledge and context-specific knowledge from other teams while maintaining its own personalized adaptability.

[0037] The federated learning-based cross-fleet incremental learning method for AMR controller models proposed in this application works by constructing a global base model library, enabling centralized management and distributed learning of AMR controller model knowledge. When a new AMR fleet joins, it can obtain an initial controller model from the global base model library that matches its operating environment and task requirements, thus avoiding the time and resource consumption of training from scratch. Each AMR fleet trains its local controller model locally using its own data and extracts local adjustment parameters representing its unique knowledge. These local adjustment parameters are updated to the global base model library after rigorous security verification. This process ensures the quality and security of shared knowledge, preventing malicious attacks or abnormal data from affecting the global model. The global base model library not only stores general knowledge but also encapsulates specialized knowledge into independent context-specific adaptation modules, better meeting the personalized needs of different fleets. Each AMR fleet can select and download relevant local adjustment parameters from the global base model library according to its current operating environment and task requirements, and fuse them with its local controller model for training. This incremental learning mechanism enables local controller models to continuously absorb the latest knowledge from other fleets while maintaining adaptability to their specific environments. This effectively solves the problems faced by traditional federated learning in AMR controller model training, such as data silos, privacy protection, balancing model generality and personalization needs, and knowledge forgetting. Through this collaborative incremental learning approach, various AMR fleets can jointly improve the performance of their controller models, forming a continuously evolving intelligent AMR ecosystem.

[0038] In some of the embodiments described above in this application, when a new AMR fleet joins, a matching controller model needs to be selected and distributed from the global basic model library as the local controller model, based on its operating environment characteristics and task requirements. However, in actual implementation, how to efficiently and accurately determine the degree of matching between the model and the fleet's operating environment and task requirements, and automatically select the optimal model, is a problem that needs further refinement. If the matching process is not precise enough or the degree of automation is not high, the newly joined fleet may receive a suboptimal initial model, thereby affecting its subsequent local training efficiency and performance.

[0039] In this regard, this application further proposes the following steps for selecting and distributing a matching controller model as a local controller model for the new AMR fleet from the global basic model library, based on the operating environment characteristics and mission requirements of the new AMR fleet: Obtain the warehouse type, channel features, lighting conditions, and main task types of the AMR fleet, and calculate the matching score between the warehouse type, channel features, lighting conditions, and main task types and the scene labels adapted to each controller model in the global basic model library. Select and distribute the controller model with the highest matching score as the local controller model for the corresponding AMR fleet.

[0040] Specifically, when acquiring the operating environment characteristics and task requirements of an AMR fleet, key information such as warehouse type, passageway characteristics, lighting conditions, and main task types can be collected. Warehouse type can refer to, for example, ambient temperature warehouses, cold storage, or hazardous materials warehouses; passageway characteristics can refer to, for example, passageway width, presence of obstacles, and ground material; lighting conditions can refer to, for example, indoor lighting, outdoor lighting, and nighttime operations; and main task types can refer to, for example, handling, sorting, and inspection. These characteristics collectively constitute the scenario profile of the AMR fleet's operation. Furthermore, each controller model in the global basic model library is assigned corresponding scenario labels during pre-training. These scenario labels describe the environment and task type best suited for the model's operation. For example, a model trained for indoor flat ground navigation scenarios might have scenario labels including: warehouse type: ambient temperature warehouse; passageway characteristics: spacious and flat; lighting conditions: indoor lighting; and main task type: handling. When calculating the matching score, the operating environment characteristics and task requirements acquired for the new AMR fleet can be compared with the scenario labels adapted to each controller model in the global basic model library. The matching score can be implemented using various algorithms, such as cosine similarity calculation based on feature vectors or a rule-based weighted scoring mechanism. For example, weights can be assigned to each feature dimension, and scores can be accumulated based on the degree of feature matching. Ultimately, the system selects and distributes the controller model with the highest matching score as the local controller model for the corresponding AMR fleet. This means that newly joined AMR fleets will receive a controller model that best suits their operating scenario from the initial stage, thus laying a solid foundation for their subsequent local incremental learning.

