A body-aware architecture and method and apparatus for environmental adaptability learning

Abstract

The present application relates to a kind of embodied intelligent architecture and environmental adaptability learning method and device, a kind of embodied intelligent architecture, the embodied intelligent architecture includes sequentially connected from bottom to top core task anchoring layer, binary decision and check layer, dynamic strategy scheduling layer, basic cognitive function unit layer, reasoning state encoding layer and dynamic revision and reconstruction layer;A kind of embodied intelligent environmental adaptability learning method is applied to a kind of embodied intelligent architecture, comprising the following steps: S1, target anchoring;S2, pre-checking;S3, strategy routing;S4, module execution and state coding;S5, real-time checking;S6, dynamic revision.The present application is through six-layer closed-loop hierarchical cognitive architecture, signal interaction between levels, data transmission and collaborative work;Realize that embodied intelligence is actively monitored to the perception-decision-execution process, real-time checking, dynamic strategy scheduling and autonomous revision, improve the environmental adaptability of embodied intelligence, generalization ability, system reliability and explainability.

Description

Technical Field

[0001] This invention relates to the field of embodied intelligence, and in particular to an embodied intelligence architecture and an environmentally adaptive learning method and apparatus. Background Technology

[0002] Embodied intelligence (such as autonomous mobile robots, service robots, and humanoid robots) is an important development direction in the field of artificial intelligence. Its core requirement is to autonomously perform tasks in unstructured, dynamically changing real-world environments and possess strong environmental adaptability and learning capabilities. However, the environmental adaptability and learning capabilities of current embodied intelligence systems still face significant bottlenecks, mainly in the following aspects: 1. The "black box" problem of perception and decision-making: Existing embodied intelligence systems mostly rely on deep reinforcement learning or end-to-end neural networks. The environmental perception and behavioral decision-making processes are highly opaque and lack metacognitive monitoring capabilities for their own perception biases and decision risks. When the environment undergoes sudden changes (such as the sudden appearance of obstacles, increased sensor noise, or drastic changes in lighting), the system is prone to unpredictable erroneous behaviors, making it difficult to meet the requirements of high-reliability scenarios such as medical care and industrial collaboration.

[0003] 2. Insufficient strategy solidification and generalization ability: Traditional embodied intelligence behavioral strategies are mostly pre-trained in specific environments, making it difficult to proactively adapt to environmental changes; for example, when switching from indoor navigation to outdoor navigation, or from a static scene to a dynamic crowd scene, the system cannot autonomously adjust its behavioral patterns, resulting in low task success rate and poor fault tolerance.

[0004] 3. Lack of proactive self-monitoring and self-correction mechanisms: Existing adaptive learning methods (such as transfer learning and online reinforcement learning) are mostly passive adjustments, lacking proactive evaluation and correction of their own cognitive processes; when behavior fails or deviates from the goal, the system has difficulty quickly backtracking the reasoning process and reconstructing the strategy, resulting in low learning efficiency and slow convergence speed.

[0005] 4. Existing interpretability methods have limitations: Although methods such as Chain of Thought (CoT) have made some progress in improving the interpretability of embodied intelligence decision-making, these methods are mostly linear and passive reasoning prompts, lacking a systematic metacognitive closed-loop architecture, and cannot provide embodied intelligence with the ability to actively monitor, evaluate and correct its own cognitive processes.

[0006] Therefore, constructing an embodied intelligence cognitive architecture that can proactively monitor its own perception and decision-making processes and make autonomous corrections, in order to solve the problems of poor environmental adaptability, insufficient generalization ability, and weak reliability and interpretability of existing systems, is a technical challenge that urgently needs to be solved in the field of embodied intelligence. Summary of the Invention

[0007] The purpose of this invention is to overcome the shortcomings of the prior art and provide an embodied intelligence architecture and an environmentally adaptive learning method and device; to realize the active monitoring, real-time verification, dynamic strategy scheduling and autonomous correction of the perception-decision-execution process by embodied intelligence, thereby improving the environmental adaptability, generalization ability, system reliability and interpretability of embodied intelligence.

[0008] The objective of this invention is achieved through the following technical solution: An embodied intelligence architecture, comprising, from bottom to top, a core task anchoring layer, a binary decision and verification layer, a dynamic strategy scheduling layer, a basic cognitive functional unit layer, a reasoning state encoding layer, and a dynamic correction and reconstruction layer; Core Task Anchoring Layer: Analyzes the task to be processed and combines multimodal perception data to identify the core task objectives and constraints; Binary decision-making and verification layer: performs binary judgment on whether the content to be verified is abnormal or not; Dynamic policy scheduling layer: Selects and schedules target policies from preset policies; Basic cognitive functional unit layer: includes multiple basic cognitive modules, which perform corresponding basic operations; Inference State Encoding Layer: The encoded data is obtained by performing layered state encoding on the encoded data source; Dynamic correction and reconstruction layer: Performs correction operations.

[0009] Furthermore, the content to be verified includes perception data, behavioral decisions, action instructions, and core task objectives and constraints; The built-in verification rule base of the binary decision and verification layer includes rules for verifying the authenticity of perceived data, rules for assessing behavioral security, and rules for verifying physical constraints. The rule for verifying the authenticity of the perceived data is as follows: the reconstruction error of the image input from the visual sensor is calculated using a self-supervised reconstruction model, and the abnormality is determined based on the reconstruction error. The behavioral safety assessment rule is to determine whether there is an anomaly based on whether there are obstacles within a certain distance of the planned path; The physical constraint verification rule is to determine whether the generated instructions violate physical constraints to determine whether they are abnormal. The verification rule base supports dynamic expansion. The expansion mechanism of the verification rule base is as follows: for abnormal situations that occur more than a predetermined number of times in the experience base, candidate rules are generated and added to the verification rule base after manual confirmation.

[0010] Furthermore, the preset strategies include a fast response mode, a step-by-step reasoning mode, a knowledge-enhanced retrieval mode, and a backtracking correction mode; The dynamic policy scheduling layer introduces a reinforcement learning-driven policy selection mechanism, which automatically optimizes policy scheduling rules based on encoded data and feedback data.

[0011] Furthermore, the basic cognitive module includes a goal analysis and planning module, a long short-term memory module, a logical reasoning and causal analysis module, an association generation and creative expansion module, a fact verification and security verification module, a language organization and expression module, a constraint control and risk avoidance module, and an interactive feedback and iterative optimization module.

[0012] Furthermore, the encoding information sources of the reasoning state encoding layer include currently active basic cognitive modules, embodied intelligence's own state data, and environmental data; The encoded data includes joint state codes and historical context vectors; The hierarchical state coding mechanism of the inference state coding layer includes three layers; The first layer is basic state coding: coding the currently active basic cognitive modules; The second layer is the encoding of the environment and the self-state: encoding the self-state data and environmental data of the embodied intelligence; The joint state code is generated by concatenating the encoding results of the basic state code and the encoding results of the environment and its own state codes; The third layer is historical state encoding: the sequence of joint state encodings from the most recent N time steps is encoded to obtain the historical context vector.

[0013] Furthermore, the correction operation includes parameter-level fine-tuning and structure-level reconstruction; The triggering conditions for the structural-level reconstruction are: the abnormal signal is still not eliminated after a preset number of parameter-level fine-tunings; or the duration of a single abnormality exceeds a set time. The dynamic correction and reconstruction layer introduces a meta-reinforcement learning mechanism to learn and optimize the effect of the correction operation.

