Cooperative AI model management system for intelligent robot with body and use method thereof
The collaborative AI model management system solves the latency and hardware compatibility issues in the deployment of embodied intelligent robots, enabling efficient and secure model acquisition and rapid iteration, and improving resource utilization and management efficiency on the robot side.
Patent Information
- Application Number
- CN202510909183.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-07-02
- Publication Date
- 2025-10-28
- Estimated Expiration
- Not applicable · inactive patent
AI Technical Summary
Embossed intelligent robots suffer from problems such as high latency, strong network dependence, poor hardware compatibility, limited resources, and insufficient security during model deployment, and lack unified management and efficient collaboration.
A collaborative AI model management system is adopted, including a cloud-based model repository and a robot-side model management agent system, to achieve centralized storage, version control, metadata management and local caching of models. It supports multiple formats, provides RESTful APIs and local interfaces, and optimizes the model acquisition and deployment process by combining role-based access control and caching strategies.
It enables low-latency and high-reliability model delivery to robots, improves hardware adaptability and resource efficiency, simplifies the development process, and supports rapid model iteration and safety management.
Smart Images
Figure CN120848975A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of artificial intelligence technology, specifically relating to a collaborative AI model management system for embodied intelligent robots and its usage method. Background Technology
[0002] Embossed intelligent robots (such as service robots and autonomous mobile robots) rely heavily on AI / ML models when performing tasks such as perception, decision-making, and control. These robots are typically deployed in the physical world, placing high demands on the real-time performance, efficiency, reliability, security, and compatibility with specific hardware of the models. Meanwhile, the robots themselves have limited computing resources (CPU, memory, power consumption), and network connections may be unstable or have limited bandwidth.
[0003] Traditional model management is mostly based on cloud repositories, with models downloaded directly from the cloud to end devices. For robotic scenarios, this leads to high latency, strong network dependence, and inability to adapt to changing edge hardware and network conditions.
[0004] Traditional methods use simple local caching, but lack unified management, version control, efficient collaboration with the cloud, and fine-grained support for robot-specific needs (such as hardware compatibility metadata).
[0005] Model training and model deployment are often disconnected. After model training is completed, a series of tedious steps such as format conversion, verification, uploading, and deployment need to be performed manually, resulting in low iteration efficiency.
[0006] There is a lack of standardized metadata management for robot deployment scenarios, which includes key information such as performance metrics, hardware compatibility, and verification status. Summary of the Invention
[0007] This invention provides a collaborative AI model management system and its usage method for embodied intelligent robots, which solves the problem that existing technologies cannot simultaneously meet the comprehensive needs of embodied intelligent robots for low latency, high reliability (weak network / offline), hardware adaptability, resource efficiency, security, and integrated training and promotion for model deployment.
[0008] This invention is achieved through the following technical solution:
[0009] A collaborative AI model management system for embodied intelligent robots, the system comprising a cloud-based model repository and a robot-side model management agent system;
[0010] The cloud-based model repository serves as a centralized storage, version control, metadata management, receiving and distribution center for models, and a connection point for the training and push processes.
[0011] The robot-side model management agent system is deployed locally on the robot, serving as a local cache for the model, a request gateway, and a runtime optimization tool.
[0012] Furthermore, the cloud-based model repository uses object storage to store model files and supports TFLite, ONNX, and TensorRTengine formats;
[0013] The cloud-based model repository associates a structured metadata record with each model version.
[0014] Furthermore, the cloud-based model repository uses RESTful API endpoints as its send / receive interface;
[0015] Authorized users can upload model files and corresponding metadata via UI or script.
[0016] Furthermore, role-based access control ensures operational permissions; a WebUI is provided for administrators and developers to manage models, view metadata, and audit logs.
[0017] Furthermore, the robot-side model management agent system uses a local interface, enabling the robot's capability program to receive model requests.
[0018] A method for using a collaborative AI model management system for embodied intelligent robots, the method comprising,
[0019] The steps for model preparation and registration;
[0020] The steps of the robot capability request model;
[0021] The steps for proxying request processing;
[0022] The steps for handling cache hits;
[0023] Steps for handling cache misses;
[0024] The document outlines steps for cache maintenance, enabling the use of a collaborative AI model management system.
[0025] Furthermore, the model preparation and registration specifically involve,
[0026] Manual registration means that users upload model files and detailed metadata through the cloud management interface;
[0027] Automatic training-promotion integrated registration means that after the model training platform completes the training task, it automatically calls the cloud repository API to register the newly generated model version and its basic metadata; the cloud repository stores the model and metadata.
[0028] Furthermore, the robot capability request model specifically refers to the process where, when the robot's capability program requires a model, it calls the interface provided by the local model management agent and passes in the model name and version number.
[0029] Specifically, the proxy request processing includes proxy request interception and proxy querying of local cache indexes.
