Robot asynchronous control method and system based on intention driving

Abstract

The invention relates to the technical field of robot control, in particular to a robot asynchronous control method and system based on intention driving. The method comprises the steps that real-time environment data and an advanced task target of a robot are acquired, the advanced task target is analyzed based on the real-time environment data, and an abstract intention set decoupled from hardware is generated; acquiring real-time hardware state data; according to the real-time hardware state data, splitting the abstract intention set to obtain a plurality of sub-intentions which can be executed in parallel, and forming an intention tree; analyzing the real-time hardware state data, and determining hardware execution capability; based on the hardware execution capability, analyzing the intention tree, and determining mutual exclusion operation; analyzing the intention tree, and determining the intention association between any two sub-intentions; and dynamically adjusting the execution sequence of the mutual exclusion operation based on the intention association. Cooperative high efficiency among sub-intentions is ensured, hardware resource conflicts are solved, asynchronous execution of mutual exclusion operation is ensured without blocking, flexibility and robustness are improved, real-time change is adapted, and high-efficiency control is realized.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of robot control, and particularly to a robot asynchronous control method and system based on intention driving. BACKGROUND

[0002] With the rapid development of service robot technology, its application in complex dynamic environment is increasingly widespread. In order to complete the advanced task goal, the robot needs to perceive the variable environmental information in real time and coordinate multiple hardware for response. However, the existing control method mostly adopts centralized or strong coupling architecture, in which the control center uniformly processes all perception data and generates bottom-level instructions, which is easy to cause overall control delay or even failure, and cannot quickly and effectively dynamically adapt to hardware heterogeneity or environmental mutation. SUMMARY The present application provides a robot asynchronous control method and system based on intention driving to solve the above problems.

[0003] In a first aspect, the present application provides a robot asynchronous control method based on intention driving, which comprises: obtaining real-time environmental data and high-level task goals of the robot, analyzing the high-level task goals based on the real-time environmental data, and generating an abstract intention set decoupled from hardware; obtaining real-time hardware state data, splitting the abstract intention set according to the real-time hardware state data to obtain a plurality of sub-intentions that can be executed in parallel, and forming an intention tree; analyzing the real-time hardware state data to determine the hardware execution capability; analyzing the intention tree based on the hardware execution capability to determine mutual exclusion operations; analyzing the intention tree to determine the intention association between any two sub-intentions; based on the intention association, dynamically adjusting the execution order of the mutual exclusion operations.

[0004] By the scheme, real-time environment data is acquired, senior task targets of the robot are acquired, the senior task targets are analyzed based on the real-time environment data, an abstract intent set decoupled from hardware is generated, the intent set is ensured to be independent of specific hardware, intent update delay caused by environment changes is avoided, and a stable and scalable intent basis is provided for processing. Real-time hardware state data is acquired, resource states are ensured to be accurately included in decision-making, and resource waste or execution failure caused by hardware neglect is avoided. The abstract intent set is split based on the real-time hardware state data, a plurality of sub-intents that can be executed in parallel are obtained, an intent tree is formed, the sub-intents are ensured to be executed asynchronously, sequential blocking is reduced, efficiency and real-time performance are improved, and hierarchical task representation is formed. The real-time hardware state data is analyzed, hardware execution capability is determined, a benchmark is provided for mutual exclusion operation identification, hardware limitations are ensured to be accurately quantified, and execution failure caused by unreasonable task allocation is avoided. Based on the hardware execution capability, the intent tree is analyzed, mutual exclusion operations are determined, a target is provided for dynamic adjustment, and resource competition problems are ensured to be effectively isolated. The intent tree is analyzed, intent association between any two sub-intents is determined, a basis is provided for execution order optimization, cooperation between the sub-intents is ensured to be efficient, and soft conflicts are reduced. Based on the intent association, the execution order of the mutual exclusion operations is dynamically adjusted, hardware resource conflicts are solved, the mutual exclusion operations are ensured to be executed asynchronously without blocking, and flexibility and robustness are improved, real-time changes are adapted to, and efficient control is implemented.

[0005] Optionally, the abstract intent set decoupled from hardware is generated based on the real-time environment data and the senior task targets, and the method comprises the following steps. The real-time environment data is analyzed, and personnel distribution and personnel movement trend are determined. The current scene feature is determined according to the personnel distribution and the personnel movement trend. The initial intent set is determined by querying a preset task-intent mapping model based on the current scene feature. An abnormal event is determined based on the current scene feature. The initial intent set is corrected according to the abnormal event, and the abstract intent set is determined.

[0006] By this scheme, real-time environmental data is analyzed to determine personnel distribution and personnel movement trend, ensuring that environmental data is converted from raw form to processable input, avoiding delay caused by data congestion. According to the personnel distribution and personnel movement trend, the current scene characteristics are determined to realize real-time capture of environmental dynamics, provide context for intent generation, and ensure adaptation to environmental changes without relying on static presets. Based on the current scene characteristics, the preset task-intent mapping model is queried to determine the initial intent set, ensuring the rapid start of intent analysis, avoiding the lack of flexibility caused by hard-coded priority, and directly providing basic input for correction. Based on the current scene characteristics, abnormal events are determined to enhance dynamic adaptation capability, ensuring timely response when the environment changes, thereby preventing intent conflict or task failure. According to the abnormal event, the initial intent set is corrected to determine the abstract intent set, ensuring the real-time and robustness of the intent set, and solving the problem of inflexible intent conflict processing.

[0007] Optionally, the analysis of the intent tree determines the intent association between any two sub-intents, including: For any sub-intent, the semantic vector of the sub-intent is analyzed and obtained; The real-time environmental data is analyzed to determine the context feature vector; the context feature vector includes: pedestrian approach speed, human posture type, environmental noise level, network delay value; The semantic vector and the context feature vector are spliced into a fusion feature vector; According to the fusion feature vector, the correlation score is calculated; Based on the correlation score and the preset association threshold, the intent association between any two sub-intents is determined.

[0008] By this scheme, for any sub-intent, the semantic vector of the sub-intent is analyzed and obtained, efficient storage and retrieval of intent information is supported, avoiding processing delay caused by instruction-level coupling, thereby supporting real-time updating of the intent tree. Analyzing real-time environmental data to determine the context feature vector solves the defect of difficulty in adapting to environmental dynamics, ensures that intent association decision is based on the latest context, and reduces intent conflict caused by environmental mutation. The semantic vector and the context feature vector are spliced into a fusion feature vector, reducing data transmission overhead and improving processing efficiency. According to the fusion feature vector, the correlation score is calculated to identify the logical mutual exclusion or cooperative relationship between intents, providing data basis for conflict processing and ensuring the efficiency of asynchronous execution. Based on the correlation score and the preset association threshold, the intent association between any two sub-intents is determined, supporting non-blocking control and improving overall robustness and real-time performance.

[0009] Optionally, the execution order of the mutually exclusive operations is dynamically adjusted based on the intent association, including: The mutually exclusive operation is analyzed to determine the mutual exclusion type; generate a dynamic priority queue according to the intention association and the mutual exclusion type; analyze the real-time environment data to determine a real-time change event; update the dynamic priority queue based on the real-time change event, and determine an execution sequence of the mutual exclusion operation according to an update result.

