Robot model scheduling method and system based on AI Agent

Abstract

The invention discloses a robot model scheduling method and system based on an AI Agent, and solves the problems of resource waste, performance reduction and work failure caused by the fact that an existing robot cannot adaptively select a proper model. The method comprises the following steps: acquiring a user task demand and hardware performance information, performing model matching through an AI Agent, acquiring user demand feedback, analyzing the demand feedback, selecting to decompose the task demand, and performing model matching according to the decomposed demand to acquire a target model combination; and loading the model to the robot. According to the method, the models can be dynamically and adaptively allocated according to user requirements and hardware performance, so that the resource utilization efficiency is improved, and the problems that the robot performance is reduced or the work is effective due to the fact that an existing method is difficult to flexibly adapt to diversified scenes and the adaptability is insufficient due to the adoption of an end-to-end model are solved. The system has self-learning and iteration capabilities, and realizes real-time data acquisition and feedback, distributed system training and adaptive task adjustment.

Description

Technical Field

[0001] The present invention relates to the field of computer technology, and more particularly to a robot model scheduling method and system based on an AI Agent. Background Art

[0002] Currently, the integration of multiple large models (such as language understanding, visual perception, motion control, and policy planning) in robots lacks a unified AI Agent "intelligent agent" - level management architecture, resulting in complex data flow, chaotic decision-making timing, and inefficient resource allocation. Simply put, currently, neither in the industrial nor academic fields does there exist a perfect model that can meet all the requirements in practical applications. Therefore, the common practice is to use multiple expert models and call the appropriate one when needed. Thus, in terms of model scheduling, a higher-level management logic is required to select the appropriate expert model. This management logic needs to be controlled by a "person" who is familiar with all models and all scenarios, and this "person" is the "AI Agent".

[0003] Currently, although end-to-end large models can simplify the interface and integration process, when new capabilities need to be extended, tasks changed, or computing resources adjusted, their overall update and maintenance complexity is relatively high, lacking adaptability and flexibility. Simply put, end-to-end models can perform well under specific tasks and interface logics. However, similarly, there needs to be an expert who knows what kind of model can perform what tasks and what interfaces are required. Since users do not understand the models, they only know their own needs, available hardware resources, and the model of the robot / robotic arm they are using. Currently, these adaptation tasks need to be completed by engineers, resulting in insufficient adaptability of end-to-end models.

[0004] In different application scenarios, the resource conditions and requirements vary greatly (such as real-time response, low power consumption, local deployment, and cloud collaboration). Existing designs usually have difficulty flexibly adapting to diverse scenarios, leading to performance degradation or work failure. For example, to complete a certain user's requirement, in fact, only a very small model that consumes very little computing resources can complete it well with a fast response speed. However, if a particularly large model is wrongly selected, since a large model means more computing resources, slower response speed, and higher operating costs. Therefore, in such requirements, a set of intelligent methods need to be designed to help users select a model that best matches the requirements and performance instead of choosing the largest model in the database.

[0005] For the invention application with the application number 202411489504.4 and the title of "A Large Model Scheduling System, Method, Server, Medium and Product", it discloses a computing power scheduling module, which is used for each large model to determine a target computing power resource node matching the large model according to the model parameter information of the large model, and mount the large model into the container of the target computing power resource node. This patent only discloses finding a matching computing power resource node to mount the large model, rather than selecting the most suitable model according to the resource node, resulting in the matching model not being the most suitable one. Summary of the Invention

[0006] The present invention mainly solves the problem that current robots cannot adaptively select a suitable model, resulting in resource waste, performance degradation, and work failure, and provides a robot model scheduling method and system based on an AI Agent.

[0007] The above technical problem of the present invention is mainly solved by the following technical solutions: A robot model scheduling method based on an AI Agent includes the following steps: Obtain user task requirements and hardware performance information, and perform model matching through an AI Agent, including, Select a first model similar to the task from the model database according to the task requirements, Select a second model matching the hardware from the first models according to the hardware information; Obtain user demand feedback, analyze the demand feedback to decompose the task requirements, and perform model matching according to the decomposed requirements to obtain a target model combination; Load the second model or the target model combination into the robot.

[0008] The AI Agent is an intelligent agent. The present invention proposes to flexibly schedule and manage models by the AI Agent, with more efficient model collaboration, and more concise data flow and resource allocation. The present invention can dynamically and adaptively allocate models according to user requirements and hardware performance, improving resource utilization efficiency, solving the problem that existing methods are difficult to flexibly adapt to diverse scenarios, resulting in performance degradation or work failure of robots, and the problem of insufficient adaptability of end-to-end models. The present invention selects the optimal model according to user input through the AI Agent, and can adjust the model in real time according to demand feedback, and obtains the optimal model or model combination through task decomposition or cloud computing. The present invention can respond to diverse demands in dynamic scenarios in real time through a flexible model scheduling strategy, such as task numbers, resource limitations, or emergencies, ensuring that the robot always maintains high performance in complex environments. In the present invention, the robot is an artificial intelligence robot, a general robot, and a general intelligent robot.

