An intelligent autonomous system architecture with active continual learning capability

Abstract

The application discloses a kind of intelligent autonomous system architecture with active continuous learning ability, including hardware unit and software unit;Hardware unit is used to realize measurement perception, job execution and information interaction between systems;Software unit is used to realize data acquisition, perception, cognition, decision and control calculation, and its core is intelligent application unit.The intelligent autonomous system architecture and learning paradigm disclosed by the application provide an effective method for the design and development of intelligent systems with self-learning and reasoning capabilities in complex unknown environments.

Description

Technical Field

[0001] This invention relates to the field of intelligent autonomous system technology, and in particular to an intelligent autonomous system architecture, composition, and learning paradigm with active and continuous learning and reasoning capabilities in complex and unknown environments. Background Technology

[0002] The autonomous capabilities of existing unmanned systems are all manually designed functions. So-called intelligent unmanned systems generally employ neural networks trained offline using large datasets and large-scale learning. They only possess online autonomous inference and limited transfer learning capabilities, lacking the ability to learn autonomously in complex environments and adapt to complex scenarios. They are still far from truly human-like intelligence. The most challenging aspect of intelligent systems with reasoning and learning capabilities is the ability to learn autonomously and continuously in response to new tasks and environments. This has attracted considerable attention, but a systematic architectural design and learning implementation method are still lacking. Summary of the Invention

[0003] The technical problem solved by this invention is to overcome the shortcomings of the prior art and provide an intelligent autonomous system architecture with active and continuous learning capabilities, thus solving the problem that existing weak intelligent systems cannot achieve active and continuous learning.

[0004] The technical solution of this invention is: an intelligent autonomous system architecture with active and continuous learning capabilities, including a measurement and calculation execution unit and an acquisition and control unit; the measurement and calculation execution unit is used to realize measurement perception, operation execution and information interaction between systems; the acquisition and control unit is used to realize data acquisition, perception, cognition, decision-making and control calculation, and its core is an intelligent application unit.

[0005] The measurement and calculation execution unit includes a measurement and sensing platform, an intelligent computing platform, an action execution platform, and a high-speed information interaction network;

[0006] The measurement and sensing platform consists of various sensors to observe the external environment, scene, and situation, as well as to sense its own state. Based on the task requirements of the intelligent autonomous system, various measurement sensors are configured for the measurement and sensing platform to realize the measurement of vision, touch, motion, hearing, taste, and smell, as well as its own state, as required by the task.

[0007] The motion execution platform is composed of a multi-functional movable structure. Based on the task requirements of the intelligent autonomous system, the motion execution platform realizes the spatial movement and various operational skills required by the intelligent unmanned system.

[0008] The intelligent computing platform consists of intelligent computer hardware, system units and middleware, as well as intelligent application units;

[0009] The high-speed information exchange network consists of high-speed cables and wireless communication equipment. It transmits the measurement information from the measurement and sensing platform to the intelligent computing platform in a timely manner, and transmits the information processed by the intelligent computing platform to the motion execution platform in a timely manner.

[0010] The data acquisition and control unit consists of a system unit and an intelligent application unit; the intelligent application unit has a professional knowledge base, in which knowledge is stored in the form of a knowledge graph; the system unit has the ability to actively and continuously learn.

[0011] The knowledge graph is represented by nodes "entities" and associated "relationships". "Entities" are divided into "source entities" and "target entities". Each "entity" is characterized by attributes that represent the entity's category and characteristics. "Relationships" are the mapping relationships from "source entities" to "target entities", and describe various mapping relationships from simple to complex, including mathematical analytical relationships, semantic relationships or deep neural networks.

[0012] The professional knowledge base includes a perception knowledge base, a cognitive knowledge base, a decision knowledge base, a manipulation knowledge base, and a meta-knowledge base, which respectively store the perception knowledge graph, the cognitive knowledge graph, the decision knowledge graph, the manipulation knowledge graph, and the meta-knowledge graph;

[0013] The perception knowledge base stores the perception knowledge graph, and the "source entities" in the perception knowledge graph are the output information of multi-source measurement sensors; the mapping relationship between entities in the perception knowledge graph represents the fusion perception processing process based on multi-source measurement information;

[0014] The cognitive knowledge base stores cognitive knowledge graphs. The "source entities" of the cognitive knowledge graphs are the perception and recognition results, and the "target entities" are the cognition and understanding of the perception results. The mapping relationship between entities in the cognitive knowledge graphs represents the process of understanding and cognizing the perception results based on the task goal and its own state.

