Intelligent system integrated with six-force intelligent model capable of understanding physical world

Abstract

The invention provides the intelligent system integrating the six-force intelligent model capable of understanding the physical world, the system integrates the six-force intelligent model, the task can be quickly responded in a low-computing-power environment, and the efficiency is ensured under the requirement of high real-time performance. As the six-force intelligent model has flexible resource scheduling and task allocation capabilities, the six-force intelligent model can be dynamically adjusted according to environmental changes or task requirements, and efficient completion of tasks is ensured. In addition, the six-force intelligent model pays attention to task target guidance, and the requirements of different tasks can be accurately responded through the synergistic effect of comprehensive capacity. The intelligent system disclosed by the invention can quickly sense, understand and execute tasks, and meets the high-efficiency requirement in real-time performance.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of intelligent systems, and particularly relates to an intelligent system integrating a six-force intelligent model of an understandable physical world. BACKGROUND

[0002] Embodied intelligence refers to an intelligent system capable of perception, decision-making, execution, and feedback in the physical world. Embodied intelligent systems not only rely on information processing and computation, but also include physical interaction, environmental perception, and real-time action. In existing technologies, the mainstream embodied intelligence methods mostly rely on deep learning architectures based on large models. These methods are characterized by relying on massive data and large-scale computing power to solve tasks. Large models usually learn through multi-layer neural networks, automatically extracting features from massive data and making inferences. Specifically, the application of large models in embodied intelligence scenarios is as follows: (1) Data-driven task learning: Traditional embodied intelligent systems use large-scale data sets for training to learn how to perform tasks in a physical environment. For example, through visual input (such as camera images) and sensor data, the system can recognize objects and control actions to perform tasks.

[0003] (2) Global activation computation: In the process of performing tasks, existing large model-based systems usually need to perform global activation computation, mobilizing a large number of parameters and computing resources. This computing method is very suitable for processing complex tasks, but when applied to embodied tasks, the real-time response capability is significantly affected, resulting in excessive delay and resource consumption.

[0004] (3) Deep feature extraction and decision inference: In existing technologies, deep learning models extract features from multi-modal inputs such as vision and hearing through convolutional neural networks (CNN) and Transformer architectures, and make task inferences. This allows the system to make action decisions based on information from the external environment.

[0005] Existing large model-based methods have made significant progress on various tasks, but their intelligent behavior mainly relies on large-scale data training and high-intensity computing power support. The core logic is to cover knowledge with data, replace rules with parameters, and approximate understanding with computing power. They have the following obvious shortcomings in low-resource, high-real-time tasks of embodied intelligence: (1) Lack of physical world modeling capability: Large models usually make decisions based on data and statistical associations, lacking an inherent understanding of the structure and causal relationships of the physical world. In embodied tasks, understanding the physical properties of objects, spatial layout, and physical requirements of tasks is crucial. Existing large models fail to provide sufficient support in this regard, especially in environmental interaction and task execution, making it difficult to achieve precise control and decision-making.

[0006] (2) Large computational resource consumption: Large models have low computational efficiency, especially during inference, where each task requires full parameter activation, resulting in a large amount of computational redundancy. In complex tasks, the amount of computation increases sharply with the size of the parameters, especially in resource-constrained or real-time feedback application scenarios, the computational overhead of large models cannot be effectively addressed, often causing response delays or low efficiency.

[0007] (3) Poor task real-time performance: Existing large models require a large amount of computational resources, regardless of task complexity, the system must mobilize a large amount of computational resources to complete the task, resulting in computational redundancy. In practical applications of embodied intelligence, such as home service robots, computational resources are often limited, so solutions based on large models cannot meet the requirements of real-time tasks.

[0008] (4) Poor adaptability and flexibility: The environment and requirements of embodied tasks often change, and existing large models often lack flexible task processing capabilities. When the task environment changes, large models often cannot quickly and dynamically adjust the processing flow and strategy, but rely on pre-trained fixed models, which makes them perform poorly in dynamic and changing task environments. SUMMARY

[0009] To solve the above problems in the prior art, the present application provides an intelligent system integrating a six-force intelligent model that can understand the physical world. The technical problem to be solved by the present application is solved by the following technical solution: An intelligent system integrating a six-force intelligent model that can understand the physical world, characterized in that the intelligent system integrates a six-force intelligent model that can understand the physical world, a perception device and an execution device, the six-force intelligent model includes an observation module, an attention module, an understanding module, a discrimination module, a memory module and an execution module; the intelligent system activates part or all of the modules in the six-force intelligent model to complete the task target according to different tasks.