[0009] The system of the present invention needs to interact with the real world in real time. In the overall architecture design, very complex nested logic will not be used to ensure low latency and fast response in the interaction between the user and the AI Agent and between the AI Agent and the system functions. Moreover, in the designed method and system, there is no dependency relationship between functions. For example, by asking the user's requirements, tasks can be classified, or tasks can be not decomposed. It can learn autonomously, or data collection can be selected not to be performed. Such a design allows the user to have more independent choices on the one hand, and on the other hand, uses a non-coupled design to have greater scalability, and can be quickly adapted for system upgrade and function expansion in the later stage.

[0010] As a preferred solution, the step of selecting a first model similar to the task from the model database according to the task requirements includes: Performing similarity matching between the task requirement information and the model task description information, and obtaining a model with high similarity as the first model.

[0011] The system of the present invention includes a model database, a task database, and a scenario database. The model database stores models with different functions, and the task database stores different tasks. Each model is set with a task id tag, and the task id is associated with one or more specific tasks in the task database. Moreover, each task is set with a task description tag, and the task description is a specific text description of the task. The first model is a model with high similarity between the model task description information and the user input task requirement description information. The user inputs a task requirement information, and each model has a task description information. Since there are many models in the model database, it is necessary to find a model with high information similarity from the task description information of all models according to the user input task requirement information. Information similarity matching can use existing algorithms. One of them can be through feature engineering similar to tf-idf, converting a piece of text into a set of vectors, and calculating the similarity between the user input vector and the vector converted from the task description of the model in the model database, such as the cosine approximation algorithm. Each model obtains a similarity through similarity calculation, and is sorted in descending order according to the similarity. A similarity threshold can be set, and models with similarity greater than the threshold are selected, that is, models with high similarity. Through similarity comparison, the first models similar to the task are initially selected. There are multiple first models, and the first model is not yet the most matching model, and the user's hardware capabilities also need to be considered to select the most matching model. The above scenario database includes task scenarios composed of multiple tasks. The scenario corresponds to multiple tasks, and the definition of the task is the smallest granularity action description.

[0012] As a preferred solution, the step of selecting a second model matching the hardware from the first models according to the hardware information includes: Obtain the model capacity adaptation range according to the hardware information, and screen the candidate model with the largest capacity that meets the capacity adaptation range from the first model as the second model.

[0013] The hardware information includes computer performance information. There is a pre-set association table between computer performance and model capacity range in the system. Obtain the required model capacity adaptation range according to the computer performance information input by the user. Select the models that meet the model capacity adaptation range from the first model, sort these models in descending order according to similarity, traverse the models from high to low according to similarity, and select the model with the largest capacity as the second model. If there are multiple models with the same capacity, further judgment is required according to the data source.

[0014] As a preferred solution, If there are multiple candidate models with the same capacity, obtain the data source information according to the candidate models, obtain the required model data source information according to the task requirement information, and select the model with the data source information consistent with the required model data source information from the multiple candidate models as the second model.

[0015] Select the model with the largest capacity that meets the model capacity matching range from the first model as the second model. When there are multiple models with the same capacity, select the model with the data source information consistent with the required model data source information in the task requirement information as the second model. The model data source information indicates what type of robot data the model uses for training. In the case of the same similarity, preferentially select the model with the same data source information as the required model, that is, select the model trained with the same robot data. Additionally, in special cases, if there is no model with the data source information consistent with the required model, select the model with the largest similarity as the second model.

[0016] As a preferred solution, if there are multiple models with the same data source information, select the model with the highest similarity as the second model.

[0017] Exclude the models with too large model capacity that cannot run locally and the models with low similarity according to the user's hardware information. Further judge according to the data source. If there are multiple models with the data source information consistent with the required model data source information, select the model with the largest similarity from them as the second model.

[0018] As a preferred solution, the steps of obtaining user demand feedback, analyzing the demand feedback to decompose the task requirements, and obtaining the target model combination according to the decomposed requirements include: Control the robot to complete the test by the second model, and ask whether the test meets the requirements. If the requirements are not met, obtain user requirement feedback, and judge whether there is a task shortage based on the requirement feedback. If there is a task shortage, decompose the user task requirements, perform model matching on each decomposed requirement, and obtain a target model combination. If the requirements are met, output the second model.