[0041] This application's solution achieves quantitative and automated selection of matching controller models by clearly defining the operating environment characteristics and task requirements of an AMR fleet and calculating matching scores between these characteristics and scene labels of each controller model in a global basic model library. Specifically, by acquiring detailed scene information such as warehouse type, channel characteristics, lighting conditions, and main task types, a unique operating profile of the new AMR fleet can be constructed. Simultaneously, each controller model in the global basic model library has pre-set scene labels for its suitability, meaning the model itself carries metadata about its applicable scope. By calculating the matching scores between these features and scene labels, the fit between each basic model and the actual needs of the new fleet can be objectively evaluated. Therefore, selecting the controller model with the highest matching score ensures that newly added AMR fleets receive a basic model that highly matches their operating environment and task requirements from the initial stage, avoiding the uncertainty and inefficiency caused by blind selection or manual judgment.

[0042] The aforementioned technical solutions significantly improve the accuracy and automation of initial controller model selection for new AMR fleets. Through refined feature matching and score calculation, it ensures that the base model best suited to the specific operating scenario is distributed to the new fleet, thereby reducing the initial adjustments required for local personalized optimization and accelerating model adaptation to the new environment. This not only improves the efficiency of AMR fleet deployment but also provides a better starting point for subsequent incremental learning, helping to achieve high-performance operation more quickly and reducing operating costs and time investment.

[0043] In some preferred embodiments, it is assumed that a new fleet of AMRs is about to be deployed in a large cold chain warehouse with narrow freezing aisles, low-light environment, and whose primary task is heavy-duty pallet handling.

[0044] First, the system will obtain the operating environment characteristics and task requirements of the AMR fleet, specifically: the warehouse type is cold storage, the passage characteristics are narrow and icy, the lighting conditions are low, and the main task type is heavy-duty handling.

[0045] Next, the system accesses the global base model library, which may contain multiple pre-trained controller models, each with its own adapted scene label. For example: Model A: Scene labels are "Ambient temperature warehouse, spacious aisle, normal lighting, light load sorting". Model B: Scene labels are "cold storage, narrow aisle, low light, heavy-duty handling". Model C: Scene labels are "Outdoor, Complex Terrain, Strong Light, Inspection" The system will match the characteristics of the new AMR fleet with the scene labels of models A, B, and C to calculate a score. For example, through a weighted scoring mechanism, model B, which closely matches the needs of the new fleet in terms of features such as cold storage, narrow passages, low light, and heavy-duty handling, will receive the highest matching score.

[0046] Ultimately, the system will select and distribute Model B as the local controller model for the new AMR fleet. Thus, the AMR fleet obtains a controller model highly matched to its demanding operating environment and heavy-load tasks from the outset, providing excellent initial conditions for subsequent local incremental learning and significantly improving model adaptability and operational efficiency.

[0047] In some of the embodiments described above in this application, each AMR fleet needs to train its local controller model based on its own operational data and extract the local adjustment parameters generated during local operation. However, in actual implementation, directly training the entire local controller model and extracting all changed parameters may lead to low training efficiency, high computational resource consumption, and a large scale of generated local adjustment parameters, which is not conducive to efficient transmission and fusion under the federated learning framework, especially when the AMR controller model itself is relatively complex.

[0048] In this regard, this application further proposes the following steps for training the local controller model based on their respective operating data and extracting the local adjustment parameters generated during training: The local controller model is trained using efficient parameter fine-tuning techniques based on their respective operational data; Only the adapter parameters or low-rank matrix parameters that change during training and represent new knowledge are retained and extracted as local adjustment parameters.

[0049] Specifically, Parameter-Efficient Fine-Tuning (PEFT) is a method for fine-tuning a pre-trained model. Its core idea is to improve model performance on a specific task by introducing only a small number of trainable parameters or making low-rank modifications to existing parameters while keeping most of the pre-trained model's parameters unchanged. This technique can significantly reduce the computational resources and storage space required for training and reduce the risk of overfitting. Adapter parameters refer to the parameters introduced by inserting small neural network modules (i.e., adapters) between or within layers of the pre-trained model. These adapter modules typically contain fewer parameters, and only these adapter parameters are trained during fine-tuning, while the main parameters of the pre-trained model remain frozen. Low-rank matrix parameters refer to the parameters obtained by performing low-rank decomposition (e.g., using LoRA) on the weight matrix in the pre-trained model, decomposing the original high-dimensional weight matrix into a product of two or more low-rank matrices, and training only the parameters of these low-rank matrices during fine-tuning. These low-rank matrix parameters can capture the adaptive changes of the model on new tasks with a smaller number of parameters. By retaining and extracting only the adapter parameters or low-rank matrix parameters that change during training and represent new knowledge as local adjustment parameters, it can be ensured that the extracted parameters are incremental knowledge learned for specific fleet operation data, rather than redundant updates to the entire model.