[0010] According to the scheme, the mutual exclusion operation is analyzed, the mutual exclusion type is determined, the conflict processing basis for generating the dynamic priority queue is laid, and the execution failure caused by resource competition or sequence error is reduced. According to the intention association and the mutual exclusion type, the dynamic priority queue is generated, and the conflict probability is preliminarily reduced. The real-time environment data is analyzed, the real-time change event is determined, the real-time interference is reflected, the basis for updating the queue is provided, and the execution sequence is ensured to respond to the sudden change. Based on the real-time change event, the dynamic priority queue is updated, and the execution sequence of the mutual exclusion operation is determined according to the update result, so as to reduce the blocking and conflict and improve the overall execution efficiency.

[0011] Optionally, the analyzing the real-time hardware state data to determine a hardware execution capability comprises: obtaining hardware performance indicators and network state data; analyzing the hardware performance indicators to determine a hardware preset processing capability, a current battery power and a sensor accuracy; analyzing the network state data to determine a network delay and a network bandwidth; determining the hardware execution capability based on the network delay and the network bandwidth, and according to the hardware preset processing capability, the current battery power and the sensor accuracy.

[0012] According to the scheme, the hardware performance indicators and the network state data are obtained, the control delay is reduced, and thus the blocking problem is alleviated. The hardware performance indicators are analyzed, the hardware preset processing capability, the current battery power and the sensor accuracy are determined, the original data is converted into comparable quantitative values, and the hardware heterogeneity problem is solved. The network state data is analyzed, the network delay and the network bandwidth are determined, the original data is converted into comparable quantitative values, and the environment adaptation deficiency problem is alleviated. Based on the network delay and the network bandwidth, the hardware execution capability is determined according to the hardware preset processing capability, the current battery power and the sensor accuracy, the overall hardware state is reflected, the dynamic task decision is supported, and the intention conflict and the blocking delay problem are solved.

[0013] Optionally, after the intention tree is formed, the method further comprises: analyzing the abnormal event to determine an abnormal event feature; determining a pedestrian intention according to the personnel distribution and the personnel movement trend; determining whether to update the intention tree according to the abnormal event feature and the pedestrian intention.

[0014] By this scheme, the abnormal event is analyzed, the abnormal event characteristics are determined, the foundation for intention update is laid, the dynamic event is avoided to be ignored due to the inability to modify the intention tree, and the environmental adaptability is improved. According to the personnel distribution and the personnel movement trend, the pedestrian intention is determined, and the control chain is avoided to be delayed due to waiting for personnel data analysis. According to the abnormal event characteristics and the pedestrian intention, it is determined whether to update the intention tree, the network burden and the control coupling degree are reduced, and the real-time performance and the expansibility in the fluctuating environment are improved.

[0015] Optionally, the method comprises: analyzing the semantic vector to determine the relevance between the sub-intention and the pedestrian intention; According to the relevance, dynamically inserting a new intention and deleting an original sub-intention in conflict with the pedestrian intention, an abstract intention set after reconstruction is obtained.

[0016] By this scheme, the semantic vector is analyzed, the relevance between the sub-intention and the pedestrian intention is determined, the intention semantic correlation is dynamically evaluated, and the delay caused by centralized processing is avoided. According to the relevance, a new intention is dynamically inserted, an original sub-intention in conflict with the pedestrian intention is deleted, an abstract intention set after reconstruction is obtained, asynchronous collaboration is supported, and robustness and real-time performance are improved.

[0017] Optionally, after the execution order of the mutually exclusive operations is dynamically adjusted based on the intention association, the method further comprises: allocating the sub-intention to a corresponding hardware execution agent; monitoring the execution state of each sub-intention to obtain a ready flag signal; According to the ready flag signal and the intention association, the execution results of multiple sub-intentions are aggregated.

[0018] By this scheme, the sub-intention is allocated to the corresponding hardware execution agent, the instruction publishing stagnation caused by network delay or hardware response timeout is avoided, and the real-time performance is improved. The execution state of each sub-intention is monitored to obtain a ready flag signal, the network burden is reduced, especially in a high-delay environment, the overhead caused by frequent data transmission is reduced. According to the ready flag signal and the intention association, the execution results of multiple sub-intentions are aggregated, the network transmission frequency is reduced, and the hardware execution details are decoupled.

[0019] Optionally, after the hardware execution capability is determined according to the hardware preset processing capability, the current battery power and the sensor accuracy based on the network delay and the network bandwidth, the method further comprises: determining network reliability according to the network delay and the network bandwidth; if the network reliability is lower than a preset level, all sub-intentions in the intention tree that depend on cloud computing are obtained; convert the cloud-computing-dependent sub-intent into a simplified sub-intent that can be executed locally; re-generate an offline execution sequence based on the hardware execution capability.

[0020] By the scheme, the network reliability is determined according to the network delay and the network bandwidth, and it is ensured that the intent adjustment is only performed when the network condition is poor, unnecessary operations are avoided, and a foundation is laid for real-time response in a dynamic environment. If the network reliability is low, all cloud-computing-dependent sub-intents in the intent tree are obtained, and it is ensured that the focus of adjustment is only on high-risk tasks, reducing the vulnerability of the overall control chain. The cloud-computing-dependent sub-intent is converted into a simplified sub-intent that can be executed locally, reducing network dependence, avoiding blocking caused by cloud computing communication, and adapting to dynamic changes in hardware. Based on the hardware execution capability, an offline execution sequence is re-generated, ensuring that the robot control chain can still run efficiently under poor network conditions.

[0021] In a second aspect, the application provides an intent-driven robot asynchronous control system, which comprises: a set generation module configured to obtain real-time environment data and a high-level task target of a robot, parse the high-level task target based on the real-time environment data, and generate an abstract intent set decoupled from hardware; an intent tree generation module configured to obtain real-time hardware state data, split the abstract intent set based on the real-time hardware state data, obtain a plurality of sub-intents that can be executed in parallel, and form an intent tree; a capability analysis module configured to analyze the real-time hardware state data and determine a hardware execution capability; an operation determination module configured to analyze the intent tree based on the hardware execution capability and determine mutually exclusive operations; a correlation analysis module configured to analyze the intent tree and determine intent correlation between any two sub-intents; a sequence adjustment module configured to dynamically adjust an execution sequence of the mutually exclusive operations based on the intent correlation. BRIEF DESCRIPTION OF DRAWINGS

[0022] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings needed to be used in the embodiment or prior art description. Obviously, the drawings in the following description are some embodiments of the present application, and those skilled in the art can also obtain other drawings according to these drawings without creative labor.

[0023] Figure 1 An application scenario diagram is provided for an embodiment of the present application.

[0024] Figure 2A flow chart of an intention-driven robot asynchronous control method provided by an embodiment of the present application.

[0025] Figure 3 A structural schematic diagram of an intention-driven robot asynchronous control system provided by an embodiment of the present application. DETAILED DESCRIPTION

[0026] To make the objectives, technical solutions and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application will be described below in connection with the drawings in the embodiments of the present application. Obviously, the described embodiments are only some of the embodiments of the present application, but not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative work fall within the protection scope of the present application.

[0027] In addition, the term “and / or” in the present document is only to describe the association relationship of the associated objects, which means that there can be three relationships, for example, A and / or B can represent the following three cases: A exists alone, A and B exist simultaneously, and B exists alone. In addition, the character “ / ” in the present document generally represents an “or” relationship between the associated objects unless otherwise specified.

[0028] The embodiments of the present application will be further described in detail below in connection with the drawings of the specification.