[0019] After obtaining the second model, the second model controls the robot to complete a task test. After the test is completed, the system asks the user through the interaction interface whether the requirements are met. The system obtains user requirement feedback, including the inquiry result and information such as images, videos, and descriptions uploaded by the user in response to the inquiry result. In the case where the requirements are met, output the second model as the finally matched model for loading. In the case where the requirements are not met, analyze the requirement feedback information to judge the reason why the current model does not meet the requirements, that is, judge whether the model has a task shortage or a generalization problem. The system has a classification module to classify the model according to information such as images, videos, and descriptions input by the user. If the task requirements proposed by the user cannot be found in the existing task database, it means there is a task shortage and a classification mark is made. Otherwise, it is a generalization problem, that is, the model can complete the task but not well, and a classification mark is made. Here, classification is performed using a classification model obtained by training with existing models. In the case where it is judged that there is a task shortage, use the large language model LLM to decompose the task requirements, and use the COT (Chain-of-thought ) mechanism to let the model decompose step by step to obtain a clear and sequential set of requirements. Match each requirement to the models in the model database according to the above model matching method, and finally obtain a target model combination, and output the target model combination to be loaded into the robot.

[0020] As a preferred solution, Judge whether there is a generalization problem according to the requirement feedback. If it is a generalization problem, select a larger model from the candidate models as the second model. If there is no larger model, use cloud computing as the second model.

[0021] The generalization problem is understood as follows. For example, in 100 tasks, the model can successfully complete 50 tasks, and fails in the other 50 tasks for various reasons. The model can complete the tasks to a certain extent, but in some cases, the model performs poorly, that is, the data is not good, there is not enough data, and the model is not large enough. Therefore, when it is found that the previously selected model can complete the tasks, but the overall completion rate is not high, a larger model based on more data will be selected. In the model matching process, if the generally selected model is already the largest model that the user can locally apply, then the cloud server computing is used as the second model. Through cloud server computing, the calculation results are transmitted to the robot through the network for robot driving. Similarly, the robot is controlled to perform tests according to the obtained calculation results until the second model that meets the requirements is obtained. This solution also designs an adaptive scheduling method for cloud and local computing resources to achieve collaborative optimization between the cloud and the local, and applies cloud computing to provide complex model calculations to ensure that the optimal model is provided for the user.

[0022] As a preferred solution, after obtaining the second model or the target model combination, it further includes a model self-learning step, specifically including: Ask whether to agree to data collection. If agreed, collect and process the robot data. Obtain the pre-trained model according to the model matching result, configure the training hyperparameters, and start training. After training, conduct remote testing, and after testing, update the model to the model database.

[0023] This solution introduces an additional data collection and model training functional stack. When the user uses the AI Agent to complete the task requirements, real-time data and user evaluations / feedbacks will be synchronously collected. After the data collection volume reaches the quota of a sending packet, it will be sent back to the remote server. The server automatically selects a suitable model on the new data for fine-tuning. The suitable model is the model obtained according to the model matching process. The model is tested through the laboratory simulation environment (real environment, real task, only for testing). This is because the small-scale robot model is unstable during training, and it may encounter a training trough during the fine-tuning process and lose the model ability in the current parameter distribution. Therefore, a testing process needs to be added during training to ensure that the final model has the due ability. After the model is tested, the new model is transmitted to the cloud. The model that the user will use for this task next time is this new model. After the user uses it without problems, the current model will be solidified in the system for long-term use in the future, while the old model will be eliminated. This entire process will run automatically, thus endowing the system with continuous self-iterative ability.

[0024] As a preferred solution, the remote testing specifically includes: Conduct remote testing after meeting the first training duration, and judge whether the test is passed. If the model passes the test, it will be added to the model database. If it fails the test, after meeting the second training duration, the training hyperparameters will be reconfigured and the model training will be carried out again.

[0025] After the model meets a certain training volume, it is tested. When the test fails and the training volume reaches the set amount, it is judged that some parameter settings may be incorrect and the parameters need to be reset for training. At this time, return to the step of configuring hyperparameters and operate again.

[0026] The present invention has self-learning and dynamic optimization functions.

[0027] Real-time data collection and feedback. Since the robot operation data is very expensive, when the system drives the model to complete real-world tasks (such as handling, sorting, collecting, decorating), the current state of the robot, the received control signals, and the inputs of other sensors (such as cameras, tactile sensors, acceleration sensors, torque sensors) will be regarded as data and collected, and sent to the remote server in packets at specific capacity intervals.

[0028] Distributed collaborative training. Through the cloud and local collaborative mechanism, the system can use the newly collected data to fine-tune the model in the cloud, automatically test the performance of the new model, and ensure that the model meets the user's needs.

[0029] Adaptive task adjustment. Based on the model updated in real time, the system can dynamically adjust the task planning and model selection to improve the task completion efficiency and accuracy.

[0030] A robot model scheduling system based on AI Agent includes, An information acquisition module that acquires user task requirements and hardware performance information; A model matching module that selects a matching model through AI Agent according to the task requirements and hardware performance information; A demand feedback processing module that decomposes the task requirements according to the demand feedback, and obtains a target model combination through model matching according to the decomposed requirements; A loading module that loads the second model or the target model combination into the robot.