[0014] An embodied intelligence environment-adaptive learning method, applied to an embodied intelligence architecture, includes the following steps: S1. Target Anchoring: Based on the tasks to be processed, identify the core task objectives and constraints; S2, Pre-verification: Verify the core task objectives and constraints, and output the verification result as the first control signal; S3. Strategy routing: Based on the first control signal, combined with the core task objective and the acquired environmental status data, select and schedule the target strategy from the preset strategies; S4. Module Execution and State Encoding: Execute the corresponding basic cognitive modules according to the target strategy, and perform hierarchical state encoding operations to obtain encoded data; S5. Real-time verification: Perform real-time verification on the data output by the basic cognitive module in step S4 and output the verification result as the second control signal. S6. Dynamic correction: If the second control signal is an abnormal signal, a correction operation is performed according to the encoded data and the abnormality type, and the process returns to step S4 after the correction operation is completed.

[0015] Furthermore, after the core task objective is completed, the following steps are performed: S7. Result convergence: Store the relevant data in the experience base.

[0016] Furthermore, in step S4, if the target strategy adopts a rapid response mode, the constraint control and risk avoidance module is first called to perform a security verification, and then the language organization and expression module is called to generate an emergency report. If the target strategy adopts a step-by-step reasoning mode, it first calls the target parsing and planning module to complete the target parsing, then calls the long short-term memory module to perform memory retrieval, and finally calls the logical reasoning and causal analysis module to perform reasoning. If the target strategy adopts a knowledge-enhanced retrieval mode, it first calls the long short-term memory module to perform memory retrieval, then calls the association generation and creative expansion module to generate candidate solutions, and finally calls the fact verification and security verification module to perform fact verification. If the target strategy adopts a backtracking correction mode, it first calls the interactive feedback and iterative optimization module for interactive feedback, then calls the logical reasoning and causal analysis module for logical reasoning and error diagnosis, and finally calls the constraint control and risk avoidance module to correct the constraints. An embodied intelligent environment-adaptive learning device includes a memory and a processor. The memory stores a computer program, which, when executed by the processor, causes the processor to perform the steps of an embodied intelligent environment-adaptive learning method.

[0017] The beneficial effects of this invention are: Through a six-layer closed-loop hierarchical cognitive architecture, signal interaction, data transmission, and collaborative work are achieved between layers; embodied intelligence can actively monitor, verify in real time, dynamically schedule strategies, and autonomously correct the perception-decision-execution process, thereby improving the environmental adaptability, generalization ability, system reliability, and interpretability of embodied intelligence. Attached Figure Description

[0018] Figure 1 A schematic diagram of a six-layer closed-loop structure for an embodied intelligence architecture; Figure 2 A closed-loop workflow flowchart for an embodied intelligent environment-adaptive learning method; Figure 3 A schematic diagram of the hierarchical state coding mechanism for the reasoning state coding layer; Figure 4 This is a flowchart of Example 2; Figure 5 This is a schematic diagram of the modular structure of an embodied intelligent environmental adaptive learning device.

[0019] In the diagram, 1-Embodied Intelligence Architecture, 101-Core Task Anchoring Layer, 102-Binary Decision and Verification Layer, 103-Dynamic Strategy Scheduling Layer, 104-Basic Cognitive Functional Unit Layer, 105-Inference State Encoding Layer, 106-Dynamic Correction and Reconstruction Layer, 301-Core Task Anchoring Module, 302-Binary Decision Verification Module, 303-Dynamic Strategy Scheduling Module, 304-Basic Cognitive Execution Module, 305-Inference State Encoding Module, 306-Dynamic Correction and Reconstruction Module, and 307-Experience Storage Module. Detailed Implementation

[0020] The following specific examples illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and various details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that, unless otherwise specified, the following embodiments and features described therein can be combined with each other.

[0021] It should be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of the present invention. Therefore, the drawings only show the components related to the present invention and are not drawn according to the actual number, shape and size of the components in the actual implementation. In the actual implementation, the form, quantity and proportion of each component can be arbitrarily changed, and the layout of the components may also be more complex.

[0022] Example 1: like Figures 1 to 3 and Figure 5 As shown, an embodied intelligence architecture 1 includes a core task anchoring layer 101, a binary decision and verification layer 102, a dynamic strategy scheduling layer 103, a basic cognitive function unit layer 104, a reasoning state encoding layer 105, and a dynamic correction and reconstruction layer 106, which are connected from bottom to top and form a closed-loop control. Core Task Anchoring Layer: Analyzes the task to be processed and combines multimodal perception data to identify the core task objectives and constraints; The constraints of the core task anchoring layer include time limits, security boundaries, and energy consumption constraints; multimodal perception data includes visual data, speech data, and environmental semantic map data.

[0023] The core task anchoring layer uses natural language understanding to parse the tasks to be processed.

[0024] Natural Language Understanding (NLU): Employs intent recognition and entity extraction techniques based on pre-trained models such as BERT to parse user instructions waiting to be processed into structured task objectives.

[0025] For multimodal perception data such as vision and speech, the core task anchoring layer uses a cross-modal attention mechanism to fuse features and then lock the core task objectives and constraints.

[0026] The core task anchoring layer is used to parse user instructions or environmental tasks waiting to be processed. It combines multimodal perception data such as vision, speech, and environmental semantic maps to lock in the core task objectives and constraints (time limits, safety boundaries, energy consumption constraints), serving as a benchmark reference for the entire environmental adaptive learning process and ensuring that subsequent cognitive processes do not deviate from the core task.

[0027] Binary decision-making and verification layer: performs binary judgment on whether the content to be verified is abnormal or not; The content to be verified includes perception data, behavioral decisions, action instructions, and core task objectives and constraints.

[0028] The built-in verification rule base of the binary decision and verification layer includes the perceived data authenticity verification rule R1, the behavior security assessment rule R2, and the physical constraint verification rule R3.

[0029] The perceptual data authenticity verification rule R1 is as follows: calculate the reconstruction error of the image input from the visual sensor through a self-supervised reconstruction model, and determine whether it is abnormal based on the reconstruction error.

[0030] If the reconstruction error exceeds the dynamic threshold θ1, a sensing anomaly is determined, and an anomaly signal is output.

[0031] The dynamic threshold θ1 is twice the historical error average; in this embodiment, it is twice the moving average of the reconstruction errors of the past 50 frames. During the system cold start phase (the first 50 frames), the dynamic threshold θ1 is initialized to a fixed value of 0.15.

[0032] The behavioral safety assessment rule R2 determines whether an abnormality exists based on whether obstacles appear within a certain distance range of the planned path.

[0033] Specifically, the planned motion trajectory is discretized into a preset number of path points, and the distance d between each path point and the nearest obstacle is calculated. i If there exists any d i If the value is less than the safety threshold safety_margin, the behavior is deemed unsafe and an abnormal signal is output.

[0034] The default value for the safety threshold, safety_margin, is 0.2m, which can be dynamically adjusted according to the task scenario. The preset number of trajectory discretizations is 10.

[0035] The physical constraint verification rule R3 is used to determine whether the joint angle of the generated joint angle command violates the physical constraints and is therefore abnormal.

[0036] Specifically, the generated joint angle commands are verified. If the joint angle exceeds the preset joint movement range, it is determined that the physical constraints are violated, and an abnormal signal is immediately output and a protective shutdown is triggered.

[0037] The verification rule base supports dynamic expansion. The expansion mechanism of the verification rule base is as follows: for abnormal situations that occur more than a predetermined number of times in the experience base, candidate rules are generated and added to the verification rule base after manual confirmation.

