Multi-module collaborative learning method of robot
Through the multi-module collaborative learning strategy, the robot system uses data interaction and self-optimization between functional partitions, and uses algorithms such as deep learning to extract causal information, solving the problems of insufficient self-optimization capabilities and causal extraction in the existing technology, achieving more efficient Independent learning and adaptive decision-making.
Patent Information
- Application Number
- CN202411060789.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-02
- Publication Date
- 2024-11-22
AI Technical Summary
Existing robot systems have limited capabilities in self-optimization, and it is difficult to automatically adjust their behavioral strategies and knowledge structures based on environmental feedback, and there are shortcomings in extracting deep causal relationships from the data, limiting their inability to unknown or dynamically changing environments. adaptability and learning efficiency.
Adopting a multi-module collaborative learning strategy, through spontaneous data interaction and self-optimization between functional partitions, imitating the human brain to integrate multiple functional partition inputs and outputs to achieve self-update, and introducing advanced information processing units to extract information using algorithms such as deep learning and reinforcement learning. Relevance and causality form a high-level understanding to guide decision-making.
It significantly improves the autonomy and adaptability of the robot, allowing it to show higher learning and decision-making abilities in complex environments, and realizes self-adjustment and optimization of the knowledge structure without relying on manual data.
Smart Images

Figure CN119002264A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of robots, and in particular to a multi-module collaborative learning method for robots. Background Art
[0002] In the field of autonomous exploration robots, multi-module collaborative learning strategies are one of the key technologies to achieve high autonomy and adaptability of robots. However, existing technologies mainly rely on supervised learning and require a large amount of manually labeled data, which limits the adaptability of robots in unknown or dynamically changing environments. In addition, the collaborative working mechanism between functional modules is not mature enough, and there is a lack of effective data sharing and integration strategies, which leads to limited learning efficiency and adaptability. Existing robot systems have limited self-optimization capabilities, and it is difficult to automatically adjust their behavioral strategies and knowledge structures according to environmental feedback, and there are deficiencies in extracting deep causal relationships from data.
[0003] Therefore, technicians in this field are committed to developing a multi-module collaborative learning method for robots. A multi-module collaborative learning strategy is proposed, which imitates the process of the human brain integrating multiple functional partition inputs and outputs to achieve self-update, and realizes self-adjustment and optimization of the knowledge structure through spontaneous data interaction and self-optimization of each functional partition, thereby improving the autonomy and adaptability of the robot. This strategy does not rely on manually organized data, and can extract the relevance and causality of information from a higher level, providing a more efficient and flexible learning mechanism for autonomous exploration robots. Summary of the invention
[0004] In view of the above-mentioned defects of the prior art, the technical problem to be solved by the present invention is to achieve self-adjustment and optimization of the knowledge structure and improve the autonomy and adaptability of the robot.
[0005] To achieve the above objectives, the present invention provides a multi-module collaborative learning method for a robot, performs functional partitioning, and each functional partition processes different types of data input; establishes a data interaction mechanism to realize information sharing and data exchange between the functional partitions.
[0006] Furthermore, the functional partitions autonomously adjust processing algorithms and knowledge representations based on interaction data and environmental feedback.
[0007] Furthermore, the functional partition includes a perception module, a decision module, an execution module and a self-optimization module;
[0008] The perception module is responsible for the collection and preliminary processing of environmental information;
[0009] The decision-making module makes behavioral decisions based on the data provided by the perception module;
[0010] The execution module performs corresponding actions according to the instructions of the decision module;
[0011] The self-optimization module adjusts the parameters and strategies of each functional partition according to the execution results and environmental feedback.
[0012] Furthermore, it includes an advanced information processing unit to identify the correlation and causality of information from the data extracted from each functional partition.
[0013] Furthermore, the advanced information processing unit adopts a machine learning algorithm.
[0014] Furthermore, the machine learning algorithm includes deep learning, reinforcement learning, and clustering.
[0015] Furthermore, the advanced information processing unit forms a high-level understanding of the environment based on correlation and causal information.
[0016] Further, the high-level understanding guides the robot to make decisions in complex environments.