[0031] The system includes an AI Agent unit, and the system also includes a scenario database, a task database and a model database. The AI ​​Agent unit is connected to the scenario database, the task database and the model database respectively, and the AI ​​Agent unit is also connected to a remote server and a remote test environment respectively. The AI ​​Agent unit further includes an information acquisition module, a model matching module, a demand feedback processing module and a loading module. The information acquisition module and the demand feedback processing module are connected to the model matching module respectively, and the demand feedback processing module is connected to the loading module. The demand feedback processing module initiates a demand inquiry to the customer and obtains user demand feedback. The demand feedback processing module also includes judging the model generalization problem according to the demand feedback, selecting a larger model from the candidate models as the second model, and in the absence of a larger model, using cloud computing as the second model.

[0032] Therefore, the advantages of the present invention are: AI Agent can flexibly schedule and manage models, making model collaboration more efficient and data flow and resource allocation simpler.

[0033] It can dynamically and adaptively allocate models according to user needs and hardware performance, improve resource utilization efficiency, and solve the problem that existing methods are difficult to flexibly adapt to diverse scenarios, resulting in reduced robot performance or work effectiveness, as well as the problem of insufficient adaptability of end-to-end models.

[0034] Through flexible scheduling model strategies, the present invention can respond to diverse demands in dynamic scenarios in real time and ensure that the robot always maintains efficient performance in complex environments.

[0035] It has self-learning and dynamic optimization functions to achieve real-time data collection and feedback. Through the cloud-local collaborative mechanism, it uses newly collected data to fine-tune the model in the cloud, automatically tests the performance of the new model, and ensures that the model meets user needs. It also has adaptive task adjustment, based on the real-time updated model, which can dynamically adjust task planning and model selection to improve task completion efficiency and accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] Figure 1 It is a schematic flow chart of the method of the present invention.

[0037] Figure 2 It is a flow chart of model matching in the method of the present invention.

[0038] Figure 3 It is a flow chart of model self-learning in the method of the present invention.

[0039] Figure 4 It is a structural schematic diagram of the system of the present invention.

[0040] 1 - Information acquisition module 2 - Model matching module 3 - Requirement feedback processing module 4 - Loading module 5 - Self - learning module 6 - Remote server 7 - Scenario database 8 - Task database 9 - Model database Detailed implementation manner

[0041] The technical solution of the present invention will be further specifically described below through embodiments in conjunction with the accompanying drawings.

[0042] Embodiment 1: As Figure 1 shown, a robot model scheduling method based on AI Agent in this embodiment includes the following steps: S1. Obtain user task requirements and hardware performance information, and perform model matching through AI Agent. As Figure 2 shown, specifically, it includes S11. Select a first model similar to the task from the model database according to the task requirements.

[0043] S111. The system asks the user for task requirements and obtains the task requirement information input by the user.

[0044] S112. Perform similarity matching between the task requirement information and the model task description information.

[0045] The system includes a model database and a task database. The model database stores models with different functions, and the task database stores different tasks. Each model is set with a task id tag, and the task id is associated with one or more specific tasks in the task database. And each task is set with a task description tag, and the task description is a specific text description of the task. The first model is a model with a high similarity between the model task description information and the user - input task requirement description information. When the user inputs a task requirement information, each model has a task description information. Since there are many models in the model database, it is necessary to find a model with a high information similarity from the task description information of all models according to the task requirement information input by the user. Information similarity matching can use existing algorithms. One of them can be through feature engineering similar to tf - idf to convert a piece of text into a set of vectors, and calculate the similarity between the input vector a of the user and the vectors b1, b2, b3... converted from the task descriptions of the models in the model database, such as the cosine approximation algorithm. Each model obtains a similarity through similarity calculation.

[0046] S113. Obtain the model with a high similarity as the first model.

[0047] Sort the models in descending order according to the similarity. Preset a similarity threshold, and select the models with similarity greater than the threshold from them, that is, the models with high acquaintance, and use this as the first model. Exclude the models with low similarity through the above selection. There are multiple selected first models, and this first model is not the most matching model yet. It is also necessary to consider the user's hardware capabilities to select the most matching model. For example, set the threshold to 60%. According to the similarity matching, three qualified models are selected, namely Model 1: similarity 99%, capacity 1B; Model 2: similarity 70%, capacity 100M; Model 3: similarity 65%, capacity 10M.

[0048] S12. Select a second model that matches the hardware from the first models according to the hardware information.

[0049] S121. The system asks the user about the hardware performance and obtains the hardware performance information input by the user.

[0050] S122. Obtain the model capacity adaptation range according to the hardware information.

[0051] The hardware information includes computer performance, call interface, and data source information. The system presets an association table between computer performance and model capacity range, and obtains the required model capacity adaptation range according to the computer performance information input by the user. For example, the user inputs: I have a 3070 computer, and the local robotic arm call interface is at / home / workspace / robots / piper / launch, where 3070 represents the user's computer performance information. There is an association table between GPU and model capacity range set in the system. For example, it is recorded that models 3050, 3060, and 3070 correspond to a model capacity of less than 100M.