[0038] The binary decision-making and verification layer constructs a real-time verification mechanism that runs through the entire perception-decision-execution process. It has a built-in verification rule base and performs binary judgments on each step of perception data, behavioral decisions, and action commands, outputting "normal / abnormal" control signals. At the same time, it introduces physical constraint verification and environmental uncertainty assessment as the safety master controller of the embodied intelligence system.

[0039] Dynamic policy scheduling layer: Selects and schedules target policies from preset policies; The preset strategy is a preset environmental adaptive cognitive reasoning strategy.

[0040] The preset strategies include a fast response mode, a step-by-step reasoning mode, a knowledge-enhanced retrieval mode, and a backtracking correction mode.

[0041] How each preset strategy works:

[0042] The triggering conditions, execution logic, and output results of the four preset strategies are as follows: Quick Response Mode: Triggering condition: The binary decision-making layer outputs an "emergency" signal (such as a collision warning); Execution logic: Skip complex reasoning and directly call the preset security action library; Output: Emergency obstacle avoidance command.

[0043] Step-by-step reasoning mode: Triggering condition: High task complexity (such as multi-step navigation + grabbing); Execution logic: Decompose the task into a sequence of subtasks and call the target parsing and logical reasoning modules step by step; Output: List of subtasks and their execution order.

[0044] Knowledge-enhanced retrieval mode: Triggering condition: Insufficient local knowledge (such as encountering an unknown object); Execution logic: Trigger experience base retrieval, match historical state codes, and call solutions for similar scenarios; Output: Retrieved experience templates.

[0045] Backtracking correction mode: Triggering conditions: Task failure or multiple verification errors; Execution logic: Backtrack to the most recent successful state and replan the subsequent path; Output: The corrected policy sequence.

[0046] The scheduling rules for the four preset strategies can be dynamically optimized through reinforcement learning, i.e., a reinforcement learning-driven strategy selection mechanism.

[0047] Basic cognitive functional unit layer: includes multiple basic cognitive modules, which perform corresponding basic operations; The basic cognitive module includes a goal analysis and planning module, a long and short-term memory module, a logical reasoning and causal analysis module, an association generation and creative expansion module, a fact verification and security verification module, a language organization and expression module, a constraint control and risk avoidance module, and an interactive feedback and iterative optimization module.

[0048] The basic cognitive functional unit layer supports multimodal perception fusion of vision, hearing, touch, and force. After fusing multi-source perception data, it outputs a unified perception result.

[0049] These eight basic cognitive modules serve as atomic units for constructing adaptive behaviors in complex environments; they also support the fusion of multimodal perception, including vision, hearing, touch, and force, thereby improving the comprehensiveness and accuracy of environmental perception.

[0050] The basic cognitive module will now be briefly introduced.

[0051] The language organization and expression module can be implemented using any of the following existing technologies: Option 1 (Template Filling): Pre-set multiple expression templates, fill in the key information in the reasoning results (such as target location, path length, task status) into the templates, and generate natural language descriptions.

[0052] The template can be designed as follows: Success report: "[Task objective] completed, time taken: [time] seconds". Anomaly report: "[Anomaly type] detected, [correction operation] being performed." Option 2 (End-to-End Generation): Based on a pre-trained language model (such as GPT, LLaMA), internal reasoning trajectories and state encoding sequences are used as input prompts to generate interpretable natural language descriptions. For specific implementation details, please refer to existing literature (such as "Language Models as Robot Controllers", 2023).

[0053] The workflow of the interactive feedback and iterative optimization module is as follows: 1. Feedback reception: Receive explicit feedback from users (such as "made a mistake" or "great") or implicit feedback from the environment (such as task success / failure, collision occurred).

[0054] 2. Feedback Analysis: If it is human language feedback, call the reverse function of the "Language Organization and Expression Module" to analyze the feedback intent; if it is environmental feedback, extract key indicators (such as success rate, time consumption, and energy consumption).

[0055] 3. Experience Update: The current task's status encoding sequence, success / failure labels, and correction operation records are packaged into an experience sample and stored in the experience storage module as a data source for the "knowledge-enhanced retrieval mode". 4. Strategy Optimization: When a certain number of experience samples are accumulated, the meta-reinforcement learning mechanism is triggered to update the weights of the dynamic policy scheduling layer.

[0056] Specific optimization algorithms can be adopted from existing reinforcement learning algorithms such as PPO and SAC.

[0057] 5. Personalized adaptation: For users who interact frequently, the language style and decision preferences of subsequent interactions can be optimized by fine-tuning the model parameters.

[0058] Inference State Encoding Layer: The encoded data is obtained by performing layered state encoding on the encoded data source; The sources of encoded information in the reasoning state encoding layer include currently active basic cognitive modules, embodied intelligence's own state data, and environmental data.

[0059] The encoded data includes joint state codes and historical context vectors.

[0060] The hierarchical state coding mechanism of the inference state coding layer includes three layers; The first layer is basic state coding: coding the currently active basic cognitive modules; Specifically, the module status codes are generated by combining the numbers of the currently active basic cognitive modules. M id ; In embodied intelligence's adaptive learning, a single basic cognitive module can hardly simultaneously meet the demands for rapid response and deep reasoning. For example, relying solely on the logical reasoning and causal analysis module (u3) may lead to slow responses, while relying solely on the constraint control and risk avoidance module (u7) may lack global planning capabilities. Theoretical analysis and simulation deduction reveal that a dual-module collaborative mechanism can achieve the optimal balance between response speed and reasoning depth.

[0061] The basic cognitive functional unit layer adopts a dual-module collaborative mechanism. At this time, there are two active basic cognitive modules.

[0062] For the case of two currently active basic cognitive modules, the module status code is... M id The calculation formula is: M id =(u_i×8+u_j) mod 64; In the formula, u_i and u_j are the numbers of two currently active basic cognitive modules.

[0063] The dual-module collaborative mechanism for generating module status codes for combinations of two currently active basic cognitive modules (a total of 64 states) has the following advantages: 1. Coding efficiency: The eight basic cognitive modules are combined in pairs, resulting in a moderate coding space and low computational overhead; 2. Expressive ability: 64 states are sufficient to cover the vast majority of cognitive scenarios; 3. Scalability: The collaboration of three or more basic cognitive modules can be expressed through the historical state encoding (LSTM) of the third layer without increasing the current encoding dimension.

[0064] Therefore, this embodiment selects a dual-module approach as the basic unit for basic state coding, which ensures the real-time performance of the system and provides sufficient adaptive capability.

[0065] In the actual operation of embodied intelligence, there may be scenarios where three or more basic cognitive modules are active simultaneously, such as multimodal perception fusion (simultaneous processing of vision, hearing, and touch) or parallel computing in complex reasoning. The embodied intelligence architecture design fully considers this requirement, supporting parallel calls and data fusion between modules in the basic cognitive functional unit layer.

[0066] For cases where three or more currently active basic cognitive modules are used: 1) Select two basic cognitive modules as the dominant modules and calculate the module status codes for the dominant modules. M id ; M id=(u_i×8+u_j) mod 64; The two core basic cognitive modules that currently dominate the cognitive process are identified. The selection of these two basic cognitive modules as the dominant modules is based on a balance between coding efficiency and expressive power, and the 64 states already cover most decision-making scenarios.

[0067] 2) Generate an auxiliary module bitmap and append the auxiliary module bitmap to the module status code. M id ; Auxiliary module bitmap: bitmap=[b1,b2,...,b8]; The auxiliary module bitmap is 8 bits, with each bit corresponding to a basic cognitive module (1 indicates that the basic cognitive module is active at the current time step, and 0 indicates that it is inactive). The bitmap can be represented by an 8-bit binary number (value range 0-255), and together with the dominant module, it constitutes a complete module state description.