[0017] Further, the following steps are included:
[0018] Step 1: Initialize each functional partition and set initial parameters and strategies;
[0019] Step 2: Collect environmental data through the perception module and perform preliminary processing;
[0020] Step 3: The processed data is interacted and shared among the functional partitions;
[0021] Step 4: The decision module makes behavioral decisions based on the shared data;
[0022] Step 5: The execution module executes actions according to the decision and feeds back the execution results to the self-optimization module;
[0023] Step 6: The self-optimization module adjusts the parameters and strategies of each functional partition according to the feedback to optimize the performance;
[0024] Step 7: The advanced information processing unit analyzes the data, extracts correlation and causal information, and updates the robot's knowledge base;
[0025] Step 8: Repeat steps 2 to 7 to achieve continuous learning and self-optimization of the robot.
[0026] Furthermore, it also includes regular performance evaluation and adjustment of the system.
[0027] The present invention has at least two functional partitions, each partition is responsible for processing different types of data input; and a data interaction mechanism is used to realize information sharing and data exchange between the functional partitions.
[0028] The functional partitions of the present invention include but are not limited to a perception module, a decision module, an execution module and a self-optimization module; the perception module is responsible for the collection and preliminary processing of environmental information; the decision module makes behavioral decisions based on the data provided by the perception module; the execution module performs corresponding actions according to the instructions of the decision module; and the self-optimization module adjusts the parameters and strategies of each partition according to the execution results and environmental feedback.
[0029] Each functional partition of the present invention has the ability of self-learning and self-updating, and can autonomously adjust its processing algorithm and knowledge representation according to interactive data and environmental feedback.
[0030] The present invention includes an advanced information processing unit for identifying the relevance and causality of information from data extracted from each functional partition; the advanced information processing unit adopts advanced machine learning algorithms, such as deep learning, reinforcement learning or cluster analysis.
[0031] The advanced information processing unit of the present invention can form a high-level understanding of the environment based on the extracted correlation and causal information; this high-level understanding can guide the robot to make more reasonable and effective decisions in a complex environment.
[0032] The present invention 1. initializes each functional partition and sets initial parameters and strategies; 2. collects environmental data through a perception module and performs preliminary processing; 3. interacts and shares the processed data between each functional partition; 4. The decision module makes behavioral decisions based on the shared data; 5. The execution module executes actions based on the decisions and feeds back the execution results to the self-optimization module; 6. The self-optimization module adjusts the parameters and strategies of each partition based on the feedback to optimize the overall performance; 7. The advanced information processing unit analyzes the data, extracts correlation and causal information, and updates the robot's knowledge base; 8. Repeats steps 2 to 7 to achieve continuous learning and self-optimization of the robot.
[0033] Existing robot systems have limited self-optimization capabilities, and it is difficult to automatically adjust their behavioral strategies and knowledge structures according to environmental feedback, and there are deficiencies in extracting deep causal relationships from data. The present invention proposes an innovative multi-module collaborative learning strategy, which imitates the process of the human brain integrating multiple functional partition inputs and outputs to achieve self-update, and realizes self-adjustment and optimization of the knowledge structure through spontaneous data interaction and self-optimization of each functional partition, thereby improving the autonomy and adaptability of the robot. The core of the present invention includes key technologies such as self-optimization modules, data interaction mechanisms, and advanced information processing units. Through the integration of these technologies, robots can achieve efficient data sharing, information extraction, and knowledge updating. This strategy does not rely on manually organized data, and can extract the relevance and causality of information from a higher level, providing a more efficient and flexible learning mechanism for autonomous exploration robots.
[0034] The purpose of the present invention is to provide a multi-module collaborative learning strategy for an autonomous exploration robot, which can overcome the limitations of the prior art and achieve efficient learning and adaptation of the robot in an unknown or dynamically changing environment.
[0035] The multi-module collaborative learning strategy provided by the present invention is implemented through the following technical solutions:
[0036] 1. Define at least two functional partitions, each responsible for processing different types of data input to achieve task specialization and modularization.
[0037] 2. Design a data interaction mechanism to ensure information sharing and data exchange between functional areas to improve learning efficiency.
[0038] 3. Introduce self-learning and self-updating capabilities so that each functional partition can autonomously adjust its processing algorithm and knowledge representation based on interaction data and environmental feedback.