[0029] The existing control method mostly adopts centralized or strong coupling architecture, in which a control center uniformly processes all perception data and generates bottom-layer instructions, which is easy to cause overall control delay or even failure, and cannot quickly and effectively dynamically adapt to hardware heterogeneity or environmental mutation. In addition, the existing technology lacks real-time analysis of the association between sub-intentions in the intention tree and dynamic adjustment mechanism of hardware mutual exclusion operation, which often causes intention conflicts (such as priority confusion when simultaneously executing “play voice” and “obstacle avoidance”), limits the parallel efficiency and real-time response capability of asynchronous control, and is difficult to ensure the robustness of non-blocking execution.

[0030] Based on this, the application provides a kind of robot asynchronous control method and system based on intention driving, obtains real-time environment data, senior task target of robot, based on real-time environment data, senior task target is parsed, and the abstract intention set decoupled with hardware is generated, ensure that intention set is independent of specific hardware, avoid the intention update delay caused by environmental change, and provide stable, scalable intention base for processing.Real-time hardware state data is obtained, ensure that resource state is accurately included in decision-making, avoid the waste of resources or execution failure caused by hardware neglect.According to real-time hardware state data, abstract intention set is split, a plurality of sub-intentions that can be executed in parallel are obtained, intention tree is formed, ensure that sub-intention can be executed asynchronously, reduce sequential blocking, and improve efficiency and real-time, form hierarchical task representation.Analyze real-time hardware state data, determine hardware execution capability, provide benchmark for mutual exclusion operation identification, ensure that hardware limit is accurately quantified, avoid the execution failure caused by unreasonable task allocation.Based on hardware execution capability, analyze intention tree, determine mutual exclusion operation, provide target for dynamic adjustment, ensure that resource competition problem is effectively isolated.Analyze intention tree, determine the intention association between any two sub-intentions, provide basis for execution order optimization, ensure the cooperation of sub-intentions is efficient, reduce soft conflict.Based on intention association, dynamically adjust the execution order of mutual exclusion operation, solve hardware resource conflict, ensure that mutual exclusion operation is executed asynchronously without blocking, and improve flexibility and robustness, adapt to real-time change, realize efficient control.

[0031] Figure 1 An application scenario is provided for the present application. When performing robot asynchronous control, the method provided by the present application is applied.

[0032] Specifically, the method provided by the present application is applied in the main control chip provided in any robot, and the main control chip interacts with a sensor module, a task management system and a hardware monitoring module. Real-time environment data is obtained through the sensor module, and at the same time, senior task target of the robot is received through the task management system. Based on real-time environment data, senior task target is parsed, and the abstract intention set decoupled with hardware is generated. Real-time hardware state data is obtained through the hardware monitoring module. According to real-time hardware state data, abstract intention set is split, a plurality of sub-intentions that can be executed in parallel are obtained, and intention tree is formed. Real-time hardware state data is analyzed, and hardware execution capability is determined. Based on hardware execution capability, intention tree is analyzed, mutual exclusion operation is determined, target for dynamic adjustment is provided, and resource competition problem is effectively isolated. Intention tree is analyzed, intention association between any two sub-intentions is determined, basis for execution order optimization is provided, cooperation of sub-intentions is ensured to be efficient, and soft conflict is reduced. Based on intention association, execution order of mutual exclusion operation is dynamically adjusted, hardware resource conflict is solved, mutual exclusion operation is ensured to be executed asynchronously without blocking, flexibility and robustness are improved, real-time change is adapted, and efficient control is realized. The specific implementation mode can refer to the following embodiments.

[0033] Figure 2 A flowchart of an intention-driven robot asynchronous control method provided for an embodiment of the present application. The method of the embodiment can be applied to the master chip in the above scenarios. As shown in the figure, the method includes: Figure 2 S201, obtaining real-time environment data and high-level task target of the robot, based on the real-time environment data, analyzing the high-level task target, and generating an abstract intention set decoupled from hardware; The real-time environment data can be dynamic environment information.

[0034] The high-level task target can be a high-level instruction describing the overall task requirement of the robot.

[0035] The abstract intention set can be a set of high-level task units decoupled from hardware.

[0036] Specifically, the real-time environment data is obtained through a sensor module, such as capturing image data through a camera to identify personnel distribution and movement trends, capturing audio data through a microphone to analyze environmental noise levels, and scanning the environment through a laser radar to generate an obstacle map (a spatial map representing obstacle positions and shapes). At the same time, the high-level task target of the robot is received through a task management system.

[0037] Based on the real-time environment data, the high-level task target is decomposed into an abstract intention set independent of specific hardware (actual physical devices such as mobile chassis, voice modules, cameras, microphones, etc.), such as resolving the abstract intention set to include moving to position B, detecting user A position, and playing guiding voice for guiding user A to position B.

[0038] S202, obtaining real-time hardware state data; according to the real-time hardware state data, splitting the abstract intention set to obtain a plurality of sub-intentions that can be executed in parallel, and forming an intention tree; The real-time hardware state data can be a hardware performance indicator obtained in real time through a hardware monitoring module.

[0039] The sub-intention can be a fine-grained, parallel-executable task unit.

[0040] The intention tree can be a tree-shaped data structure formed by organizing sub-intentions.

[0041] ​Specifically, real-time hardware state data is acquired by a hardware monitoring module. According to the real-time hardware state data, the parallelism of each abstract intent (such as playing a voice and moving a chassis which can be executed in parallel) is analyzed; and then according to the parallelism, an intent splitting engine (for splitting an abstract intent set) is used to split the abstract intent set into a plurality of sub-intents which can be executed in parallel (such as splitting moving to position B into path planning, real-time obstacle avoidance, and speed control); and then the sub-intents are organized into a tree structure to form an intent tree, wherein a root node represents a high-level task goal, and a branch node represents a sub-intent.

[0042] S203, analyzing real-time hardware state data to determine hardware execution capability; The hardware execution capability can be the performance upper limit of the hardware in the current state.

[0043] Specifically, key indicators (core performance parameters such as battery power and network delay) in the real-time hardware state data are analyzed; and according to the key indicators, a pre-defined rule set (for converting the key indicators into an operable hardware execution capability) set according to expert experience is applied to determine the hardware execution capability, such as when the battery is sufficient, the task can be executed on the local processing unit of the robot, and then the hardware execution capability is determined as the local execution capability; when the network is stable, the task can be offloaded to the cloud, and then the hardware execution capability is determined as the cloud-dependent capability.

[0044] S204, analyzing the intent tree based on the hardware execution capability to determine mutual exclusion operations; The mutual exclusion operation can be an operation between sub-intents that cannot be executed in parallel due to the need for the same hardware resource and insufficient hardware execution capability.

[0045] Specifically, according to the hardware execution capability, all sub-intents of the intent tree are traversed; and then a resource conflict detection algorithm is used to check the hardware requirements (the type of hardware resource required when the sub-intent is executed and its constraint conditions) of each sub-intent, and when multiple sub-intents require the same resource (such as multiple sub-intents requiring audio hardware) and the hardware execution capability is insufficient (such as the audio hardware only supporting a single task), the operation is marked as a mutual exclusion operation.

[0046] S205, analyzing the intent tree to determine the intent association between any two sub-intents; The intent association can be a logical relationship between any two sub-intents.

[0047] Specifically, all pairs of sub-intents in the intent tree are traversed; and according to the intent description (a text attribute on each sub-intent, used to describe the intent relationship between the sub-intent and other sub-intents), the intent association between any two sub-intents is determined, such as when planning a path must be executed after locating a user, it is marked as a dependent association; and when playing a voice and recording require the same device, it is marked as a mutual exclusion association.

[0048] S206, based on the intention association, dynamically adjusting the execution order of the mutual exclusion operation.