[0050] This application's solution employs a parameter-efficient fine-tuning technique to train the local controller model, extracting only the adapter parameters or low-rank matrix parameters that change during training and represent new knowledge as local adjustment parameters. This effectively solves the efficiency and resource consumption problems associated with traditional full-scale fine-tuning. Specifically, when an AMR fleet trains its local controller model locally based on its operational data, the parameter-efficient fine-tuning technique keeps most of the model's pre-trained parameters unchanged, updating only a small number of introduced adapter parameters or low-rank matrix parameters. These few parameters are specifically designed to capture new knowledge arising from the fleet's unique operating environment and task requirements. This significantly reduces the computational and memory usage during training, accelerating the training speed. Furthermore, because only these few parameters representing new knowledge are extracted as local adjustment parameters, rather than the parameters of the entire model, the amount of data that needs to be uploaded to the global base model library is significantly reduced, communication bandwidth requirements are lowered, and the overall efficiency of federated learning is improved.

[0051] Through the aforementioned technical solutions, this application significantly improves the incremental learning efficiency of AMR controller models within a federated learning framework. Specifically, by employing efficient parameter fine-tuning techniques, the computational resources and time required for local training can be drastically reduced, enabling AMR teams to update their models more frequently and economically. Furthermore, extracting only a small number of locally adjusted parameters representing new knowledge greatly reduces data transmission volume, alleviates network burden, and helps protect the privacy of local data, as what is transmitted is not the original data or the complete model, but highly compressed parameters focused on incremental knowledge. This allows the global base model library to more quickly and effectively aggregate personalized learning results from different teams, thereby achieving continuous optimization of the AMR controller model and rapid adaptation to new environments.

[0052] In some preferred embodiments, assume an AMR fleet operates in a new warehouse environment with unique aisle widths and obstacle distributions. The fleet obtains a generic AMR controller model from a global base model library as its local controller model. To adapt this local controller model to the new environment, the fleet locally collects its operational data in the new warehouse. During training, LoRA (Low-Rank Adaptation) can be used as an efficient parameter fine-tuning technique. Specifically, in the key layers of the local controller model (e.g., neural network layers used for feature extraction or decision output), a low-rank update matrix ΔW is added to each weight matrix W, where ΔW is decomposed into the product of two smaller matrices A and B (ΔW = BA). During training, only the parameters of matrices A and B are updated, while the original weight matrix W remains frozen. After training, only the parameters of these trained matrices A and B are extracted as local adjustment parameters. For example, if the original model has millions of parameters, LoRA may only need to extract tens of thousands or even thousands of parameters. These local adjustment parameters are then securely verified and uploaded to the global base model library. When other AMR teams need to adapt to similar new environments, they can select and download these low-rank matrix parameters from the global base model library and fuse them with their own local controller models, thereby quickly achieving personalized optimization of the model without having to train or download the entire model from scratch.

[0053] In some embodiments described above in this application, each AMR fleet collects its own operational data during local operation, trains its local controller model based on this data, extracts the local adjustment parameters generated during training, and performs safety verification on these local adjustment parameters. Specifically, the safety verification steps described above can be further refined.

[0054] Security verification includes: Check whether the value range of the local adjustment parameters is within the preset reasonable range; And / or, analyze whether there are any anomalies in the gradient statistics of the locally adjusted parameters; And / or, compare local adjustment parameters with local adjustment parameters from historical global model updates to identify potentially malicious data.

[0055] Specifically, upon receiving the local adjustment parameters submitted by each AMR team, a preliminary numerical check is first performed. This involves iterating through each value in the local adjustment parameters and comparing it with a pre-defined reasonable value range. For example, the weight parameters of a neural network are typically neither too large nor too small; if extremely large or small values ​​appear outside the normal range, it may indicate anomalies in the training process or data contamination. This pre-defined reasonable range can be set based on model type, training experience, and domain knowledge.