[0039] 4. Implement advanced information processing units and use advanced algorithms to extract correlation and causal information from data to enhance the robot's decision-making ability.
[0040] 5. Develop a series of steps, including initialization, data collection and processing, data interaction and sharing, behavioral decision-making, execution and feedback, knowledge base update, etc., to form a cycle of continuous learning and self-optimization of the robot.
[0041] Through the above technical solutions, the multi-module collaborative learning strategy of the present invention can significantly improve the autonomy and adaptability of the autonomous exploration robot, enabling it to demonstrate higher learning and decision-making capabilities in complex environments.
[0042] Compared with the prior art, the present invention has the following obvious substantial features and significant advantages:
[0043] 1. Autonomous learning and self-update mechanism: The strategy of the present invention imitates the self-update process of the human brain, enabling each functional partition of the robot to learn and update autonomously without relying on external manually labeled data.
[0044] 2. Inter-module collaboration and data interaction: By establishing an efficient data interaction mechanism, information sharing and data exchange between functional partitions are realized, enhancing the robot's ability to handle complex tasks.
[0045] 3. Self-optimization module: The strategy of the present invention includes a self-optimization module, which can automatically adjust the parameters and strategies of each partition according to the execution results and environmental feedback, thereby improving the self-optimization capability of the robot.
[0046] 4. Advanced Information Processing Unit: Introduce an advanced information processing unit and use advanced machine learning algorithms, such as deep learning, reinforcement learning, or cluster analysis, to extract the correlation and causality of information from the data and form a high-level understanding of the environment.
[0047] 5. Cycle of continuous learning and self-optimization: The robot's continuous learning and self-optimization are achieved through a series of steps, including initialization of functional partitions, collection and processing of environmental data, interaction and sharing of data, behavioral decision-making, execution and feedback, and updating of the knowledge base.
[0048] The concept, specific structure and technical effects of the present invention will be further described below in conjunction with the accompanying drawings to fully understand the purpose, characteristics and effects of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS
[0049] Figure 1 It is a diagram of implementation steps of a preferred embodiment of the present invention;
[0050] Figure 2 It is a perceptual LTM module architecture based on the attention mechanism of a preferred embodiment of the present invention;
[0051] Figure 3 It is a brain speech information alignment architecture of a preferred embodiment of the present invention;
[0052] Figure 4 It is a diagram of the “world model LTM” architecture of a preferred embodiment of the present invention. DETAILED DESCRIPTION
[0053] The following describes several preferred embodiments of the present invention with reference to the drawings in the specification, so that the technical content is clearer and easier to understand. The present invention can be embodied in many different forms of embodiments, and the protection scope of the present invention is not limited to the embodiments mentioned in the text.
[0054] In the drawings, components with the same structure are indicated by the same numerical reference numerals, and components with similar structures or functions are indicated by similar numerical reference numerals. The size and thickness of each component shown in the drawings are arbitrarily shown, and the present invention does not limit the size and thickness of each component. In order to make the illustration clearer, the thickness of the components is appropriately exaggerated in some places in the drawings.
[0055] Example 1: Autonomous Exploration Robot
[0056] 1. System Overview
[0057] The specific steps of autonomous exploration robot implementation. Figure 1 As shown, the multi-module collaborative learning strategy of the present invention includes the following key components: perception module, decision module, execution module, self-optimization module and advanced information processing unit. These modules are interconnected through a data interaction mechanism to form a closed-loop collaborative learning system.
[0058] 2. Initialization of functional partitions
[0059] Step 2.1 Initialize the perception module and set its parameters to adapt to the perception requirements of the specific environment, such as vision, sound, touch, etc.
[0060] Step 2.2 initializes the decision module and loads the preset decision algorithm and rule base.
[0061] Step 2.3 initializes the execution module and configures the parameters of the actuator according to the physical characteristics of the robot.
[0062] Step 2.4 initializes the self-optimization module, sets the learning rate, adaptive adjustment strategy, etc.
[0063] Step 2.5 initializes the advanced information processing unit and loads algorithm models such as deep learning and reinforcement learning.
[0064] 3. Collection and preliminary processing of environmental data
[0065] Step 3.1 The perception module collects environmental data in real time through the sensor array.