[0052] S123. Screen the candidate model that meets the capacity adaptation range and has the largest capacity from the first models as the second model.

[0053] Select the models that meet the model capacity adaptation range from the first models, sort these models in descending order according to the similarity, and select the model with the largest capacity as the second model. According to the example of selecting the first model, the models with a capacity greater than or equal to 100M, that is, Model 2, can be excluded, and Model 1 and Model 3 are obtained. Sort them in descending order of similarity, and select the candidate model with the largest capacity, that is, Model 3: capacity 10M, as the second model. If there are multiple models with the same capacity, further judgment is required according to the data source.

[0054] S124. Whether there are multiple candidate models with the same capacity. If not, use the candidate model as the second model. If so, obtain the data source information based on the candidate model, obtain the data source information of the required model based on the task requirement information, and select the model with the same data source information as the required model from multiple candidate models as the second model.

[0055] The system model database is a relational database. The model can trace back to all the information related to this model in the database through the model id. Therefore, as long as there is a model id, the capacity of this model can be obtained, as well as which dataset it is trained with and which robotic arm the data of this dataset comes from. For example, for the obtained model 1: capacity 1B, data source Piper, FrankaPanda robotic arm, the data source information of model 1 can be obtained as Piper. Obtain the data source information of the required model based on the task requirement information. For example, for the above user input: the local robotic arm call interface is in / home / workspace / robots / piper / launch, the data source information of the required model can be obtained as piper. When there are multiple models with the same similarity, select the model with the same data source information as the required model in the task requirement information from them as the second model. The data source information of the model indicates what type of robotic arm data the model uses for training. In the case of the same similarity, preferentially select the model with the same data source information as the required model, that is, select the model trained with the same robotic arm data. Suppose two models with the same capacity are obtained in the previous step, one is model 1 with a data source of piper, and the other is model 3 with a data source of UR5, then select model 1 as the second model. Additionally, if the data sources of both models are piper, further judgment is required based on the model similarity.

[0056] S125. Whether there are multiple models with the same data source information. If not, use the model selected in step S124 as the second model. If so, select the model with the maximum similarity from multiple models as the second model.

[0057] In this step, models that are too large in capacity to run locally and models with low similarity are excluded based on the user's hardware information. Further, after judging based on the data source among candidate models with the same capacity, if there are multiple models with the same data source information as the required model, select the model with the maximum similarity from them as the second model.

[0058] After obtaining the second model, the second model controls the robotic arm to complete a task test, and the system then initiates a user requirement inquiry to further adjust the model based on the user requirement feedback.

[0059] S2. Obtain user requirement feedback, analyze the requirement feedback to select task requirement decomposition, and perform model matching according to the decomposed requirements to obtain the target model combination.

[0060] S21. Obtain user requirement feedback.

[0061] After the test is completed, the system asks the user through the interaction interface whether the requirements are met. The user inputs the result information of whether the requirements are met. If the requirements are not met, the user uploads information such as images, videos, and descriptions for the model failure. The system obtains the user input requirement feedback information.

[0062] S22. According to the result of the requirement feedback, if so, proceed to the next step, step S3; if not, enter the step of analyzing the requirement feedback to adjust the model, step S23.

[0063] S23. Analyze the requirement feedback information to determine whether the reason for the current model not meeting the requirements is task missing or generalization problem.

[0064] By analyzing the requirement feedback information, determine the reason why the current model does not meet the requirements, that is, determine whether the model has a task missing or a generalization problem. The system has a classification module to classify the model according to the information such as images, videos, and descriptions input by the user. If the task requirements proposed by the user cannot be found in the existing task database, it means that the task is missing and a classification mark is made. If it is not a task missing, it is a generalization problem, that is, the model can complete the task but not well, and a classification mark is made. Here, the classification is performed using a classification model obtained by training with the existing model.

[0065] S24. If the task is missing, decompose the user task requirements, perform model matching for each decomposed requirement, and obtain the target model combination.

[0066] In the case of task absence, the task requirements are decomposed. For example, the user inputs the task requirement: "Prepare dishes: Scrambled eggs with tomatoes". There is a kitchen dish preparation scenario in the scenario database, but there is no accumulated data on scrambled eggs with tomatoes in the tasks, and no model is trained based on the dish preparation data of scrambled eggs with tomatoes. In such a situation, it can be considered that the task is absent, that is, all current models are 100% unable to complete the task. At this time, the large language model can be used to decompose the task first. Using the COT (Chain-of-thought) mechanism, the model is decomposed step by step, and finally a clear sequence of requirements is obtained, such as (Requirement 1: Place the bowls and chopsticks; Requirement 2: Cut the tomatoes into pieces and put them on a plate; Requirement 3: Crack the eggs into a bowl; Requirement 4: Discard the broken eggshells; Requirement 5: Grab the egg beater; Requirement 6: Stir). The above classification is operated using a classification model obtained by training with existing models. Each requirement is matched with the models in the model database according to the above model matching method, and finally a target model combination is obtained.