[0056] The analysis of gradient statistics can be understood as evaluating the update trend and stability of locally adjusted parameters during training. Specifically, this involves calculating the L2 norm, mean, and variance of the locally adjusted parameters and comparing them to their historical distribution or preset thresholds under normal training conditions. For example, an abnormally large L2 norm may indicate gradient explosion, usually caused by an unstable training process or malicious data injection. Conversely, an abnormally small L2 norm may indicate vanishing gradients, preventing the model from learning effectively.

[0057] In practical applications, to more comprehensively identify potentially malicious data, the currently submitted local adjustment parameters can be compared with historical local adjustment parameters stored in the global base model library. For example, the similarity between the current local adjustment parameters and historical parameters (such as cosine similarity) can be calculated, or it can be detected whether it is similar to known malicious patterns (such as backdoor attack patterns). If the current parameter has extremely low similarity to historical normal parameters, or highly matches a malicious pattern, it can be marked as a suspicious parameter. This comparison process helps to discover malicious behaviors that attempt to compromise the global model or steal information by submitting abnormal parameters.

[0058] In some of the embodiments described above in this application, AMR fleets need to select relevant local adjustment parameters from the global basic model library based on their current operating environment and task requirements to achieve continuous personalized optimization of the local controller model. However, in actual implementation, how to efficiently and accurately identify and select local adjustment parameters that highly match the current AMR fleet operating environment and task requirements is a problem that needs to be solved. If the selection mechanism is not refined enough, it may lead to downloading unnecessary parameters, increasing communication overhead, or failing to fully utilize the expertise in the global basic model library, thereby affecting the effectiveness and efficiency of personalized optimization.

[0059] In this regard, this application further proposes that the steps for selecting relevant local adjustment parameters from the global basic model library based on the current operating environment and task requirements include: Maintain tags that represent the current operating environment and mission requirements of each AMR fleet; The matching degree of the above labels with the association descriptions of each local adjustment parameter in the global basic model library is calculated; Only download and apply one or more local adjustment parameters whose matching degree exceeds the preset threshold.

[0060] Specifically, maintaining tags representing the current operating environment and task requirements of each AMR fleet involves dynamically or statically assigning each AMR fleet a set of identifiers describing its current working scenario, task type, environmental characteristics, etc. For example, an AMR fleet might be tagged as an indoor warehouse, high-density shelving, picking task, low-light operation, etc. These tags can be automatically generated by the system or manually configured by operators, with the aim of accurately depicting the personalized needs of the AMR fleet.

[0061] The matching degree calculation, which involves comparing the aforementioned tags with the associated descriptions of various local adjustment parameters in the global basic model library, can be understood as quantifying the correlation between the current needs tags of the AMR fleet and the applicable scenarios or the knowledge represented by each local adjustment parameter stored in the global basic model library, using a certain algorithm or rule. For example, each local adjustment parameter, when uploaded to the global basic model library, can be accompanied by its applicable scenario tag or functional description. The matching degree calculation can employ cosine similarity, Jaccard similarity, or other semantic or feature vector-based similarity algorithms, with the aim of identifying the local adjustment parameter that best matches the current needs of the AMR fleet.

[0062] In practical applications, downloading and applying only one or more local adjustment parameters whose matching degree exceeds a preset threshold means that after completing the matching degree calculation, the system will filter according to a pre-set matching degree threshold. Only those local adjustment parameters with a matching degree higher than this threshold will be downloaded to the AMR fleet and used for fusion training with the local controller model. The purpose of this is to ensure that the downloaded parameters are highly relevant, avoiding the introduction of unnecessary or mismatched knowledge, thereby improving the efficiency and effectiveness of model optimization. The preset threshold can be adjusted according to the needs and experience of the actual application scenario to balance the coverage and relevance of the parameters.

[0063] However, in practical applications, the local adjustment parameters generated by different AMR fleets during their respective operations may contain general knowledge or specialized knowledge specific to particular scenarios. If these local adjustment parameters are not differentiated and a unified update strategy is not adopted, the general capabilities of the global base model library may be diluted, or it may be difficult to efficiently support personalized optimization by AMR fleets, thereby affecting the overall adaptability and performance of the model.