[0068] Mechanisms to ensure the uniqueness of codes:

[0069] The encoding design in this embodiment guarantees the mathematical uniqueness of the status code: 1) M id (0-63) and bitmap (0-255) are combined by concatenation rather than mixed operations, and different combinations will necessarily correspond to different complete codes; 2) Environmental data (pose, progress, label) further increases the distinguishability of the coding.

[0070] Even in extreme situations in a real-world system (such as hardware errors causing code duplication), the system can efficiently resolve the issue through the following mechanisms: 1. Real-time detection: The binary decision and verification layer can be extended with the R4 encoding verification rule to detect encoding conflicts and anomalies; 2. Dynamic correction: Fine-tuning of encoding parameters at the parameter level, or switching to an alternative encoding scheme at the structure level; 3. Meta-reinforcement learning: Incorporating conflict resolution strategies into the learning framework for continuous optimization.

[0071] The implicit representation of the third-level historical state encoding:

[0072] Furthermore, the third-layer historical state encoding (LSTM encoding the joint state encoding sequence of the most recent N time steps) can also indirectly express information about multi-module collaboration. For example, when a system requires the collaboration of three basic cognitive modules, they may be active sequentially in time. Historical state encoding captures this temporal relationship, providing context for multi-module collaboration in decision-making.

[0073] Therefore, even without using auxiliary module bitmap extensions, the system can still perceive the timing patterns of multi-module collaboration through historical context vectors, ensuring the integrity of basic functions; while the auxiliary module bitmap provides a more refined real-time state perception capability and can be used as a preferred implementation scheme.

[0074] The second layer is the encoding of the environment and the self-state: encoding the self-state data and environmental data of the embodied intelligence; The embodied intelligence's own state data includes the embodied intelligence's pose data and task progress; the environmental data is environmental semantic information, that is, environmental semantic tags.

[0075] Environment and self-state coding normalizes the self-state data and environment data of embodied intelligence.

[0076] The encoding result of the basic state encoding (i.e., the module state code) M id The joint state code is generated by concatenating the encoding results of the environment and its own state codes. When there are two currently active basic cognitive modules, the joint state encoding is: S id =[ M id ||pose_x_norm||pose_y_norm||theta_norm||task_progress||env_label]; Among them, pose_x_norm, pose_y_norm, and theta_norm are the normalized data of the x-axis, y-axis, and angle data in the pose data of embodied intelligence; task_progress is the normalized data of the task progress; and env_label is the normalized data of the environmental semantic label.

[0077] For cases where three or more currently active basic cognitive modules are used: The joint state coding is as follows: S id =[ M id ||bitmap||pose_x_norm||pose_y_norm||theta_norm||task_progress||env_label]; The coded components of the auxiliary module bitmap and the original module status code have equal weight and together serve as input features for policy scheduling and correction decisions.

[0078] Example: If the current leading modules are u3 (logical reasoning and causal analysis module) and u5 (fact verification and security verification module), and u2 (long short-term memory module) is also assisting, then M id =29, bitmap=00000100 (decimal 4), joint state encoding is [29,4,0.52,0.38,0.25,0.3,2] (robot current pose (x=5.2m, y=3.8m, θ=90°, current task progress (0.3), env_label=2).

[0079] The pose data of the embodied intelligence is smoothed by Kalman filtering, and the environmental semantic labels are generated in real time by a pre-trained semantic segmentation model.

[0080] The third layer is historical state encoding: the sequence of joint state encodings of the most recent N time steps is encoded to obtain the historical context vector; The third layer's historical state encoding uses an LSTM network to jointly encode the state of the most recent N time steps. S id The sequence is encoded to generate a historical context vector. H t It is used to capture the temporal characteristics of environmental changes.

[0081] The parameters of the LSTM network for the inference state coding layer are: time step N=10, hidden layer dimension d. hidden =128, output historical context vector H t ∈R128; The LSTM network is randomly initialized and fine-tuned online using historical successful trajectories after each task. The loss function is the mean square error of predicting the next state encoding.

[0082] The reasoning state coding layer adopts a three-layer coding mechanism of basic state coding, environment and self-state coding, and historical state coding. It encodes the current reasoning and behavior process as a unique "environment-self" state identifier, while capturing the temporal characteristics of environmental changes, providing a complete decision basis for strategy scheduling and correction.

[0083] The dynamic policy scheduling layer introduces a reinforcement learning-driven policy selection mechanism, which automatically optimizes policy scheduling rules based on encoded data and feedback data.

[0084] The feedback data consists of real-time data returned by environmental and internal sensors after the embodied intelligence performs an action. This feedback data includes task success rate, energy consumption data, and safety scores.

[0085] The modeling of the reinforcement learning-driven policy selection mechanism is as follows: State space: Joint state encoding in encoded dataS id With historical context vectors H t The concatenated vector; Action space: Discrete selection of multiple preset strategies (fast response mode, step-by-step reasoning mode, knowledge-enhanced retrieval mode, backtracking correction mode); Reward function: R = w 1 R success + w 2 R safety + w 3 R energy ,in R success This is the success indicator for the task (success = 1, failure = 0). R safety The normalized value of the minimum safe distance. R energy The reciprocal of the normalized energy consumption value; weight w 1 = 0.5, w 2 = 0.3, w 3 = 0.2; Learning algorithm: DQN (Deep Q-Network) is used for policy optimization, the experience replay pool capacity is 10000, and the target network update frequency is 200 steps.

[0086] The dynamic strategy scheduling layer dynamically selects and schedules the target strategy from four preset strategies—fast response mode, step-by-step reasoning mode, knowledge-enhanced retrieval mode, and backtracking correction mode—based on the control signals from the binary decision and verification layer. At the same time, a reinforcement learning-driven strategy selection mechanism is introduced to automatically optimize the strategy scheduling rules based on task success rate, energy consumption, and safety score feedback from the environment.

[0087] Dynamic correction and reconstruction layer: Performs correction operations; The correction operations include parameter-level fine-tuning and structure-level reconstruction; Parameter-level fine-tuning: Adjust the model weights of the perception module and the path planning parameters of the decision module to adapt to the perception bias and decision requirements of the current environment; Structural-level restructuring: Switching the combination of behavioral strategies, or restructuring the combination of basic cognitive modules, and increasing the module calls of the target perception modality; The triggering conditions for structural-level reconstruction are: the abnormal signal is still not eliminated after a preset number of parameter-level fine-tunings; or the duration of a single abnormality exceeds the set time.

[0088] The dynamic correction and reconstruction layer introduces a meta-reinforcement learning mechanism to learn and optimize the effects of correction operations, thereby improving the efficiency and autonomy of environmental adaptive learning.

[0089] When the binary decision and verification layer determines that the perception or decision-making process is abnormal or the behavior fails, it triggers the dynamic correction and reconstruction layer to perform correction operations. The correction operations include parameter-level fine-tuning (adjusting the parameters of the perception / decision module) and structure-level reconstruction (switching the strategy / reconstruction module combination). At the same time, a meta-reinforcement learning mechanism is introduced to enable the correction operations to learn and optimize autonomously.

[0090] The embodied intelligence architecture is implemented based on a hierarchical state machine driven autonomous cognitive architecture (HSM-ACA).