[0066] Step 3.2 performs preliminary processing such as noise filtering and feature extraction on the collected data.
[0067] 4. Data interaction and sharing
[0068] Step 4.1 encapsulates the preliminarily processed data into information packets and shares them among modules through the data interaction mechanism.
[0069] Step 4.2 Each module parses the information packet according to its own needs and extracts useful data.
[0070] 5. Behavioral Decision Making
[0071] Step 5.1 The decision module uses the decision algorithm to make behavioral decisions based on the shared data.
[0072] Step 5.2 The decision result is sent to the execution module in the form of instructions.
[0073] 6. Execution and Feedback
[0074] Step 6.1 After receiving the instruction, the execution module performs the corresponding action.
[0075] The execution result of step 6.2 is fed back to the self-optimization module through the sensor.
[0076] 7. Self-optimization
[0077] Step 7.1 The self-optimization module adjusts the parameters and strategies of each module according to the execution results and environmental feedback.
[0078] After the parameters are adjusted in step 7.2, the environmental data are collected and processed again to form an optimization cycle.
[0079] 8. Advanced Information Processing
[0080] Step 8.1 The high-level information processing unit analyzes the data provided by each module and extracts correlation and causal information.
[0081] Step 8.2 uses deep learning and other algorithms to further analyze and understand the extracted information.
[0082] 9. Knowledge base update
[0083] Step 9.1 integrates the analysis results and updates the knowledge base to support higher-level decision-making.
[0084] 10. Continuous learning and optimization cycle
[0085] Step 10.1 The system repeats steps 3 to 9 based on the updated knowledge base and adjusted parameters to achieve continuous learning and self-optimization.
[0086] 11. System evaluation and adjustment
[0087] Step 11.1 Regularly evaluate system performance, including autonomy, adaptability, decision accuracy, etc.
[0088] Step 11.2: Make necessary adjustments to the system structure or parameters based on the evaluation results.
[0089] 12. Application Cases
[0090] Step 12.1 describes a specific application scenario, such as environmental monitoring or disaster relief.
[0091] Through the above implementation, the multi-module collaborative learning strategy of the present invention can enable the autonomous exploration robot to achieve efficient autonomous learning and adaptation in an unknown or dynamically changing environment, significantly improving its autonomy and adaptability.
[0092] Example 2: System architecture of autonomous exploration robot
[0093] Detailed description of the autonomous exploration robot architecture. Through the multimodal information alignment technology based on the attention mechanism, the robot system architecture based on the brain-like behavior decision-making computing model is realized, which includes: the LTM module architecture based on the attention mechanism and the multi-module information alignment architecture based on collaborative learning.
[0094] The collaborative activities between multiple functional modules of the system architecture rely on the semantic information encoding transmitted between the functional modules. Based on the multimodal information alignment technology, the present invention constructs a unified feature information encoding between multiple modules, and uses continuous feedback learning to align the brain language information and weight information of multiple modules, providing an information expression basis for competitive behavior decision-making; at the same time, in the process of extracting brain language information from the long-term memory (LTM) module, the attention mechanism is combined to fully consider the influence of the LTM modules on each other, so that different functional modules produce associated decision evaluations, and enhance the predictive ability of the brain-like behavior decision architecture. The multimodal information alignment technology based on the attention mechanism ensures that the information extracted from different modules has a unified and aligned expression and more accurate predictive ability, thereby realizing the competitive decision-making and collaborative learning of the brain-like behavior decision architecture more comprehensively and accurately, making the system more adaptable and intelligent, and better able to adapt to the needs of different tasks and environments.
[0095] In the brain-like behavioral decision-making computational model, the communication information of different LTM modules has a similar structure, which is called Chunk. Chunk is a five-tuple:
[0096] <address,t,gist,weight,intensity> ,
[0097] Where address is the address of the LTM processor that generates the information block, t is the time when the block is generated, gist is the brain language information encoded in the LTM module, and is also the core communication information for aligning multiple LTM modules, weight is the weight of the processor to measure the importance of the current brain language, and intensity is the cumulative amount used to record the weights in the competing communication paths, and its initial (time t0) are |weight| and weight respectively.