[0067] S25. If it is a generalization problem, use cloud computing as the second model.

[0068] The generalization problem is understood as, for example, in 100 tasks, the model can successfully complete 50 times, and the other 50 times fail for various reasons. The model can complete the task to a certain extent, but in some cases, the model performs poorly, that is, the data is not good, there is not much data, and the model is not large enough. So when it is found that the previously selected model can complete the task, but the overall completion rate is not high, a larger model based on larger data will be selected. If the model selected during the model matching process is already the largest model that the user can locally apply, then cloud server computing is used as the second model. Through cloud server computing, the calculation results are transmitted to the robot through the network for robot driving. Similarly, the robot is controlled to perform tests according to the obtained calculation results until a second model that meets the requirements is obtained. This solution also designs an adaptive scheduling method for cloud and local computing resources to achieve cloud-local collaborative optimization, applies cloud computing to provide complex model calculations, and ensures to provide the best model for users.

[0069] S3. Model self-learning steps. As Figure 3 shown, S31. Ask whether to agree to data collection. If not, go to step S4; if so, perform model self-learning.

[0070] S32. Collect and process robot data.

[0071] The user agrees to data collection and inputs the sensor call interface and task description. The system collects data according to the provided call interface. After the data collection volume reaches the quota of a transmission packet, it will be sent back to the remote server. The collected data is processed, including timestamp matching, data correction, and null value handling.

[0072] S33. Obtain the pre-trained model.

[0073] Use the model obtained after adjustment through model matching and requirement feedback as the pre-trained model.

[0074] S34. Configure the training hyperparameters. This step is configured using existing algorithms.

[0075] S35. Start the training script to perform training.

[0076] S36. Conduct remote testing after training. Specifically including: S361. Conduct remote testing after meeting the first training duration.

[0077] S362. Determine whether there is a real test environment. If so, start remote testing; if not, start simulation testing.

[0078] S362. Determine whether the test is passed. If the test is passed, add the model to the model database. If the test is not passed, return to step S34 after meeting the second training duration, reconfigure the training hyperparameters, and then perform model training again.

[0079] The present invention introduces additional data collection and a functional stack for model training. When the user uses the AI Agent to complete task requirements, real-time data and user evaluations / feedbacks will be synchronously collected. After the data collection volume reaches the quota of a transmission packet, it will be sent back to the remote server. The server automatically selects a suitable model on the new data for fine-tuning. The suitable model is the model obtained through the model matching process. Model testing is carried out through a laboratory simulation environment (real environment, real task, only for testing). This is because the small-scale robot model may be unstable during training, and it is possible to encounter a training trough during the fine-tuning process, resulting in the loss of model capabilities in the current parameter distribution. Therefore, a testing process needs to be added during training to ensure that the final model has the required capabilities. After model testing, the new model is transmitted to the cloud. The model that the user will use for this task next time is this new model. After the user uses it without problems, the current model will be solidified in the system for long-term use in the future, while the old model will be phased out. This entire process will run automatically, thereby endowing the system with continuous self-iterative capabilities.

[0080] S4. Load the second model or the target model combination onto the robot.

[0081] The present invention proposes to flexibly schedule and manage models by an AI Agent, enabling more efficient model collaboration, simpler data flow, and resource allocation. The present invention can dynamically and adaptively allocate models according to user requirements and hardware performance, improving resource utilization efficiency, solving the problems that existing methods are difficult to flexibly adapt to diverse scenarios, resulting in a decline in robot performance or work effectiveness, and the problem of insufficient adaptability of end-to-end models. The present invention selects the optimal model according to user input through the AI Agent and can adjust the model in real time according to demand feedback, obtaining the optimal model or model combination through task decomposition or cloud computing. By flexibly scheduling model strategies, the present invention can respond in real time to diverse demands in dynamic scenarios, such as task numbers, resource limitations, or emergencies, ensuring that the robot always maintains high efficiency in complex environments.

[0082] The system of the present invention needs to interact with the real world in real time. In the overall architecture design, very complex nested logic will not be used to ensure low latency and fast response in the interaction between the user and the AI Agent and between the AI Agent and the system functions. Moreover, in the designed method and system, there is no dependency relationship between functions. For example, by asking the user's needs, tasks can be classified, or tasks can not be decomposed; it can self-learn, or data collection can be selected not to be performed. Such a design gives users more autonomy on the one hand, and on the other hand, uses a non-coupled design to have greater scalability, and can be quickly adapted for system upgrades and function expansions in the later stage.

[0083] Embodiment 2: This embodiment also discloses another implementation manner of a robot model scheduling method based on an AI Agent, including the following steps: S1. Obtain user task requirements and hardware performance information, and perform model matching through the AI Agent. Specifically, it includes, S11. Select the first model similar to the task from the model database according to the task requirements.

[0084] S111. The system asks the user for task requirements and obtains the task requirement information input by the user.