[0015] The decision knowledge base stores decision knowledge graphs, and based on goals and cognitive results, makes decisions on the next action of the system, generating actions or action sequences;

[0016] Manipulate the knowledge base to store the knowledge graph, and manipulate the system's next action based on task objectives and decision results;

[0017] The meta-knowledge base stores the learning knowledge graph and generates deep learning or reinforcement learning strategies based on task objectives, scenario understanding results, and its own state.

[0018] The perception knowledge base, cognitive knowledge graph, decision knowledge graph, and manipulation knowledge graph are connected in series. The "target node" of the previous graph is the "source node" of the next graph. Intensive design can be carried out at any stage, from "perception-cognition-decision-manipulation" to "perception-decision-manipulation" or "perception-cognition-manipulation", with the highest level being "perception-manipulation".

[0019] The system unit has the ability to actively and continuously learn: when the perception result exceeds the scope of the cognitive knowledge graph, the learning machine is triggered to call the meta-knowledge graph to start active learning and update the cognitive knowledge graph; when the perception result exceeds the scope of the decision knowledge graph, the learning machine is triggered to call the meta-knowledge graph to start active learning and update the decision knowledge graph; when the output of the decision knowledge graph exceeds the scope of the control knowledge graph, the learning machine is triggered to call the meta-knowledge graph to start active learning and update the control knowledge graph.

[0020] The system unit has active and continuous learning capabilities, which are realized through a learning machine; the learning machine consists of two parts: a learning initiator and a learning prediction and evaluation machine.

[0021] The learning initiator triggers active learning; the learning initiator autonomously or in response to commands determines whether the current state of the perceptual knowledge graph, cognitive knowledge graph, decision-making knowledge graph, and manipulation knowledge graph meets the requirements. If the system encounters new tasks, new skills, new behavioral requirements, enters a new environment or scenario, or its own state changes, the learning initiator starts learning; during the autonomous survival or operation phase, learning is automatically triggered by new tasks, new environments, new scenarios, or new state changes, or it can be actively triggered by its own new skill requirements or new cognitive intentions; during the training or testing phase, the learning initiator starts learning in response to external commands.

[0022] The learning prediction and evaluation machine enables cautious knowledge updates and solves the problem of knowledge forgetting. The learning prediction and evaluation machine predicts and evaluates the results after the learning is initiated, and decides whether to update the knowledge graph and the update strategy of the knowledge graph.

[0023] The system unit continuously learns and reasons according to the following paradigm: data -> perception and recognition -> cognitive understanding -> planning and decision-making -> control and execution.

[0024] The initial knowledge graph in the intelligent autonomous system is constructed by the designers based on prior knowledge, according to the characteristics of the object and its task.

[0025] The advantages of this invention compared to the prior art are:

[0026] (1) Most of the existing publicly available intelligent system architectures are geared towards the needs of unmanned driving or intelligent robotic arms. The architecture design does not reflect the ability to learn actively and continuously. This invention adopts human-like intelligent learning and reasoning mechanisms to construct a system architecture, providing the necessary software and hardware foundation for achieving intelligent improvement in complex and unknown environments.

[0027] (2) Existing publicly available intelligent system architectures generally use neural networks as the main modules for reasoning and learning. Neural networks mainly rely on nodes and weights to represent knowledge. During continuous learning, catastrophic forgetting is a fatal problem currently existing in neural networks. This invention uses a combination of classical knowledge representation and neural networks to realize knowledge representation and storage, including a composite knowledge graph composed of mathematical analytical relations, semantic relations, or deep neural networks, thus avoiding the forgetting problem caused by simple neural network knowledge representation and storage.