[0098] In the robot architecture, external information with different characteristics is processed by different LTM modules respectively and shared between LTM modules. Although each LTM processing module needs to transmit consciousness information in Chunk format, the "language" used in different LTMs is not the same. For example, the voice module processes one-dimensional vector information, while the image module needs to process three-dimensional vector information, and LTMs need to compete with each other for consciousness information to form a unified decision. Therefore, it is necessary to extract features and align information of the input perception information, so as to facilitate the construction of unified aligned information encoding between multiple LTM modules for collaborative communication. At the same time, different LTMs need to be related to each other, so as to autonomously and dynamically adjust their respective brain language information generation, so as to have a more comprehensive and consistent cognition of the environment and to accurately and collaboratively complete the prediction of complex tasks, so as to realize autonomous competitive decision-making. Therefore, the present invention proposes a multimodal information alignment technology based on an attention mechanism in view of the above problems, which mainly includes an LTM module architecture based on an attention mechanism and a multi-module information alignment architecture based on collaborative learning.
[0099] 1.1 LTM module architecture based on attention mechanism
[0100] In the brain-like behavior decision architecture, since the perceptual information processed by different LTM modules may have different modalities, the encoder's attention will vary greatly due to the different modalities, thus affecting the information alignment of multiple modules. This project intends to combine the attention mechanism to fully consider the mutual influence between LTM modules in the process of brain-speech information alignment, so as to improve the overall architecture's ability to predict perceptual inputs of different modalities. The corresponding perceptual LTM module architecture based on the attention mechanism is as follows: Figure 2 shown.
[0101] The core dynamic mechanism of the attention-based perceptual LTM module architecture includes bottom-up attention and top-down attention. Among them, the perceptual LTM generates information blocks Chunks that need to be published to the STM based on information from the short-term memory (STM) and the information observed by itself. The core information includes brain language elements (gist) and importance indicators (weight). The brain language element gist is the feature encoding vector output by the encoder, and the weight weight is randomly initialized and adjusted and learned based on the Hebbian learning-based sleep expert algorithm according to the continuous information feedback of the STM. By publishing the information block Chunk to the STM to achieve the "consciousness" change of the STM, bottom-up attention is achieved. At the same time, the perceptual LTM module receives information broadcast from the STM to form top-down attention, making it sensitive to specific information to affect the encoder's attention range or area of the input information, thereby affecting the output content of the perceptual encoder and improving the correctness of the encoding. By combining the attention mechanism, the information correlation between the LTM modules can be fully considered, so that different functional modules can generate associated decision evaluations, thereby improving the predictive ability of the brain-like behavior decision architecture.
[0102] 1.2 Multi-module information alignment architecture based on collaborative learning
[0103] The brain-like behavior decision-making architecture mainly realizes the dynamic process of multiple LTM modules competing to form STM for decision-making and STM feeding back information to LTM modules for collaborative learning. The core of the dynamic stream of consciousness that forms LTM-STM is the brain language information and weight information in the information block Chunk, that is, the gist information and weight information in the information block. Based on the above attention mechanism, this project intends to implement the information alignment architecture at the level of brain language information and weight information, and at the same time use the world model LTM to improve the predictive ability of the brain-like behavior decision-making architecture, thereby supporting the autonomous decision-making and learning of the overall architecture.
[0104] ●Brain speech information alignment
[0105] Brain language is expressed as vectors in different LTM modules. Its specific value is obtained by LTM encoding the input information. For example, the image encoding LTM assigns a unique vector as its brain language to each category of image. Different LTMs have different encoding methods, which makes the same object output inconsistent vectors after being encoded by different LTMs. Therefore, a mechanism is needed to align brain language information to ensure that different LTM modules can effectively transmit and understand information. One way to achieve brain language alignment is through the "brain language translator" set inside each LTM module. It can be regarded as a vector mapping table, which translates the vector obtained from STM (brain language from other LTM) into the vector corresponding to its own encoder. When the brain language of an individual LTM module can match the brain language of the surrounding environment or other LTMs, the system can more easily capture changes in the environment and better adapt to changing needs and challenges. The brain speech information alignment architecture based on collaborative learning to be implemented in this project fully combines the LTM-STM dynamics of the brain-like behavior decision-making model. The perception information encoding of different LTM modules forms brain speech information to participate in STM competition, and at the same time receives other LTM brain speech information encoding broadcast by STM. Each LTM learns the correspondence between the brain speech vectors output by different LTMs by comparing the vector of its own output brain speech with the vector of other LTM brain speech received from STM, and realizes brain speech alignment, such as Figure 3 shown.