[0010] Advantages: The present application provides an intelligent system integrating a six-force intelligent model that can understand the physical world, which integrates a six-force intelligent model and can quickly respond to tasks in a low computational power environment and ensure efficiency under high real-time requirements. Due to the flexible resource scheduling and task allocation capabilities of the six-force intelligent model, it can be dynamically adjusted according to environmental changes or task requirements to ensure efficient task completion. In addition, the six-force intelligent model emphasizes task goal orientation and can accurately respond to the needs of different tasks through the synergistic effect of comprehensive capabilities. The intelligent system of the present application can quickly perceive, understand and execute tasks, achieving high efficiency in real-time.

[0011] The present application will be further described in detail below with reference to the accompanying drawings and examples. Attached Figure Description

[0012] Figure 1 This is a schematic diagram of an intelligent system that integrates a six-force intelligent model that can understand the physical world, as provided by this invention. Figure 2 This is a schematic diagram of the six-force intelligent model provided by the present invention. Detailed Implementation

[0013] The present invention will be further described in detail below with reference to specific embodiments, but the implementation of the present invention is not limited thereto.

[0014] The present invention aims to solve the following technical problems: (1) High-efficiency execution in low-resource environments: How to build an efficient embodied intelligence system in low-computing-power and low-bandwidth environments, avoid large-model computational redundancy, and ensure that the system can execute tasks quickly under low-resource conditions.

[0015] (2) Physical world modeling and task causal reasoning: How to accurately understand the state changes of the physical world and the causal relationship of tasks in embodied intelligence systems, thereby improving the accuracy and flexibility of task execution.

[0016] (3) Dynamic adjustment and adaptive capability: How to dynamically adjust the task processing flow, resource allocation and decision-making strategy under changing environment and task requirements, so as to achieve high adaptability of embodied intelligence system.

[0017] (4) Task response under high real-time requirements: Embodied intelligent tasks require the system to be able to respond to environmental changes in milliseconds. The goal of this invention is to ensure that the system can quickly perceive, understand and execute tasks, achieving high efficiency in real-time performance.

[0018] like Figure 1 As shown, the present invention provides an intelligent system that integrates a six-force intelligent model that can understand the physical world. The intelligent system integrates the six-force intelligent model, sensing devices, and execution devices. The six-force intelligent model includes an observation module, an attention module, a comprehension module, a discrimination module, a memory module, and an execution module. The intelligent system activates some or all of the modules in the six-force intelligent model according to different tasks to complete the task objectives.

[0019] Compared with data-driven intelligence of large models, biological intelligent systems exhibit more flexible, efficient and adaptive forms of intelligence. Humans and other animals exhibit multi-level intelligent capabilities in perception, understanding, reasoning and action, and can dynamically adjust processing procedures and resource allocation according to task requirements. Even in low-power organisms such as ants, bees and other small animals, task decomposition and action execution capabilities based on environmental understanding can be observed. Therefore, intelligence should be a capability system that can be flexibly combined and dynamically scheduled in different structures, different computing power and different task backgrounds, rather than a single mode of scale stacking.

[0020] Based on this understanding, the present application proposes a six-force intelligent model theory system, aiming to build a new artificial intelligence theory architecture with clear structure, adjustable function, embodied adaptation and task-oriented, so as to realize a clear structure, adjustable function and strong adaptability embodied intelligent system. Referring to the structure and function of biological intelligence, the six-force intelligent model gives the embodied agent six-dimensional capabilities: observation, attention, discrimination, understanding, memory and execution, with the state transition of things as the core, to build an embodied system with a closed loop of "perception-cognition-decision-execution". The system not only focuses on the mapping relationship between input and output, but also emphasizes state modeling, dynamic control and multi-module collaboration under task guidance.