[0028] (3) Existing publicly available intelligent system architectures do not elaborate on the learning triggering and updating methods. This invention adopts a "dual-loop knowledge generation" design, dividing the learning machine into a starter and a prediction and evaluation machine. The starter triggers learning, and the prediction and evaluation machine judges the learning effect. The dual-loop judgment determines whether to update the knowledge. This further avoids the forgetting problem in existing intelligent system architectures during active learning.

[0029] (4) Existing publicly available intelligent system architectures clearly involve the system's initial intelligence or starting point intelligence. This invention explicitly proposes that the initial source knowledge graph in an intelligent system is constructed by the designer based on prior knowledge according to the characteristics of the object and its task. The knowledge architecture of the source knowledge graph determines the starting point intelligence level of the intelligent system, and the knowledge graph can be updated during the active and continuous learning process to improve the intelligence level. Attached Figure Description

[0030] Figure 1 This is a diagram of the system's physical architecture.

[0031] Figure 2 This is a diagram illustrating the learning and reasoning process. Detailed Implementation

[0032] An intelligent autonomous system architecture with active and continuous learning capabilities is proposed. The system comprises a hardware system consisting of a measurement and sensing platform, an intelligent computing platform, an action execution platform, and a high-speed information interaction network; and a physical architecture consisting of system units and a software system comprising a perception database, a cognition database, a decision-making database, a control database, a meta-knowledge base, a learning machine (including a learning initiator and a learning effect prediction and evaluation machine), and an inference machine. The physical architecture of the system is as follows: Figure 1 As shown.

[0033] The measurement and sensing platform consists of various sensors, enabling the observation of the external environment, scene, and situation, as well as the sensing of its own state. Based on the task requirements of the intelligent autonomous system, various measurement sensors are configured for the measurement and sensing platform to achieve the required measurements of vision (cameras, LiDAR, etc.), touch (force, heat, hardness, etc.), kinesthetic sense (speed, distance, attitude, etc.), hearing (speech recognition), taste and smell (chemical sensors), and the user's own state.

[0034] The motion execution platform consists of a multifunctional movable structure. Based on the task requirements of the intelligent autonomous system, the motion execution platform can realize the spatial movement (including surface movement) and various operational skills required by the intelligent unmanned system.

[0035] The intelligent computing platform consists of intelligent computer hardware, system units and middleware, as well as intelligent application units.

[0036] The intelligent application unit of the intelligent autonomous system possesses a professional knowledge base, which stores knowledge in the form of knowledge graphs, including perceptual knowledge graphs, cognitive knowledge graphs, decision-making knowledge graphs, manipulation knowledge graphs, and meta-knowledge graphs. The knowledge graph is represented by nodes "entities" and associated "relationships." "Entities" are divided into "source entities" and "target entities," with each entity represented by attributes indicating its category and characteristics. "Relationships" are mappings from "source entities" to "target entities," denoted by o = fs_o(t, s), where t is time, s is a combination of source entities, and o is the target entity or combination. The mapping relationship o = fs_o(t, s) can represent various mapping relationships from simple to complex, including mathematical analytical relationships, semantic relationships, or deep neural networks. The initial source knowledge graph in the intelligent autonomous system is constructed by the designers based on prior knowledge and the characteristics of the object and its task. The knowledge architecture of the source knowledge graph determines the initial intelligence level of the intelligent autonomous system, and it can be updated during active and continuous learning to improve the intelligence level.

[0037] The perception library stores a perception knowledge graph. In this graph, "source entities" represent the output information from multi-source measurement sensors, while "target entities" represent the results of perception and recognition, such as various identified targets and scenes. The mapping relationships between entities in the perception knowledge graph represent the fusion perception processing process based on multi-source measurement information.

[0038] The cognitive library stores a cognitive knowledge graph. The "source entities" in the cognitive knowledge graph represent the perception and recognition results, while the "target entities" represent the cognition and understanding of these results, such as suitable scenarios or suitable sampling targets. The mapping relationships between entities in the cognitive knowledge graph represent the process of understanding and cognizing the perception results based on the task objective and the entity's own state.

[0039] The decision base stores decision knowledge graphs, which make decisions on the next action of the system based on the goals and cognitive results, and generate actions or action sequences.

[0040] Manipulate the knowledge base to store the knowledge graph, and manipulate the system's next action based on task objectives and decision results.