[0049] The execution order can be a sub-intention execution queue order determined by dynamically adjusting the mutual exclusion operation.

[0050] Specifically, a conflict pair (a combination of any two sub-intentions marked as a mutual exclusion operation) is selected from the mutual exclusion operation; then the execution order is reordered according to the intention association, such as when the mutual exclusion operation involves playing an emergency warning and playing a regular guide, and the intention association shows a warning intention association (such as related to obstacle avoidance), the order is dynamically adjusted to prioritize the warning execution.

[0051] By this scheme, real-time environment data and high-level task goals of the robot are obtained, based on the real-time environment data, the high-level task goals are analyzed, an abstract intention set decoupled from hardware is generated, ensuring that the intention set is independent of specific hardware, avoiding intention update delay caused by environmental changes, and providing a stable and scalable intention basis for processing. Real-time hardware state data is obtained to ensure that resource status is accurately included in decision-making, avoiding resource waste or execution failure caused by hardware neglect. According to the real-time hardware state data, the abstract intention set is split to obtain a number of sub-intentions that can be executed in parallel, forming an intention tree, ensuring that sub-intentions can be executed asynchronously, reducing sequential blocking, and improving efficiency and real-time performance, forming a hierarchical task representation. Analyzing real-time hardware state data, determining hardware execution capability, providing a benchmark for mutual exclusion operation identification, ensuring that hardware limitations are accurately quantified, and avoiding execution failure caused by unreasonable task allocation. Based on the hardware execution capability, analyze the intention tree to determine the mutual exclusion operation, provide the target for dynamic adjustment, and ensure that resource competition problems are effectively isolated. Analyzing the intention tree, determining the intention association between any two sub-intentions, providing a basis for execution order optimization, ensuring the efficient cooperation between sub-intentions, and reducing soft conflicts. Based on the intention association, dynamically adjusting the execution order of the mutual exclusion operation, solving hardware resource conflicts, ensuring asynchronous execution of mutual exclusion operations without blocking, and improving flexibility and robustness, adapting to real-time changes, and achieving efficient control.

[0052] In some embodiments, real-time environment data is analyzed to determine personnel distribution and personnel movement trend; based on the personnel distribution and personnel movement trend, the current scene characteristics are determined; based on the current scene characteristics, the preset task-intention mapping model is queried to determine the initial intention set; based on the current scene characteristics, the abnormal event is determined; according to the abnormal event, the initial intention set is corrected to determine the abstract intention set.

[0053] The personnel distribution can be structured data of the positions of the personnel in the environment.

[0054] The personnel movement trend can be a dynamic behavior pattern of the personnel detected in the environment.

[0055] The current scene feature can be a high-level abstract environment description generated based on personnel distribution and personnel movement trend.

[0056] The preset task-intent mapping model can be a model predefining a mapping relationship of the current scene feature to an initial intent set, pre-stored in the master chip and called when used.

[0057] The initial intent set can be an initial abstract task target obtained from the preset task-intent mapping model.

[0058] The abnormal event can be a sudden event identifier checked from the current scene feature.

[0059] Specifically, real-time environment data is analyzed, personnel positions (specific coordinate points of each personnel) in the environment are recognized through computer vision algorithms, and personnel distribution data is generated; at the same time, a motion tracking algorithm constructed by applying motion estimation theory is used to calculate the moving direction and speed of each personnel and determine the personnel movement trend.

[0060] Cluster analysis is performed on the personnel distribution data to identify crowd gathering areas (high-density personnel concentration areas) and densities; according to the crowd gathering areas and densities, the dynamic characteristics (describing the dynamic behavior of the scene) are calculated in combination with the personnel movement trend, and the current scene feature is comprehensively generated.

[0061] The current scene feature is input into a preset task-intent mapping model (storing a mapping relationship of the current scene feature to an initial intent set) established according to expert knowledge, and a matching initial intent set is retrieved, for example, when the current scene feature is a high-dynamic pedestrian flow scene, the initial intent set mapped is obstacle avoidance and path planning.

[0062] According to the current scene feature, outliers in the personnel distribution (such as a single person standing still in a non-passing area) or mutations in the personnel movement trend (such as sudden stop or reverse movement) are checked to determine abnormal events, for example, when a personnel hand-raising gesture is detected, it is marked as a help-seeking event; when a crowd dense movement is detected, it is marked as a congestion event.

[0063] According to the abnormal event type, a sub-intent in the initial intent set is modified (such as adding a response help-seeking intent when the abnormal event is a help-seeking event; removing the path planning intent and replacing it with a waiting avoidance intent when the abnormal event is a congestion event); then a simple rule engine (such as if-else logic, used to handle the matching of abnormal event types and intents and update the intent description) is used to handle the matching of abnormal event types and intents (such as deleting or adding an intent node when the abnormal event type and the intent conflict), and the intent description is updated (such as adjusting the functional semantics), thereby determining an abstract intent set.

[0064] By the scheme, real-time environment data is analyzed to determine personnel distribution and personnel movement trend, to ensure that environment data is converted from original form to processable input, to avoid delay caused by data blocking. According to the personnel distribution and the personnel movement trend, the current scene feature is determined to realize real-time capture of environment dynamics, to provide context for intent generation, to ensure adaptation to environment change without relying on static preset. Based on the current scene feature, the preset task-intent mapping model is queried to determine the initial intent set, to ensure quick start of intent analysis, to avoid insufficient flexibility caused by hard-coded priority, and to directly provide basic input for correction. Based on the current scene feature, the abnormal event is determined to enhance dynamic adaptation capability, to ensure timely response when the environment changes, thereby preventing intent conflict or task failure. According to the abnormal event, the initial intent set is corrected to determine the abstract intent set, to ensure real-time and robustness of the intent set, and to solve the problem of inflexible intent conflict processing.

[0065] In some embodiments, for any sub-intent, a semantic vector of the sub-intent is parsed and obtained; real-time environment data is parsed to determine a context feature vector; the context feature vector includes: pedestrian approaching speed, human posture type, environment noise level, network delay value; the semantic vector and the context feature vector are spliced into a fusion feature vector; according to the fusion feature vector, a correlation score is calculated; based on the correlation score and a preset correlation threshold, the intent correlation between any two sub-intents is determined.

[0066] The semantic vector can be a numerical vector representation of the sub-intent.

[0067] The context feature vector can be a feature vector extracted from real-time environment data, including pedestrian approaching speed, human posture type, environment noise level, and network delay value.

[0068] The pedestrian approaching speed can be a moving speed value of a target pedestrian relative to the robot.

[0069] The human posture type can be a posture category of the pedestrian.

[0070] The environment noise level can be a discrete numerical value representing the level of environmental noise.

[0071] The network delay value can be a numerical value representing the current network round-trip time.

[0072] The fusion feature vector can be a feature vector formed by splicing the semantic vector and the context feature vector.

[0073] The correlation score can be a correlation score between the fusion feature vectors of any two sub-intents.

[0074] The association threshold value can be a preset numerical threshold value for comparing the association degree score to determine the intention association, which is pre-stored in the master chip and called when used.

[0075] Specifically, the text description (sub-intention name) of any sub-intention is parsed, and semantic information is extracted according to an intention semantic mapping table (storing a mapping relationship between a text description and corresponding semantic information) constructed based on expert knowledge; then, the semantic information is directly converted into a semantic vector of the sub-intention.