[0106] In the brain language information alignment architecture, each LTM spontaneously provides training data through STM to achieve self-organized collaborative brain language alignment learning. Specifically, in addition to the encoder, the perception LTM in the brain language information alignment architecture also contains a brain language translator, which has a brain language translation table for comparing the difference between its own encoding information and external feedback information. The perception LTM module updates its own encoder parameters by comparing its own encoding (brain language) output with the brain language of other LTMs received on the STM, and through the difference intensity signal provided in the brain language translation table in the brain language translator, so that it is consistent with the brain language encoding received from the STM. By adopting this self-organized collaborative brain language alignment architecture, this project builds a unified feature information encoding vector between multiple modules based on the neural network encoder, and uses continuous feedback learning to align the brain language information of multiple modules, ensuring the effectiveness and universality of brain language information transmission between different LTMs, and providing an information expression basis for competitive behavior decision-making. At the same time, there is an information gap between the brain language in the brain-like behavior decision-making architecture and the semantics understandable to humans. In order to achieve the "explainability" of the brain-like behavior decision-making architecture, it is only necessary to add a "language encoding LTM module" to match human-readable language and perceptually encoded brain language, thereby realizing the translation process from brain language to natural language.
[0107] ●Weight information alignment
[0108] In addition to brain-speech information alignment, the weight information alignment of the LTM module also affects the final behavioral decision-making results. The weight of an information block is the LTM module's estimate of the importance of putting the data block into the STM. This project intends to use a distributed "Hebbian learning" sleep expert algorithm to adjust the weight of the information block, that is, a separate "Hebbian learning" sleep expert algorithm is run in each LTM processor. The LTM module uses the "Hebbian learning" sleep expert algorithm to modify the intensity of the set weights, which helps to define the dynamic process of competition and feedback. In the sleep expert algorithm, the corresponding LTM function module p performs the following operations:
[0109] (1) If the contents of the STM are correct, then p is a no-op.
[0110] (2) If the content in the STM is wrong: If p is correct, then p promotes itself, that is, increases the strength of the information block p it generates, and p marks this submission as verified. If p is wrong: p corrects its error; then, p demotes itself, that is, reduces the strength of the information block it generates; p marks this submission as verified.
[0111] In the above-mentioned sleep expert algorithm, how the LTM module determines whether it has made an error is a key difficulty. This project considers the use of a combination of unsupervised and supervised signals. The unsupervised method is that the LTM module finds out that it has made an error by discovering that the received STM information block contradicts its own brain language information with a sufficiently high confidence (strength). Therefore, each broadcast enables each LTM module to check and judge whether it has made an error, that is, the LTM module modifies the intensity based on the feedback from the STM broadcast. At the same time, the feedback information learned in the supervised learning algorithm can be expanded to include supervisory signals from the environment feedback of the input mapping, and the weight information of the self-generated information block is modified according to the external supervisory signal prompt.
[0112] World Model LTM
[0113] In order to improve the predictive ability of brain-like behavioral decision-making architecture and system, this project intends to implement a special "world model LTM" module, such as Figure 4As shown. The "world model LTM" is mainly composed of a learnable predictor, a prediction error comparator, and a prediction information storage. The "world model LTM" simulates the process of humans relying on sensory organs to encode information for environmental cognition. The "world model LTM" does not directly receive external environmental information, but indirectly obtains information processed by the "sensory coding LTM" (brain language) through STM, and then uses the predictor to predict the information on the STM. This LTM predicts the direction of information on the STM, evaluates the possible consequences of the current direction of consciousness, and publishes this prediction information to the STM, thereby affecting other LTMs, thereby guiding the future direction of consciousness. In this process, assuming any time t, the state of the brain-like behavior decision-making computing system can be expressed as The “world model LTM” is denoted by W t ,Right now Where e t:0 is all the environmental information up to time t. In addition to evaluating the current direction of consciousness, the world model LTM also predicts itself And it is updated through the learning algorithm L(·), that is, W t+1 =L(C t:0 ,e t+1:0 ). This ability distinguishes it from a world model with limited capacity and inability to grow, i.e. The world model LTM regards the time-varying information on the observed STM as a "time series" and trains it in a self-supervised manner to recursively improve its predictive ability and achieve synchronization and consistency with the external environment.