[0041] The perceptual knowledge base, cognitive knowledge graph, decision-making knowledge graph, and manipulation knowledge graph are interconnected. The "target node" of the previous graph is the "source node" of the next graph. Therefore, intensive design can be carried out at any stage, from "perception-cognition-decision-manipulation" to "perception-decision-manipulation" or "perception-cognition-manipulation", with the highest level being "perception-manipulation".

[0042] The meta-knowledge base stores the learning knowledge graph and generates learning strategies such as deep learning or reinforcement learning based on task objectives, scenario understanding results, and its own state.

[0043] The system has the ability to learn actively and continuously: when the perceived result exceeds the scope of the cognitive knowledge graph, the learning machine is triggered to call the meta-knowledge graph to start active learning and update the cognitive knowledge graph; when the perceived result exceeds the scope of the decision knowledge graph, the learning machine is triggered to call the meta-knowledge graph to start active learning and update the decision knowledge graph; when the output of the decision knowledge graph exceeds the scope of the control knowledge graph, the learning machine is triggered to call the meta-knowledge graph to start active learning and update the control knowledge graph.

[0044] The learning machine consists of two parts: a learning initiator and a learning prediction and evaluation machine. The learning initiator autonomously or in response to commands determines whether the current state of the perceptual knowledge graph, cognitive knowledge graph, decision-making knowledge graph, and manipulation knowledge graph meets the requirements. If the system encounters new tasks, new skills, new behavioral requirements, enters a new environment or scenario, or its own state changes, the learning machine initiates learning. During the autonomous survival / operation phase, learning can be automatically triggered by new tasks / environments / scenarios / state changes, or it can be proactively triggered by its own desire to improve its skills / behaviors. During the training or testing phase, learning can be initiated in response to external commands.

[0045] The learning machine consists of two parts: a learning initiator and a learning prediction and evaluation machine. The learning prediction and evaluation machine predicts and evaluates the results after the learning is initiated, and decides whether to update the knowledge graph and the update strategy. The knowledge graph is only updated when it is determined that the new knowledge learned will not destroy the original knowledge, thus avoiding the problem of catastrophic forgetting caused by active learning.

[0046] Continuous proactive learning and reasoning proceed in the following paradigm: data -> perception and recognition -> cognitive understanding -> planning and decision-making -> manipulation and execution.

[0047] The core of an intelligent autonomous system lies in the professional knowledge graph and its scalability, encompassing the perception database, cognition database, decision database, meta-learning database, meta-cognition database, learning initiator, and learning prediction and evaluation machine. These factors determine the system's learning and reasoning capabilities and its level of intelligence. The source knowledge graph can be designed according to specific object and task requirements, and it determines the system's initial level of intelligence.

[0048] Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make possible changes and modifications to the technical solutions of the present invention based on the above-disclosed technical content without departing from the spirit and scope of the present invention. Therefore, any simple modifications, equivalent changes and modifications made to the above embodiments based on the technical essence of the present invention without departing from the content of the technical solutions of the present invention shall fall within the protection scope of the technical solutions of the present invention.