[0076] Based on real-time environmental data, a moving speed value of a target pedestrian relative to the robot is extracted from the personnel movement trend data as a pedestrian approaching speed; then, an image frame (a single static image used to identify the posture of the pedestrian to determine the human posture type) is analyzed to identify the posture of the pedestrian (such as raising hands, walking, and standing still) and determine the human posture type; further, a sound pressure level (sound wave pressure amplitude used to determine the environmental noise level) is extracted from an audio sensor installed on the robot body to determine the environmental noise level; and a current network round-trip time (the time required for a data packet to be sent to the master chip and returned, reflecting the current network communication state) is extracted from a network monitoring module to determine the network delay value.

[0077] The pedestrian approaching speed, human posture type, environmental noise level, and network delay value are concatenated into a multi-dimensional vector to form a context feature vector. All elements of the semantic vector are arranged in sequence, followed by all elements of the context feature vector, to form a fusion feature vector.

[0078] According to the fusion feature vector, a preset association degree calculation function constructed using vector similarity theory is used to determine the association degree score. Then, the association degree score is compared with a preset association threshold value (used to compare the association degree score to determine the intention association) determined according to experimental calibration (such as testing the critical value of intention conflict and cooperation in different scenarios to select the score with the highest conflict determination accuracy as the association threshold value). According to the comparison result, the intention association between any two sub-intentions is determined, for example, when the association degree score is greater than or equal to the association threshold value, any two sub-intentions are marked as associated and can be cooperatively executed, such as obstacle avoidance and path planning not conflicting in a low-noise environment; when the association degree score is less than the association threshold value, any two sub-intentions are marked as not associated and potential conflict exists, such as playing voice and obstacle avoidance needing to adjust the execution order under high delay.

[0079] By the scheme, for any sub-intention, the semantic vector of the sub-intention is parsed and obtained, the intention information is efficiently stored and retrieved, the processing delay caused by instruction-level coupling is avoided, and the real-time update of the intention tree is supported. The real-time environment data is parsed, the context feature vector is determined, the defect of difficulty in adapting to environmental dynamics is solved, the intention association decision is ensured to be based on the latest context, and the intention conflict caused by environmental mutation is reduced. The semantic vector and the context feature vector are spliced into a fusion feature vector, the data transmission overhead is reduced, and the processing efficiency is improved. According to the fusion feature vector, the correlation score is calculated, the logical mutual exclusion or cooperative relationship between intentions is identified, data basis is provided for conflict processing, and the efficiency of asynchronous execution is ensured. Based on the correlation score and the preset correlation threshold, the intention association between any two sub-intentions is determined, non-blocking control is supported, and the overall robustness and real-time performance are improved.

[0080] In some embodiments, the mutually exclusive operations are analyzed to determine the mutual exclusion type; a dynamic priority queue is generated according to the intention association and the mutual exclusion type; real-time environment data is analyzed to determine real-time change events; the dynamic priority queue is updated based on the real-time change events, and the execution sequence of the mutually exclusive operations is determined according to the update result.

[0081] The mutual exclusion type can be a conflict classification between mutually exclusive operations.

[0082] The dynamic priority queue can be a data structure for storing mutually exclusive operations and their priority values.

[0083] The real-time change event can be a different event detected from the real-time environment data.

[0084] The update result can be the result of reordering the priority values in the dynamic priority queue after adjustment.

[0085] The execution sequence can be a final execution order list of mutually exclusive operations.

[0086] Specifically, for a sub-intention, a predefined resource occupation table (storing hardware resource requirements of each sub-intention) established according to the robot hardware specifications is accessed to determine the hardware resource requirements (a list of hardware resources required when the sub-intention is executed); then the hardware resource requirements of the mutually exclusive operations are compared to determine the mutual exclusion type, such as when the operations compete for the same exclusive resource (e.g., multiple sub-intentions simultaneously request audio output), the hard mutual exclusion type is determined; when the operations have no resource exclusivity but have logical sequence dependency (e.g., obstacle avoidance must be performed before path planning to avoid collision), the soft mutual exclusion type is determined.

[0087] An initial priority value is assigned for the intention association, such as, when the intention association is associated, the initial priority value is set as high, indicating that the operations can be executed cooperatively; when the intention association is not associated, the initial priority value is set as low, indicating potential conflict. Then, the initial priority value is adjusted in combination with the mutual exclusion type, such as, for the hard mutual exclusion type, the initial priority value is further reduced, forcing strict serialization execution (forced sequential execution, such as operation A can be executed only after operation B is completed); for the soft mutual exclusion type, the initial priority value remains the same or is slightly adjusted, allowing flexible order (dynamically adjusting the execution order, such as operation A and operation B can be exchanged in order). Subsequently, the adjusted priority values are sorted from high to low, generating a dynamic priority queue (such as the obstacle avoidance operation with high priority is arranged at the front end of the queue).

[0088] Real-time environmental data is analyzed to determine real-time change events, such as, when the human posture type changes to hand raising, the real-time change event is determined as a help-seeking event (indicating a change in the intention of the pedestrian); when the network delay value exceeds the preset delay threshold according to the experimental calibration (such as simulating a high network delay scenario) (used to determine whether to generate a high delay event), the real-time change event is determined as a high delay event (indicating deterioration of network state); when the environmental noise level rises to high, the real-time change event is determined as a noisy event (indicating environmental interference); when the speed of the pedestrian nears suddenly increases, the real-time change event is determined as an emergency approach event (indicating that the pedestrian is approaching quickly).

[0089] According to the real-time change event, the priority value in the dynamic priority queue is updated, such as, when the real-time change event is a help-seeking event, the priority value of the related sub-intention (such as responding to the help-seeking) is raised; when the real-time change event is a high delay event, the priority value of the sub-intention dependent on the network (such as cloud voice processing) is lowered; when the real-time change event is a noisy event, the priority value of the audio-related intention (such as playing voice) is lowered; when the real-time change event is an emergency approach event, the priority value of the safety-related intention (such as obstacle avoidance) is raised.

[0090] According to the update result, the execution sequence of the mutually exclusive operations is determined (such as the operation with the highest priority value is arranged at the front).

[0091] Through the scheme, mutually exclusive operations are analyzed, the mutual exclusion type is determined, the conflict processing basis for generating a dynamic priority queue is laid, and execution failures caused by resource competition or sequence errors are reduced. According to the intention association and the mutual exclusion type, a dynamic priority queue is generated, and the conflict probability is preliminarily reduced. Real-time environmental data is analyzed, real-time change events are determined, real-time interference is reflected, and a basis for updating the queue is provided to ensure that the execution order responds to sudden changes. Based on the real-time change event, the dynamic priority queue is updated, the execution sequence of the mutually exclusive operations is determined according to the update result, blocking and conflict are reduced, and the overall execution efficiency is improved.

[0092] In some embodiments, the hardware performance indicators and the network state data are acquired; the hardware performance indicators are analyzed to determine hardware preset processing capability, current battery power and sensor accuracy; the network state data are analyzed to determine network delay and network bandwidth; and the hardware execution capability is determined based on the network delay and the network bandwidth, according to the hardware preset processing capability, the current battery power and the sensor accuracy.

[0093] The hardware performance indicators can be a set of real-time state parameters of the hardware module.

[0094] The network state data can be a set of real-time state parameters of the network interface.

[0095] The hardware preset processing capability can be a static value representing the theoretical processing capability of the hardware.

[0096] The current battery power can be a real-time acquired remaining battery power value.

[0097] The sensor accuracy can be a real-time accuracy value of the sensor.

[0098] The network delay can be a real-time communication delay value, which can be in units of milliseconds.