[0091] The six-layer structure of the embodied intelligence architecture constitutes a complete closed-loop control workflow, enabling the embodied intelligence architecture to form closed-loop control.

[0092] The embodied intelligence architecture is an autonomous cognitive control system (HSM-ACA) driven by a hierarchical state machine.

[0093] An embodied intelligence environment-adaptive learning method, applied to an embodied intelligence architecture, includes the following steps: S1. Target Anchoring: Based on the tasks to be processed, identify the core task objectives and constraints; Step S1 is completed through the core task anchoring layer, and the core task objectives and constraints obtained therefrom serve as a benchmark reference system for environmental adaptive learning. S2, Pre-verification: Verify the core task objectives and constraints, and output the verification result as the first control signal; Step S2 is completed through a binary decision and verification layer.

[0094] The binary decision-making and verification layer verifies the legality, executability, and security constraints of the core task objectives and constraints, and outputs the first control signal; S3. Strategy routing: Based on the first control signal, combined with the core task objective and the acquired environmental status data, select and schedule the target strategy from the preset strategies; Step S3 involves dynamically selecting and scheduling the target strategy; Step S3 is completed through the dynamic strategy scheduling layer. Based on the first control signal, combined with the task characteristics of the core task objective and the acquired environmental state data, the target inference strategy is dynamically selected and scheduled. S4. Module Execution and State Encoding: Execute the corresponding basic cognitive modules according to the target strategy, and perform hierarchical state encoding operations to obtain encoded data; The module execution in step S4 is completed through the basic cognitive functional unit layer. Specifically, the basic cognitive functional unit layer calls the corresponding basic cognitive module to perform environmental perception, task planning, and behavioral decision-making operations, and outputs perception data, decision instructions, and action signals.

[0095] The calling relationship between the target strategy and the basic cognitive module follows these rules: 1. Each strategy can invoke one or more basic cognitive modules; 2. The module call order is determined by the strategy logic and can be sequential, parallel, or conditional branching; 3. Modules interact with each other through standardized input / output interfaces.

[0096] Typical call relationships for the four target strategies: Strategy: Rapid Response Calling modules: u6 Language Organization and Expression Module (Language Expression), u7 Constraint Control and Risk Avoidance Module (Constraint Control); Call order: u7 constraint control and risk avoidance module (security verification) → u6 language organization and expression module (emergency report); If the target strategy adopts a rapid response mode, it first calls the constraint control and risk avoidance module to perform security verification, and then calls the language organization and expression module to generate an emergency report.

[0097] Strategy: Step-by-step reasoning:

[0098] Modules called: u1 Goal Analysis and Planning Module (Goal Analysis), u3 Logical Reasoning and Causal Analysis Module (Logical Reasoning), u2 Long Short-Term Memory Module (Memory Retrieval);

[0099] Calling order: u1 Goal analysis and planning module (task decomposition) → u2 Long Short-Term Memory module (experience retrieval) → u3 Logical reasoning and causal analysis module (reasoning);

[0100] The target strategy adopts a step-by-step reasoning mode, which first calls the target parsing and planning module to complete the target parsing, then calls the long short-term memory module to perform memory retrieval, and finally calls the logical reasoning and causal analysis module to perform reasoning.

[0101] Strategy: Knowledge-enhanced retrieval

[0102] Calling modules: u2 Long Short-Term Memory Module (Memory Retrieval), u4 Association Generation and Creative Extension Module (Association Generation), u5 Fact Verification and Security Verification Module (Fact Verification);

[0103] Calling order: u2 Long Short-Term Memory Module (retrieval) → u4 Association Generation and Creative Extension Module (generating candidates) → u5 Fact Check and Security Verification Module (verification);

[0104] If the target strategy adopts a knowledge-enhanced retrieval mode, it first calls the long short-term memory module for memory retrieval, then calls the association generation and creative expansion module for association generation to obtain candidate solutions, and finally calls the fact verification and security verification module for fact verification.

[0105] Strategy: Backtracking and Correction

[0106] Modules called: u8 Interactive Feedback and Iterative Optimization Module (Interactive Feedback), u3 Logical Reasoning and Causal Analysis Module (Logical Reasoning), u7 Constraint Control and Risk Avoidance Module (Constraint Control);

[0107] Call order: u8 interactive feedback and iterative optimization module (receiving feedback) → u3 logical reasoning and causal analysis module (diagnosing errors) → u7 constraint control and risk avoidance module (correcting constraints);

[0108] If the target strategy adopts a backtracking correction mode, it first calls the interactive feedback and iterative optimization module for interactive feedback, then calls the logical reasoning and causal analysis module for logical reasoning and error diagnosis, and finally calls the constraint control and risk avoidance module to correct the constraints.

[0109] Dynamic adaptation: In actual execution, the strategy scheduling layer adjusts the calling order based on real-time feedback. For example, if a step in "step-by-step reasoning" fails, "knowledge enhancement" can be temporarily inserted to retrieve similar experiences.

[0110] Regardless of the calling relationship, the currently active combination of basic cognitive modules will be encoded into a unique state identifier in real time by the "inference state encoding layer", forming a traceable cognitive trajectory.

[0111] The hierarchical state encoding operation in step S4 is completed through the inference state encoding layer. The encoded data serves as the basis for decision-making and correction.

[0112] S5. Real-time verification: Perform real-time verification on the data output by the basic cognitive module in step S4 and output the verification result as the second control signal. The data output by the basic cognitive module in step S4 includes perception data, decision instructions, and action signals.

[0113] Step S5 is completed through a binary decision and verification layer, which performs real-time verification of the module execution in step S4 throughout the entire process. S6. Dynamic correction: If the second control signal is an abnormal signal, a correction operation is performed according to the encoded data and the abnormality type, and the process returns to step S4 after the correction operation is completed.

[0114] The correction operation in step S6 is completed through the dynamic correction and reconstruction layer. The dynamic correction and reconstruction layer performs parameter-level fine-tuning or structure-level reconstruction correction operations based on the joint state encoding, historical context vector and anomaly type in the encoded data. After correction, it returns to step S4 to be executed again. If the second control signal is a normal signal, the current operation continues to be executed. After the core mission objective is completed, perform the following steps: S7. Result convergence: Store the relevant data in the experience base.

[0115] After completing the core task objective, output stable behavioral results consistent with the core task objective, and store the state encoding sequence of this task, correction operation records and successful experiences into the experience base to complete one environmental adaptive learning iteration.

[0116] An embodied intelligent environment-adaptive learning device includes a memory and a processor. The memory stores a computer program, which, when executed by the processor, causes the processor to perform the steps of an embodied intelligent environment-adaptive learning method.

[0117] The computer program includes: a core task anchoring module 301, a binary decision verification module 302, a dynamic strategy scheduling module 303, a basic cognitive execution module 304, a reasoning state encoding module 305, and a dynamic correction and reconstruction module 306. Each module corresponds to a six-layer structure of an embodied intelligence architecture, and they achieve signal interaction and data transmission through a data bus to form a closed-loop control system.

