An AI agent adaptive decision-making method, system and storage medium
By combining multimodal perception and dynamic intent generation with reinforcement learning, the problem of insufficient fusion of unstructured information in AI Agent platforms has been solved, enabling a comprehensive understanding of complex environments and adaptive decision-making, thereby improving the robustness and adaptability of task completion.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- SHANGHAI EMBEDWAY INFORMATION TECH
- Filing Date
- 2026-03-26
- Publication Date
- 2026-06-23
AI Technical Summary
Existing AI Agent platforms cannot effectively integrate unstructured information such as voice, video images, and sensor data, resulting in incomplete understanding of the external environment, neglect of user history preferences and emotional states in intent recognition, rigid decision-making strategies, lack of dynamic adjustment capabilities, and difficulty in adapting to dynamic environments.
A multimodal perception method is adopted to directly map text, speech, image, and sensor data to a unified semantic space. Features are fused through a cross-modal attention mechanism to generate a unified perception vector. Dynamic intent representations are generated by combining user profiles and historical memory. Decision-making strategies are optimized through reinforcement learning to build an end-to-end closed-loop architecture.
It achieves a comprehensive understanding and accurate decision-making in complex environments, enhances the AI Agent's adaptability and task completion robustness in dynamic environments, and has the ability to continuously optimize.
Smart Images
Figure CN122264115A_ABST
Abstract
Description
Technical Field
[0001] This application relates to the fields of artificial intelligence and intelligent agent technology, specifically to an AI Agent adaptive decision-making method, system, and storage medium. Background Technology
[0002] With the development of large model technology, the industry urgently needs intelligent agents that can process multi-source information from the real world (such as text, voice, video, images, and sensors) and continuously evolve to support highly reliable automated tasks.
[0003] The existing technology mainly has the following problems: 1. Existing AI Agent platforms (such as LangChain and Coze) mainly rely on text input and cannot effectively integrate multimodal signals of unstructured information such as voice, video images, and sensor data, resulting in an incomplete understanding of the external environment.
[0004] 2. Intent recognition only parses explicit instructions, ignoring user history preferences, emotional state, and contextual implicit needs, leading to task deviation and easily producing "irrelevant answers".
[0005] 3. The decision-making strategy is rigid and lacks closed-loop capability. Strategy generation relies on preset toolchains or fixed prompt templates, lacking the ability to dynamically adjust and optimize online based on execution feedback, making it difficult to adapt to dynamic environments. Execution results do not effectively feed back into the perception and decision-making modules, hindering continuous optimization.
[0006] The root cause of the above problems lies in the fragmentation of the perception, intent, and decision-making modules, and the lack of an end-to-end adaptive closed-loop learning mechanism. Summary of the Invention
[0007] To help solve the above-mentioned technical problems, this application provides an AI Agent adaptive decision-making method, system, and storage medium.
[0008] In a first aspect, this application provides an AI Agent adaptive decision-making method based on multimodal perception, which includes the following steps: S1, receiving multimodal input signals from the external environment, wherein the multimodal input signals include at least one or more of text data, voice data, video image data, or IoT sensor data; S2, skip the intermediate processing step of converting non-text modal data into text, independently encode each modal data in the multimodal input signal, and extract the corresponding original feature vector to retain the fine-grained semantic features in the original modality; S3, a cross-modal attention mechanism is used to process all the extracted original feature vectors, mapping the feature vectors of different modalities to a unified semantic space and fusing them to generate a unified perceptual vector; S4. Generate a dynamic intent representation based on the unified perception vector, and generate a decision strategy based on the dynamic intent representation to control the AI Agent to perform corresponding actions.
[0009] S2 includes: For text input, a Tokenizer encoder is used for word segmentation and vectorization, outputting text feature vectors; for speech input, a Whisper model is used for acoustic feature extraction, outputting speech feature vectors; for image or video input, a ViT encoder is used for visual feature extraction, outputting image or video feature vectors; for sensor data, an MLP network is used for normalization and numerical feature extraction, outputting sensor feature vectors.
[0010] S3 includes: By calculating the interaction weights between modalities through a cross-modal attention mechanism, the features of each modality are weighted and fused to generate a unified perceptual representation.