[0099] The network bandwidth can be a real-time data transmission rate value, which can be in units of megabits per second.

[0100] Specifically, the hardware performance indicators are acquired by hardware sensors installed on the robot hardware (such as near the CPU module, on the surface of the battery pack, and on the sensor module interface circuit); at the same time, the network state data are acquired by performing a network detection operation (active detection action performed by the network protocol stack, used to acquire network state data) through the network protocol stack. The hardware preset processing capability is directly extracted from the hardware performance indicators; the current battery power is acquired in real time from the hardware performance indicators as the remaining battery power percentage; and the sensor accuracy is directly acquired from the hardware performance indicators.

[0101] The network state data are analyzed to extract an average round-trip time value (average value of response times of multiple detections in the network detection operation) as the network delay; and a current available bandwidth rate (real-time available data transmission rate of the network connection) is extracted as the network bandwidth.

[0102] The network delay and the network bandwidth, the hardware preset processing capability, the current battery power and the sensor accuracy are integrated, and the hardware execution capability is determined by a predefined rule logic (used to determine the hardware comprehensive capability) established according to expert experience, such as reducing the hardware execution capability evaluation when the network delay is high or the network bandwidth is low; reducing the hardware execution capability evaluation when the hardware preset processing capability is low; reducing the hardware execution capability evaluation when the current battery power is low; and reducing the hardware execution capability evaluation when the sensor accuracy is low.

[0103] By the scheme, hardware performance indicators and network state data are acquired to reduce control delay and thus relieve congestion problems. Hardware performance indicators are analyzed to determine hardware preset processing capacity, current battery power and sensor accuracy, to convert raw data into comparable quantitative values and solve hardware heterogeneity problems. Network state data are analyzed to determine network delay and network bandwidth, to convert raw data into comparable quantitative values and relieve insufficient environmental adaptation problems. Based on network delay and network bandwidth, hardware execution capacity is determined according to hardware preset processing capacity, current battery power and sensor accuracy, to reflect overall hardware state, support dynamic task decision-making and solve intent conflict and congestion delay problems.

[0104] In some embodiments, an abnormal event is analyzed to determine abnormal event features; a pedestrian intent is determined according to personnel distribution and personnel movement trend; and it is determined whether to update an intent tree according to the abnormal event features and the pedestrian intent.

[0105] The abnormal event features can be key quantitative parameters extracted from the abnormal event, including event type, event location, severity level, etc.

[0106] The pedestrian intent can be a pedestrian behavior target for identifying personnel interaction intent.

[0107] Specifically, abnormal event features are directly extracted from the abnormal event, including event type (category of the abnormal event, such as help-seeking, obstacle, noise), event location (spatial location of the event occurrence), and severity level (event impact degree, such as high, medium, and low).

[0108] According to the personnel distribution and the personnel movement trend, a preset rule logic (storing a mapping relationship between the personnel distribution and the personnel movement trend and the pedestrian intent, and used to determine the pedestrian intent) constructed according to expert field knowledge is applied to determine the pedestrian intent. For example, when the moving direction points to the robot and the speed is high, the pedestrian intent is marked as a help-seeking intent (such as help-seeking near the robot); when the moving direction is away from the robot or the speed is low, the pedestrian intent is marked as a passing-by intent (such as passing by); and when the personnel distribution density is high and the personnel movement trend is chaotic, the pedestrian intent is marked as a group intent (such as a potential interference event).

[0109] According to the abnormal event feature and the pedestrian intention, a preset rule logic (mapping relationship between the abnormal event feature, the pedestrian intention, and whether to update the intention tree is stored, and is used to determine whether to update the intention tree) constructed according to expert experience is applied to determine whether to update the intention tree. For example, when the abnormal event feature indicates an emergency event (such as the event type is help and the severity level is high) and the pedestrian intention is a help intention, the intention tree is updated (such as switching from path movement to responding to help). When there is no emergency event (such as the abnormal event type is noise and the severity level is low) and the pedestrian intention is a passing intention, the intention tree is not updated.

[0110] Through the scheme, the abnormal event is analyzed, the abnormal event feature is determined, the foundation for intention update is laid, the dynamic event is avoided from being ignored due to the inability to correct the intention tree, and thus the environmental adaptability is improved. According to the personnel distribution and the personnel movement trend, the pedestrian intention is determined, and the control chain is avoided from being delayed due to waiting for personnel data analysis. According to the abnormal event feature and the pedestrian intention, whether to update the intention tree is determined, the network burden and the control coupling degree are reduced, and the real-time performance and the expansibility in a fluctuating environment are improved.

[0111] In some embodiments, the semantic vector is analyzed, the relevance between the sub-intention and the pedestrian intention is determined, a new intention is dynamically inserted according to the relevance, and an original sub-intention in conflict with the pedestrian intention is deleted, to obtain a reconstructed abstract intention set.

[0112] The relevance can be the association strength between the sub-intention and the pedestrian intention.

[0113] The new intention can be a sub-task target dynamically generated and inserted into the abstract intention set according to the pedestrian intention.

[0114] The original sub-intention can be an existing sub-task target in the abstract intention set before reconstruction.

[0115] Specifically, the semantic vector is analyzed, the semantic vector of the sub-intention and the semantic vector of the pedestrian intention are read one by one (for example, when the sub-intention is to play a guide voice, the semantic vector of the sub-intention is resource dependency = audio hardware; when the pedestrian intention is a help intention, the semantic vector of the pedestrian intention is intention type = help), and then the relevance between the sub-intention and the pedestrian intention is determined according to the semantic vector of the sub-intention and the semantic vector of the pedestrian intention using a preset condition rule (used to determine the relevance) constructed according to expert knowledge. For example, when the resource dependency of the sub-intention is the same as the resource dependency of the pedestrian intention (such as both depending on audio hardware), it is marked as high relevance (indicating potential resource conflict); when the intention type of the sub-intention and the intention type of the pedestrian intention are logically mutually exclusive (such as the sub-intention is path movement and the pedestrian intention is a help intention, and the help needs to pause movement), it is marked as low relevance (indicating conflict).

[0116] The pedestrian intention is directly mapped to a new intention, such as, when the pedestrian intention is a help-seeking intention (such as raising a hand for help), the new intention is a response to the help-seeking (such as pausing movement and broadcasting information); when the pedestrian intention is a passing intention (such as a pedestrian approaching without interaction), the new intention is to avoid or standby (such as adjusting the path to avoid). Then, according to the relevance, the new intention is added to the abstract intention set; then, a preset rule (for judging conflicts) constructed according to expert experience is applied to judge conflicts, and the original sub-intention conflicting with the pedestrian intention is deleted, to obtain a reconstructed abstract intention set, such as, when the play guide voice of the sub-intention conflicts with the help-seeking intention of the pedestrian intention, the original sub-intention is deleted due to resource competition.

[0117] By the scheme, the semantic vector is analyzed, the relevance between the sub-intention and the pedestrian intention is determined, the intention semantic correlation is dynamically evaluated, and the delay caused by centralized processing is avoided. According to the relevance, the new intention is dynamically inserted, and the original sub-intention conflicting with the pedestrian intention is deleted, to obtain a reconstructed abstract intention set, supporting asynchronous collaboration, improving robustness and real-time performance.

[0118] In some embodiments, the sub-intention is assigned to a corresponding hardware execution agent; the execution state of each sub-intention is monitored to obtain a ready flag signal; and the execution results of multiple sub-intentions are aggregated according to the ready flag signal and the intention correlation.

[0119] The hardware execution agent can be a hardware module unit in the robot responsible for executing a specific sub-intention.