[0118] The core task anchoring module corresponds to the core task anchoring layer of an embodied intelligence architecture. It is used to parse user instructions or environmental tasks and combine multimodal perception data to lock the core task objectives and constraints. The binary decision verification module corresponds to a binary decision and verification layer of an embodied intelligent architecture. It has a built-in verification rule library and is used to perform pre-verification and real-time verification of tasks, perception data, decision instructions and action signals, and output control signals. The dynamic strategy scheduling module corresponds to a dynamic strategy scheduling layer of an embodied intelligence architecture. It is used to dynamically select and schedule target inference strategies based on control signals, task characteristics and environmental states, and optimize strategy scheduling rules through reinforcement learning. The basic cognitive execution module corresponds to the basic cognitive functional unit layer of an embodied intelligence architecture, which contains eight basic cognitive sub-modules for performing operations such as multimodal perception fusion, task planning, and behavioral decision-making. The reasoning state coding module corresponds to a reasoning state coding layer of an embodied intelligence architecture, which is used to generate module state codes, joint state codes and historical context vectors through a hierarchical coding mechanism. The eight basic cognitive sub-modules of the basic cognitive execution module correspond one-to-one with the core capabilities of embodied intelligence, respectively realizing the functions of task decomposition, environmental map construction, environmental causal relationship inference, behavior innovation in unknown scenarios, perception data verification, human-computer interaction, risk avoidance, and strategy iteration and optimization.

[0119] The dynamic correction and reconstruction module corresponds to a dynamic correction and reconstruction layer of an embodied intelligent architecture. It is used to perform parameter-level fine-tuning or structure-level reconstruction based on abnormal signals, and optimizes the correction operation through meta-reinforcement learning. The modules interact with each other and transmit data via a data bus, forming a closed-loop control system.

[0120] The computer program also includes an experience storage module 307, which stores the task state encoding sequence of embodied intelligence, correction operation records and successful experiences, providing data support for knowledge-enhanced retrieval mode; Data structure and retrieval mechanism of experience storage module 307: The experience storage module 307 uses a graph database (such as Neo4j) for storage. Each node represents a joint state code Sid, and node attributes include task progress, environmental semantic labels, and execution timestamps. Edges represent state transitions, and edge attributes include the executed action, correction operation type, and task success flag. During retrieval, the system uses the current Sid as the query vector and employs cosine similarity to retrieve the K=5 nodes with the highest similarity in historical experiences, returning their corresponding successful experiences and correction records. When the storage capacity exceeds the limit, an LRU (Least Recently Used) strategy is used to evict infrequently accessed historical nodes.

[0121] The basic cognitive execution module contains eight basic cognitive sub-modules, each implementing a corresponding basic cognitive function; the experience storage module is used to store task status encoding sequences, correction operation records, and successful experiences; the device supports plug-and-play integration of new perceptual modalities and new basic cognitive sub-modules, and can be adapted to various embodied intelligent carriers such as autonomous mobile robots, service robots, and humanoid robots.

[0122] The embodied intelligent environment-adaptive learning device adopts a modular design; it supports plug-and-play integration of new perceptual modalities and new basic cognitive sub-modules, and can be adapted to various embodied intelligent carriers such as autonomous mobile robots, service robots, and humanoid robots.

[0123] Through a six-layer closed-loop hierarchical cognitive architecture, signal interaction, data transmission, and collaborative work are achieved between layers. This enables embodied intelligence to proactively monitor, verify in real time, dynamically schedule strategies, and autonomously correct the perception-decision-execution process, thereby improving the environmental adaptability, generalization ability, system reliability, and interpretability of embodied intelligence. At the same time, the modularity and scalability of the architecture are guaranteed.

[0124] Simultaneously, by combining hierarchical state coding, two-level dynamic correction, and reinforcement learning-driven strategy optimization mechanisms, it significantly improves the environmental adaptability, generalization ability, system reliability, and learning efficiency of embodied intelligence. Moreover, its architecture is highly modular and scalable, and can be adapted to various embodied intelligence carriers such as autonomous mobile robots, service robots, and humanoid robots. It is suitable for high-reliability scenarios such as medical care, industrial collaboration, and indoor and outdoor navigation. It solves the technical problems of existing embodied intelligence systems, such as black-box perception and decision-making, rigid strategies, insufficient metacognitive ability, and poor environmental adaptability.

[0125] By applying an embodied intelligence architecture to embodied intelligence environment adaptive learning, and through core technologies such as a six-layer closed-loop hierarchical design, real-time binary verification, hierarchical state encoding, and dual-level dynamic correction, the following significant beneficial effects are achieved: 1. Significantly improves environmental adaptability and generalization ability: Through dynamic policy scheduling and full-state perception of "environment-self-history", embodied intelligence can actively adapt to different environments such as indoor / outdoor and static / dynamic. It is expected to improve the success rate of tasks in dynamic environments by more than 30% (compared to traditional end-to-end reinforcement learning methods, the dynamic policy scheduling and full-state perception mechanism of this invention can adapt to environmental changes more quickly, with a theoretical gain of up to 30%).

[0126] 2. Enhanced System Reliability and Security: The binary decision-making and verification layer's triple-rule real-time verification and physical constraint control effectively avoid dangerous behaviors such as collisions and overloads, improving the security of embodied intelligence applications in high-reliability scenarios such as medical care and industrial collaboration. It is expected to improve application security in high-reliability scenarios by more than 50% (through the triple real-time verification of the binary decision-making and verification layer, dangerous behaviors such as collisions and overloads can be effectively avoided, with an expected security improvement of over 50%).

[0127] 3. Accelerate adaptive learning efficiency: The metacognitive closed-loop architecture and the dynamic correction mechanism driven by meta-reinforcement learning enable embodied intelligence to quickly learn from failures and correct its strategies, improving the learning convergence speed and significantly reducing the learning cost.

[0128] 4. Achieve explainable perception and decision-making: The hierarchical mechanism of reasoning state coding and the modular design of cognitive units make the environmental perception and behavioral decision-making process of embodied intelligence explainable and traceable, which facilitates system debugging and optimization.

[0129] 5. Excellent scalability and compatibility: The architecture and devices adopt a modular design, supporting plug-and-play integration of new sensing modalities and new cognitive modules. It can be directly adapted to various existing embodied intelligent carriers, significantly reducing system upgrade and iteration costs.

[0130] An embodied intelligence architecture and environmentally adaptive learning method and device can be seamlessly integrated with existing embodied intelligence sensors, actuators, and control systems without requiring large-scale modifications to existing equipment, and has good engineering practicality and industrialization prospects.

[0131] An embodied intelligent architecture and environmentally adaptive learning method and device can be directly applied to the research and development and upgrading of various embodied intelligent carriers such as autonomous mobile robots, service robots, humanoid robots, and industrial collaborative robots. It significantly improves their task execution capabilities and environmental adaptability in unstructured and dynamic environments. It is applicable to multiple fields such as medical care, industrial manufacturing, intelligent buildings, and smart cities, and has significant industrial applicability and industrialization value.

[0132] Through inter-level collaborative work and signal interaction, embodied intelligence enables proactive monitoring, real-time verification, dynamic policy scheduling, and autonomous correction of the perception-decision-execution process. Simultaneously, by combining hierarchical state coding, reinforcement learning-driven policy optimization, and a two-level dynamic correction mechanism driven by meta-reinforcement learning, it addresses the problems of black-box perception and decision-making, policy rigidity, and insufficient metacognitive capabilities in existing embodied intelligence systems. This significantly improves the environmental adaptability, generalization ability, system reliability, and learning efficiency of embodied intelligence. Furthermore, its modular architecture and good scalability allow it to adapt to various embodied intelligence carriers, making it suitable for high-reliability scenarios and possessing excellent industrial applicability.

[0133] Example 2: like Figures 1 to 5 As shown, Embodiment 2 has all the features of Embodiment 1, except that: This embodiment uses an embodied intelligent architecture and environmental adaptive learning method and device from Embodiment 1 to complete autonomous navigation and obstacle avoidance in dynamic crowd environments.