[0011] S4 includes: S41: Retrieve long-term preference data from user profiles and historical memory banks, and convert it into user long-term profile vectors and historical interaction memory vectors; S42: Concatenate the unified perception vector, the user long-term profile vector, the historical interaction memory vector, and the current task target to construct a joint input sequence; S43: Input the joint input sequence into the large language model and force the output of a dynamic intent representation that conforms to a preset structured format; S43 includes: calling the Function Calling interface of the large language model to force the output of a dynamic intent representation in JSON format, wherein the dynamic intent representation includes at least an action type field and a corresponding execution parameter field; S44: Parse the dynamic intent representation, match and call the corresponding executable action sequence from the pre-built tool library to execute the task, and output the execution result; S45: Collect execution result logs in real time, and construct a multi-dimensional reward signal based on the execution result logs and user feedback data. With the goal of maximizing cumulative reward, a reinforcement learning algorithm is used to simultaneously optimize the parameters of two objects: the prompt word template parameters of the large language model called in S43 and the policy network parameters of the matching action sequence in S44, and update the optimized parameters.
[0012] When using a proximal strategy optimization algorithm to simultaneously fine-tune parameters, the same reward signal is shared for joint gradient updates.
[0013] When calling the corresponding tools to execute tasks, secure communication with cloud services is conducted via HTTPS combined with the OAuth2.0 protocol or the national cryptographic SM2 / SM9 protocol.
[0014] Secondly, this application provides an AI Agent adaptive decision-making system based on multimodal perception, comprising: A multimodal sensing module is used to acquire multimodal input signals; A unified embedding module, connected to a multimodal perception module, is used to skip the intermediate processing step of converting non-textual modal data into text, independently encode each modal data in the multimodal input signal, extract the corresponding original feature vectors, and use a cross-modal attention mechanism to process all the extracted original feature vectors, mapping the feature vectors of different modalities to a unified semantic space and fusing them to generate a unified perception vector. A dynamic intent recognition engine, connected to a unified embedding module, is used to generate dynamic intent representations based on the unified perception vector; An adaptive decision-making and execution engine, connected to a dynamic intent recognition engine, is used to generate decision strategies based on the dynamic intent representation to control the AI Agent to perform corresponding actions.
[0015] The dynamic intent recognition engine is also used to retrieve long-term preference data from user profiles and historical memory banks, convert it into user long-term profile vectors and historical interaction memory vectors, and concatenate the unified perception vector, user long-term profile vector, and historical interaction memory vector with the current task target to construct a joint input sequence. The joint input sequence is then input into the large language model and forced to output a dynamic intent representation that conforms to a preset structured format. The adaptive decision and execution engine is also used to parse the dynamic intent representation, match and call the corresponding executable action sequence from the pre-built tool library to execute the task, and output the execution result; The system also includes: The execution feedback module, built into the adaptive decision and execution engine, is used to collect execution result logs and environmental state changes in real time. The online learning module, connected to the execution feedback module, is used to construct a multi-dimensional reward signal based on execution result logs and user feedback data. With the goal of maximizing cumulative rewards, it uses a reinforcement learning algorithm to simultaneously optimize the parameters of two objects: the prompt word template parameters of the large language model called in the dynamic intent recognition engine and the policy network parameters of the action sequence matching in the adaptive decision and execution engine, and updates the optimized parameters. The multimodal perception module and the unified embedding module are integrated into the local terminal device or edge computing node; The dynamic intent recognition engine runs on an edge server with GPU acceleration, and the user's long-term profile vector and historical interaction memory vector are stored in a local encrypted database; The adaptive decision and execution engine is deployed on an edge server, including a decision engine and a tool call interface for communicating with cloud services; the decision engine matches executable action sequences based on dynamic intent representations and calls cloud services to execute tasks through the tool call interface; The execution feedback module is built into the decision engine and is used to monitor execution result logs and user interaction events in real time, and transmit them asynchronously to the online learning module via a message queue. The online learning module is deployed on a centralized training server, and the optimized model parameters are encrypted and pushed to edge nodes for incremental updates. The tool calls the API to communicate securely with the cloud service via HTTPS combined with the OAuth2.0 protocol or the national cryptographic SM2 / SM9 protocol.
[0016] Thirdly, this application provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the AIAgent adaptive decision-making method based on multimodal perception as described in any one of the first aspects.