[0030] Furthermore, the cache hit processing specifically involves the agent returning the local path of the model file to the capability program if a cache hit occurs and the model is valid.
[0031] Capability loading model;
[0032] The process is complete.
[0033] Furthermore, the cache miss handling specifically involves:
[0034] The proxy constructs a request to the cloud-based model repository;
[0035] The agent sends requests to the cloud repository over the network;
[0036] The cloud repository processes requests, locates the appropriate model file, and returns a download link or directly transfers the file;
[0037] Securely download model files via proxy;
[0038] The proxy verifies the integrity of downloaded files;
[0039] The agent stores the validated model files in the local cache directory and updates the cache index;
[0040] The agent returns the local path of the model file to the capability program;
[0041] Capability loading model;
[0042] The process is complete.
[0043] The beneficial effects of this invention are:
[0044] This invention can provide robots with the required AI models efficiently and with low latency.
[0045] This invention can improve the robustness of model acquisition in unstable network or offline environments.
[0046] This invention can optimize model distribution by providing adapted model versions based on the characteristics of robot hardware.
[0047] This invention can simplify the development of robot-side capabilities and shield the complexity of model acquisition.
[0048] This invention streamlines the basic process from model training to deployment, supporting rapid model iteration.
[0049] This invention provides centralized management and tracking of model version, source, and verification status. Attached Figure Description
[0050] Figure 1 This is the overall system architecture diagram of the present invention.
[0051] Figure 2 This is a sequence diagram of the model request processing in this invention (distinguishing between cache hits and misses). Detailed Implementation
[0052] In the following description, specific details such as particular system architectures and techniques are set forth for illustrative purposes and not for limitation, in order to provide a thorough understanding of the embodiments of this application. However, those skilled in the art will understand that this application may also be implemented in other embodiments without these specific details. In other instances, detailed descriptions of well-known systems, apparatuses, circuits, and methods are omitted so as not to obscure the description of this application with unnecessary detail.
[0053] It will be understood that when used in this specification and the appended claims, the term "comprising" indicates the presence of described features, integers, steps, operations, elements and / or components, but does not preclude the presence or addition of one or more other features, integers, steps, operations, elements, components and / or groups thereof.
[0054] It should also be understood that the terminology used in this application specification is for the purpose of describing particular embodiments only and is not intended to limit the application. As used in this application specification and the appended claims, the singular forms “a,” “an,” and “the” are intended to include the plural forms unless the context clearly indicates otherwise.
[0055] The following is in conjunction with the appendix to this application specification. Figure 1-2 The technical solutions in the embodiments of this application are clearly and completely described. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of the embodiments. Based on the embodiments in this application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of this application.
[0056] Many specific details are set forth in the following description in order to provide a full understanding of this application. However, this application may also be implemented in other ways different from those described herein. Those skilled in the art can make similar extensions without departing from the spirit of this application. Therefore, this application is not limited to the specific embodiments disclosed below.
[0057] Implementation Method 1
[0058] This embodiment provides a collaborative AI model management system for embodied intelligent robots, such as... Figure 1 As shown, the system is a collaborative system of a cloud-based model repository and a robot-side model management agent; the system includes a cloud-based model repository and a robot-side model management agent system.
[0059] The cloud-based model repository serves as a centralized storage, version control, metadata management, receiving and distribution center for models, and a connection point for the training and push processes.
[0060] The robot-side model management agent system is deployed locally on the robot, serving as a local cache for the model, a request gateway, and a runtime optimizer. It typically runs as a background service or daemon within the robot's operating system (such as Linux + ROS / ROS2).
[0061] Furthermore, the cloud-based model repository uses object storage to store model files, supporting multiple formats such as TFLite, ONNX, and TensorRTengine; it enables fine-grained version control (such as major.minor.patch or timestamp) and clearly distinguishes the model source (manual_upload, training_generated);
[0062] To achieve fine-grained version control: A Git service is introduced, requiring users to use version numbers to mark the uploaded models, thus enabling fine-grained version control.
[0063] Distinguishing Model Sources: During the model upload process, the metadata of the unit to which the model belongs must be uploaded, and account login is required for the upload process. Only the model content uploaded by the corresponding account can be managed and modified. Other users only have the right to download or apply for editing.
[0064] The cloud-based model repository associates a structured metadata record (such as one stored in a database) with each model version.
[0065] Example of key metadata fields:
[0066] model_name:"object_detection_yolov5"
[0067] version: "2.1.0"
[0068] source:"training_generated"
[0069] training_task_id:"train_job_abc123" (optional)
[0070] target_hardware:["NvidiaJetsonAGX","RaspberryPi4+CoralTPU"]
[0071] performance_metrics:[{"hw":"JetsonAGX","latency_ms":50,"memory_mb":250,"power_w_avg":5},...]