[0021] Reference Figure 2 , Figure 2 is the framework diagram of the six-force intelligent model. In the six-force intelligent model described in the Figure 2 , the six-force intelligent model includes an observation module, an attention module, an understanding module, a discrimination module, a memory module and an execution module.

[0022] (1) Attention is the core ability of the perception stage of the embodied intelligent agent. Its task is to perceive multi-source heterogeneous environmental information through visual sensors, and extract task-related key inputs such as target regions in images and key instructions in semantics through neural networks (such as CNN, Transformer, etc.). It plays a role in information filtering and focusing, and is the "entrance" of the intelligent chain. The efficiency of attention directly affects the computational redundancy and judgment accuracy of the subsequent system.

[0023] (2) Understanding is responsible for connecting perception input and internal knowledge structure to achieve abstract modeling of things state, task target and physical law. It is the central ability of intelligent systems for task reasoning, structure analysis and target planning, and is the key bridge from perception to cognition.

[0024] (3) Discrimination is used to judge the difference between the current state and the target state, which is an important basis for decision-making. This module compares the "understood world" with the "task target" to assist in generating specific "operation instructions" or "behavior suggestions". In other words, discrimination is the key to the transition from cognition to action, and is the key generator of the task execution path.

[0025] (4) Memory is the ability to store and retrieve knowledge, experience, and historical state, and is the core module that supports generalization and transfer learning. It not only accumulates long-term knowledge, but also schedules historical experience according to the context during task execution to guide current behavior. The memory module can also be used for dynamic rule generation to assist understanding and discrimination.

[0026] (5) Execution transforms cognitive results into embodied behavior, manifested as motion control, operation planning, and environmental interaction. This module can call robot execution modules, API controllers, or be abstracted as an action path generator, control signal scheduler, etc., and is the actual execution unit to achieve the task target.

[0027] (6) Observation, as a feedback mechanism, is responsible for monitoring the matching degree of behavior results and task targets, evaluating execution effectiveness and generating the next round of perception tasks. It is the key channel for system evolution, self-correction, and iterative optimization. Observation realizes the "internal reflection" and "external detection" of the system, and is the endpoint of the intelligent closed loop and the starting point of the new round.

[0028] The six-force intelligent model is a closed-loop control model centered on the state space, which is constructed around the transformation process of the task from the initial state to the target state, i.e. a closed-loop control system. The state space is used to describe the execution state of the task and to predict the next state according to the current state; in the formalization layer, the state space S is defined as a set containing all possible states, i.e. ; where each state is symbolically expressed to describe the current state of things. The state of things refers to a set of physical quantities that reflect the properties of things, such as position, velocity, acceleration, shape, color, temperature, etc. State changes are predicted using state transition functions, so the state transition function is defined as: , where represents the set of execution sequences that the model can take, represents the state transitioned to after the model takes execution action in state . In the six-force intelligent model, each intelligent task can be mapped to the state space and abstracted as a state transition process:​ , where, is the initial state, i.e. the initial state of things and environment when accepting the task; is the target state, i.e. the final state of things and environment when completing the task; is the current state, i.e. the current actual state of things and environment perceived and cognized. The goal of state transition is to achieve the state transition path from to through a series of optimal or feasible execution sequences . Each capability module in the Six Force Intelligent Model plays the following roles in this state transition process: (1) Observation: Capture environmental signals , through network computing, the signal expression of environmental signals is converted to symbolic expression of thing state, and feedback to the system to activate the state transition process, realize the cycle of state recognition and cognitive update.