[0134] This embodiment uses the autonomous navigation and obstacle avoidance of a service robot in a dynamic crowd environment as an example to explain in detail the working process of an embodied intelligence architecture.

[0135] Task settings: The service robot navigates from starting point A to destination point B, passing through a hall area with a large number of dynamic crowds. It needs to complete the navigation task within 30 seconds, and the safety boundary with pedestrians must be no less than 0.3m to avoid collisions.

[0136] Specific execution steps:

[0137] 1. Target Anchoring (Core Task Anchoring Layer): The robot receives the voice command "Navigate to target point B", confirms the physical coordinates of target point B through visual recognition, and locks the core target vector t0={target coordinates (x,y), time constraint 30s, safety boundary 0.3m, energy consumption constraint ≤50W} as the benchmark for this navigation task.

[0138] 2. Pre-verification (binary decision and verification layer): Verify whether target point B is within the robot's navigable area based on the environmental semantic map, and verify whether the 30-second time constraint is reasonable in combination with the robot's motion performance. If all verifications meet the requirements, the first control signal is output as a normal signal (+1).

[0139] Explanation of the dynamic expansion mechanism of the validation rule base: To enhance the system's adaptability to new anomaly patterns, the binary decision and verification layer supports dynamic expansion of the rule base. Specifically, when the frequency of occurrence of the same anomaly pattern (such as similar sensor noise or similar collision risk) in the experience storage module exceeds a preset threshold M=5, the system automatically extracts the feature vector of that anomaly pattern, generates a candidate rule, and prompts the operator for confirmation via a human-computer interaction interface. After manual confirmation, the rule is added to the rule base and can be directly used for real-time verification. This mechanism enables the system to continuously learn and evolve.

[0140] 3. Strategy Routing (Dynamic Strategy Scheduling Layer): The system detects that the current environment is a dynamic, densely populated area through visual sensors. Combined with the first control signal, it selects a combined target strategy of "step-by-step reasoning mode + knowledge-enhanced retrieval mode". The step-by-step reasoning mode is used to break down the navigation task into multiple sub-tasks, and the knowledge-enhanced retrieval mode is used to retrieve path templates of historically successful crowd crossings from the experience storage module.

[0141] 4. Module Execution (Basic Cognitive Functional Unit Layer): Based on the combined goal strategy, the following basic cognitive modules are invoked sequentially: - Target Analysis and Planning Module: Decomposes the navigation task into three sub-tasks: "Starting Point A → Corridor → Lobby → Target Point B", and initially plans the reference path for each segment; - Long Short-Term Memory Module: Retrieves historical path templates for crowd navigation and crowd trajectory prediction patterns from the experience storage module; -Logical Reasoning and Causal Analysis Module: Combines the current location and movement speed of the crowd to predict the crowd's trajectory in the next 5 seconds and marks high-risk movement areas; -Association Generation and Creative Extension Module: Generates 3 alternative local paths for the lobby area based on the crowd trajectory prediction results; - Fact Check and Security Verification Module: Performs a preliminary collision risk assessment on 3 alternative paths and selects 2 low-risk paths; - Constraint control and risk avoidance module: Verifies the selected low-risk paths to ensure that the paths do not exceed the robot's drivable area and meet energy consumption constraints.

[0142] The final output includes the execution path and robot motion commands (wheel speed, steering angle) for this navigation.

[0143] 5. State Coding (Inference State Coding Layer): -Basic status coding: The currently active basic cognitive modules are the logical reasoning and causal analysis module (u3) and the fact verification and security verification module (u5). The module status code M_id=29 is generated by the formula M_id=(3×8+5)mod64. -Environment-Self-State Encoding: The robot's current pose (x=5.2m, y=3.8m, θ=90°) is smoothed and normalized using Kalman filtering. Combined with the current task progress (0.3), environmental semantic labels (crowd_density=high), and module status codes, a joint state code is generated. S id =[29||0.52||0.38||0.5||0.3||high]; -Historical state encoding: Using an LSTM network to encode the most recent 10 time steps S id The sequence is encoded to generate a historical context vector. H t To capture the changing trends in crowd density in the lobby area.

[0144] Mechanism for generating and updating environmental semantic tags: In this embodiment, the environmental semantic labels are generated in real time by the pre-trained semantic segmentation model DeepLabV3+. This model is pre-trained on the Cityscapes dataset and can be fine-tuned through online learning during task execution. Specifically, when the binary decision and verification layer determines that the perceived data is real and reliable, the frame image and its semantic segmentation result are used as training samples, and the model is updated online with a learning rate of 0.001, allowing the model to gradually adapt to the visual features of the target deployment environment. The fine-tuning process is executed in a background thread and does not affect real-time inference.

[0145] 6. Real-time Verification (Binary Decision and Verification Layer): As the robot begins to move according to the execution path, the binary decision and verification layer performs real-time triple verification of the perception data and motion commands for each step. -Perception data authenticity verification rule (R1): The image input from the visual sensor is calculated by the self-supervised reconstruction model, and the reconstruction error is 0.08, which is less than the dynamic threshold θ1 (historical error mean 0.05×2=0.1), so the verification is passed; - Behavioral safety assessment rule (R2): Discretize the real-time motion trajectory into 10 path points, calculate that the distance between each path point and the nearest pedestrian is ≥0.35m, which is greater than the safety threshold of 0.3m, and the verification is passed; -Physical constraint verification rule (R3): The robot's wheel speed and steering angle commands do not exceed the preset motion range, and the energy consumption is 35W≤50W. Verification passed. The second control signal is continuously output as a normal signal (+1).

[0146] Supplementary explanation of the dynamic threshold calculation process: In this embodiment, the first 50 frames of images after the robot starts are used to initialize the historical error mean μ. e Initially, the dynamic threshold θ1 is fixed at 0.15. Starting from frame 51, the system calculates the moving average of the reconstruction error over the past 50 frames in real time. The "reconstruction error of 0.08" mentioned in the perceptual data authenticity verification rule (R1) is less than the dynamic threshold of 0.1. The calculation process is as follows: the average reconstruction error over the past 50 frames is 0.05, so θ1 = 2 × 0.05 = 0.1, therefore the verification passes.

[0147] 7. Dynamic Correction (Dynamic Correction and Reconstruction Layer): When the robot moves to the center of the hall, a child suddenly runs out from the blind spot. The robot's vision sensor detects the dynamic obstacle, and the real-time distance between the robot and the child is only 0.2m. - The binary decision and verification layer triggers the R2 rule, determines that the behavior is unsafe, and outputs the second control signal as an abnormal signal (-1). - The dynamic correction and reconstruction layer is activated, based on the joint state encoding. S id With historical context vectors H t The anomaly type is determined to be "sudden dynamic obstacle". First, parameter-level fine-tuning is performed: the safety threshold is tightened from 0.3m to 0.5m, the robot wheel speed is reduced by 50%, and a local path for emergency obstacle avoidance is generated. -Since the child is still moving in front of the robot and cannot escape within 5 seconds, the dynamic correction and reconstruction layer performs a structural-level reconstruction: the current strategy is switched from "step-by-step reasoning mode + knowledge-enhanced retrieval mode" to "fast response mode + backtracking correction mode", the original navigation path is paused, the association generation and creative extension module is called to generate a new obstacle avoidance path, and the perception data of the last 3 time steps are backtracked to confirm the obstacle movement trend.