[0085] S112. Perform similarity matching between the task requirement information and the model task description information, Judge whether there is a situation of no similarity. If so, perform task missing processing and enter step S13. If not, enter step S113.

[0086] The system includes a model database and a task database. The model database stores models with different functions, and the task database stores different tasks. Each model is set with a task ID tag, and the task ID is associated with one or more specific tasks in the task database. Moreover, each task is set with a task description tag, and the task description is a specific text description of the task. The first model is the one with a high similarity between the model task description information and the user input task requirement description information. The user inputs a task requirement information, and each model has a task description information. Since there are many models in the model database, it is necessary to find the model with a high information similarity from the task description information of all models according to the user input task requirement information. Information similarity matching can adopt existing algorithms. One of them can convert a piece of text into a set of vectors through feature engineering similar to tf-idf, and calculate the similarity between the input vector a of the user and the vectors b1, b2, b3... converted from the task descriptions of the models in the model database, such as the cosine approximation algorithm. Each model obtains a similarity through the similarity calculation. Determine whether there is a situation of no similarity. The situation of no similarity means that there is no task in the existing task database with a task description similar to the task requirement proposed by the user. For example, the user inputs the task requirement: "Prepare dishes: Scrambled eggs with tomatoes". There is a kitchen dish preparation scenario in the scenario database, but there is no accumulated data on scrambled eggs with tomatoes in the tasks, and no model is trained based on the dish preparation data of scrambled eggs with tomatoes.

[0087] S113. Obtain the model with a high similarity as the first model.

[0088] S12. Select the second model that matches the hardware from the first models according to the hardware information.

[0089] S121. The system asks the user about the hardware performance and obtains the hardware performance information input by the user.

[0090] S122. Obtain the model capacity adaptation range according to the hardware information.

[0091] S123. Screen the candidate model that meets the capacity adaptation range and has the largest capacity from the first models as the second model.

[0092] S124. Whether there are multiple candidate models with the same capacity. If not, use the candidate model as the second model; if so, obtain the data source information according to the candidate model, obtain the required model data source information according to the task requirement information, and select the model with the same data source information as the required model data source information from the multiple candidate models as the second model.

[0093] S125. Whether there are multiple models with the same data source information. If not, use the model selected in step S124 as the second model; if so, select the model with the largest similarity from the multiple models as the second model.

[0094] S13. Decompose the user task requirements, perform model matching on each decomposed requirement, and obtain the target model combination. In the process of performing model matching on the decomposed requirements, use the above steps S11 - S12.

[0095] S2. Obtain user requirement feedback, analyze the requirement feedback, select to decompose the task requirements, and perform model matching according to the decomposed requirements to obtain the target model combination.

[0096] S21. Obtain user requirement feedback.

[0097] S22. According to the result of the requirement feedback, if so, proceed to the next step, enter step S3; if not, enter the step of analyzing the requirement feedback to adjust the model, enter step S23.

[0098] S23. Analyze the requirement feedback information and determine whether the reason why the current model fails to meet the requirements is a generalization problem.

[0099] S24. If it is a generalization problem, use cloud computing as the second model.

[0100] S3. Model self - learning step.

[0101] S31. Ask whether to agree to data collection. If not, enter step S4; if yes, perform model self - learning.

[0102] S32. Collect and process robot data.

[0103] The user agrees to data collection, and the user inputs the sensor call interface and the task description. The system performs data collection according to the provided call interface. After the data collection volume reaches the quota of a sending packet, it will be sent back to the remote server. Process the collected data, including timestamp matching, data correction, and null value processing.

[0104] S33. Obtain the pre - trained model.

[0105] Use the model obtained after model matching and requirement feedback adjustment as the pre - trained model.

[0106] S34. Configure the training hyperparameters. This step uses existing algorithms for configuration.

[0107] S35. Start the training script for training.

[0108] S36. Conduct remote testing after training. Specifically include: S361. Conduct remote testing after meeting the first training duration.

[0109] S362. Determine whether there is a real test environment. If so, start remote testing. If not, start simulation testing.

[0110] S362. Determine whether the test is passed. If so, add the model to the model database. If not, return to step S34 after the second training time is met, reconfigure the training hyperparameters, and then conduct model training.