[0114] Example 3: Post-disaster environmental monitoring robot
[0115] 1. System architecture design
[0116] The robot system consists of the following modules:
[0117] ●Perception module: integrates high-resolution camera, infrared thermal imager, LiDAR, temperature and humidity sensor, etc.
[0118] ●Data processing unit: Equipped with a high-performance computing platform for data fusion and feature extraction.
[0119] ●Decision-making module: A hybrid decision-making system based on rules and machine learning.
[0120] ●Execution module: Contains the drive and mechanical structure to achieve precise movement and operation.
[0121] ●Self-optimization module: Integrates reinforcement learning algorithms to achieve adaptive behavior adjustment.
[0122] ●Advanced information processing unit: uses deep learning architecture to perform complex scenario analysis.
[0123] ●Communication system: supports data exchange with external base stations and remote command reception.
[0124] 2. Detailed initialization of functional partitions
[0125] ● Perception module initialization: calibrate all sensors, ensure synchronous data collection, and set initial thresholds and parameters.
[0126] ●Decision module initialization: load predefined behavior rule library and machine learning model.
[0127] ●Execute module initialization: perform self-inspection of the mechanical structure to ensure the range of motion and accuracy.
[0128] ●Self-optimization module initialization: setting learning rate and exploration / exploitation balance parameters.
[0129] ●Advanced information processing unit initialization: Select and load the deep learning network suitable for environmental monitoring.
[0130] 3. Detailed collection and preliminary processing of environmental data
[0131] ●Data collection: The perception module collects visual, infrared, distance and other data in real time.
[0132] ●Data preprocessing: including denoising, data standardization, and feature extraction (such as edge detection and texture analysis).
[0133] 4. Implementation of data interaction and sharing
[0134] ●Information encapsulation: Encapsulate the preprocessed data into information packets in a unified format.
[0135] ●Data transmission: Securely transmit information packets between modules through wireless communication or wired connection.
[0136] ●Data analysis: Each module analyzes the information package according to the requirements and extracts key data.
[0137] 5. Detailed formulation of behavioral decisions
[0138] ●Environmental assessment: The decision-making module analyzes environmental characteristics and assesses safe areas and potential dangers.
[0139] ●Path planning: Plan the optimal exploration path based on environmental assessment results.
[0140] ●Task allocation: Assign specific tasks to the execution module based on the path and monitoring objectives.
[0141] 6. Detailed process of execution and feedback
[0142] ●Precise execution: The execution module moves precisely to the specified location based on the decision results.
[0143] ●Real-time monitoring: The perception module provides real-time feedback on robot status and environmental changes.
[0144] ●Feedback processing: The self-optimization module analyzes execution results and environmental feedback to evaluate task completion.
[0145] 7. Detailed strategies for self-optimization
[0146] ●Parameter adjustment: Dynamically adjust sensor sensitivity, decision threshold, etc. based on feedback.
[0147] ●Strategy optimization: Use reinforcement learning to optimize behavioral strategies to improve task completion rate and efficiency.
[0148] ●Model update: Regularly update machine learning models to adapt to environmental changes.
[0149] 8. Detailed algorithms for advanced information processing
[0150] ●Image recognition: Use CNN to classify images and identify objects.
[0151] ●Scene understanding: Use deep learning models to analyze scenes and identify damaged buildings, road blockages, etc.
[0152] ●Causal analysis: Use models such as graph neural networks to explore the causal relationship between events.
[0153] 9. Detailed operations for knowledge base update
[0154] ●Knowledge representation: Convert the extracted features, recognition results and analysis conclusions into entries in the knowledge base.
[0155] ●Knowledge fusion: Integrate knowledge from different modules to update environmental models and behavioral strategies.
[0156] ●Knowledge reasoning: Use the knowledge base to perform reasoning, assist decision making and prediction.