[0017] In summary, the method of this application has the following advantages: 1. It pioneers a multimodal unified embedding and dynamic intent joint modeling mechanism, skipping the intermediate conversion between multimodal and text, and directly integrating original features such as vision, voice, and IoT to generate structured intent. This breaks through the limitations of the traditional AI Agent system's "text-centric" approach (such as the conversion of multimodal inputs such as voice, video, and images into text, which leads to the loss of a large amount of fine-grained semantics in the original modality, such as visual anomalies, voiceprint emotions, and sensor temporal features). This enables intelligent agents to achieve more comprehensive contextual understanding and more accurate task decision-making in real and complex environments. 2. Introduce a dual-path approach driven by online reinforcement learning (dynamic intent generation + policy execution) to enable the AI Agent to continuously evolve and improve its performance in an edge cloud collaborative environment; 3. Construct an end-to-end closed-loop architecture of "perception-decision-execution-learning" to drive model updates in real time through execution feedback, which significantly improves the robustness and adaptability of task completion in complex and dynamic environments. Attached Figure Description
[0018] Figure 1 This is the overall architecture diagram of the AI Agent system in this application.
[0019] Figure 2 This is a schematic diagram of the multimodal perception and unified embedding process of this application; Figure 3 This is a flowchart illustrating the interaction between the dynamic intent recognition and strategy generation module of this application. Figure 4 This is a flowchart of the online feedback-driven policy update mechanism of this application. Detailed Implementation
[0020] The present application will be further described below with reference to the accompanying drawings. The principles of the present application are very clear to those skilled in the art. It should be understood that the specific embodiments described herein are merely illustrative of the present application and are not intended to limit the present application.
[0021] like Figure 1 As shown, the overall architecture of the AI Agent system in this solution includes a multimodal perception module, a unified embedding module, a dynamic intent recognition engine, an adaptive decision-making and execution engine, an execution feedback module, and an online learning module.
[0022] The modules in this solution work together, and the specific implementation steps are as follows: 1. Multimodal perception and unified embedding process (corresponding to) Figure 1 For details on the "Multimodal Perception Module" and "Unified Embedding Module", please refer to [link / reference]. Figure 2 ): The system receives multimodal input signals from the environment, including text, voice, images, and IoT sensor data; Text is encoded using a Tokenizer, speech is converted into feature vectors using a Whisper model, video images are processed by a ViT encoder, and sensor data is normalized using an MLP. All modal vectors are fused through the Cross-Attention mechanism to generate a unified perception vector P.
[0023] The multimodal perception module and the unified embedding module are integrated into local terminal devices or edge computing nodes (such as smart speakers and in-vehicle computers) to ensure low-latency real-time processing.
[0024] 2. Dynamic intent recognition and strategy generation (corresponding to) Figure 1 The "Dynamic Intent Recognition Engine" in the middle, and details of its interaction with the policy generation module, can be found in [link to relevant documentation]. Figure 3 ): Retrieve long-term preference data from user profiles and historical memory databases; By combining the current task objective with the unified perception vector P, a unified structured intent object is formed, and a structured prompt word that conforms to the input specification of Large Language Model (LLM) is dynamically constructed. Call the Function Calling interface of the Large Language Model (LLM) to force the output of a dynamic intent representation in JSON format (e.g., {"action": "adjust_light", "brightness": 25%}).
[0025] The generated intent object is transmitted in real time to Figure 1 The "Adaptive Decision and Execution Engine" triggers the policy generation process (see details of the interaction mechanism). Figure 3 The dynamic intent recognition engine runs on a GPU-accelerated edge server, while the user profile database is stored in a local encrypted database to ensure privacy and security.