[0029] (2) Attention: Through the relevance calculation of thing state and current task, perceive and focus on the semantic feature information related to the task in the current state , provide high-quality input for state modeling (3) Understanding: Compress semantic information, extract pragmatic information related to the task in the current state , such as state variables or physical laws, etc., and can reconstruct without distortion of semantics; (4) Discrimination: Distance measurement between current state and target state, obtain the difference between the two states ; (5) Memory: Quickly store the current state, and schedule past experience to provide state-action pairs or state transition patterns in historical experience to generate next action suggestions; (6) Execution: According to the strategy instruction to carry out state transition calculation, convert the strategy into specific operation to promote state transformation. Actually call the action , through the state transition function to promote state change . The whole process revolves around the exploration and update of state space, forming a dynamic closed loop process:

[0030] . . The "observation" of this intelligent system captures the initial state of things through environmental signals ​​​Activate the "attention" and "understanding" modules, update the perception and cognition of the environment and task, and obtain semantic information and pragmatic information Then, the "discernment" determines the state error between the current state and the target state The "memory" module analyzes past experiences to develop a strategy for the next action The "execution" module executes the action strategy , and obtains the state of the changed thing . Then enter the next round of state transition cycle until the current thing state reaches the target state .

[0031] In addition, in order to realize the efficient operation of the intelligent system under different task complexity, the Six Force Model has a built-in ability adaptive activation mechanism based on task characteristics. The basic principle of this mechanism is "activation on demand": activate local path for simple tasks, and activate global path for complex tasks, to ensure that the system uses the minimum necessary resource path in different tasks. Therefore, the task is divided into simple task, medium complexity and high complexity task; For simple tasks, activate the attention module, understanding module and execution module in order; Simple task processing flow (lightweight path): For tasks such as target recognition, semantic matching, etc., only activate "attention → understanding → execution", at this time there is no need to call the memory library and feedback system, saving computing resources and responding to tasks quickly.

[0032] For medium complexity tasks, activate the attention module, understanding module, discernment module, execution module and observation module in order to complete the task target; Medium complexity task (standard path): Tasks involving state changes and operation instruction generation require activation of "attention → understanding → discernment → execution → observation", such as path planning, operation strategy generation, etc., which require real-time feedback and iterative updates.

[0033] For high complexity tasks, activate the attention module, understanding module, discernment module, memory module, execution module and observation module in order, and then feedback to the attention module to complete the task target.

[0034] High complexity task (complete closed loop path): For example, long-term autonomous tasks, complex assembly, etc., require activation of the full path "attention → understanding → discernment → memory → execution → observation → attention update". Such tasks have characteristics such as high uncertainty, long time span, and fast dynamic changes, and require the participation of memory scheduling and feedback reconstruction.

[0035] Through this adaptive mechanism, the Six-Force Intelligent Model not only improves the system's running efficiency and intelligent performance, but also provides a solid theoretical support and system foundation for complex application scenarios such as "energy-saving operation", "edge deployment", "multi-task concurrency" and so on.

[0036] Embodied intelligence tasks usually involve highly dynamic environments and complex interactive behaviors, requiring agents to not only perceive the external world but also make quick decisions and perform corresponding actions. The characteristics of such tasks require low resource consumption, high real-time response, and must be executed in the physical world, such as object grasping, navigation obstacle avoidance, and human-robot collaboration. To address these challenges, the Six-Force Intelligent Model in the embodied intelligence scenario follows a closed-loop process of "Observation → Attention → Understanding → Discrimination → Memory → Execution", specifically: The attention module is configured to extract semantic information related to the task from the environmental information. Attention is the core module of the perception stage, responsible for extracting semantic information related to the task from environmental information, avoiding information overload, and improving computational efficiency.

[0037] In a specific embodiment of the present application, the attention module is specifically configured to: If the task is a multi-modal input task, all environmental information in the current state obtained from the perception device is fused to obtain fused information, and invalid information unrelated to the task in the fused information is filtered to obtain semantic information. Multi-modal information fusion: For embodied intelligence tasks with multi-modal input, the attention module can fuse visual, auditory, and tactile information to effectively filter irrelevant information.

[0038] If the task is a single-modal input task, the semantic information related to the task is selected from the environmental information in the current state obtained from the perception device. Information filtering and focusing: In embodied intelligence tasks, the attention module selects the most important semantic information for task execution from sensor data such as camera images, depth maps, and sounds. For example, when a robot is grasping an object in a complex environment, the attention module focuses on the features of the target object, such as shape, position, or color, and ignores other irrelevant backgrounds.

[0039] According to the progress of task execution, the attention area during task execution is dynamically adjusted.