[0072] input_spec:{"type":"image","format":"RGB","size":[640,640]}
[0073] output_spec:{"type":"bounding_boxes","classes":["person","car"]}
[0074] dependencies:["cuda_11.4","tensorrt_8.0"]
[0075] validation_status:"Passed_Simulation_Level_2"
[0076] security_review: "Approved"
[0077] model_checksum:"sha256:abcdef..."
[0078] The key metadata sequence, from top to bottom, is as follows: model name, version, model source, training task ID, target hardware limitations, model safety metrics, input format, output format, model software environment dependencies, model validity, security check results, and model checksum.
[0079] Furthermore, the cloud-based model repository uses secure RESTful API endpoints as its send / receive interface; (POST / models, POST / models / {name} / versions) allows:
[0080] Authorized users can upload model files and corresponding metadata via UI or script.
[0081] After successful training, the automated model training platform / service calls this API to automatically register the newly generated model file and its basic metadata (including at least the model name, version, source identifier, checksum, and optional training task ID and basic performance information).
[0082] (GET / models / {name} / {version}, GET / models?query=...) allows the robot agent to query and download models based on conditions such as model name, version, and target hardware; the server can intelligently select and return the most suitable optimized version (if it exists) based on the hardware information in the request.
[0083] Furthermore, role-based access control (RBAC) ensures operational permissions; a WebUI is provided for administrators and developers to manage models, view metadata, and audit logs.
[0084] Furthermore, the robot-side model management agent system uses a local interface (IPC / API) to enable the robot's capability programs (such as ROS nodes and applications) to receive model requests (such as gRPC calls, local sockets, shared memory queries, etc.) and parse out the required model name and version.
[0085] Local cache lookup: Maintain a local file system directory as a cache and may use an index file (such as an SQLite database or JSON file) to record metadata of the models in the cache (including at least name, version, file path, last used time, and checksum). Upon receiving a request, this index is queried first.
[0086] Cache hit: If a matching and valid model is found in the cache index (e.g., checksum validation passes), the local path of the model file is returned directly to the requesting capability program.
[0087] Cache miss (Miss) and cloud interaction: If a cache miss occurs or the model is invalid, the agent uses the robot's network connection to call the query / distribution API of the cloud model repository to request the download of the specified model version. The request can include local hardware information (?robot_hw=jetson_agx) so that the cloud can return the most suitable version. Network errors and retries should be handled during the download process.
[0088] Secure Download and Storage: HTTPS is used when downloading the model. After download, the file checksum is verified to ensure it matches the cloud metadata. Once verification is successful, the model file is stored in the local cache directory, and a record is added / updated in the cache index.
[0089] Cache management (Eviction): Implements cache eviction policies to manage storage space. A common policy is LRU (Least Recently Used): When the cache reaches a preset size limit, the model file and its index record that have not been used for the longest time are deleted.
[0090] (Optional) Model Prefetching: Based on a pre-configured list or task instructions received from the cloud, models that may be needed can be downloaded to the cache in advance.
[0091] (Optional) Status synchronization and update check: Can be synchronized with the cloud repository periodically or on demand, check whether the version of the cached model has been updated, or receive update notifications pushed from the cloud.
[0092] Efficiency Improvement: Local caching significantly reduces model loading latency and improves robot response speed.
[0093] Robustness enhancement: The robot can still use the cached model when the network is poor or offline, improving the reliability of task execution.
[0094] Resource optimization: The cloud can provide an optimized model based on the robot's hardware metadata; local caching avoids repeated downloads and saves bandwidth.
[0095] Simplified development: Capability developers do not need to deal with the complex logic of model downloading, caching, and version management.
[0096] Integrated training and deployment: Automated model registration connects the training to deployment path, accelerating iteration.
[0097] Controllable management: Centralized cloud management of model versions and metadata (including verification / security status) enhances traceability and deployment security.
[0098] Highly adaptable: The system design takes into account the specific constraints and requirements of embodied intelligent robots.
[0099] Implementation Method 2
[0100] This embodiment provides a method for using a collaborative AI model management system for embodied intelligent robots. The method uses a system as described in Embodiment 1, such as... Figure 2 As shown, the method of use includes,
[0101] The steps for model preparation and registration;
[0102] The steps of the robot capability request model;
[0103] The steps for proxying request processing;
[0104] The steps for handling cache hits;
[0105] Steps for handling cache misses;
[0106] The process also includes cache maintenance steps. The agent executes cache eviction policies as needed (such as when space is insufficient), enabling the use of a collaborative AI model management system.
[0107] Furthermore, the model preparation and registration specifically involve,
[0108] Manual registration means that users upload model files and detailed metadata through the cloud management interface;
[0109] Automatic training-promotion integrated registration means that after the model training platform completes the training task, it automatically calls the cloud repository API to register the newly generated model version and its basic metadata; the cloud repository stores the model and metadata.