[0026] It's important to note that the "unified structured intent object" mentioned above is not directly used as model input, but rather as an intermediate semantic representation. Based on this intent object, the system dynamically constructs a "structured cue word" that conforms to the Large Language Model (LLM) input specification, which can be understood as a structured "joint input sequence." This cue word is entirely composed of discrete text tokens and can seamlessly call any standard LLM API that supports Function Calling (such as GPT, GLM, etc.), outputting a structured intent in JSON format. The specific process is explained below: Generate a unified structured intent object: The system first integrates multimodal perception vectors, user profiles, historical memories, and the current task objective, outputting an internal non-textual structured intent object. This object is a semantic framework containing key-value pairs (this object exists only internally, guiding cue word generation, and is not directly exposed to the LLM), for example: { "perception_summary": {#Summary (derived from vector parsing)} "visual": "crowded_scene", "audio": "high_stress_vocal_tone,XXX", "sensor": "rapid_movement" }, "user_context": "prefers_quiet,XXXX", "current_task": "optimize_workspace " } Constructing Structured Prompts: The system has a built-in Prompt Template Engine. Based on the structured intent object mentioned above, it fills in a predefined text template to generate the final structured prompt (composed entirely of natural language and preset keywords, containing no floating-point numbers or raw vectors, therefore directly parsable by any plain text LLM), which is a joint input sequence. The format of this sequence is as follows: [ROLE] You are an adaptive AI agent with real-time multimodalawareness. [CONTEXT] - Visual Perception: crowded_scene, anomaly_detected - Audio Perception: high_stress_vocal_tone - Sensor Data: rapid_movement - User Profile: prefers quiet environments - Recent Memory: user was frustrated by loud notifications 5 minutes ago [TASK] Optimize workspace for focus. [INSTRUCTION] Analyze the above context and output a JSON object withkeys: "primary_intent", "confidence_score", "triggering_factors", "recommended_actions".
[0027] Calling the LLM Function Calling interface: The system uses the structured prompts constructed above as the user role content in the `messages` parameter to call the LLM Function Calling function through the standard API. Simultaneously, the expected JSON Schema output format is declared in the `functions` parameter, for example: response = llm_client.chat.completions.create( model="gpt-4-turbo", messages=[ {"role": "system", "content": "You are XXX"}, {"role": "user", "content": "[the complete prompt word constructed above]"} ], functions=[{ "name": "parse_user_intent", "parameters": { "type": "object", "properties": { "primary_intent": {"type": "string"}, "confidence_score": {"type": "number"}, "triggering_factors": {"type": "array", "items": {"type": "string"}}, "recommended_actions": {"type": "array", "items": {"type": "string"}} } } }] ) Outputting structured intents in JSON format: After receiving structured cue words in plain text, LLM, based on its understanding of the context, forces the return of a dynamic structured intent representation that conforms to the preset JSON format through the Function Calling mechanism.
[0028] 3. Adaptive decision-making and execution (corresponding to) Figure 1 (Adaptive Decision and Execution Engine) The strategy generator matches a sequence of executable actions from tools / API libraries based on dynamic intents; The execution engine calls the corresponding tools (such as smart home APIs and database query interfaces) to complete the action; The execution results are output to a physical or digital environment.
[0029] The decision engine is deployed on edge servers, while tool / API calls can be distributed across cloud services (such as smart home scenarios: Alibaba Cloud Living IoT or Tencent LianLian; industrial equipment control: Huawei Cloud IoTDA (IoT Device Access) or Alibaba Cloud Industrial Brain; and general service calls: Tencent Cloud API Gateway or Alibaba Cloud API Gateway). Each service communicates through security protocols (such as HTTPS+OAuth2.0 or the national cryptographic SM2 / SM9 protocol).
[0030] 4. Perform feedback collection (corresponding) Figure 1 (Execution Feedback Module) The system collects two types of feedback: a) Execution result log (such as API return status, execution time); b) Explicit / implicit user feedback (such as ratings, subsequent actions); The feedback data is structured and then transmitted to the online learning module.
[0031] Deployment and division of labor: The feedback module is built into the execution engine, which listens for execution results and user interaction events in real time, and transmits them asynchronously to the learning module through message queues (such as Kafka / MQ).
[0032] Online feedback strategy update mechanism (corresponding) Figure 1 For details on the mechanism and process of the "Online Learning Module" in the Chinese version, please refer to [link / reference]. Figure 4 ) Rewards are based on task completion rate, execution efficiency, and user satisfaction. The PPO (Proximal Policy Optimization) algorithm is used to simultaneously fine-tune the cue word template parameters of the Large Language Model (LLM) for dynamic intent recognition, as well as the action selection policy network parameters of the policy generator. The updated model is used for the next round of decision-making, forming a closed-loop optimization.
[0033] The online learning module runs on a centralized training server (which may be located in a private cloud) and periodically (e.g., every 24 hours) pushes the optimized model parameters to edge nodes in an encrypted manner to achieve incremental updates.
[0034] Specifically, this application provides an AI Agent adaptive decision-making method based on multimodal perception and dynamic intent recognition. This method can be applied to scenarios requiring real-time interaction with complex physical / digital environments, such as smart homes, autonomous driving, industrial inspection, and intelligent customer service. Figure 1 As shown, the system architecture of this application includes a multimodal perception module, a unified embedding module, a dynamic intent recognition engine, an adaptive decision-making and execution engine, an execution feedback module, and an online learning module. The following section combines... Figure 1-4 The methods of the embodiments of this application will be described in detail.