[0120] The execution state can be a state event published in real time during the execution of the sub-intention.

[0121] The ready flag signal can be a flag indicating whether the execution of the sub-intention is completed.

[0122] The execution result can be output data of the sub-intention after being completed by the hardware execution agent.

[0123] Specifically, according to a preset hardware mapping table (storing the correspondence between the sub-intention and the hardware execution agent) constructed according to expert experience, the sub-intention is assigned to a corresponding hardware execution agent, such as, when the sub-intention is audio hardware, it is assigned to an audio hardware execution agent (such as a hardware unit responsible for playing voice); when the sub-intention is a mobile chassis, it is assigned to a mobile chassis hardware execution agent (such as a hardware unit responsible for path movement); when the sub-intention is a sensor, it is assigned to a sensor hardware execution agent (such as a hardware unit responsible for obstacle avoidance or environment detection).

[0124] The execution status of each sub-intention is periodically polled by the state monitoring module, including busy (currently executing other tasks), idle (no task execution, but not ready), ready (hardware has completed initialization or previous tasks, and can accept new sub-intention execution); when the execution status is ready, a ready flag signal is automatically generated.

[0125] According to the ready flag signal and the intention association, a preset rule (for aggregating the execution results of multiple sub-intentions according to the ready flag signal and the intention association) set by expert experience is applied to aggregate the execution results of multiple sub-intentions. For example, the ready flag signals of multiple sub-intentions are all ready (obstacle avoidance and path movement), and the intention association is medium correlation (no resource conflict), then they are aggregated in parallel (integrated navigation instructions are generated by combining path coordinates and obstacle avoidance data); when the ready flag signals of sub-intentions conflict (such as high correlation intention Figure 1 ready, the other is not ready), only the ready sub-intention is aggregated, and the conflicting intention is ignored or delayed; when the ready flag signals of multiple sub-intentions are all ready (path movement and response to help), and the intention association is low correlation (path movement and response to help logic are mutually exclusive), the execution results are aggregated in sequence based on the ready flag signals (the response to help result is aggregated first, and then the path movement is processed).

[0126] By this scheme, sub-intentions are assigned to corresponding hardware execution agents, avoiding instruction publishing stagnation caused by network delay or hardware response timeout, and improving real-time performance. The execution status of each sub-intention is monitored to obtain a ready flag signal, reducing network burden, especially in high delay environment, reducing the overhead caused by frequent data transmission. According to the ready flag signal and the intention association, the execution results of multiple sub-intentions are aggregated, reducing network transmission frequency and decoupling hardware execution details.

[0127] In some embodiments, according to network delay and network bandwidth, network reliability is determined; if the network reliability is low, all sub-intentions dependent on cloud computing in the intention tree are obtained; the sub-intentions dependent on cloud computing are converted into simplified sub-intentions that can be executed locally; based on hardware execution capability, an offline execution sequence is regenerated.

[0128] The network reliability can be a reliability level for evaluating the stability of the current network environment.

[0129] The simplified sub-intention can be a simplified version of the sub-intention dependent on cloud computing that can be executed locally after conversion.

[0130] The offline execution sequence can be an ordered list of sub-intentions for local hardware execution.

[0131] Specifically, the network delay is compared with a preset threshold (for determining the network delay) set according to the hardware parameters, and the network bandwidth is compared with a preset threshold (for determining the network bandwidth) set according to the hardware parameters; and the network reliability is determined according to the comparison results, for example, when the network delay is greater than the preset threshold or the network bandwidth is less than the preset threshold, the network reliability is determined to be low level; otherwise, the network reliability is determined to be high level.

[0132] If the network reliability is low level, all sub-intent nodes of the intent tree are traversed to check the resource dependency attribute (for identifying the core resource type, such as cloud computing, on which the execution of the sub-intent depends) of each sub-intent; when the resource dependency attribute is cloud computing, the sub-intent is marked as a sub-intent dependent on cloud computing; and then all sub-intents marked as dependent on cloud computing in the intent tree are collected.

[0133] The sub-intents dependent on cloud computing are traversed, and a preset conversion rule table (storing the mapping relationship of each sub-intent dependent on cloud computing to a simplified sub-intent that can be executed locally, such as mapping the sub-intent cloud voice recognition to local keyword detection) is applied to each sub-intent to determine the simplified sub-intent that can be executed locally.

[0134] According to the hardware execution capability and the semantic vector of the sub-intent, the simplified sub-intents are sorted (when the hardware execution capability value is low, the low-load sub-intent (such as simple obstacle avoidance) is prioritized over the high-load sub-intent (such as voice broadcast)); and then the sorted sub-intents are encapsulated into an offline execution sequence.

[0135] According to the network delay and the network bandwidth, the network reliability is determined to ensure that the intent adjustment is only performed when the network condition is poor, unnecessary operations are avoided, and a foundation is laid for real-time response in a dynamic environment. If the network reliability is low level, all sub-intents dependent on cloud computing in the intent tree are obtained to ensure that the focus of adjustment is only on high-risk tasks, and the vulnerability of the overall control chain is reduced. The sub-intents dependent on cloud computing are converted into simplified sub-intents that can be executed locally to reduce network dependence, avoid blocking caused by cloud computing communication, and adapt to dynamic changes in hardware. Based on the hardware execution capability, the offline execution sequence is regenerated to ensure that the robot control chain can still operate efficiently under poor network conditions.

[0136] Figure 3 The structure schematic diagram of a robot asynchronous control system based on intent driving provided by an embodiment of the present application is shown in FIG. 3. Figure 3 As shown in FIG. 3, the robot asynchronous control system 300 based on intent driving of the present embodiment includes a set generation module 301, an intent tree generation module 302, a capability analysis module 303, an operation determination module 304, a correlation analysis module 305, and a sequence adjustment module 306.

[0137] The collection generation module 301 is configured to acquire real-time environment data and a high-level task target of the robot, parse the high-level task target based on the real-time environment data, and generate an abstract intention set decoupled from hardware. The intention tree generation module 302 is configured to acquire real-time hardware state data, split the abstract intention set based on the real-time hardware state data to obtain a plurality of sub-intentions that can be executed in parallel, and form an intention tree. The capability analysis module 303 is configured to analyze the real-time hardware state data and determine a hardware execution capability. The operation determination module 304 is configured to analyze the intention tree based on the hardware execution capability and determine mutually exclusive operations. The correlation analysis module 305 is configured to analyze the intention tree and determine intention correlations between any two sub-intentions. The sequence adjustment module 306 is configured to dynamically adjust an execution sequence of the mutually exclusive operations based on the intention correlations.

[0138] Optionally, when the collection generation module 301 parses the high-level task target based on the real-time environment data and generates an abstract intention set decoupled from hardware, the collection generation module 301 is configured to: analyze the real-time environment data to determine a personnel distribution and a personnel movement trend; determine a current scene feature based on the personnel distribution and the personnel movement trend; query a preset task-intention mapping model based on the current scene feature to determine an initial intention set; determine an abnormal event based on the current scene feature; correct the initial intention set based on the abnormal event to determine the abstract intention set.

[0139] Optionally, when the correlation analysis module 305 analyzes the intention tree and determines intention correlations between any two sub-intentions, the correlation analysis module 305 is configured to: parse a semantic vector of any sub-intention; analyze the real-time environment data to determine a context feature vector; the context feature vector includes a pedestrian approaching speed, a human posture type, an environment noise level, and a network delay value; splice the semantic vector and the context feature vector into a fusion feature vector; calculate a correlation score based on the fusion feature vector; determine intention correlations between any two sub-intentions based on the correlation score and a preset correlation threshold.