[0148] The structural-level reconstruction in this embodiment is not triggered arbitrarily, but is based on explicit judgment logic: The triggering condition for this embodiment is: when the abnormal signal is not eliminated after three consecutive parameter-level fine-tunings (T=3 times), or when the duration of a single abnormality exceeds Δt=5s, structural-level reconstruction is triggered. Refactoring operation: Pause the current execution path, save the current system state (including active modules, strategy types, and historical context vectors), and switch from the strategy library to the target strategy combination (e.g., switch from "step-by-step reasoning" to "rapid response"). State preservation and hot switching: During the reconstruction process, inactive modules in the basic cognitive functional unit layer remain in standby mode, while intermediate calculation results of active modules are cached in shared memory; after the switch is completed, the new strategy can directly read the cached data to achieve hot switching of modules and avoid system cold start.

[0149] Supplementary explanation of the triggering process for structure-level refactoring: When a child suddenly appears, bringing the distance to within 0.2m, the system initially triggers parameter-level fine-tuning (tightening the safety threshold and reducing speed). Thereafter, the system assesses the abnormal state every 0.5 seconds. After three consecutive fine-tunings (cumulatively 1.5 seconds), if the abnormality persists (the minimum distance remains below 0.5m), the system determines that escape is impossible and triggers structural-level reconstruction. During reconstruction, the system pauses the original path planning, calls the associative generation module to generate a new path, and simultaneously reviews the most recent three frames of perception data to confirm the child's movement trend, ultimately completing obstacle avoidance.

[0150] 8. Result Convergence: After three rounds of parameter-level fine-tuning and one round of structure-level reconstruction, the robot successfully bypassed dynamic crowds and sudden obstacles, ultimately reaching the target point B within 28 seconds, satisfying time, safety, and energy constraints, and completing the navigation task; simultaneously, the results of this task were... S id Sequences, correction operation records, and dynamic obstacle avoidance experience are stored in the experience storage module for reuse in subsequent similar tasks.

[0151] Example of the optimization process of meta-reinforcement learning: Upon completion of this task, the dynamic correction and reconstruction layer records the correction operation sequence: [fine-tuning (threshold tightening) → fine-tuning (deceleration) → fine-tuning (deceleration) → reconstruction (policy switching)]. The meta-reinforcement learning module associates this sequence with the task success flag (success = 1) and updates the value evaluation model of the correction operation. The next time a similar dynamic obstacle-prone scenario occurs, the system will prioritize "reconstruction" rather than multiple fine-tunings, thereby shortening the obstacle avoidance response time. This mechanism enables the correction operation to have autonomous optimization capabilities.

[0152] The embodied intelligence architecture and environmentally adaptive learning method and device are not limited to the above-mentioned dynamic navigation and obstacle avoidance embodiments, but can also be applied to various embodied intelligence task scenarios such as precision assembly of industrial robots, path planning of medical care robots, and dynamic interaction of humanoid robots; the parameters at each level (such as safety threshold, dynamic threshold, time step N of LSTM network, etc.) can be adaptively adjusted according to specific application scenarios, all of which are within the protection scope of this invention.

[0153] The embodiments described above are merely illustrative of specific implementations of the present invention, and while the descriptions are detailed, they should not be construed as limiting the scope of the present invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these modifications and improvements all fall within the scope of protection of the present invention.

Claims

1. An embodied intelligence system, characterized in that, The embodied intelligence system comprises, from bottom to top, a core task anchoring layer, a binary decision-making and verification layer, a dynamic strategy scheduling layer, a basic cognitive functional unit layer, a reasoning state encoding layer, and a dynamic correction and reconstruction layer. The core task anchoring layer: analyzes the task to be processed and combines multimodal perception data to lock in the core task objectives and constraints; The binary decision and verification layer performs a binary judgment on whether the content to be verified is abnormal or not. The built-in verification rule base includes perceptual data authenticity verification rules, behavioral security assessment rules, and physical constraint verification rules. The dynamic policy scheduling layer selects and schedules target policies from preset policies, and introduces a reinforcement learning-driven policy selection mechanism to automatically optimize policy scheduling rules based on encoded data and feedback data. The basic cognitive functional unit layer includes several basic cognitive modules, which perform corresponding basic operations. The reasoning state encoding layer: performs hierarchical state encoding on the encoding data source to obtain encoded data. The hierarchical state encoding includes three layers: the first layer is the basic state encoding that encodes the currently active basic cognitive modules; the second layer is the environment and self-state encoding that encodes the self-state data and environment data of the embodied intelligence; and the third layer is the historical state encoding that encodes the sequence of joint state encodings of the most recent N time steps to obtain the historical context vector. The dynamic correction and reconstruction layer performs correction operations, which include parameter-level fine-tuning and structure-level reconstruction, and introduces a meta-reinforcement learning mechanism to learn and optimize the effect of the correction operations.

2. The embodied intelligent system according to claim 1, characterized in that: The basic cognitive module includes a goal analysis and planning module, a long and short term memory module, a logical reasoning and causal analysis module, an association generation and creative expansion module, a fact verification and security verification module, a language organization and expression module, a constraint control and risk avoidance module, and an interactive feedback and iterative optimization module.

3. The embodied intelligent system according to claim 1, characterized in that: The triggering conditions for the structural-level reconstruction are: the abnormal signal is still not eliminated after a preset number of parameter-level fine-tunings; or the duration of a single abnormality exceeds a set time.

4. An embodied intelligence environment-adaptive learning method, applied to an embodied intelligence system as described in any one of claims 1-3, characterized in that: Includes the following steps: S1. Target Anchoring: Based on the tasks to be processed, identify the core task objectives and constraints; S2, Pre-verification: Verify the core task objectives and constraints, and output the verification result as the first control signal; S3. Strategy routing: Based on the first control signal, combined with the core task objective and the acquired environmental status data, select and schedule the target strategy from the preset strategies; S4. Module Execution and State Encoding: Execute the corresponding basic cognitive modules according to the target strategy, and perform hierarchical state encoding operations to obtain encoded data; S5. Real-time verification: Perform real-time verification on the data output by the basic cognitive module in step S4 and output the verification result as the second control signal. S6. Dynamic correction: If the second control signal is an abnormal signal, a correction operation is performed according to the encoded data and the abnormality type, and the process returns to step S4 after the correction operation is completed.

5. The embodied intelligent environment adaptive learning method according to claim 4, characterized in that: After the core mission objective is completed, perform the following steps: S7. Result convergence: Store the relevant data in the experience base.

6. The embodied intelligent environment adaptive learning method according to claim 4, characterized in that: In step S4, if the target strategy adopts a rapid response mode, the constraint control and risk avoidance module is first called to perform a security verification, and then the language organization and expression module is called to generate an emergency report. If the target strategy adopts a step-by-step reasoning mode, it first calls the target parsing and planning module to complete the target parsing, then calls the long short-term memory module to perform memory retrieval, and finally calls the logical reasoning and causal analysis module to perform reasoning. If the target strategy adopts a knowledge-enhanced retrieval mode, it first calls the long short-term memory module to perform memory retrieval, then calls the association generation and creative expansion module to generate candidate solutions, and finally calls the fact verification and security verification module to perform fact verification. If the target strategy adopts a backtracking correction mode, it first calls the interactive feedback and iterative optimization module for interactive feedback, then calls the logical reasoning and causal analysis module for logical reasoning and error diagnosis, and finally calls the constraint control and risk avoidance module to correct the constraints.

7. An embodied intelligent environment-adaptive learning device, characterized in that: It includes a memory and a processor, the memory storing a computer program that, when executed by the processor, causes the processor to perform the steps of the embodied intelligent environment adaptive learning method of claim 4.

Images