[0035] Multimodal unified perception.
[0036] like Figure 2 As shown, this step is implemented through a multimodal perception module and a unified embedding module.
[0037] First, the multimodal sensing module acquires multimodal input signals from the environment. These signals include at least text, voice, images or video, and IoT sensor data. Specifically: text input can be natural language instructions or document content entered by the user; voice input can be voice commands from the user; image or video input can be real-time footage captured by a camera; sensor data can be numerical data such as temperature, humidity, and distance.
[0038] Then, the unified embedding module extracts the original feature vectors of each modality through the corresponding feature extraction network: for text input, the Tokenizer encoder is used for word segmentation and vectorization, and the text feature vector is output; for speech input, the Whisper model is used for acoustic feature extraction, and the speech feature vector is output; for image or video input, the ViT (Vision Transformer) encoder is used for visual feature extraction, and the image or video feature vector is output; for sensor data, the MLP (Multilayer Perceptron) network is used for normalization and numerical feature extraction, and the sensor feature vector is output.
[0039] Next, the feature vectors from all modalities are input into the cross-modal attention layer. The cross-modal attention layer performs feature fusion by calculating the attention weights between the original feature vectors of different modalities. Specifically, a multi-head attention mechanism is used to calculate the interaction weights between modalities, and the features of each modality are weighted and fused to generate a unified perceptual representation that preserves the original fine-grained semantics of each modality. This unified perceptual representation contains a comprehensive and accurate description of the environment, avoiding the information loss caused by converting multimodal data into text in traditional methods.
[0040] This step achieves the feature of skipping the intermediate processing step of converting non-textual modal data into text by independently encoding each modality data and using a cross-modal attention mechanism for fusion, thereby preserving the fine-grained semantic features in the original modality.
[0041] Dynamic intent generation.
[0042] like Figure 3 As shown, this step is performed by the dynamic intent recognition engine.
[0043] First, long-term preference data from user profiles and the historical memory bank are retrieved and converted into long-term user profile vectors and historical interaction memory vectors. User profiles include user preferences, habits, and personal information, while the historical memory bank records the history of user interactions with the system.
[0044] Then, the generated unified perceptual representation, user long-term profile vector, historical interaction memory vector, and current task objective are concatenated to construct a structured prompt. This prompt describes the current context, user background, and task requirements in natural language or a specific format.
[0045] Next, the Function Calling interface of the Large Language Model (LLM) is invoked to force the output of a dynamic intent representation conforming to a preset structured format. In this embodiment, the preset structured format is JSON, and the dynamic intent representation includes at least an action type field and a corresponding execution parameter field, for example: {"action": "adjust_light", "brightness": 25%}. This structured output allows subsequent decision-making steps to directly parse and use it.
[0046] Finally, the generated dynamic intent representation is transmitted to the adaptive decision-making and execution engine in real time, triggering the policy generation process. Based on the dynamic intent representation, the policy generator matches executable action sequences from a pre-built tool / API library and prepares to call the corresponding tools to execute the task.
[0047] This step retrieves long-term preference data from user profiles and historical memory banks, concatenates the unified perception vector with the user's long-term profile vector and historical interaction memory vector into a joint input sequence, and calls the FunctionCalling interface of the large language model to force the output of a dynamic intent representation in JSON format, thus realizing dynamic intent generation based on unified perception vectors.
[0048] Adaptive decision execution.
[0049] This step is completed by the adaptive decision and execution engine.
[0050] First, the dynamic intent representation is parsed to extract the action type and its corresponding execution parameters. Then, based on the action type, matching executable action sequences are retrieved from a pre-built tool / API library. This tool / API library includes local tool interfaces and cloud service interfaces, such as smart home APIs, database query interfaces, and industrial equipment control interfaces.
[0051] Based on the matching results, specific invocation instructions are generated, and the corresponding tools are invoked to execute the task via a secure protocol. The security protocol can be HTTPS combined with OAuth 2.0, or the national cryptographic standard SM2 / SM9 protocol to ensure secure communication. During execution, changes in the execution status are monitored in real time, and execution result logs such as API return status and execution time are recorded.