[0064] In this regard, this application further proposes that the step of updating the locally adjusted parameters that have passed security verification to the global base model library also includes: For each local adjustment parameter that passes security verification, perform a contextual attribute assessment to determine whether it is a general local adjustment parameter or a professional local adjustment parameter; If it is determined to be a general local adjustment parameter, it will be integrated into the core general capability module of the global basic model library; If it is determined to be a professional local adjustment parameter, it is encapsulated as an independent context-specific adaptation module and stored in the global basic model library.

[0065] Specifically, context attribute evaluation refers to analyzing locally adjusted parameters that have passed safety verification to determine whether the knowledge they embody is a general capability applicable to various AMR operating scenarios or a specialized capability specific to a particular environment or task. For example, this evaluation can compare the performance improvement of the model on general tasks and specific extreme situation tasks based on parameters. General-purpose locally adjusted parameters refer to those parameters that can improve the performance of the AMR controller model in a wide range of scenarios, such as optimization of the basic navigation algorithm or improvement of general obstacle avoidance logic. When identified as general-purpose locally adjusted parameters, they will be integrated into the core general-purpose capability module of the global basic model library. The core general-purpose capability module is the part of the global basic model library that carries universal and fundamental knowledge. By integrating these general parameters, the overall generalization ability and basic performance of the global model can be continuously enhanced. Furthermore, specialized locally adjusted parameters refer to those parameters optimized for specific operating environments (such as narrow passages or specific lighting conditions) or specific task requirements (such as heavy-load handling or high-precision positioning). When identified as specialized locally adjusted parameters, they will be encapsulated into independent context-specific adaptation modules. Context-specific adaptation modules are additional components specifically designed to handle particular complex or special scenarios. They are stored independently in the global basic model library, allowing AMR fleets to selectively download and apply these modules according to their own needs, thereby achieving more refined and personalized optimization.

[0066] In some preferred embodiments, suppose an AMR convoy operates in a narrow passageway of a specific warehouse, generating a set of local tuning parameters in the process. These parameters may include optimizations for fine-tuning and obstacle avoidance of the AMRs in confined spaces. Before updating these local tuning parameters to the global base model library, they are evaluated for contextual attributes. If the evaluation shows that these parameters primarily improve the model's performance in the specific context of narrow passage navigation, and have little improvement on general navigation tasks, then these parameters are identified as specialized local tuning parameters. Subsequently, these specialized local tuning parameters are encapsulated into a separate context-specific adaptation module, for example named the Narrow Passage Fine-Track Navigation Module, and stored in the global base model library. When another AMR convoy operates in a similar narrow passageway environment, it can select and download the Narrow Passage Fine-Track Navigation Module from the global base model library according to its own needs, and fuse it with its own local controller model for training, thereby quickly obtaining optimized performance in narrow passageways without having to learn from scratch. Conversely, if a set of local adjustment parameters is evaluated as generally improving the basic path planning capabilities of AMRs in various terrains, it will be judged as a general local adjustment parameter and integrated into the core general capability module of the global basic model library to enhance the basic navigation capabilities of all AMR fleets.

[0067] Specifically, the aforementioned contextual attribute assessment includes: The local adjustment parameters are applied to a preset simulated global model copy, and the performance improvement is evaluated on a general task test set in the preset simulated global model copy, and the performance improvement is evaluated on a preset specific extreme situation task set. If the performance improvement on a pre-set specific extreme scenario task set is greater than the performance improvement on a pre-set general task test set, it is judged as a professional local adjustment parameter; otherwise, it is judged as a general local adjustment parameter.

[0068] Contextual attribute evaluation refers to the process of classifying locally adjusted parameters that have passed safety verification, aiming to determine whether their main contribution is improving the general capabilities of the AMR controller model or its specialized capabilities for specific scenarios. A pre-set simulated global model copy can be understood as an isolated test version of the current state of the global base model library. Its purpose is to provide a stable and controllable environment for evaluating the impact of locally adjusted parameters on model performance without directly affecting the running global model. The general task test set contains a collection of tasks commonly encountered by AMR fleets in routine operations, such as cargo handling or path navigation in a standard warehouse environment. The specific extreme situation task set contains special or challenging tasks that AMR fleets may encounter, such as obstacle avoidance and path planning in complex environments with low light, slippery ground, or dense obstacles. The performance improvement refers to the extent to which the AMR controller model's performance on a specific task improves after applying the locally adjusted parameters compared to when those parameters are not applied. This can be quantified using metrics such as task completion time, navigation accuracy, and obstacle avoidance success rate.