[0157] 10. Detailed steps for continuous learning and optimization cycle
[0158] ●Loop trigger: trigger a new round of learning cycle based on task completion and environmental changes.
[0159] ●Data re-collection: Re-collect environmental data for a new round of learning and optimization.
[0160] ●Performance evaluation: Regularly evaluate learning outcomes and system performance to ensure continuous improvement.
[0161] 11. Detailed methods for system evaluation and adjustment
[0162] ●Performance indicators: Define key performance indicators such as autonomy, adaptability, and decision accuracy.
[0163] ●Evaluation method: Evaluate system performance through simulation environment and actual operation data.
[0164] ●System adjustment: Adjust module parameters, optimize algorithms or upgrade hardware based on evaluation results.
[0165] 12. Detailed description of the application case
[0166] ●Scene setting: Specific disaster area environment, including terrain, buildings, potential dangers, etc.
[0167] ●Implementation process: the entire process from robot deployment to task completion.
[0168] ●Expected effect: The robot effectively monitors the environment, assesses damage and generates reports.
[0169] The multi-module collaborative learning strategy of the present invention can enable the post-disaster environmental monitoring robot to have a high degree of autonomy and adaptability, and effectively complete monitoring tasks in complex environments.
[0170] The preferred specific embodiments of the present invention are described in detail above. It should be understood that ordinary technicians in the field can make many modifications and changes based on the concept of the present invention without creative work. Therefore, all technical solutions that can be obtained by technicians in the technical field based on the concept of the present invention through logical analysis, reasoning or limited experiments on the basis of the prior art should be within the scope of protection determined by the claims.
Claims
1. A multi-module collaborative learning method for robots, It is characterized in that Carry out functional partitioning, with each functional partition processing different types of data input; establish a data interaction mechanism to achieve information sharing and data exchange between functional partitions.
2. The multi-module collaborative learning method of a robot as claimed in claim 1, It is characterized in that The functional partitions autonomously adjust processing algorithms and knowledge representations based on interaction data and environmental feedback.
3. The multi-module collaborative learning method of a robot as claimed in claim 1, It is characterized in that The functional partitions include a perception module, a decision module, an execution module and a self-optimization module; The perception module is responsible for the collection and preliminary processing of environmental information; The decision-making module makes behavioral decisions based on the data provided by the perception module; The execution module performs corresponding actions according to the instructions of the decision module; The self-optimization module adjusts the parameters and strategies of each functional partition according to the execution results and environmental feedback.
4. The multi-module collaborative learning method of a robot as claimed in claim 3, It is characterized in that Includes an advanced information processing unit that identifies the relevance and causality of information extracted from each functional partition.
5. The multi-module collaborative learning method of a robot as claimed in claim 4, It is characterized in that The advanced information processing unit uses a machine learning algorithm.
6. The multi-module collaborative learning method of a robot as claimed in claim 5, It is characterized in that The machine learning algorithms include deep learning, reinforcement learning, and clustering.
7. The multi-module collaborative learning method of a robot as claimed in claim 4, It is characterized in that The high-level information processing unit forms a high-level understanding of the environment based on associative and causal information.
8. The multi-module collaborative learning method of a robot as claimed in claim 7, It is characterized in that This high-level understanding guides the robot to make decisions in complex environments.
9. The multi-module collaborative learning method of a robot as claimed in claim 4, It is characterized in that The following steps are involved: Step 1: Initialize each functional partition and set initial parameters and strategies; Step 2: Collect environmental data through the perception module and perform preliminary processing; Step 3: The processed data is interacted and shared among the functional partitions; Step 4: The decision module makes behavioral decisions based on the shared data; Step 5: The execution module executes actions according to the decision and feeds back the execution results to the self-optimization module; Step 6: The self-optimization module adjusts the parameters and strategies of each functional partition according to the feedback to optimize the performance; Step 7: The advanced information processing unit analyzes the data, extracts correlation and causal information, and updates the robot's knowledge base; Step 8: Repeat steps 2 to 7 to achieve continuous learning and self-optimization of the robot.
10. The multi-module collaborative learning method of a robot as claimed in claim 9, It is characterized in that It also includes regular performance evaluation and adjustment of the system.
Citation Information
Cited By
Brain-like behavior decision calculation model of autonomous robot
CN119692382A