[0052] Meanwhile, the execution feedback module is built into the decision engine to collect execution result logs and environmental state changes in real time. In addition, it collects explicit and implicit user feedback data, such as user ratings and subsequent actions. The collected data is asynchronously transmitted to the online learning module via message queues (such as Kafka or RabbitMQ) for further optimization.
[0053] This step executes tasks by parsing dynamic intent representations and matching executable action sequences from a pre-built tool library, while simultaneously collecting execution result logs in real time and communicating with cloud services via a security protocol, thus enabling the execution of decision-making strategies.
[0054] Closed-loop online optimization.
[0055] like Figure 4 As shown, this step is performed by the online learning module.
[0056] First, the online learning module receives data from the execution feedback module, including execution result logs and user feedback data. The execution result logs include API return status, execution time, etc.; user feedback includes ratings, subsequent actions, etc.
[0057] Then, the collected data is structured and a multi-dimensional reward signal is constructed based on task completion rate, execution efficiency, and user satisfaction. Task completion rate reflects whether a task is successfully completed, execution efficiency measures response time or resource consumption, and user satisfaction is indirectly reflected through user ratings or behavior.
[0058] Next, the Proximal Policy Optimization (PPO) algorithm is used to simultaneously fine-tune the parameters of two objects with the goal of maximizing cumulative reward: the cue word template parameters of the large language model called in the dynamic intent generation step; and the policy network parameters of the matching action sequence in the adaptive decision execution step.
[0059] By sharing the same reward signal for joint gradient updates, intent recognition and decision execution can work together to improve system performance.
[0060] Finally, the updated model parameters are encrypted and pushed to the edge nodes for the next round of decision-making, forming a closed-loop optimization. In this way, the system can continuously evolve with use, achieving an adaptive capability of "getting smarter with use".
[0061] This step constructs a multi-dimensional reward signal based on execution result logs and user feedback data, with task completion rate, execution efficiency, and user satisfaction as dimensions. It adopts the Proximal Policy Optimization (PPO) algorithm, shares the same reward signal for joint gradient updates, and synchronously fine-tunes the prompt word template parameters and policy network parameters of the large language model, thus realizing closed-loop optimization and model parameter updates.
[0062] This application also provides an AI Agent adaptive decision-making system based on multimodal perception and dynamic intent recognition, which is used to implement the above-mentioned method. For example... Figure 1 As shown, the system includes the following modules: The multimodal sensing module is used to acquire multimodal input signals. This module can integrate various hardware devices such as sensors, microphones, and cameras, or receive data through a software interface.
[0063] The unified embedding module, connected to the multimodal perception module, is used to extract the original feature vectors of each modality through the corresponding feature extraction network. The original feature vectors of each modality are input into the cross-modal attention layer. By calculating the attention weights between the original feature vectors of different modalities, feature fusion is performed to generate a unified perception representation that retains the original fine-grained semantics of each modality.
[0064] The dynamic intent recognition engine, connected to the unified embedding module, is used to concatenate the unified perception representation with the pre-stored long-term user profile vector and historical interaction memory vector to construct a joint input sequence. The joint input sequence is then input into the large language model, and a dynamic intent representation conforming to a preset structured format is forcibly output.
[0065] The adaptive decision and execution engine, connected to the dynamic intent recognition engine, is used to parse dynamic intent representations, match and call corresponding executable action sequences from a pre-built tool library to execute tasks, and output execution results.
[0066] The execution feedback module, built into the adaptive decision-making and execution engine, is used to collect execution result logs and environmental state changes in real time. This module is typically implemented in the form of listeners or hook functions, capable of capturing key events during the execution process.
[0067] The online learning module, connected to the execution feedback module, is used to construct a multi-dimensional reward signal based on execution result logs and user feedback data. With the goal of maximizing cumulative rewards, it uses a reinforcement learning algorithm to simultaneously optimize the parameters of two objects: the prompt word template parameters of the large language model called in the dynamic intent recognition engine and the policy network parameters of the action sequence matching in the adaptive decision and execution engine, and updates the optimized parameters.
[0068] In practical applications, the system adopts an edge-cloud collaborative architecture to improve real-time performance: The multimodal perception module and the unified embedding module are integrated into local terminal devices or edge computing nodes, such as smart speakers and in-vehicle computers. This allows preliminary processing to be completed near the data source, reducing latency.