[0069] Specifically, the aforementioned different AMR fleet controller models were pre-trained for different AMR operating scenarios.

[0070] In some embodiments of this application, multiple different AMR fleet controller models are pre-trained for different AMR operating scenarios. These multiple different AMR fleet controller models refer to a series of pre-built AMR controller models stored in a global basic model library, each with different capability focuses or applicable to different operating environments. These models are not trained from scratch but have already undergone preliminary training and optimization. Different AMR operating scenarios can be understood as various operating environments and task types that AMRs may encounter in practical applications, such as, but not limited to, indoor warehouse navigation, obstacle avoidance in complex outdoor terrain, heavy-duty handling, and precision assembly. Pre-training means that these AMR fleet controller models have been fully trained using a large dataset representing specific scenarios before being included in the global basic model library, enabling them to possess basic and reliable control capabilities in their respective target scenarios. In this way, each controller model in the global basic model library has a certain degree of specialization and applicability, better meeting the initial needs of different AMR fleets.

[0071] Specifically, the different MR operating scenarios mentioned above can include a variety of typical operating environments to ensure that the pre-trained MR controller model can cover a wide range of application needs.

[0072] Based on the aforementioned federated learning-based cross-fleet incremental learning method for MR controller models, multiple different MR fleet controller models are pre-trained for different MR operating scenarios. Furthermore, these different MR operating scenarios include indoor flat ground navigation, outdoor complex terrain obstacle avoidance, and heavy-load transport path planning.

[0073] The indoor flat-ground navigation scenario refers to the navigation and operation of MR vehicle fleets in indoor environments with flat ground and structured layouts, such as warehouses and factory workshops. In this scenario, the MR vehicle fleet controller model mainly focuses on the efficiency of path planning, the accuracy of positioning, and obstacle avoidance capabilities in known obstacle environments. The outdoor complex terrain obstacle avoidance scenario refers to the operation of MR vehicle fleets in outdoor environments with complex environmental factors such as irregular ground, dynamic obstacles, and changing lighting, such as construction sites and unstructured outdoor areas. In this scenario, the MR vehicle fleet controller model needs to have stronger environmental perception capabilities, robust obstacle avoidance strategies, and adaptability to unknown or dynamic obstacles. The heavy-load transport path planning scenario refers to the path planning and motion control of MR vehicle fleets when transporting heavy or large items. This scenario involves not only navigation and obstacle avoidance but also the impact of load on the MR vehicle's motion performance, stability, and energy consumption. Therefore, the controller model needs to optimize the path to ensure the safe and efficient completion of the transport task.

[0074] Specifically, the aforementioned AMR fleet controller model can be implemented using a navigation policy network based on deep reinforcement learning.

[0075] The AMR fleet controller model includes a navigation policy network based on deep reinforcement learning, which outputs motion control commands for the AMR fleet based on sensor data.

[0076] Specifically, the AMR fleet controller model is the core decision-making unit of the AMR (Autonomous Mobile Robot), responsible for receiving environmental information and generating corresponding motion commands to guide the AMR in navigation, obstacle avoidance, and task execution. The navigation policy network based on deep reinforcement learning is an intelligent decision-making model trained using deep learning techniques and reinforcement learning paradigms. This network can learn the optimal navigation strategy through interaction with the environment, thereby achieving efficient and robust path planning and motion control in complex dynamic environments. In practical applications, the sensor data of the AMR fleet can include, but is not limited to, LiDAR data, visual image data, ultrasonic data, odometry data, and IMU (Inertial Measurement Unit) data. This data is used to perceive environmental information around the AMR fleet, such as obstacle positions, its own attitude, and speed. Motion control commands refer to the instructions given to the AMR fleet to execute specific actions, such as speed commands (linear velocity, angular velocity), steering commands, and acceleration / deceleration commands. These commands are sent to the underlying actuators of the AMR fleet to drive the AMR fleet to move according to preset or learned strategies. The above are merely embodiments of this application and are not intended to limit the scope of protection of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of protection of this application.