Claims

1. An intelligent autonomous system with active continual learning capability, characterized by: The system comprises a measurement calculation execution unit and a collection control unit; the measurement calculation execution unit is used for completing measurement perception, operation execution and information interaction between systems of the whole system, and comprises a measurement perception platform, an intelligent calculation platform, an operation execution platform and a high-speed information interaction network; the collection control unit completes data collection, perception, cognition, decision and operation calculation process; The measurement perception platform is composed of various sensors, and realizes observation on external environment, scene and situation and perception on self state; various measurement sensors are configured according to the task demand of the intelligent autonomous system, so as to realize vision, touch, motion sense, hearing, taste and smell and measurement of self state required by the task; The operation execution platform is composed of multifunctional movable structure mechanism, and realizes space movement and various operation skills required by the intelligent unmanned system according to the task demand of the intelligent autonomous system; The intelligent calculation platform is composed of intelligent computer hardware, system unit, middleware and intelligent application unit; The high-speed information interaction network is composed of high-speed cable and wireless communication equipment, and is used for timely transmitting measurement information of the measurement perception platform to the intelligent calculation platform and timely transmitting information processed by the intelligent calculation platform to the operation execution platform; The collection control unit is composed of the system unit and the intelligent application unit; the intelligent application unit has a professional knowledge base, and the professional knowledge base stores knowledge in the form of a knowledge graph; the system unit has active continuous learning ability; The knowledge graph is represented by nodes "entity" and relations "relation", wherein the "entity" is divided into "source entity" and "target entity", each "entity" is characterized by attributes of the entity category and characteristics; the "relation" is a mapping relation from the "source entity" to the "target entity", and expresses various mapping relations from simple to complex, including mathematical analytical relation, semantic relation or deep neural network; The professional knowledge base comprises a perception knowledge base, a cognition knowledge base, a decision knowledge base, a manipulation knowledge base and a meta-knowledge base, and respectively stores a perception knowledge graph, a cognition knowledge graph, a decision knowledge graph, a manipulation knowledge graph and a meta-knowledge graph; The perception knowledge base stores the perception knowledge graph, the "source entity" of the perception knowledge graph is multi-source measurement sensor output information; the mapping relation between entities in the perception knowledge graph represents a fusion perception processing process based on multi-source measurement information; The cognition knowledge base stores the cognition knowledge graph, the "source entity" of the cognition knowledge graph is a sensing recognition result, and the "target entity" is a cognition and understanding of the sensing result; the mapping relation between entities in the cognition knowledge graph represents an understanding and cognition process of the sensing result based on the task target and self state; The decision knowledge base stores the decision knowledge graph, and based on the target and the cognition result, the next step behavior of the system is decided to generate a behavior or a behavior sequence; The manipulation knowledge base stores the manipulation knowledge graph, and based on the task target and the decision result, the next step behavior of the system is manipulated; The meta-knowledge base stores the learning knowledge graph, and based on the task target, the scene understanding result and the self state, a deep learning or reinforcement learning strategy is generated. The perception knowledge base, cognitive knowledge graph, decision knowledge graph and manipulation knowledge graph are in series, and the "target node" of the previous graph is the "source node" of the next graph. Any link can be intensively designed, from "perception-cognition-decision-manipulation" to "perception-decision-manipulation" or "perception-cognition-manipulation", and the highest level is "perception-manipulation"; The system unit has active continuous learning capability: when the perception result exceeds the scope of the cognitive knowledge graph, the learning machine is triggered to call the meta-knowledge graph to start active learning and update the cognitive knowledge graph; When the cognitive result exceeds the scope of the decision knowledge graph, the learning machine calls the meta-knowledge graph to start active learning and update the decision knowledge graph; When the decision knowledge graph output exceeds the scope of the manipulation knowledge graph, the learning machine calls the meta-knowledge graph to start active learning and update the manipulation knowledge graph; The system unit has active continuous learning capability and is realized through a learning machine; the learning machine is divided into a learning start machine and a learning prediction and evaluation machine; The learning start machine triggers active learning; the learning start machine judges whether the current state of the perception knowledge graph, the cognitive knowledge graph, the decision knowledge graph and the manipulation knowledge graph meets the demand autonomously or in response to an instruction, and if a new task, a new skill, a new behavior demand appears, or a new environment, a new scene or a state change occurs, the learning machine starts learning; In the autonomous survival or operation stage, the system is automatically triggered to learn by a new task or a new environment or a new scene or a new state change, and can also be actively triggered to learn by a new skill demand or a new cognitive intention; In the training or testing stage, the learning machine starts learning in response to an external instruction; The learning prediction and evaluation machine realizes cautious knowledge updating and solves the problem of knowledge forgetting; the learning prediction and evaluation machine predicts and evaluates the result after starting learning to decide whether to update the knowledge graph and the updating strategy of the knowledge graph.

2. The intelligent autonomous system with active continuous learning capability according to claim 1, characterized in that: the continuous active learning and reasoning of the system unit are carried out according to the following paradigm: data -> perception identification -> cognitive understanding -> planning decision -> manipulation execution.

3. The intelligent autonomous system with active continuous learning capability according to any one of claims 1-2, characterized in that: the original knowledge graph in the intelligent autonomous system is constructed by a designer according to the characteristics of the object and its task based on prior knowledge.

Images