[0069] The dynamic intent recognition engine runs on a GPU-accelerated edge server to quickly process large language model inference tasks. Long-term user profile vectors and historical interaction memory vectors are stored in a local encrypted database to ensure user privacy and security.
[0070] The adaptive decision-making and execution engine is deployed on an edge server and includes a decision engine and a tool invocation interface for communicating with cloud services. The decision engine matches executable action sequences based on dynamic intent representations and invokes cloud services to execute tasks through the tool invocation interface. Cloud services can be smart home platforms (such as Alibaba Cloud's IoT platform and Tencent Cloud), industrial IoT platforms (such as Huawei Cloud IoTDA and Alibaba Cloud Industrial Brain), or general API gateways (such as Tencent Cloud API Gateway and Alibaba Cloud API Gateway).
[0071] The execution feedback module is built into the decision engine. It is used to monitor execution result logs and user interaction events in real time and transmit them asynchronously to the online learning module through a message queue (such as Kafka) to avoid blocking the main process.
[0072] The online learning module is deployed on a centralized training server, utilizing powerful computing resources for model training. Optimized model parameters are periodically (e.g., every 24 hours) and encryptedly pushed to edge nodes for incremental updates.
[0073] To ensure secure communication, the tool calls the API to communicate securely with the cloud service via HTTPS combined with the OAuth2.0 protocol or the national cryptographic SM2 / SM9 protocol.
[0074] Furthermore, this application also provides a computer-readable storage medium storing a computer program thereon. When executed by a processor, this computer program can implement the aforementioned adaptive decision-making method for AI agents based on multimodal perception. Those skilled in the art will understand that all or part of the steps in the above method can be implemented by hardware related to program instructions, and the program can be stored in a computer-readable storage medium, such as ROM, RAM, disk, optical disk, etc. When executed by a processor, this computer-readable storage medium enables the AI agent to achieve functions such as multimodal perception, dynamic intent recognition, adaptive decision execution, and closed-loop online optimization, achieving the same technical effects as the aforementioned method.
[0075] In summary, the method presented in this application addresses the problems of existing technologies by constructing a four-step closed loop: unified multimodal perception, dynamic intent modeling, adaptive policy generation, and online feedback learning. This application employs a cross-modal attention mechanism to map text, speech, image, and sensor data to a unified semantic space, achieving comprehensive environmental perception. Furthermore, it constructs a dynamic intent engine based on a Large Language Model (LLM), integrating real-time input and long-term user profiles to output a structured intent representation. Additionally, it introduces a reinforcement learning-driven policy update mechanism, enabling the agent to continuously optimize its decision-making logic based on task completion rate and user satisfaction. Ultimately, this gives the agent an adaptive capability that becomes "smarter with use."
[0076] The technical scope of this application is not limited to the contents of the above specification. Those skilled in the art can make various modifications and variations to the above embodiments without departing from the technical concept of this application, and all such modifications and variations should fall within the scope of this application.
Claims
1. An adaptive decision-making method for AI agents based on multimodal perception, characterized in that, Includes the following steps: S1, receiving multimodal input signals from the external environment, wherein the multimodal input signals include at least one or more of text data, voice data, video image data, or IoT sensor data; S2, skip the intermediate processing step of converting non-text modal data into text, independently encode each modal data in the multimodal input signal, and extract the corresponding original feature vector to retain the fine-grained semantic features in the original modality; S3, a cross-modal attention mechanism is used to process all the extracted original feature vectors, mapping the feature vectors of different modalities to a unified semantic space and fusing them to generate a unified perceptual vector; S4. Generate a dynamic intent representation based on the unified perception vector, and generate a decision strategy based on the dynamic intent representation to control the AI Agent to perform corresponding actions.
2. The method according to claim 1, characterized in that, S2 include: For text input, a Tokenizer encoder is used for word segmentation and vectorization, outputting text feature vectors; for speech input, a Whisper model is used for acoustic feature extraction, outputting speech feature vectors; for image or video input, a ViT encoder is used for visual feature extraction, outputting image or video feature vectors; for sensor data, an MLP network is used for normalization and numerical feature extraction, outputting sensor feature vectors.
3. The method according to claim 1, characterized in that, S3 includes: By calculating the interaction weights between modalities through a cross-modal attention mechanism, the features of each modality are weighted and fused to generate a unified perceptual representation.