[0110] Furthermore, the robot capability request model specifically refers to the process where, when a certain capability program on the robot requires a model, it calls the interface provided by the local model management agent and passes in the model name and version number.
[0111] Specifically, for example, with a face recognition model, we can learn from the model capability list in the model repository that model A can perform face recognition. Then, we request the download of the specific model parameters of model A from the model repository through the proxy access interface already available on the hardware side. After downloading model A to the robot hardware, we can use it locally.
[0112] Specifically, the proxy request processing includes proxy request interception and proxy querying of local cache indexes.
[0113] Furthermore, the cache hit processing specifically involves the agent returning the local path of the model file to the capability program if a cache hit occurs and the model is valid.
[0114] Capability loading model;
[0115] The process is complete.
[0116] Furthermore, the cache miss handling specifically involves:
[0117] The agent constructs a request to the cloud model repository (including model identifiers and optional hardware information);
[0118] The agent sends requests to the cloud repository over the network;
[0119] The cloud repository processes requests, locates the appropriate model file, and returns a download link or directly transfers the file;
[0120] Securely download model files via proxy;
[0121] The proxy verifies the integrity of downloaded files (such as checksums);
[0122] The agent stores the validated model files in the local cache directory and updates the cache index;
[0123] The agent returns the local path of the model file to the capability program;
[0124] Capability loading model;
[0125] The process is complete.
[0126] The cloud repository can be implemented based on AWS S3 / GCP Cloud Storage + PostgreSQL / MongoDB. The API can be built using Python Flask / Django or GoGin.
[0127] The robot-side agent can be implemented as a Linux systemd service or a ROS node, written in Python / C++. Caching can directly utilize the file system + SQLite. IPC can be either gRPC or D-Bus.
[0128] Cache eviction using LRU can be achieved by maintaining an ordered list of access times or by using timestamps.
[0129] The model checksum can be SHA256 encoded. Downloads are encrypted using TLS.
Claims
1. A collaborative AI model management system for embodied intelligent robots, characterized in that, The system includes a cloud-based model repository and a robot-side model management agent system; The cloud-based model repository serves as a centralized storage, version control, metadata management, receiving and distribution center for models, and a connection point for the training and push processes. The robot-side model management agent system is deployed locally on the robot, serving as a local cache for the model, a request gateway, and a runtime optimization tool.
2. The system according to claim 1, characterized in that, The cloud-based model repository uses object storage to store model files and supports TFLite, ONNX and TensorRTengine formats. The cloud-based model repository associates a structured metadata record with each model version.
3. The system according to claim 1, characterized in that, The cloud-based model repository uses RESTful API endpoints as its send and receive interfaces. Authorized users can upload model files and corresponding metadata via UI or script.
4. The system according to claim 3, characterized in that, Role-based access control ensures operational permissions; a WebUI is provided for administrators and developers to manage models, view metadata, and audit logs.
5. The system according to claim 1, characterized in that, The robot-side model management agent system uses a local interface, enabling the robot's capability program to receive model requests.
6. A method for using a collaborative AI model management system for embodied intelligent robots according to any one of claims 1-5, characterized in that, The method of use includes, The steps for model preparation and registration; The steps of the robot capability request model; The steps for proxying request processing; The steps for handling cache hits; Steps for handling cache misses; The document outlines steps for cache maintenance, enabling the use of a collaborative AI model management system.
7. The method of use according to claim 6, characterized in that, The model preparation and registration specifically involve, Manual registration means that users upload model files and detailed metadata through the cloud management interface; Automatic training and push integrated registration means that after the model training platform completes the training task, it automatically calls the cloud repository API to register the newly generated model version and its basic metadata. Cloud-based warehouse storage model and metadata.
8. The method of use according to claim 6, characterized in that, Specifically, the robot capability request model is as follows: when the robot's capability program needs a model, it calls the interface provided by the local model management agent and passes in the model name and version number. Specifically, the proxy request processing includes proxy request interception and proxy querying of local cache indexes.
9. The method of use according to claim 8, characterized in that, The cache hit processing specifically involves the agent returning the local path of the model file to the capability program if the cache hit is successful and the model is valid. Capability loading model; The process is complete.
10. The method of use according to claim 6, characterized in that, The cache miss handling specifically involves... The proxy constructs a request to the cloud-based model repository; The agent sends requests to the cloud repository over the network; The cloud repository processes requests, locates the appropriate model file, and returns a download link or directly transfers the file; Securely download model files via proxy; The proxy verifies the integrity of downloaded files; The agent stores the validated model files in the local cache directory and updates the cache index; The agent returns the local path of the model file to the capability program; Capability loading model; The process is complete.