[0111] Embodiment 3: This embodiment discloses a robot model scheduling system based on AI Agent, which is used to implement the robot model scheduling method based on AI Agent, such as the method in Embodiment 1 and Embodiment 2. Figure 4 As shown, the system includes, Information acquisition module 1, obtaining user task requirements and hardware performance information; Model matching module 2, selects a matching model based on task requirements and hardware performance information through AI Agent; Demand feedback processing module 3, decomposing task requirements according to demand feedback, and performing model matching according to the decomposed requirements to obtain a target model combination; Loading module 4, the second model or the target model combination is loaded into the robot; Self-learning module 5 collects and processes robot data and trains the pre-selected modules. After training, remote testing is performed and the model database is updated after testing. The system includes an AI Agent unit, and the system also includes a scene database 7, a task database 8, and a model database 9. The AI ​​Agent unit is connected to the scene database, the task database, and the model database respectively, and the AI ​​Agent unit is also connected to a remote server 6 and a remote test environment respectively. The AI ​​Agent unit specifically includes an information acquisition module 1, a model matching module 2, a demand feedback processing module 3, a loading module 4, and a self-learning module 5. The information acquisition module and the demand feedback processing module are connected to the model matching module respectively, the demand feedback processing module is connected to the loading module and the self-learning module, the scene database 7, the task database 8, and the model database 9 are specifically connected to the model matching module, and the remote server is connected to the demand feedback processing module. The demand feedback processing module initiates a demand inquiry to the customer and obtains user demand feedback. The demand feedback processing module also includes judging the model generalization problem according to the demand feedback, selecting a larger model from the candidate models as the second model, and using cloud computing as the second model when there is no larger model.

[0112] The specific embodiments described herein are merely illustrative of the spirit of the present invention. Those skilled in the art to which the present invention pertains may make various modifications or supplements to the described specific embodiments or use similar means for substitution, but will not deviate from the spirit of the present invention or exceed the scope defined by the appended claims.

[0113] Although terms such as information acquisition module, model matching module, demand feedback processing module, loading module, etc. are used more frequently herein, the possibility of using other terms is not excluded. The use of these terms is only for more convenient description and explanation of the essence of the present invention; interpreting them as any additional limitation is contrary to the spirit of the present invention.

Claims

1. A robot model scheduling method based on AI Agent, characterized in that: The following steps are involved: Obtain user task requirements and hardware performance information, and perform model matching through AI Agent, including: According to the task requirements, select the first model similar to the task from the model database, Selecting a second model matching the hardware from the first model according to the hardware information; Obtain user demand feedback, analyze the demand feedback and select task demand decomposition, perform model matching based on the decomposed demand to obtain the target model combination; Load the second model or target model combination into the robot.

2. The robot model scheduling method based on AI Agent according to claim 1 is characterized in that: The step of selecting a first model similar to the task from a model database according to task requirements includes: The task requirement information is matched with the model task description information by similarity, and a model with high similarity is obtained as the first model.

3. The robot model scheduling method based on AI Agent according to claim 1 is characterized in that: The step of selecting a second model matching the hardware from the first model according to the hardware information includes: The model capacity adaptation range is obtained according to the hardware information, and a candidate model that meets the capacity adaptation range and has the largest capacity is selected from the first model as the second model.

4. The robot model scheduling method based on AI Agent according to claim 3 is characterized by: If there are multiple candidate models with the same capacity, data source information is obtained according to the candidate models, data source information of the requirement model is obtained according to the task requirement information, and a model whose data source information is consistent with the data source information of the requirement model is selected from the multiple candidate models as the second model.

5. The robot model scheduling method based on AI Agent according to claim 4 is characterized by: If there are multiple models with the same data source information, the model with the highest similarity is selected as the second model.

6. The robot model scheduling method based on AI Agent according to claim 5 is characterized in that: The steps of obtaining user demand feedback, analyzing demand feedback, selecting task demand decomposition, and matching models according to the decomposed demand to obtain the target model combination include: The second model controls the robot to complete the test and asks whether the test meets the requirements. If the requirements are not met, obtain user feedback and determine whether there are any missing tasks based on the feedback. If there are any missing tasks, decompose the user's task requirements, match each decomposed requirement with a model, and obtain the target model combination. If the requirements are met, output the second model.

7. The robot model scheduling method based on AI Agent according to any one of claims 3 to 6, characterized in that: Determine whether it is a generalization problem based on demand feedback. If it is a generalization problem, use cloud computing as the second model.

8. The robot model scheduling method based on AI Agent according to claim 1 is characterized by: After obtaining the second model or the target model combination, a model self-learning step is also included, specifically including: Ask whether you agree to data collection. If you agree, the robot data will be collected and processed. Obtain the pre-trained model based on the model matching results, configure the training hyperparameters, and start training; After training, remote testing is performed and the model database is updated with the model after testing.

9. The robot model scheduling method based on AI Agent according to claim 8 is characterized in that: Remote testing specifically includes: After the first training duration is met, a remote test is conducted to determine whether the test is passed. If the test is passed, the model is added to the model database. If the test is not passed, the training hyperparameters are reconfigured after the second training time is met and the model training is then performed again.

10. A robot model scheduling system based on AI Agent, implementing the method described in any one of claims 1 to 9, characterized in that: include, Information acquisition module, which obtains user task requirements and hardware performance information; Model matching module, which uses AI Agent to select matching models based on task requirements and hardware performance information; The demand feedback processing module decomposes the task requirements according to the demand feedback, and performs model matching according to the decomposed requirements to obtain the target model combination; The loading module, the second model or the target model combination is loaded into the robot.

Images