4. The method according to claim 1, characterized in that, S4 include: S41: Retrieve long-term preference data from user profiles and historical memory banks, and convert it into user long-term profile vectors and historical interaction memory vectors; S42: Concatenate the unified perception vector, the user long-term profile vector, the historical interaction memory vector, and the current task target to construct a joint input sequence; S43: Input the joint input sequence into the large language model and force the output of a dynamic intent representation that conforms to a preset structured format; S43 includes: calling the Function Calling interface of the large language model to force the output of a dynamic intent representation in JSON format, wherein the dynamic intent representation includes at least an action type field and a corresponding execution parameter field; S44: Parse the dynamic intent representation, match and call the corresponding executable action sequence from the pre-built tool library to execute the task, and output the execution result; S45: Collect execution result logs in real time, and construct a multi-dimensional reward signal based on the execution result logs and user feedback data. With the goal of maximizing cumulative reward, a reinforcement learning algorithm is used to simultaneously optimize the parameters of two objects: the prompt word template parameters of the large language model called in S43 and the policy network parameters of the matching action sequence in S44, and update the optimized parameters.
5. The method according to claim 4, characterized in that, When using a proximal strategy optimization algorithm to simultaneously fine-tune parameters, the same reward signal is shared for joint gradient updates.
6. The method according to claim 4, characterized in that, When calling the corresponding tools to execute tasks, secure communication with cloud services is conducted via HTTPS combined with the OAuth2.0 protocol or the national cryptographic SM2 / SM9 protocol.
7. An AI Agent adaptive decision-making system based on multimodal perception, characterized in that, include: A multimodal sensing module is used to acquire multimodal input signals; A unified embedding module, connected to a multimodal perception module, is used to skip the intermediate processing step of converting non-textual modal data into text, independently encode each modal data in the multimodal input signal, extract the corresponding original feature vectors, and use a cross-modal attention mechanism to process all the extracted original feature vectors, mapping the feature vectors of different modalities to a unified semantic space and fusing them to generate a unified perception vector. A dynamic intent recognition engine, connected to a unified embedding module, is used to generate dynamic intent representations based on the unified perception vector; An adaptive decision-making and execution engine, connected to a dynamic intent recognition engine, is used to generate decision strategies based on the dynamic intent representation to control the AI Agent to perform corresponding actions.
8. The system according to claim 7, characterized in that, The dynamic intent recognition engine is also used to retrieve long-term preference data from user profiles and historical memory banks, convert it into user long-term profile vectors and historical interaction memory vectors, and concatenate the unified perception vector, user long-term profile vector, and historical interaction memory vector with the current task target to construct a joint input sequence. The joint input sequence is then input into the large language model and forced to output a dynamic intent representation that conforms to a preset structured format. The adaptive decision and execution engine is also used to parse the dynamic intent representation, match and call the corresponding executable action sequence from the pre-built tool library to execute the task, and output the execution result; The system also includes: The execution feedback module, built into the adaptive decision and execution engine, is used to collect execution result logs and environmental state changes in real time. The online learning module, connected to the execution feedback module, is used to construct a multi-dimensional reward signal based on execution result logs and user feedback data. With the goal of maximizing cumulative rewards, it uses a reinforcement learning algorithm to simultaneously optimize the parameters of two objects: the prompt word template parameters of the large language model called in the dynamic intent recognition engine and the policy network parameters of the action sequence matching in the adaptive decision and execution engine, and updates the optimized parameters. The multimodal perception module and the unified embedding module are integrated into the local terminal device or edge computing node; The dynamic intent recognition engine runs on an edge server with GPU acceleration, and the user's long-term profile vector and historical interaction memory vector are stored in a local encrypted database; The adaptive decision and execution engine is deployed on an edge server, including a decision engine and a tool call interface for communicating with cloud services; the decision engine matches executable action sequences based on dynamic intent representations and calls cloud services to execute tasks through the tool call interface; The execution feedback module is built into the decision engine and is used to monitor execution result logs and user interaction events in real time, and transmit them asynchronously to the online learning module via a message queue. The online learning module is deployed on a centralized training server, and the optimized model parameters are encrypted and pushed to edge nodes for incremental updates. The tool calls the API to communicate securely with the cloud service via HTTPS combined with the OAuth2.0 protocol or the national cryptographic SM2 / SM9 protocol.
9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the AI Agent adaptive decision-making method based on multimodal perception as described in any one of claims 1 to 6.