Claims

1. A cross-fleet incremental learning method for AMR controller models based on federated learning, characterized in that, include: Establish and maintain a global basic model library that includes multiple different AMR fleet controller models; When a new AMR fleet joins, a matching controller model is selected and distributed from the global basic model library as a local controller model, based on the operating environment characteristics and task requirements of the new AMR fleet. Each AMR fleet collects its own operational data during local operation, trains the local controller model based on the operational data, extracts the local adjustment parameters generated during training, performs safety verification on the local adjustment parameters, and updates the local adjustment parameters that have passed the safety verification to the global basic model library. Each AMR fleet selects relevant local adjustment parameters from the global basic model library based on its current operating environment and mission requirements, and uses the relevant local adjustment parameters to perform fusion training with the local controller model to achieve continuous personalized optimization of the local controller model.

2. The method for incremental learning across fleets of AMR controller models based on federated learning according to claim 1, characterized in that, The step of selecting and distributing a matching controller model as a local controller model from the global basic model library for the new AMR fleet based on the operating environment characteristics and task requirements of the new AMR fleet includes: Obtain the warehouse type, channel features, lighting conditions, and main task type of the AMR fleet, and calculate the matching score between the warehouse type, channel features, lighting conditions, and main task type and the scene labels adapted to each controller model in the global basic model library. The controller model with the highest matching score is selected and distributed as the local controller model for the corresponding AMR fleet.

3. The method for incremental learning across fleets of AMR controller models based on federated learning according to claim 1, characterized in that, The steps of training the local controller model based on the respective operating data and extracting the local adjustment parameters generated during training include: The local controller model is trained using efficient parameter fine-tuning techniques based on the respective operating data. Only the adapter parameters or low-rank matrix parameters that change during training and represent new knowledge are retained and extracted as the local adjustment parameters.

4. The method for incremental learning across fleets of AMR controller models based on federated learning according to claim 1, characterized in that, The security verification includes: Check whether the value range of the local adjustment parameter is within the preset reasonable range; And / or, analyze whether there are any anomalies in the gradient statistical characteristics of the local adjustment parameters; And / or, compare the local adjustment parameters with the local adjustment parameters of the historical global model update to identify potentially malicious data.

5. The method for incremental learning across fleets of AMR controller models based on federated learning according to claim 1, characterized in that, The step of selecting the relevant local adjustment parameters from the global basic model library based on its current operating environment and task requirements includes: Maintain tags representing the current operating environment and mission requirements of each AMR fleet; The matching degree of the label is calculated by associating it with the description of each local adjustment parameter in the global basic model library; Only download and apply one or more of the local adjustment parameters whose matching degree exceeds a preset threshold.

6. The method for incremental learning across fleets of AMR controller models based on federated learning according to claim 1, characterized in that, The step of updating the locally adjusted parameters that have passed security verification to the global basic model library further includes: For each local adjustment parameter that passes security verification, a contextual attribute assessment is performed to determine whether it is a general local adjustment parameter or a specialized local adjustment parameter. If it is determined to be the general local adjustment parameter, then it is integrated into the core general capability module of the global basic model library; If it is determined to be a professional local adjustment parameter, it is encapsulated as an independent context-specific adaptation module and stored in the global basic model library.

7. The method for incremental learning across fleets of AMR controller models based on federated learning according to claim 6, characterized in that, The contextual attribute assessment includes: The local adjustment parameters are applied to a preset simulated global model copy, and the performance improvement is evaluated on a general task test set in the preset simulated global model copy, and the performance improvement is evaluated on a preset specific extreme situation task set. If the performance improvement on the preset specific extreme situation task set is greater than the performance improvement on the preset general task test set, it is determined to be the professional local adjustment parameter; otherwise, it is determined to be the general local adjustment parameter.

8. The method for incremental learning across fleets of AMR controller models based on federated learning according to claim 1, characterized in that, The multiple different AMR fleet controller models were pre-trained for different AMR operating scenarios.

9. The method for incremental learning across fleets of AMR controller models based on federated learning according to claim 8, characterized in that, The different AMR operating scenarios include indoor flat ground navigation scenarios, outdoor complex terrain obstacle avoidance scenarios, and heavy-load transport path planning scenarios.

10. The method for incremental learning across fleets of AMR controller models based on federated learning according to claim 8, characterized in that, The AMR fleet controller model includes a navigation policy network based on deep reinforcement learning, which is used to output motion control commands for the AMR fleet based on sensor data of the AMR fleet.

Images