[0140] Optionally, when the sequence adjustment module 306 dynamically adjusts an execution sequence of the mutually exclusive operations based on the intention correlations, the sequence adjustment module 306 is configured to: analyzing the mutual exclusion operation, determining a mutual exclusion type; generating a dynamic priority queue according to the intention association and the mutual exclusion type; analyzing the real-time environment data, determining a real-time change event; updating the dynamic priority queue based on the real-time change event, and determining an execution sequence of the mutual exclusion operation according to an update result.

[0141] Optionally, when the capability analysis module 303 analyzes the real-time hardware state data to determine a hardware execution capability, the capability analysis module 303 is configured to: obtain hardware performance indicators and network state data; analyze the hardware performance indicators to determine a hardware preset processing capability, a current battery power, and a sensor accuracy; analyze the network state data to determine a network delay and a network bandwidth; determine the hardware execution capability based on the network delay and the network bandwidth, and according to the hardware preset processing capability, the current battery power, and the sensor accuracy.

[0142] Optionally, the intention-driven robot asynchronous control system further includes an intention tree determination module 307, configured to: analyze the abnormal event to determine an abnormal event feature; determine a pedestrian intention according to the personnel distribution and the personnel movement trend; determine whether to update the intention tree according to the abnormal event feature and the pedestrian intention.

[0143] Optionally, the intention-driven robot asynchronous control system further includes an intention set determination module 308, configured to: analyze the semantic vector to determine a correlation between the sub-intention and the pedestrian intention; dynamically insert a new intention and delete an original sub-intention conflicting with the pedestrian intention according to the correlation, to obtain a reconstructed abstract intention set.

[0144] Optionally, the intention-driven robot asynchronous control system further includes a result aggregation module 309, configured to: assign a sub-intention to a corresponding hardware execution agent; monitor an execution state of each sub-intention to obtain a ready flag signal; aggregate execution results of multiple sub-intentions according to the ready flag signal and the intention association.

[0145] Optionally, the intention-driven robot asynchronous control system further includes a sequence generation module 310, configured to: determine a network reliability according to the network delay and the network bandwidth; if the network reliability is lower than a preset level, acquire all sub-intents in the intent tree that depend on cloud computing; convert the sub-intents that depend on cloud computing into simplified sub-intents that can be executed locally; re-generate an offline execution sequence based on the hardware execution capability.

[0146] The system of the embodiment can be used to execute the method of any of the above embodiments, and has similar implementation principles and technical effects, which will not be described here again.

Claims

1. An intent-driven robot asynchronous control method, characterized by, The method comprises the following steps: acquiring real-time environment data, high-level task goals of a robot, based on the real-time environment data, analyzing the high-level task goals, and generating an abstract intent set decoupled from hardware; acquiring real-time hardware state data; based on the real-time hardware state data, splitting the abstract intent set to obtain a plurality of sub-intents that can be executed in parallel, and forming an intent tree; analyzing the real-time hardware state data to determine hardware execution capability; based on the hardware execution capability, analyzing the intent tree to determine mutually exclusive operations; analyzing the intent tree to determine intent association between any two sub-intents; based on the intent association, dynamically adjusting the execution order of the mutually exclusive operations.

2. The method of claim 1, wherein, The step of based on the real-time environment data, analyzing the high-level task goals, and generating an abstract intent set decoupled from hardware comprises the following steps: analyzing the real-time environment data to determine personnel distribution and personnel movement trend; based on the personnel distribution and the personnel movement trend, determining current scene characteristics; based on the current scene characteristics, querying a preset task-intent mapping model to determine an initial intent set; based on the current scene characteristics, determining an abnormal event; based on the abnormal event, revising the initial intent set to determine the abstract intent set.

3. The method of claim 2, wherein, The step of analyzing the intent tree to determine intent association between any two sub-intents comprises the following steps: for any sub-intent, analyzing and obtaining a semantic vector of the sub-intent; analyzing the real-time environment data to determine a context feature vector; the context feature vector comprises pedestrian approaching speed, human posture type, environmental noise level, and network delay value; concatenating the semantic vector and the context feature vector into a fusion feature vector; based on the fusion feature vector, calculating an association score; based on the association score and a preset association threshold, determining intent association between any two sub-intents.

4. The method of claim 1, wherein, The step of based on the intent association, dynamically adjusting the execution order of the mutually exclusive operations comprises the following steps: analyzing the mutually exclusive operations to determine mutually exclusive types; based on the intent association and the mutually exclusive types, generating a dynamic priority queue; analyzing the real-time environment data to determine real-time change events; based on the real-time change events, updating the dynamic priority queue, and based on the update result, determining an execution sequence of the mutually exclusive operations.

5. The method of claim 1, wherein, The step of analyzing the real-time hardware state data to determine hardware execution capability comprises the following steps: acquiring hardware performance indicators and network state data; analyzing the hardware performance indicators to determine hardware preset processing capability, current battery capacity, and sensor accuracy; analyzing the network state data to determine network delay and network bandwidth; based on the network delay and the network bandwidth, determining hardware execution capability based on the hardware preset processing capability, current battery capacity, and sensor accuracy.

6. The method of claim 3, wherein, After forming the intent tree, the method further comprises the following steps: analyzing the abnormal event to determine abnormal event characteristics; based on the personnel distribution and the personnel movement trend, determining pedestrian intent; based on the abnormal event characteristics and the pedestrian intent, determining whether to update the intent tree.

7. The method of claim 6, wherein, The method comprises the following steps: analyzing the semantic vector to determine relevance between the sub-intent and the pedestrian intent; According to the correlation, a new intention is dynamically inserted, and an original sub-intention conflicting with the pedestrian intention is deleted, to obtain a reconstructed abstract intention set.

8. The method of claim 1, wherein, After dynamically adjusting the execution order of the mutually exclusive operations based on the intention association, the method further includes: allocating the sub-intentions to corresponding hardware execution agents; monitoring execution states of each sub-intention to obtain a ready flag signal; according to the ready flag signal and the intention association, aggregating execution results of the multiple sub-intentions.

9. The method of claim 5, wherein, After determining the hardware execution capability according to the hardware preset processing capability, the current battery power and the sensor accuracy based on the network delay and the network bandwidth, the method further includes: determining network reliability according to the network delay and the network bandwidth; if the network reliability is lower than a preset level, obtaining all sub-intentions in the intention tree that depend on cloud computing; converting the sub-intentions that depend on cloud computing into simplified sub-intentions that can be executed locally; regenerating an offline execution sequence based on the hardware execution capability.

10. An intent-driven robot asynchronous control system, comprising: The method is applied to any one of claims 1-9, and includes: a set generation module configured to obtain real-time environment data and a high-level task target of a robot, parse the high-level task target based on the real-time environment data, and generate an abstract intention set decoupled from hardware; an intention tree generation module configured to obtain real-time hardware state data, split the abstract intention set based on the real-time hardware state data to obtain a plurality of sub-intentions that can be executed in parallel, and form an intention tree; a capability analysis module configured to analyze the real-time hardware state data and determine a hardware execution capability; an operation determination module configured to analyze the intention tree based on the hardware execution capability and determine mutually exclusive operations; an association analysis module configured to analyze the intention tree and determine an intention association between any two sub-intentions; a sequence adjustment module configured to dynamically adjust an execution order of the mutually exclusive operations based on the intention association.

Images