[0040] Dynamic adjustment of attention area: According to the progress of the task, the attention module dynamically adjusts the attention area. For example, when a robot approaches an obstacle during navigation, the attention module will prioritize the obstacle in front of it, adjusting the path planning in real time.

[0041] The understanding module is configured to compress the semantic information into task-oriented pragmatic information. The semantic information includes the category, position, and motion state of an object, and the pragmatic information is an information format matched with the task execution form of the intelligent system.

[0042] The understanding is responsible for connecting the perception input with the internal knowledge structure, compressing the semantic information obtained by the attention module into task-oriented pragmatic information, and realizing the abstract modeling of the environmental state, task target, and physical law. The understanding compresses the semantic information such as the category, position, and motion state of an object into pragmatic information, i.e., a form that can be directly applied to task execution by the system. The compressed information is usually high-dimensional and abstract, facilitating fast processing under limited computing power.

[0043] For example, in a embodied intelligence task, the robot needs to identify the target object from visual signals. The understanding module converts the visual information of the object into pragmatic information containing parameters such as target position, size, and shape. These pragmatic information are low-dimensional embedding vectors that retain the importance and relevance of semantics while removing redundant parts, which facilitates subsequent task execution and decision-making.

[0044] The discrimination module is configured to compare the pragmatic information with the task target, determine whether there is a difference between the two, and generate an adjustment instruction if there is a difference. The discrimination is used to determine the difference between the current state and the target state, and generate specific action decisions or behavior suggestions. The discrimination module compares the pragmatic information (i.e., the current state) obtained by the understanding module with the target state, analyzes the difference between them, and generates an adjustment instruction if there is a gap. For example, the robot determines whether the object has been placed at the target position, and if there is a gap, the discrimination generates a corresponding adjustment instruction.

[0045] The memory module is configured to store key information during task execution and activate historical experience to develop specific strategies according to the adjustment instruction. The memory module is responsible for the storage and retrieval of knowledge, experience, and historical state. This module realizes the storage of historical experience and the generation of strategies. During task execution, the memory module continuously records and stores key information such as successful grasping actions and navigation routes. These historical data provide a reference for subsequent tasks, improving the efficiency and success rate of task execution. According to the difference between the task target and the current state obtained by the discrimination module, the memory module activates relevant historical experience to develop specific strategies, such as "adjust the grasping angle".

[0046] The execution module is configured to convert the specific strategy into a specific action and execute the task through an execution device according to the specific action until the task target is achieved. The execution force module converts the strategy into specific actions, such as controlling the mobile chassis or mechanical arm, to achieve the completion of the task goal. Behavior generation and action control: the execution force module converts the operation instructions generated by the memory force module into specific mechanical operations. For example, when the robot performs a grasping operation, the execution force module controls the movement of the mechanical arm to ensure accurate grasping of the target object.

[0047] During task execution, the six modules form a closed loop, with observation constantly feeding back task results, attention, understanding, discrimination, memory, and execution adjusting the task execution path according to the feedback information, thus achieving dynamic optimization of the task. Each module not only works independently, but also closely cooperates with other modules to ensure the accuracy and efficiency of embodied task execution.

[0048] The observation module is used to monitor the state changes during task execution and evaluate the execution results of the task according to the task target, thereby generating feedback information; The observation module is specifically used for: Monitoring the state changes during task execution and feeding back the execution results of the task in real time; Evaluating the execution results and generating feedback information to the attention module.

[0049] Observation is the feedback mechanism of the intelligent system, responsible for monitoring the state changes during task execution and evaluating the execution effect according to the task target. This module mainly realizes environment monitoring and result evaluation, execution effect evaluation and iterative feedback. Through sensors, cameras and other perception devices, real-time monitoring of environmental changes is realized. For example, when the robot performs a grasping task, the observation module monitors whether the target object is successfully grasped and feeds back the task execution results in real time. Observation not only monitors the completion of the task, but also evaluates the difference between the task target and the execution results. For example, if the robot mistakenly grasps the object, the observation module will perceive this difference in time and generate feedback to guide the system to make the next adjustment. The observation module provides feedback information according to the deviation between the current state and the target, guides the allocation of attention in the next round, and strengthens the effective execution of the task.

[0050] The present application provides an intelligent agent loaded with the intelligent system integrating the six-force intelligent model of the understandable physical world.

[0051] It is worth noting that the terms "first" and "second" in the present application are only for descriptive purposes and cannot be understood as indicating or implying relative importance or implicitly indicating the number of the indicated technical features. Therefore, the features defined as "first" and "second" can explicitly or implicitly include one or more of the features. In the description of the present application, the meaning of "multiple" is two or more, unless otherwise specifically limited.

[0052] The above description is further detailed in connection with specific preferred embodiments of the present application, and it is not to be construed that the specific implementation of the present application is limited to these descriptions. For those skilled in the art to which the present application belongs, without departing from the concept of the present application, a number of simple deductions or substitutions can be made, and all of them should be considered as falling within the protection scope of the present application.

Claims

1. An intelligent system integrating intelligible physical world's six-force intelligent model, characterized in that, The intelligent system is integrated with an integrated six-force intelligent model capable of understanding the physical world, a perception device and an execution device, the six-force intelligent model includes an observation force module, an attention force module, an understanding force module, a discrimination force module, a memory force module and an execution force module; the intelligent system activates part of the modules or all of the modules in the six-force intelligent model according to different tasks to complete the task target.

2. The intelligent system integrating intelligible physical world's six-force intelligent model of claim 1, wherein, The six-force intelligent model is a closed-loop control model constructed around the transformation process from the initial state to the target state of the task with the state space as the core, the state space is used to describe the execution state of the task, and the next state is predicted according to the current state; the state change is predicted by using the state transition function.

3. The intelligent system integrating intelligible physical world's six-force intelligent model of claim 1, wherein, The task is divided into simple task, medium complex task and high complex task; For a simple task, the attention force module, the understanding force module and the execution force module are activated in sequence; For a medium complex task, the attention force module, the understanding force module, the discrimination force module, the execution force module and the observation force module are activated in sequence to complete the task target; For a high complex task, the attention force module, the understanding force module, the discrimination force module, the memory force module, the execution force module and the observation force module are activated in sequence, and then the attention force module is fed back to complete the task target.

4. The intelligent system of the integrated six-force intelligent model capable of understanding the physical world according to claim 1 or 3, characterized in that: The attention force module is used to extract semantic information related to the task from environmental information; The understanding force module is used to compress the semantic information into pragmatic information under the guidance of the task; The discrimination force module is used to compare the pragmatic information and the task target to determine whether there is a difference, and generate an adjustment instruction if there is a difference; The memory force module is used to store key information in the task execution process and activate historical experience according to the adjustment instruction to formulate a specific strategy; The execution force module is used to convert the specific strategy into specific actions, and execute the task according to the specific actions through the execution device until the task target is completed; The observation force module is used to monitor the state change in the task execution process, and evaluate the execution result of the task according to the task target, thereby generating feedback information.

5. The intelligent system integrating intelligible physical world's six-force intelligent model of claim 4, wherein, The attention force module is specifically used for: If the task is a multi-modal input task, all the environmental information in the current state obtained from the perception device is fused to obtain fusion information, and invalid information unrelated to the task in the fusion information is filtered to obtain semantic information; If the task is a single-modal input task, the semantic information related to the task is selected from the environmental information in the current state obtained from the perception device; According to the execution progress of the task, the attention area during task execution is dynamically adjusted.

6. The intelligent system integrating intelligible physical world's six-force intelligent model of claim 5, wherein, The semantic information includes the type, position and motion state of the object, and the pragmatic information is an information format matched with the task execution form of the intelligent system.

7. The intelligent system integrating intelligible physical world's six-force intelligent model of claim 5, wherein, The observation force module is specifically used for: Monitoring the state change in the task execution process and feeding back the execution result of the task in real time; Evaluating the execution result and generating feedback information to the attention force module.

8. An agent, characterized in that The intelligent agent is loaded with the intelligent system of the integrated intelligible physical world six-force intelligent model according to any one of claims 1 to 7. The intelligent agent is loaded with the intelligent system of the integrated intelligible physical world six-force intelligent model according to any one of claims 1 to 7.

Images